wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-05-28 01:47:32 +08:00
Code Issues Actions 29 Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/master' into lens

Browse Source
This commit is contained in:
kijai 2026-05-26 07:45:40 +03:00
commit 5b102121df
101 changed files with 27460 additions and 2805 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
app/assets/api/routes.py
View File

@ -160,10 +160,12 @@ def _build_asset_response(result: schemas.AssetDetailResult | schemas.UploadResu
preview_url = None preview_url = None
else: else:
preview_url = _build_preview_url_from_view(result.tags, result.ref.user_metadata) preview_url = _build_preview_url_from_view(result.tags, result.ref.user_metadata)
asset_content_hash = result.asset.hash if result.asset else None
return schemas_out.Asset( return schemas_out.Asset(
id=result.ref.id, id=result.ref.id,
name=result.ref.name, name=result.ref.name,
asset_hash=result.asset.hash if result.asset else None, hash=asset_content_hash,
asset_hash=asset_content_hash,
size=int(result.asset.size_bytes) if result.asset else None, size=int(result.asset.size_bytes) if result.asset else None,
mime_type=result.asset.mime_type if result.asset else None, mime_type=result.asset.mime_type if result.asset else None,
tags=result.tags, tags=result.tags,

1
app/assets/api/schemas_out.py
View File

@ -10,6 +10,7 @@ class Asset(BaseModel):
id: str id: str
name: str name: str
hash: str | None = None
asset_hash: str | None = None asset_hash: str | None = None
size: int | None = None size: int | None = None
mime_type: str | None = None mime_type: str | None = None

1
app/assets/services/metadata_extract.py
View File

@ -4,7 +4,6 @@ Tier 1: Filesystem metadata (zero parsing)
Tier 2: Safetensors header metadata (fast JSON read only) Tier 2: Safetensors header metadata (fast JSON read only)
""" """
from __future__ import annotations
import json import json
import logging import logging

2
app/custom_node_manager.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import os import os
import folder_paths import folder_paths
import glob import glob

1
app/frontend_management.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import argparse import argparse
import logging import logging
import os import os

40
app/logger.py
View File

@ -5,6 +5,40 @@ import logging
import sys import sys
import threading import threading
ANSI_NAMED_COLORS = {
'black': '\033[30m',
'red': '\033[31m',
'green': '\033[32m',
'yellow': '\033[33m',
'blue': '\033[34m',
'magenta': '\033[35m',
'cyan': '\033[36m',
'white': '\033[37m',
}
ANSI_LEVEL_COLORS = {
'DEBUG': ANSI_NAMED_COLORS['cyan'],
'INFO': ANSI_NAMED_COLORS['green'],
'WARNING': ANSI_NAMED_COLORS['yellow'],
'ERROR': ANSI_NAMED_COLORS['red'],
'CRITICAL': ANSI_NAMED_COLORS['magenta'],
}
ANSI_RESET = '\033[0m'
ANSI_BOLD = '\033[1m'
class ColoredFormatter(logging.Formatter):
def format(self, record):
color = ANSI_LEVEL_COLORS.get(record.levelname, '')
bold = ANSI_BOLD if record.levelno >= logging.WARNING else ''
level_tag = f"{bold}{color}[{record.levelname}]{ANSI_RESET} "
message = super().format(record)
line_color = ANSI_NAMED_COLORS.get(getattr(record, 'color', ''), '')
if line_color:
return f"{level_tag}{line_color}{message}{ANSI_RESET}"
return level_tag + message
logs = None logs = None
stdout_interceptor = None stdout_interceptor = None
stderr_interceptor = None stderr_interceptor = None
@ -68,8 +102,10 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool
logger = logging.getLogger() logger = logging.getLogger()
logger.setLevel(log_level) logger.setLevel(log_level)
formatter = ColoredFormatter("%(message)s")
stream_handler = logging.StreamHandler() stream_handler = logging.StreamHandler()
stream_handler.setFormatter(logging.Formatter("%(message)s")) stream_handler.setFormatter(formatter)
if use_stdout: if use_stdout:
# Only errors and critical to stderr # Only errors and critical to stderr
@ -77,7 +113,7 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool
# Lesser to stdout # Lesser to stdout
stdout_handler = logging.StreamHandler(sys.stdout) stdout_handler = logging.StreamHandler(sys.stdout)
stdout_handler.setFormatter(logging.Formatter("%(message)s")) stdout_handler.setFormatter(formatter)
stdout_handler.addFilter(lambda record: record.levelno < logging.ERROR) stdout_handler.addFilter(lambda record: record.levelno < logging.ERROR)
logger.addHandler(stdout_handler) logger.addHandler(stdout_handler)

2
app/model_manager.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import os import os
import base64 import base64
import json import json

1
app/user_manager.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import json import json
import os import os
import re import re

2091
blueprints/Audio Generation (Stable Audio 3 Medium Base).json Normal file
View File

File diff suppressed because one or more lines are too long

2091
blueprints/Audio Generation (Stable Audio 3 Medium).json Normal file
View File

File diff suppressed because one or more lines are too long

2
blueprints/Canny to Image (Z-Image-Turbo).json
View File

@ -1553,7 +1553,7 @@
"VHS_MetadataImage": true, "VHS_MetadataImage": true,
"VHS_KeepIntermediate": true "VHS_KeepIntermediate": true
}, },
"category": "Image generation and editing/Canny to image", "category": "Image generation and editing/Conditioned",
"description": "Generates an image from a Canny edge map using Z-Image-Turbo, with text conditioning." "description": "Generates an image from a Canny edge map using Z-Image-Turbo, with text conditioning."
} }
] ]

2
blueprints/Canny to Video (LTX 2.0).json
View File

@ -3600,7 +3600,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Video generation and editing/Canny to video", "category": "Video generation and editing/Conditioned",
"description": "Generates video from Canny edge maps using LTX-2, with optional synchronized audio." "description": "Generates video from Canny edge maps using LTX-2, with optional synchronized audio."
} }
] ]

2
blueprints/ControlNet (Z-Image-Turbo).json
View File

@ -1401,7 +1401,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Image generation and editing/ControlNet", "category": "Image generation and editing/Conditioned",
"description": "Generates images from a text prompt and ControlNet conditioning (e.g. depth, canny) using Z-Image-Turbo." "description": "Generates images from a text prompt and ControlNet conditioning (e.g. depth, canny) using Z-Image-Turbo."
} }
] ]

2
blueprints/Depth to Image (Z-Image-Turbo).json
View File

@ -1579,7 +1579,7 @@
"VHS_MetadataImage": true, "VHS_MetadataImage": true,
"VHS_KeepIntermediate": true "VHS_KeepIntermediate": true
}, },
"category": "Image generation and editing/Depth to image", "category": "Image generation and editing/Conditioned",
"description": "Generates an image from a depth map using Z-Image-Turbo with text conditioning." "description": "Generates an image from a depth map using Z-Image-Turbo with text conditioning."
}, },
{ {

2
blueprints/Depth to Video (ltx 2.0).json
View File

@ -4233,7 +4233,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Video generation and editing/Depth to video", "category": "Video generation and editing/Conditioned",
"description": "Generates depth-controlled video with LTX-2: motion and structure follow a depth-reference video alongside text prompting, optional first-frame image conditioning, with optional synchronized audio." "description": "Generates depth-controlled video with LTX-2: motion and structure follow a depth-reference video alongside text prompting, optional first-frame image conditioning, with optional synchronized audio."
}, },
{ {

2
blueprints/First-Last-Frame to Video (LTX-2.3).json
View File

@ -3350,7 +3350,7 @@
} }
], ],
"extra": {}, "extra": {},
"category": "Video generation and editing/First-Last-Frame to Video", "category": "Video generation and editing/Conditioned",
"description": "Generates a video interpolating between first and last keyframes using LTX-2.3." "description": "Generates a video interpolating between first and last keyframes using LTX-2.3."
} }
] ]

2
blueprints/First-Last-Frame to Video.json
View File

@ -3350,7 +3350,7 @@
} }
], ],
"extra": {}, "extra": {},
"category": "Video generation and editing/First-Last-Frame to Video", "category": "Video generation and editing/FLF2V",
"description": "Generates a video that interpolates between the first and last keyframes using LTX-2.3, including optional audio." "description": "Generates a video that interpolates between the first and last keyframes using LTX-2.3, including optional audio."
} }
] ]

1266
blueprints/Geometry Estimation (MoGe).json Normal file
View File

File diff suppressed because it is too large Load Diff

2
blueprints/Image Captioning (gemini).json
View File

@ -310,7 +310,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Text generation/Image Captioning", "category": "Image Tools",
"description": "Generates descriptive captions for images using Google's Gemini multimodal LLM." "description": "Generates descriptive captions for images using Google's Gemini multimodal LLM."
} }
] ]

150
blueprints/Image to Depth Map (Lotus).json → blueprints/Image Depth Estimation (Lotus Depth).json
View File

@ -1,19 +1,18 @@
{ {
"id": "6af0a6c1-0161-4528-8685-65776e838d44",
"revision": 0, "revision": 0,
"last_node_id": 75, "last_node_id": 76,
"last_link_id": 245, "last_link_id": 0,
"nodes": [ "nodes": [
{ {
"id": 75, "id": 76,
"type": "488652fd-6edf-4d06-8f9f-4d84d3a34eaf", "type": "96338968-1242-4f02-b6a1-d496af4bcffe",
"pos": [ "pos": [
600, 670,
830 1280
], ],
"size": [ "size": [
400, 400,
110 201.3125
], ],
"flags": {}, "flags": {},
"order": 0, "order": 0,
@ -59,47 +58,44 @@
"links": [] "links": []
} }
], ],
"title": "Image Depth Estimation (Lotus Depth)",
"properties": { "properties": {
"proxyWidgets": [ "proxyWidgets": [
[ [
"-1", "28",
"sigma" "sigma"
], ],
[ [
"-1", "10",
"unet_name" "unet_name"
], ],
[ [
"-1", "14",
"vae_name" "vae_name"
] ]
], ],
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.14.1" "ver": "0.14.1"
}, },
"widgets_values": [ "widgets_values": []
999.0000000000002,
"lotus-depth-d-v1-1.safetensors",
"vae-ft-mse-840000-ema-pruned.safetensors"
]
} }
], ],
"links": [], "links": [],
"groups": [], "version": 0.4,
"definitions": { "definitions": {
"subgraphs": [ "subgraphs": [
{ {
"id": "488652fd-6edf-4d06-8f9f-4d84d3a34eaf", "id": "96338968-1242-4f02-b6a1-d496af4bcffe",
"version": 1, "version": 1,
"state": { "state": {
"lastGroupId": 1, "lastGroupId": 1,
"lastNodeId": 75, "lastNodeId": 76,
"lastLinkId": 245, "lastLinkId": 245,
"lastRerouteId": 0 "lastRerouteId": 0
}, },
"revision": 0, "revision": 0,
"config": {}, "config": {},
"name": "Image to Depth Map (Lotus)", "name": "Image Depth Estimation (Lotus Depth)",
"inputNode": { "inputNode": {
"id": -10, "id": -10,
"bounding": [ "bounding": [
@ -191,12 +187,12 @@
"id": 10, "id": 10,
"type": "UNETLoader", "type": "UNETLoader",
"pos": [ "pos": [
108.05555555555557, 110,
-253.05555555555557 -250
], ],
"size": [ "size": [
254.93706597222226, 260,
82 90
], ],
"flags": {}, "flags": {},
"order": 4, "order": 4,
@ -234,9 +230,9 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "UNETLoader",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "UNETLoader",
"models": [ "models": [
{ {
"name": "lotus-depth-d-v1-1.safetensors", "name": "lotus-depth-d-v1-1.safetensors",
@ -255,12 +251,12 @@
"id": 18, "id": 18,
"type": "DisableNoise", "type": "DisableNoise",
"pos": [ "pos": [
607.0641494069639, 610,
-268.33337840371513 -270
], ],
"size": [ "size": [
175, 180,
33.333333333333336 40
], ],
"flags": {}, "flags": {},
"order": 0, "order": 0,
@ -278,26 +274,25 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "DisableNoise",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "DisableNoise",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 23, "id": 74,
"type": "VAEEncode", "type": "VAEEncode",
"pos": [ "pos": [
620, 620,
160 160
], ],
"size": [ "size": [
175, 180,
50 50
], ],
"flags": {}, "flags": {},
"order": 10, "order": 11,
"mode": 0, "mode": 0,
"inputs": [ "inputs": [
{ {
@ -325,12 +320,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "VAEEncode",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "VAEEncode",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 21, "id": 21,
@ -341,7 +335,7 @@
], ],
"size": [ "size": [
210, 210,
58 60
], ],
"flags": {}, "flags": {},
"order": 1, "order": 1,
@ -369,9 +363,9 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "KSamplerSelect",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "KSamplerSelect",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, },
"widgets_values": [ "widgets_values": [
@ -386,7 +380,7 @@
-170 -170
], ],
"size": [ "size": [
175, 180,
50 50
], ],
"flags": {}, "flags": {},
@ -418,12 +412,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "BasicGuider",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "BasicGuider",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 16, "id": 16,
@ -433,8 +426,8 @@
-130 -130
], ],
"size": [ "size": [
295.99609375, 300,
271.65798611111114 280
], ],
"flags": {}, "flags": {},
"order": 6, "order": 6,
@ -490,12 +483,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "SamplerCustomAdvanced",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "SamplerCustomAdvanced",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 28, "id": 28,
@ -506,10 +498,10 @@
], ],
"size": [ "size": [
210, 210,
58 60
], ],
"flags": {}, "flags": {},
"order": 11, "order": 10,
"mode": 0, "mode": 0,
"inputs": [ "inputs": [
{ {
@ -540,9 +532,9 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "SetFirstSigma",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "SetFirstSigma",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, },
"widgets_values": [ "widgets_values": [
@ -557,7 +549,7 @@
-120 -120
], ],
"size": [ "size": [
175, 180,
50 50
], ],
"flags": {}, "flags": {},
@ -589,12 +581,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "VAEDecode",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "VAEDecode",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 22, "id": 22,
@ -604,8 +595,8 @@
-220 -220
], ],
"size": [ "size": [
175, 180,
33.333333333333336 40
], ],
"flags": {}, "flags": {},
"order": 9, "order": 9,
@ -630,12 +621,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "ImageInvert",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "ImageInvert",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 14, "id": 14,
@ -645,8 +635,8 @@
-90 -90
], ],
"size": [ "size": [
254.93706597222226, 260,
58 60
], ],
"flags": {}, "flags": {},
"order": 5, "order": 5,
@ -675,9 +665,9 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "VAELoader",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "VAELoader",
"models": [ "models": [
{ {
"name": "vae-ft-mse-840000-ema-pruned.safetensors", "name": "vae-ft-mse-840000-ema-pruned.safetensors",
@ -692,15 +682,15 @@
] ]
}, },
{ {
"id": 68, "id": 75,
"type": "LotusConditioning", "type": "LotusConditioning",
"pos": [ "pos": [
400, 400,
-150 -150
], ],
"size": [ "size": [
175, 180,
33.333333333333336 40
], ],
"flags": {}, "flags": {},
"order": 2, "order": 2,
@ -718,12 +708,11 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "LotusConditioning",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "LotusConditioning",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, }
"widgets_values": []
}, },
{ {
"id": 20, "id": 20,
@ -734,7 +723,7 @@
], ],
"size": [ "size": [
210, 210,
106 110
], ],
"flags": {}, "flags": {},
"order": 8, "order": 8,
@ -786,9 +775,9 @@
} }
], ],
"properties": { "properties": {
"Node name for S&R": "BasicScheduler",
"cnr_id": "comfy-core", "cnr_id": "comfy-core",
"ver": "0.3.34", "ver": "0.3.34",
"Node name for S&R": "BasicScheduler",
"widget_ue_connectable": {} "widget_ue_connectable": {}
}, },
"widgets_values": [ "widgets_values": [
@ -850,7 +839,7 @@
}, },
{ {
"id": 201, "id": 201,
"origin_id": 23, "origin_id": 74,
"origin_slot": 0, "origin_slot": 0,
"target_id": 16, "target_id": 16,
"target_slot": 4, "target_slot": 4,
@ -866,7 +855,7 @@
}, },
{ {
"id": 238, "id": 238,
"origin_id": 68, "origin_id": 75,
"origin_slot": 0, "origin_slot": 0,
"target_id": 19, "target_id": 19,
"target_slot": 1, "target_slot": 1,
@ -892,7 +881,7 @@
"id": 38, "id": 38,
"origin_id": 14, "origin_id": 14,
"origin_slot": 0, "origin_slot": 0,
"target_id": 23, "target_id": 74,
"target_slot": 1, "target_slot": 1,
"type": "VAE" "type": "VAE"
}, },
@ -908,7 +897,7 @@
"id": 37, "id": 37,
"origin_id": -10, "origin_id": -10,
"origin_slot": 0, "origin_slot": 0,
"target_id": 23, "target_id": 74,
"target_slot": 0, "target_slot": 0,
"type": "IMAGE" "type": "IMAGE"
}, },
@ -948,12 +937,11 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Image generation and editing/Depth to image", "category": "Conditioning & Preprocessors/Depth",
"description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model." "description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model."
} }
] ]
}, },
"config": {},
"extra": { "extra": {
"ds": { "ds": {
"scale": 1.3589709866044692, "scale": 1.3589709866044692,
@ -961,8 +949,6 @@
-138.53613935617864, -138.53613935617864,
-786.0629126022195 -786.0629126022195
] ]
}, }
"workflowRendererVersion": "LG" }
},
"version": 0.4
} }

1154
blueprints/Image Depth Estimation (MoGe).json Normal file
View File

File diff suppressed because it is too large Load Diff

779
blueprints/Image Face Detection (Mediapipe).json Normal file
View File

@ -0,0 +1,779 @@
{
"revision": 0,
"last_node_id": 33,
"last_link_id": 0,
"nodes": [
{
"id": 33,
"type": "6062babb-b649-4a71-be9e-20ebce567744",
"pos": [
-450,
4240
],
"size": [
420,
400
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "image",
"name": "image",
"type": "IMAGE",
"link": null
},
{
"name": "face_landmarker",
"type": "FACE_LANDMARKER",
"link": null
},
{
"name": "detector_variant",
"type": "COMBO",
"widget": {
"name": "detector_variant"
},
"link": null
},
{
"name": "num_faces",
"type": "INT",
"widget": {
"name": "num_faces"
},
"link": null
},
{
"label": "custom_face_oval",
"name": "regions.face_oval",
"type": "BOOLEAN",
"widget": {
"name": "regions.face_oval"
},
"link": null
},
{
"label": "custom_lips",
"name": "regions.lips",
"type": "BOOLEAN",
"widget": {
"name": "regions.lips"
},
"link": null
},
{
"label": "custom_left_eye",
"name": "regions.left_eye",
"type": "BOOLEAN",
"widget": {
"name": "regions.left_eye"
},
"link": null
},
{
"label": "custom_right_eye",
"name": "regions.right_eye",
"type": "BOOLEAN",
"widget": {
"name": "regions.right_eye"
},
"link": null
},
{
"label": "custom_irises",
"name": "regions.irises",
"type": "BOOLEAN",
"widget": {
"name": "regions.irises"
},
"link": null
},
{
"name": "model_name",
"type": "COMBO",
"widget": {
"name": "model_name"
},
"link": null
}
],
"outputs": [
{
"localized_name": "face_landmarks",
"name": "face_landmarks",
"type": "FACE_LANDMARKS",
"links": []
},
{
"localized_name": "bboxes",
"name": "bboxes",
"type": "BOUNDING_BOX",
"links": []
},
{
"label": "mask",
"name": "MASK_1",
"type": "MASK",
"links": []
}
],
"title": "Image Face Detection (Mediapipe)",
"properties": {
"proxyWidgets": [
[
"11",
"detector_variant"
],
[
"11",
"num_faces"
],
[
"20",
"regions.face_oval"
],
[
"20",
"regions.lips"
],
[
"20",
"regions.left_eye"
],
[
"20",
"regions.right_eye"
],
[
"20",
"regions.irises"
],
[
"2",
"model_name"
]
],
"cnr_id": "comfy-core",
"ver": "0.22.0",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": []
}
],
"links": [],
"version": 0.4,
"definitions": {
"subgraphs": [
{
"id": "6062babb-b649-4a71-be9e-20ebce567744",
"version": 1,
"state": {
"lastGroupId": 2,
"lastNodeId": 158,
"lastLinkId": 140,
"lastRerouteId": 0
},
"revision": 0,
"config": {},
"name": "Image Face Detection (Mediapipe)",
"description": "Detects facial landmarks from an image using MediaPipe, outputting landmark data, face bounding boxes, and an optional face-region mask.",
"inputNode": {
"id": -10,
"bounding": [
-710,
4300,
148.880859375,
248
]
},
"outputNode": {
"id": -20,
"bounding": [
140,
4480,
137.677734375,
108
]
},
"inputs": [
{
"id": "705dc1ae-6dc9-4155-92df-52f816ad451e",
"name": "image",
"type": "IMAGE",
"linkIds": [
60
],
"localized_name": "image",
"pos": [
-585.119140625,
4324
]
},
{
"id": "d6277190-732c-4604-b7cd-d3a9588bf761",
"name": "face_landmarker",
"type": "FACE_LANDMARKER",
"linkIds": [
74
],
"pos": [
-585.119140625,
4344
]
},
{
"id": "ac473a08-6a86-42a7-b460-e70c6c5e1e2b",
"name": "detector_variant",
"type": "COMBO",
"linkIds": [
75
],
"pos": [
-585.119140625,
4364
]
},
{
"id": "1bec2252-ca2d-496e-8a33-33a61d21f897",
"name": "num_faces",
"type": "INT",
"linkIds": [
76
],
"pos": [
-585.119140625,
4384
]
},
{
"id": "17994fa2-0ea0-4c9b-a70a-19789c459c80",
"name": "regions.face_oval",
"type": "BOOLEAN",
"linkIds": [
77
],
"label": "custom_face_oval",
"pos": [
-585.119140625,
4404
]
},
{
"id": "1c6c5893-2aee-4c37-b702-15ef2e20d863",
"name": "regions.lips",
"type": "BOOLEAN",
"linkIds": [
78
],
"label": "custom_lips",
"pos": [
-585.119140625,
4424
]
},
{
"id": "f353fcea-4b6f-42a1-8fdd-32b3aa1e1f09",
"name": "regions.left_eye",
"type": "BOOLEAN",
"linkIds": [
79
],
"label": "custom_left_eye",
"pos": [
-585.119140625,
4444
]
},
{
"id": "1387e121-c1fb-4522-8f0d-43459e11dd86",
"name": "regions.right_eye",
"type": "BOOLEAN",
"linkIds": [
80
],
"label": "custom_right_eye",
"pos": [
-585.119140625,
4464
]
},
{
"id": "14acb0a0-d1f4-48f3-ba31-811b26236ef9",
"name": "regions.irises",
"type": "BOOLEAN",
"linkIds": [
81
],
"label": "custom_irises",
"pos": [
-585.119140625,
4484
]
},
{
"id": "25a82859-87de-42c8-8431-09948665546e",
"name": "model_name",
"type": "COMBO",
"linkIds": [
86
],
"pos": [
-585.119140625,
4504
]
}
],
"outputs": [
{
"id": "d2ba3f92-e8b1-49c3-9590-cfad56c54cf4",
"name": "face_landmarks",
"type": "FACE_LANDMARKS",
"linkIds": [
44
],
"localized_name": "face_landmarks",
"pos": [
164,
4504
]
},
{
"id": "4f356bb0-d4c4-4f93-b4cf-0845a65c4e6d",
"name": "bboxes",
"type": "BOUNDING_BOX",
"linkIds": [
25
],
"localized_name": "bboxes",
"pos": [
164,
4524
]
},
{
"id": "f6309e1d-6397-4363-b38f-778a122abc51",
"name": "MASK_1",
"type": "MASK",
"linkIds": [
83
],
"label": "mask",
"pos": [
164,
4544
]
}
],
"widgets": [],
"nodes": [
{
"id": 11,
"type": "MediaPipeFaceLandmarker",
"pos": [
-280,
4280
],
"size": [
350,
220
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [
{
"localized_name": "face_detection_model",
"name": "face_detection_model",
"type": "FACE_DETECTION_MODEL",
"link": 66
},
{
"localized_name": "image",
"name": "image",
"type": "IMAGE",
"link": 60
},
{
"localized_name": "detector_variant",
"name": "detector_variant",
"type": "COMBO",
"widget": {
"name": "detector_variant"
},
"link": 75
},
{
"localized_name": "num_faces",
"name": "num_faces",
"type": "INT",
"widget": {
"name": "num_faces"
},
"link": 76
},
{
"localized_name": "min_confidence",
"name": "min_confidence",
"type": "FLOAT",
"widget": {
"name": "min_confidence"
},
"link": null
},
{
"localized_name": "missing_frame_fallback",
"name": "missing_frame_fallback",
"type": "COMBO",
"widget": {
"name": "missing_frame_fallback"
},
"link": null
},
{
"name": "face_landmarker",
"type": "FACE_LANDMARKER",
"link": 74
}
],
"outputs": [
{
"localized_name": "face_landmarks",
"name": "face_landmarks",
"type": "FACE_LANDMARKS",
"links": [
44,
46
]
},
{
"localized_name": "bboxes",
"name": "bboxes",
"type": "BOUNDING_BOX",
"links": [
25
]
}
],
"properties": {
"Node name for S&R": "MediaPipeFaceLandmarker",
"cnr_id": "comfy-core",
"ver": "0.22.0",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": [
"full",
0,
0.5,
"empty"
]
},
{
"id": 2,
"type": "LoadMediaPipeFaceLandmarker",
"pos": [
-290,
4060
],
"size": [
350,
140
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "model_name",
"name": "model_name",
"type": "COMBO",
"widget": {
"name": "model_name"
},
"link": 86
}
],
"outputs": [
{
"localized_name": "FACE_DETECTION_MODEL",
"name": "FACE_DETECTION_MODEL",
"type": "FACE_DETECTION_MODEL",
"links": [
66
]
}
],
"properties": {
"Node name for S&R": "LoadMediaPipeFaceLandmarker",
"cnr_id": "comfy-core",
"ver": "0.22.0",
"models": [
{
"name": "mediapipe_face_fp32.safetensors",
"url": "https://huggingface.co/Comfy-Org/mediapipe/resolve/main/detection/mediapipe_face_fp32.safetensors",
"directory": "detection"
}
],
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": [
"mediapipe_face_fp32.safetensors"
]
},
{
"id": 20,
"type": "MediaPipeFaceMask",
"pos": [
-290,
4560
],
"size": [
360,
180
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"localized_name": "face_landmarks",
"name": "face_landmarks",
"type": "FACE_LANDMARKS",
"link": 46
},
{
"localized_name": "regions",
"name": "regions",
"type": "COMFY_DYNAMICCOMBO_V3",
"widget": {
"name": "regions"
},
"link": null
},
{
"localized_name": "regions.face_oval",
"name": "regions.face_oval",
"type": "BOOLEAN",
"widget": {
"name": "regions.face_oval"
},
"link": 77
},
{
"localized_name": "regions.lips",
"name": "regions.lips",
"type": "BOOLEAN",
"widget": {
"name": "regions.lips"
},
"link": 78
},
{
"localized_name": "regions.left_eye",
"name": "regions.left_eye",
"type": "BOOLEAN",
"widget": {
"name": "regions.left_eye"
},
"link": 79
},
{
"localized_name": "regions.right_eye",
"name": "regions.right_eye",
"type": "BOOLEAN",
"widget": {
"name": "regions.right_eye"
},
"link": 80
},
{
"localized_name": "regions.irises",
"name": "regions.irises",
"type": "BOOLEAN",
"widget": {
"name": "regions.irises"
},
"link": 81
}
],
"outputs": [
{
"localized_name": "MASK",
"name": "MASK",
"type": "MASK",
"links": [
83
]
}
],
"properties": {
"Node name for S&R": "MediaPipeFaceMask",
"cnr_id": "comfy-core",
"ver": "0.22.0",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": [
"custom",
true,
false,
false,
false,
false
]
}
],
"groups": [],
"links": [
{
"id": 66,
"origin_id": 2,
"origin_slot": 0,
"target_id": 11,
"target_slot": 0,
"type": "FACE_DETECTION_MODEL"
},
{
"id": 46,
"origin_id": 11,
"origin_slot": 0,
"target_id": 20,
"target_slot": 0,
"type": "FACE_LANDMARKS"
},
{
"id": 60,
"origin_id": -10,
"origin_slot": 0,
"target_id": 11,
"target_slot": 1,
"type": "IMAGE"
},
{
"id": 44,
"origin_id": 11,
"origin_slot": 0,
"target_id": -20,
"target_slot": 0,
"type": "FACE_LANDMARKS"
},
{
"id": 25,
"origin_id": 11,
"origin_slot": 1,
"target_id": -20,
"target_slot": 1,
"type": "BOUNDING_BOX"
},
{
"id": 74,
"origin_id": -10,
"origin_slot": 1,
"target_id": 11,
"target_slot": 6,
"type": "FACE_LANDMARKER"
},
{
"id": 75,
"origin_id": -10,
"origin_slot": 2,
"target_id": 11,
"target_slot": 2,
"type": "COMBO"
},
{
"id": 76,
"origin_id": -10,
"origin_slot": 3,
"target_id": 11,
"target_slot": 3,
"type": "INT"
},
{
"id": 77,
"origin_id": -10,
"origin_slot": 4,
"target_id": 20,
"target_slot": 2,
"type": "BOOLEAN"
},
{
"id": 78,
"origin_id": -10,
"origin_slot": 5,
"target_id": 20,
"target_slot": 3,
"type": "BOOLEAN"
},
{
"id": 79,
"origin_id": -10,
"origin_slot": 6,
"target_id": 20,
"target_slot": 4,
"type": "BOOLEAN"
},
{
"id": 80,
"origin_id": -10,
"origin_slot": 7,
"target_id": 20,
"target_slot": 5,
"type": "BOOLEAN"
},
{
"id": 81,
"origin_id": -10,
"origin_slot": 8,
"target_id": 20,
"target_slot": 6,
"type": "BOOLEAN"
},
{
"id": 83,
"origin_id": 20,
"origin_slot": 0,
"target_id": -20,
"target_slot": 2,
"type": "MASK"
},
{
"id": 86,
"origin_id": -10,
"origin_slot": 9,
"target_id": 2,
"target_slot": 0,
"type": "COMBO"
}
],
"extra": {},
"category": "Conditioning & Preprocessors/Face Detection"
}
]
},
"extra": {}
}

2
blueprints/Image Segmentation (SAM3).json
View File

@ -703,7 +703,7 @@
} }
], ],
"extra": {}, "extra": {},
"category": "Image Tools/Image Segmentation", "category": "Conditioning & Preprocessors/Segmentation & Mask",
"description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes." "description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes."
} }
] ]

2
blueprints/Image Upscale(Z-image-Turbo).json
View File

@ -1302,7 +1302,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Image generation and editing/Enhance", "category": "Image generation and editing/Upscale",
"description": "Upscales images to higher resolution using Z-Image-Turbo." "description": "Upscales images to higher resolution using Z-Image-Turbo."
} }
] ]

1206
blueprints/Image to Pose Map (SDPose Multi-Person).json Normal file
View File

File diff suppressed because it is too large Load Diff

888
blueprints/Image to Pose Map (SDPose-OOD).json Normal file
View File

@ -0,0 +1,888 @@
{
"revision": 0,
"last_node_id": 675,
"last_link_id": 0,
"nodes": [
{
"id": 675,
"type": "01b6a731-fb78-4070-9a38-c87146da9604",
"pos": [
-2480,
3400
],
"size": [
360,
433.3125
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"localized_name": "input",
"name": "input",
"type": "IMAGE,MASK",
"link": null
},
{
"label": "resize_target_longer_size",
"name": "resize_type.longer_size",
"type": "INT",
"widget": {
"name": "resize_type.longer_size"
},
"link": null
},
{
"name": "scale_method",
"type": "COMBO",
"widget": {
"name": "scale_method"
},
"link": null
},
{
"name": "draw_body",
"type": "BOOLEAN",
"widget": {
"name": "draw_body"
},
"link": null
},
{
"name": "draw_hands",
"type": "BOOLEAN",
"widget": {
"name": "draw_hands"
},
"link": null
},
{
"name": "draw_face",
"type": "BOOLEAN",
"widget": {
"name": "draw_face"
},
"link": null
},
{
"name": "draw_feet",
"type": "BOOLEAN",
"widget": {
"name": "draw_feet"
},
"link": null
},
{
"name": "stick_width",
"type": "INT",
"widget": {
"name": "stick_width"
},
"link": null
},
{
"name": "face_point_size",
"type": "INT",
"widget": {
"name": "face_point_size"
},
"link": null
},
{
"name": "score_threshold",
"type": "FLOAT",
"widget": {
"name": "score_threshold"
},
"link": null
},
{
"name": "ckpt_name",
"type": "COMBO",
"widget": {
"name": "ckpt_name"
},
"link": null
},
{
"name": "bboxes",
"shape": 7,
"type": "BOUNDING_BOX",
"link": null
}
],
"outputs": [
{
"localized_name": "IMAGE",
"name": "IMAGE",
"type": "IMAGE",
"links": []
},
{
"name": "keypoints",
"type": "POSE_KEYPOINT",
"links": null
}
],
"properties": {
"proxyWidgets": [
[
"674",
"resize_type.longer_size"
],
[
"674",
"scale_method"
],
[
"672",
"draw_body"
],
[
"672",
"draw_hands"
],
[
"672",
"draw_face"
],
[
"672",
"draw_feet"
],
[
"672",
"stick_width"
],
[
"672",
"face_point_size"
],
[
"672",
"score_threshold"
],
[
"673",
"ckpt_name"
]
],
"cnr_id": "comfy-core",
"ver": "0.15.1",
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [],
"title": "Image to Pose Map (SDPose-OOD)"
}
],
"links": [],
"version": 0.4,
"definitions": {
"subgraphs": [
{
"id": "01b6a731-fb78-4070-9a38-c87146da9604",
"version": 1,
"state": {
"lastGroupId": 0,
"lastNodeId": 676,
"lastLinkId": 1715,
"lastRerouteId": 0
},
"revision": 0,
"config": {},
"name": "Image to Pose Map (SDPose-OOD)",
"inputNode": {
"id": -10,
"bounding": [
-3290,
3590,
190.8984375,
288
]
},
"outputNode": {
"id": -20,
"bounding": [
-1756.2451602089645,
3366,
128,
88
]
},
"inputs": [
{
"id": "e24699c3-1356-4634-9eb4-19bb58e5c0b0",
"name": "input",
"type": "IMAGE,MASK",
"linkIds": [
1700
],
"localized_name": "input",
"pos": [
-3123.1015625,
3614
]
},
{
"id": "088eefc1-cd8a-4573-993f-9e4da008a12d",
"name": "resize_type.longer_size",
"type": "INT",
"linkIds": [
1704
],
"label": "resize_target_longer_size",
"pos": [
-3123.1015625,
3634
]
},
{
"id": "b6449bd3-73d4-41c8-b81f-cf8d33f76a2e",
"name": "scale_method",
"type": "COMBO",
"linkIds": [
1705
],
"pos": [
-3123.1015625,
3654
]
},
{
"id": "4cff52ad-ed07-4c97-8803-fcbd89554fd0",
"name": "draw_body",
"type": "BOOLEAN",
"linkIds": [
1706
],
"pos": [
-3123.1015625,
3674
]
},
{
"id": "7af63dce-f7df-4d7e-8215-d7c7f60bf81c",
"name": "draw_hands",
"type": "BOOLEAN",
"linkIds": [
1707
],
"pos": [
-3123.1015625,
3694
]
},
{
"id": "af3a9bce-61f9-4aca-b530-9f65e028b35e",
"name": "draw_face",
"type": "BOOLEAN",
"linkIds": [
1708
],
"pos": [
-3123.1015625,
3714
]
},
{
"id": "4620f6a3-2c85-4b79-ad8f-35d0326b568f",
"name": "draw_feet",
"type": "BOOLEAN",
"linkIds": [
1709
],
"pos": [
-3123.1015625,
3734
]
},
{
"id": "fee5d0c9-8d4b-4934-81d8-ba2206dc56cb",
"name": "stick_width",
"type": "INT",
"linkIds": [
1710
],
"pos": [
-3123.1015625,
3754
]
},
{
"id": "aafdd060-ba81-4324-a9cc-b656e1ebc133",
"name": "face_point_size",
"type": "INT",
"linkIds": [
1711
],
"pos": [
-3123.1015625,
3774
]
},
{
"id": "514c5503-f9e6-4d23-b1ae-1d3291acb2a3",
"name": "score_threshold",
"type": "FLOAT",
"linkIds": [
1712
],
"pos": [
-3123.1015625,
3794
]
},
{
"id": "ae46de61-2cc6-483e-8ee9-87e4144a2ffa",
"name": "ckpt_name",
"type": "COMBO",
"linkIds": [
1713
],
"pos": [
-3123.1015625,
3814
]
},
{
"id": "41bec0c6-dffa-4c78-9289-ee678715ae54",
"name": "bboxes",
"type": "BOUNDING_BOX",
"linkIds": [
1714
],
"pos": [
-3123.1015625,
3834
]
}
],
"outputs": [
{
"id": "f05ed8cc-9403-4f14-8085-4364b06f8a48",
"name": "IMAGE",
"type": "IMAGE",
"linkIds": [
1701
],
"localized_name": "IMAGE",
"pos": [
-1732.2451602089645,
3390
]
},
{
"id": "29a6584e-4685-4986-8ffd-e6d8539953fd",
"name": "keypoints",
"type": "POSE_KEYPOINT",
"linkIds": [
1715
],
"pos": [
-1732.2451602089645,
3410
]
}
],
"widgets": [],
"nodes": [
{
"id": 671,
"type": "SDPoseKeypointExtractor",
"pos": [
-2470,
3250
],
"size": [
270,
180
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "model",
"name": "model",
"type": "MODEL",
"link": 1696
},
{
"localized_name": "vae",
"name": "vae",
"type": "VAE",
"link": 1697
},
{
"localized_name": "image",
"name": "image",
"type": "IMAGE",
"link": 1698
},
{
"localized_name": "bboxes",
"name": "bboxes",
"shape": 7,
"type": "BOUNDING_BOX",
"link": 1714
},
{
"localized_name": "batch_size",
"name": "batch_size",
"type": "INT",
"widget": {
"name": "batch_size"
},
"link": null
}
],
"outputs": [
{
"localized_name": "keypoints",
"name": "keypoints",
"type": "POSE_KEYPOINT",
"links": [
1699,
1715
]
}
],
"properties": {
"Node name for S&R": "SDPoseKeypointExtractor",
"cnr_id": "comfy-core",
"ver": "0.15.0",
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
16
]
},
{
"id": 674,
"type": "ResizeImageMaskNode",
"pos": [
-2960,
3490
],
"size": [
270,
110
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [
{
"localized_name": "input",
"name": "input",
"type": "IMAGE,MASK",
"link": 1700
},
{
"localized_name": "resize_type",
"name": "resize_type",
"type": "COMFY_DYNAMICCOMBO_V3",
"widget": {
"name": "resize_type"
},
"link": null
},
{
"localized_name": "resize_type.longer_size",
"name": "resize_type.longer_size",
"type": "INT",
"widget": {
"name": "resize_type.longer_size"
},
"link": 1704
},
{
"localized_name": "scale_method",
"name": "scale_method",
"type": "COMBO",
"widget": {
"name": "scale_method"
},
"link": 1705
}
],
"outputs": [
{
"localized_name": "resized",
"name": "resized",
"type": "*",
"links": [
1698
]
}
],
"properties": {
"Node name for S&R": "ResizeImageMaskNode",
"cnr_id": "comfy-core",
"ver": "0.15.0",
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"scale longer dimension",
1024,
"area"
]
},
{
"id": 672,
"type": "SDPoseDrawKeypoints",
"pos": [
-2120,
3260
],
"size": [
270,
280
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [
{
"localized_name": "keypoints",
"name": "keypoints",
"type": "POSE_KEYPOINT",
"link": 1699
},
{
"localized_name": "draw_body",
"name": "draw_body",
"type": "BOOLEAN",
"widget": {
"name": "draw_body"
},
"link": 1706
},
{
"localized_name": "draw_hands",
"name": "draw_hands",
"type": "BOOLEAN",
"widget": {
"name": "draw_hands"
},
"link": 1707
},
{
"localized_name": "draw_face",
"name": "draw_face",
"type": "BOOLEAN",
"widget": {
"name": "draw_face"
},
"link": 1708
},
{
"localized_name": "draw_feet",
"name": "draw_feet",
"type": "BOOLEAN",
"widget": {
"name": "draw_feet"
},
"link": 1709
},
{
"localized_name": "stick_width",
"name": "stick_width",
"type": "INT",
"widget": {
"name": "stick_width"
},
"link": 1710
},
{
"localized_name": "face_point_size",
"name": "face_point_size",
"type": "INT",
"widget": {
"name": "face_point_size"
},
"link": 1711
},
{
"localized_name": "score_threshold",
"name": "score_threshold",
"type": "FLOAT",
"widget": {
"name": "score_threshold"
},
"link": 1712
}
],
"outputs": [
{
"localized_name": "IMAGE",
"name": "IMAGE",
"type": "IMAGE",
"links": [
1701
]
}
],
"properties": {
"Node name for S&R": "SDPoseDrawKeypoints",
"cnr_id": "comfy-core",
"ver": "0.15.0",
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
true,
true,
true,
true,
4,
2,
0.5
]
},
{
"id": 673,
"type": "CheckpointLoaderSimple",
"pos": [
-2960,
3250
],
"size": [
390,
190
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"localized_name": "ckpt_name",
"name": "ckpt_name",
"type": "COMBO",
"widget": {
"name": "ckpt_name"
},
"link": 1713
}
],
"outputs": [
{
"localized_name": "MODEL",
"name": "MODEL",
"type": "MODEL",
"links": [
1696
]
},
{
"localized_name": "CLIP",
"name": "CLIP",
"type": "CLIP",
"links": []
},
{
"localized_name": "VAE",
"name": "VAE",
"type": "VAE",
"links": [
1697
]
}
],
"properties": {
"Node name for S&R": "CheckpointLoaderSimple",
"cnr_id": "comfy-core",
"ver": "0.15.0",
"models": [
{
"name": "sdpose_wholebody_fp16.safetensors",
"url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors",
"directory": "checkpoints"
}
],
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"sdpose_wholebody_fp16.safetensors"
]
}
],
"groups": [],
"links": [
{
"id": 1696,
"origin_id": 673,
"origin_slot": 0,
"target_id": 671,
"target_slot": 0,
"type": "MODEL"
},
{
"id": 1697,
"origin_id": 673,
"origin_slot": 2,
"target_id": 671,
"target_slot": 1,
"type": "VAE"
},
{
"id": 1698,
"origin_id": 674,
"origin_slot": 0,
"target_id": 671,
"target_slot": 2,
"type": "IMAGE"
},
{
"id": 1699,
"origin_id": 671,
"origin_slot": 0,
"target_id": 672,
"target_slot": 0,
"type": "POSE_KEYPOINT"
},
{
"id": 1700,
"origin_id": -10,
"origin_slot": 0,
"target_id": 674,
"target_slot": 0,
"type": "IMAGE,MASK"
},
{
"id": 1701,
"origin_id": 672,
"origin_slot": 0,
"target_id": -20,
"target_slot": 0,
"type": "IMAGE"
},
{
"id": 1704,
"origin_id": -10,
"origin_slot": 1,
"target_id": 674,
"target_slot": 2,
"type": "INT"
},
{
"id": 1705,
"origin_id": -10,
"origin_slot": 2,
"target_id": 674,
"target_slot": 3,
"type": "COMBO"
},
{
"id": 1706,
"origin_id": -10,
"origin_slot": 3,
"target_id": 672,
"target_slot": 1,
"type": "BOOLEAN"
},
{
"id": 1707,
"origin_id": -10,
"origin_slot": 4,
"target_id": 672,
"target_slot": 2,
"type": "BOOLEAN"
},
{
"id": 1708,
"origin_id": -10,
"origin_slot": 5,
"target_id": 672,
"target_slot": 3,
"type": "BOOLEAN"
},
{
"id": 1709,
"origin_id": -10,
"origin_slot": 6,
"target_id": 672,
"target_slot": 4,
"type": "BOOLEAN"
},
{
"id": 1710,
"origin_id": -10,
"origin_slot": 7,
"target_id": 672,
"target_slot": 5,
"type": "INT"
},
{
"id": 1711,
"origin_id": -10,
"origin_slot": 8,
"target_id": 672,
"target_slot": 6,
"type": "INT"
},
{
"id": 1712,
"origin_id": -10,
"origin_slot": 9,
"target_id": 672,
"target_slot": 7,
"type": "FLOAT"
},
{
"id": 1713,
"origin_id": -10,
"origin_slot": 10,
"target_id": 673,
"target_slot": 0,
"type": "COMBO"
},
{
"id": 1714,
"origin_id": -10,
"origin_slot": 11,
"target_id": 671,
"target_slot": 3,
"type": "BOUNDING_BOX"
},
{
"id": 1715,
"origin_id": 671,
"origin_slot": 0,
"target_id": -20,
"target_slot": 1,
"type": "POSE_KEYPOINT"
}
],
"extra": {
"workflowRendererVersion": "LG"
},
"category": "Conditioning & Preprocessors/Pose",
"description": "Extracts human pose keypoints and stick-figure visuals from an image using SDPose-OOD, with optional bounding-box input per subject."
}
]
},
"extra": {
"ue_links": []
}
}

1219
blueprints/Merge Videos.json Normal file
View File

File diff suppressed because it is too large Load Diff

2
blueprints/Pose to Image (Z-Image-Turbo).json
View File

@ -1298,7 +1298,7 @@
"VHS_MetadataImage": true, "VHS_MetadataImage": true,
"VHS_KeepIntermediate": true "VHS_KeepIntermediate": true
}, },
"category": "Image generation and editing/Pose to image", "category": "Image generation and editing/Conditioned",
"description": "Generates an image from pose keypoints using Z-Image-Turbo with text conditioning." "description": "Generates an image from pose keypoints using Z-Image-Turbo with text conditioning."
} }
] ]

2
blueprints/Pose to Video (LTX 2.0).json
View File

@ -3870,7 +3870,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Video generation and editing/Pose to video", "category": "Video generation and editing/Conditioned",
"description": "Generates video from pose reference frames using LTX-2, with optional synchronized audio." "description": "Generates video from pose reference frames using LTX-2, with optional synchronized audio."
} }
] ]

2
blueprints/Prompt Enhance.json
View File

@ -270,7 +270,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Text generation/Prompt enhance", "category": "Text Tools",
"description": "Expands short text prompts into detailed descriptions using a text generation model for better generation quality." "description": "Expands short text prompts into detailed descriptions using a text generation model for better generation quality."
} }
] ]

2
blueprints/Remove Background (BiRefNet).json
View File

@ -389,7 +389,7 @@
} }
], ],
"extra": {}, "extra": {},
"category": "Image generation and editing/Background Removal" "category": "Image Tools/Background Removal"
} }
] ]
}, },

485
blueprints/Select Per-Line Text by Index.json Normal file
View File

@ -0,0 +1,485 @@
{
"revision": 0,
"last_node_id": 10,
"last_link_id": 0,
"nodes": [
{
"id": 10,
"type": "3fb7557a-470d-4983-9d8c-6d5caa9788f0",
"pos": [
-250,
8590
],
"size": [
280,
360
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "text_per_line",
"name": "text_per_line",
"type": "STRING",
"widget": {
"name": "text_per_line"
},
"link": null
},
{
"localized_name": "index",
"name": "index",
"type": "INT",
"widget": {
"name": "index"
},
"link": null
}
],
"outputs": [
{
"localized_name": "selected_line",
"name": "selected_line",
"type": "STRING",
"links": []
}
],
"properties": {
"proxyWidgets": [
[
"2",
"string"
],
[
"3",
"value"
]
],
"cnr_id": "comfy-core",
"ver": "0.19.0",
"ue_properties": {
"widget_ue_connectable": {},
"input_ue_unconnectable": {}
}
},
"widgets_values": [],
"title": "Select Per-Line Text by Index"
}
],
"links": [],
"version": 0.4,
"definitions": {
"subgraphs": [
{
"id": "3fb7557a-470d-4983-9d8c-6d5caa9788f0",
"version": 1,
"state": {
"lastGroupId": 0,
"lastNodeId": 10,
"lastLinkId": 14,
"lastRerouteId": 0
},
"revision": 0,
"config": {},
"name": "Select Per-Line Text by Index",
"inputNode": {
"id": -10,
"bounding": [
-990,
8595,
128,
88
]
},
"outputNode": {
"id": -20,
"bounding": [
710,
8585,
128,
68
]
},
"inputs": [
{
"id": "75417d82-a934-4ac9-b667-d8dcd5a3bfb3",
"name": "text_per_line",
"type": "STRING",
"linkIds": [
13
],
"localized_name": "text_per_line",
"pos": [
-886,
8619
]
},
{
"id": "46e69a73-1804-4ca6-9175-31445bf0be96",
"name": "index",
"type": "INT",
"linkIds": [
14
],
"localized_name": "index",
"pos": [
-886,
8639
]
}
],
"outputs": [
{
"id": "e34e8ad1-84d2-4bd2-a460-eb7de6067c10",
"name": "selected_line",
"type": "STRING",
"linkIds": [
10
],
"localized_name": "selected_line",
"pos": [
734,
8609
]
}
],
"widgets": [],
"nodes": [
{
"id": 1,
"type": "PreviewAny",
"pos": [
-500,
8400
],
"size": [
230,
180
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "source",
"name": "source",
"type": "*",
"link": 1
}
],
"outputs": [
{
"localized_name": "STRING",
"name": "STRING",
"type": "STRING",
"links": [
6
]
}
],
"properties": {
"Node name for S&R": "PreviewAny",
"cnr_id": "comfy-core",
"ver": "0.19.0",
"ue_properties": {
"widget_ue_connectable": {},
"input_ue_unconnectable": {}
}
},
"widgets_values": [
null,
null,
null
]
},
{
"id": 2,
"type": "RegexExtract",
"pos": [
-240,
8740
],
"size": [
470,
460
],
"flags": {},
"order": 1,
"mode": 0,
"showAdvanced": false,
"inputs": [
{
"localized_name": "string",
"name": "string",
"type": "STRING",
"widget": {
"name": "string"
},
"link": 13
},
{
"localized_name": "regex_pattern",
"name": "regex_pattern",
"type": "STRING",
"widget": {
"name": "regex_pattern"
},
"link": 9
},
{
"localized_name": "mode",
"name": "mode",
"type": "COMBO",
"widget": {
"name": "mode"
},
"link": null
},
{
"localized_name": "case_insensitive",
"name": "case_insensitive",
"type": "BOOLEAN",
"widget": {
"name": "case_insensitive"
},
"link": null
},
{
"localized_name": "multiline",
"name": "multiline",
"type": "BOOLEAN",
"widget": {
"name": "multiline"
},
"link": null
},
{
"localized_name": "dotall",
"name": "dotall",
"type": "BOOLEAN",
"widget": {
"name": "dotall"
},
"link": null
},
{
"localized_name": "group_index",
"name": "group_index",
"type": "INT",
"widget": {
"name": "group_index"
},
"link": null
}
],
"outputs": [
{
"localized_name": "STRING",
"name": "STRING",
"type": "STRING",
"links": [
10
]
}
],
"properties": {
"Node name for S&R": "RegexExtract",
"cnr_id": "comfy-core",
"ver": "0.19.0",
"ue_properties": {
"widget_ue_connectable": {},
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"",
"",
"First Group",
false,
false,
false,
1
]
},
{
"id": 3,
"type": "PrimitiveInt",
"pos": [
-810,
8400
],
"size": [
270,
110
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"localized_name": "value",
"name": "value",
"type": "INT",
"widget": {
"name": "value"
},
"link": 14
}
],
"outputs": [
{
"localized_name": "INT",
"name": "INT",
"type": "INT",
"links": [
1
]
}
],
"title": "Int (line index)",
"properties": {
"Node name for S&R": "Int (line index)",
"cnr_id": "comfy-core",
"ver": "0.19.0",
"ue_properties": {
"widget_ue_connectable": {},
"input_ue_unconnectable": {}
}
},
"widgets_values": [
0,
"fixed"
]
},
{
"id": 8,
"type": "StringReplace",
"pos": [
-240,
8400
],
"size": [
400,
280
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [
{
"localized_name": "string",
"name": "string",
"type": "STRING",
"widget": {
"name": "string"
},
"link": null
},
{
"localized_name": "find",
"name": "find",
"type": "STRING",
"widget": {
"name": "find"
},
"link": null
},
{
"localized_name": "replace",
"name": "replace",
"type": "STRING",
"widget": {
"name": "replace"
},
"link": 6
}
],
"outputs": [
{
"localized_name": "STRING",
"name": "STRING",
"type": "STRING",
"links": [
9
]
}
],
"properties": {
"Node name for S&R": "StringReplace",
"cnr_id": "comfy-core",
"ver": "0.19.0",
"ue_properties": {
"widget_ue_connectable": {},
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"^(?:[^\\n]*\\n){index}([^\\n]*)(?:\\n|$)",
"index",
""
]
}
],
"groups": [],
"links": [
{
"id": 1,
"origin_id": 3,
"origin_slot": 0,
"target_id": 1,
"target_slot": 0,
"type": "INT"
},
{
"id": 9,
"origin_id": 8,
"origin_slot": 0,
"target_id": 2,
"target_slot": 1,
"type": "STRING"
},
{
"id": 6,
"origin_id": 1,
"origin_slot": 0,
"target_id": 8,
"target_slot": 2,
"type": "STRING"
},
{
"id": 10,
"origin_id": 2,
"origin_slot": 0,
"target_id": -20,
"target_slot": 0,
"type": "STRING"
},
{
"id": 13,
"origin_id": -10,
"origin_slot": 0,
"target_id": 2,
"target_slot": 0,
"type": "STRING"
},
{
"id": 14,
"origin_id": -10,
"origin_slot": 1,
"target_id": 3,
"target_slot": 0,
"type": "INT"
}
],
"extra": {},
"category": "Text Tools",
"description": "Selects one line from multiline text by zero-based index for batch or list-driven prompt workflows."
}
]
},
"extra": {
"ue_links": [],
"links_added_by_ue": []
}
}

714
blueprints/Split Image Grid to Tiles.json Normal file
View File

@ -0,0 +1,714 @@
{
"revision": 0,
"last_node_id": 251,
"last_link_id": 0,
"nodes": [
{
"id": 251,
"type": "609e1fd1-b731-4b78-89ac-d19b1156b025",
"pos": [
-1490,
130
],
"size": [
230,
164
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [
{
"localized_name": "source_image",
"name": "source_image",
"type": "IMAGE",
"link": null
},
{
"localized_name": "columns",
"name": "columns",
"type": "INT",
"widget": {
"name": "columns"
},
"link": null
},
{
"localized_name": "rows",
"name": "rows",
"type": "INT",
"widget": {
"name": "rows"
},
"link": null
}
],
"outputs": [
{
"localized_name": "tiles",
"name": "tiles",
"type": "IMAGE",
"links": []
}
],
"properties": {
"proxyWidgets": [
[
"228",
"value"
],
[
"252",
"value"
]
],
"cnr_id": "comfy-core",
"ver": "0.20.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": [],
"title": "Split Image Grid to Tiles"
}
],
"links": [],
"version": 0.4,
"definitions": {
"subgraphs": [
{
"id": "609e1fd1-b731-4b78-89ac-d19b1156b025",
"version": 1,
"state": {
"lastGroupId": 9,
"lastNodeId": 252,
"lastLinkId": 429,
"lastRerouteId": 0
},
"revision": 0,
"config": {},
"name": "Split Image Grid to Tiles",
"inputNode": {
"id": -10,
"bounding": [
-1690,
260,
128,
108
]
},
"outputNode": {
"id": -20,
"bounding": [
-510,
590,
128,
68
]
},
"inputs": [
{
"id": "866ac798-cfbc-450a-b755-e704f86404d9",
"name": "source_image",
"type": "IMAGE",
"linkIds": [
386,
389
],
"localized_name": "source_image",
"pos": [
-1586,
284
]
},
{
"id": "bc37b1f8-8ab2-4f19-bd00-75d4fbc4feb3",
"name": "columns",
"type": "INT",
"linkIds": [
427
],
"localized_name": "columns",
"pos": [
-1586,
304
]
},
{
"id": "d45915da-e848-43dd-9ccc-e3161e9c99d9",
"name": "rows",
"type": "INT",
"linkIds": [
428
],
"localized_name": "rows",
"pos": [
-1586,
324
]
}
],
"outputs": [
{
"id": "18bc780f-064b-4038-87c6-67dba71deb08",
"name": "tiles",
"type": "IMAGE",
"linkIds": [
394
],
"localized_name": "tiles",
"shape": 6,
"pos": [
-486,
614
]
}
],
"widgets": [],
"nodes": [
{
"id": 225,
"type": "SplitImageToTileList",
"pos": [
-1010,
620
],
"size": [
290,
170
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [
{
"localized_name": "image",
"name": "image",
"type": "IMAGE",
"link": 386
},
{
"localized_name": "tile_width",
"name": "tile_width",
"type": "INT",
"widget": {
"name": "tile_width"
},
"link": 403
},
{
"localized_name": "tile_height",
"name": "tile_height",
"type": "INT",
"widget": {
"name": "tile_height"
},
"link": 404
},
{
"localized_name": "overlap",
"name": "overlap",
"type": "INT",
"widget": {
"name": "overlap"
},
"link": null
}
],
"outputs": [
{
"localized_name": "IMAGE",
"name": "IMAGE",
"shape": 6,
"type": "IMAGE",
"links": [
394
]
}
],
"properties": {
"Node name for S&R": "SplitImageToTileList",
"cnr_id": "comfy-core",
"ver": "0.20.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65
},
"widgets_values": [
1024,
1024,
0
]
},
{
"id": 231,
"type": "ComfyMathExpression",
"pos": [
-1080,
330
],
"size": [
370,
190
],
"flags": {},
"order": 4,
"mode": 0,
"inputs": [
{
"label": "a",
"localized_name": "values.a",
"name": "values.a",
"type": "FLOAT,INT,BOOLEAN",
"link": 390
},
{
"label": "b",
"localized_name": "values.b",
"name": "values.b",
"shape": 7,
"type": "FLOAT,INT,BOOLEAN",
"link": 429
},
{
"label": "c",
"localized_name": "values.c",
"name": "values.c",
"shape": 7,
"type": "FLOAT,INT,BOOLEAN",
"link": null
},
{
"localized_name": "expression",
"name": "expression",
"type": "STRING",
"widget": {
"name": "expression"
},
"link": null
}
],
"outputs": [
{
"localized_name": "FLOAT",
"name": "FLOAT",
"type": "FLOAT",
"links": null
},
{
"localized_name": "INT",
"name": "INT",
"type": "INT",
"links": [
404
]
},
{
"localized_name": "BOOL",
"name": "BOOL",
"type": "BOOLEAN",
"links": null
}
],
"title": "Math Expression （Height）",
"properties": {
"Node name for S&R": "ComfyMathExpression",
"cnr_id": "comfy-core",
"ver": "0.18.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65,
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"max(1, (int(a) + int(b) - 1) // int(b))"
]
},
{
"id": 229,
"type": "ComfyMathExpression",
"pos": [
-1090,
-30
],
"size": [
370,
190
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"label": "a",
"localized_name": "values.a",
"name": "values.a",
"type": "FLOAT,INT,BOOLEAN",
"link": 387
},
{
"label": "b",
"localized_name": "values.b",
"name": "values.b",
"shape": 7,
"type": "FLOAT,INT,BOOLEAN",
"link": 388
},
{
"label": "c",
"localized_name": "values.c",
"name": "values.c",
"shape": 7,
"type": "FLOAT,INT,BOOLEAN",
"link": null
},
{
"localized_name": "expression",
"name": "expression",
"type": "STRING",
"widget": {
"name": "expression"
},
"link": null
}
],
"outputs": [
{
"localized_name": "FLOAT",
"name": "FLOAT",
"type": "FLOAT",
"links": null
},
{
"localized_name": "INT",
"name": "INT",
"type": "INT",
"links": [
403
]
},
{
"localized_name": "BOOL",
"name": "BOOL",
"type": "BOOLEAN",
"links": null
}
],
"title": "Math Expression （Width）",
"properties": {
"Node name for S&R": "ComfyMathExpression",
"cnr_id": "comfy-core",
"ver": "0.18.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65,
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
"max(1, (int(a) + int(b) - 1) // int(b))"
]
},
{
"id": 228,
"type": "PrimitiveInt",
"pos": [
-1380,
90
],
"size": [
230,
110
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [
{
"localized_name": "value",
"name": "value",
"type": "INT",
"widget": {
"name": "value"
},
"link": 427
}
],
"outputs": [
{
"localized_name": "INT",
"name": "INT",
"type": "INT",
"links": [
388
]
}
],
"title": "Int (grid columns)",
"properties": {
"Node name for S&R": "Int (grid columns)",
"cnr_id": "comfy-core",
"ver": "0.18.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65,
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
2,
"fixed"
]
},
{
"id": 230,
"type": "GetImageSize",
"pos": [
-1380,
290
],
"size": [
230,
100
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [
{
"localized_name": "image",
"name": "image",
"type": "IMAGE",
"link": 389
}
],
"outputs": [
{
"localized_name": "width",
"name": "width",
"type": "INT",
"links": [
387
]
},
{
"localized_name": "height",
"name": "height",
"type": "INT",
"links": [
390
]
},
{
"localized_name": "batch_size",
"name": "batch_size",
"type": "INT",
"links": null
}
],
"properties": {
"Node name for S&R": "GetImageSize",
"cnr_id": "comfy-core",
"ver": "0.18.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65,
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
}
},
{
"id": 252,
"type": "PrimitiveInt",
"pos": [
-1380,
470
],
"size": [
230,
110
],
"flags": {},
"order": 5,
"mode": 0,
"inputs": [
{
"localized_name": "value",
"name": "value",
"type": "INT",
"widget": {
"name": "value"
},
"link": 428
}
],
"outputs": [
{
"localized_name": "INT",
"name": "INT",
"type": "INT",
"links": [
429
]
}
],
"title": "Int (grid rows)",
"properties": {
"Node name for S&R": "Int (grid rows)",
"cnr_id": "comfy-core",
"ver": "0.18.1",
"enableTabs": false,
"tabWidth": 65,
"tabXOffset": 10,
"hasSecondTab": false,
"secondTabText": "Send Back",
"secondTabOffset": 80,
"secondTabWidth": 65,
"ue_properties": {
"widget_ue_connectable": {},
"version": "7.7",
"input_ue_unconnectable": {}
}
},
"widgets_values": [
3,
"fixed"
]
}
],
"groups": [],
"links": [
{
"id": 403,
"origin_id": 229,
"origin_slot": 1,
"target_id": 225,
"target_slot": 1,
"type": "INT"
},
{
"id": 404,
"origin_id": 231,
"origin_slot": 1,
"target_id": 225,
"target_slot": 2,
"type": "INT"
},
{
"id": 390,
"origin_id": 230,
"origin_slot": 1,
"target_id": 231,
"target_slot": 0,
"type": "INT"
},
{
"id": 387,
"origin_id": 230,
"origin_slot": 0,
"target_id": 229,
"target_slot": 0,
"type": "INT"
},
{
"id": 388,
"origin_id": 228,
"origin_slot": 0,
"target_id": 229,
"target_slot": 1,
"type": "INT"
},
{
"id": 386,
"origin_id": -10,
"origin_slot": 0,
"target_id": 225,
"target_slot": 0,
"type": "IMAGE"
},
{
"id": 389,
"origin_id": -10,
"origin_slot": 0,
"target_id": 230,
"target_slot": 0,
"type": "IMAGE"
},
{
"id": 394,
"origin_id": 225,
"origin_slot": 0,
"target_id": -20,
"target_slot": 0,
"type": "IMAGE"
},
{
"id": 427,
"origin_id": -10,
"origin_slot": 1,
"target_id": 228,
"target_slot": 0,
"type": "INT"
},
{
"id": 428,
"origin_id": -10,
"origin_slot": 2,
"target_id": 252,
"target_slot": 0,
"type": "INT"
},
{
"id": 429,
"origin_id": 252,
"origin_slot": 0,
"target_id": 231,
"target_slot": 1,
"type": "INT"
}
],
"extra": {},
"category": "Image Tools/Crop",
"description": "Splits an image into a configurable columns×rows grid of equal tiles for tiled generation or processing."
}
]
},
"extra": {}
}

1085
blueprints/Text to Image (Anima).json Normal file
View File

File diff suppressed because it is too large Load Diff

2
blueprints/Video Captioning (Gemini).json
View File

@ -307,7 +307,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Text generation/Video Captioning", "category": "Video Tools",
"description": "Generates descriptive captions for video input using Google's Gemini multimodal LLM." "description": "Generates descriptive captions for video input using Google's Gemini multimodal LLM."
} }
] ]

1226
blueprints/Video Depth Estimation (MoGe).json Normal file
View File

File diff suppressed because it is too large Load Diff

1109
blueprints/Video Face Detection (Mediapipe).json Normal file
View File

File diff suppressed because it is too large Load Diff

4340
blueprints/Video Inpaint (VOID).json Normal file
View File

File diff suppressed because it is too large Load Diff

2388
blueprints/Video Inpaint(Wan2.1 VACE).json
View File

File diff suppressed because it is too large Load Diff

4196
blueprints/Video Inpainting (Wan2.1 VACE).json Normal file
View File

File diff suppressed because it is too large Load Diff

2
blueprints/Video Segmentation (SAM3).json
View File

@ -818,7 +818,7 @@
} }
], ],
"extra": {}, "extra": {},
"category": "Video Tools", "category": "Conditioning & Preprocessors/Segmentation & Mask",
"description": "Segments video into temporally consistent masks using Meta SAM3 from text or interactive prompts." "description": "Segments video into temporally consistent masks using Meta SAM3 from text or interactive prompts."
} }
] ]

2
blueprints/Video Upscale(GAN x4).json
View File

@ -412,7 +412,7 @@
"extra": { "extra": {
"workflowRendererVersion": "LG" "workflowRendererVersion": "LG"
}, },
"category": "Video generation and editing/Enhance video", "category": "Video generation and editing/Upscale",
"description": "Upscales video to 4× resolution using a GAN-based upscaling model." "description": "Upscales video to 4× resolution using a GAN-based upscaling model."
} }
] ]

1323
blueprints/Video to Pose Map (SDPose Multi-Person).json Normal file
View File

File diff suppressed because it is too large Load Diff

4
comfy/cli_args.py
View File

@ -49,7 +49,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.") parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.") parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.") parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.") parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use, as a comma-separated list (e.g. '0' or '0,1'). All other devices will not be visible.")
parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.") parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")
cm_group = parser.add_mutually_exclusive_group() cm_group = parser.add_mutually_exclusive_group()
cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).") cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
@ -111,7 +111,7 @@ parser.add_argument("--preview-method", type=LatentPreviewMethod, default=Latent
parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.") parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")
cache_group = parser.add_mutually_exclusive_group() cache_group = parser.add_mutually_exclusive_group()
cache_group.add_argument("--cache-ram", nargs='*', type=float, default=[], metavar="GB", help="Use RAM pressure caching with the specified headroom thresholds. This is the default caching mode. The first value sets the active-cache threshold; the optional second value sets the inactive-cache/pin threshold. Defaults when no values are provided: active 25%% of system RAM (min 4GB, max 32GB), inactive 75%% of system RAM (min 12GB, max 96GB).") cache_group.add_argument("--cache-ram", nargs='*', type=float, default=[], metavar="GB", help="Use RAM pressure caching with the specified headroom thresholds. This is the default caching mode. The first value sets the active-cache threshold; the optional second value sets the inactive-cache/pin threshold. Defaults when no values are provided: active 10%% of system RAM (min 2GB, max 10GB), inactive 100%% of system RAM (max 96GB).")
cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.") cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.") cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.") cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")

1
comfy/comfy_types/node_typing.py
View File

@ -1,6 +1,5 @@
"""Comfy-specific type hinting""" """Comfy-specific type hinting"""
from __future__ import annotations
from typing import Literal, TypedDict, Optional from typing import Literal, TypedDict, Optional
from typing_extensions import NotRequired from typing_extensions import NotRequired
from abc import ABC, abstractmethod from abc import ABC, abstractmethod

65
comfy/controlnet.py
View File

@ -15,13 +15,14 @@
You should have received a copy of the GNU General Public License You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>. along with this program. If not, see <https://www.gnu.org/licenses/>.
""" """
from __future__ import annotations
import torch import torch
from enum import Enum from enum import Enum
import math import math
import os import os
import logging import logging
import copy
import comfy.utils import comfy.utils
import comfy.model_management import comfy.model_management
import comfy.model_detection import comfy.model_detection
@ -38,7 +39,7 @@ import comfy.ldm.hydit.controlnet
import comfy.ldm.flux.controlnet import comfy.ldm.flux.controlnet
import comfy.ldm.qwen_image.controlnet import comfy.ldm.qwen_image.controlnet
import comfy.cldm.dit_embedder import comfy.cldm.dit_embedder
from typing import TYPE_CHECKING from typing import TYPE_CHECKING, Union
if TYPE_CHECKING: if TYPE_CHECKING:
from comfy.hooks import HookGroup from comfy.hooks import HookGroup
@ -64,6 +65,18 @@ class StrengthType(Enum):
CONSTANT = 1 CONSTANT = 1
LINEAR_UP = 2 LINEAR_UP = 2
class ControlIsolation:
'''Temporarily set a ControlBase object's previous_controlnet to None to prevent cascading calls.'''
def __init__(self, control: ControlBase):
self.control = control
self.orig_previous_controlnet = control.previous_controlnet
def __enter__(self):
self.control.previous_controlnet = None
def __exit__(self, *args):
self.control.previous_controlnet = self.orig_previous_controlnet
class ControlBase: class ControlBase:
def __init__(self): def __init__(self):
self.cond_hint_original = None self.cond_hint_original = None
@ -77,7 +90,7 @@ class ControlBase:
self.compression_ratio = 8 self.compression_ratio = 8
self.upscale_algorithm = 'nearest-exact' self.upscale_algorithm = 'nearest-exact'
self.extra_args = {} self.extra_args = {}
self.previous_controlnet = None self.previous_controlnet: Union[ControlBase, None] = None
self.extra_conds = [] self.extra_conds = []
self.strength_type = StrengthType.CONSTANT self.strength_type = StrengthType.CONSTANT
self.concat_mask = False self.concat_mask = False
@ -85,6 +98,7 @@ class ControlBase:
self.extra_concat = None self.extra_concat = None
self.extra_hooks: HookGroup = None self.extra_hooks: HookGroup = None
self.preprocess_image = lambda a: a self.preprocess_image = lambda a: a
self.multigpu_clones: dict[torch.device, ControlBase] = {}
def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]): def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]):
self.cond_hint_original = cond_hint self.cond_hint_original = cond_hint
@ -111,17 +125,38 @@ class ControlBase:
def cleanup(self): def cleanup(self):
if self.previous_controlnet is not None: if self.previous_controlnet is not None:
self.previous_controlnet.cleanup() self.previous_controlnet.cleanup()
for device_cnet in self.multigpu_clones.values():
with ControlIsolation(device_cnet):
device_cnet.cleanup()
self.cond_hint = None self.cond_hint = None
self.extra_concat = None self.extra_concat = None
self.timestep_range = None self.timestep_range = None
def get_models(self): def get_models(self):
out = [] out = []
for device_cnet in self.multigpu_clones.values():
out += device_cnet.get_models_only_self()
if self.previous_controlnet is not None: if self.previous_controlnet is not None:
out += self.previous_controlnet.get_models() out += self.previous_controlnet.get_models()
return out return out
def get_models_only_self(self):
'Calls get_models, but temporarily sets previous_controlnet to None.'
with ControlIsolation(self):
return self.get_models()
def get_instance_for_device(self, device):
'Returns instance of this Control object intended for selected device.'
return self.multigpu_clones.get(device, self)
def deepclone_multigpu(self, load_device, autoregister=False):
'''
Create deep clone of Control object where model(s) is set to other devices.
When autoregister is set to True, the deep clone is also added to multigpu_clones dict.
'''
raise NotImplementedError("Classes inheriting from ControlBase should define their own deepclone_multigpu funtion.")
def get_extra_hooks(self): def get_extra_hooks(self):
out = [] out = []
if self.extra_hooks is not None: if self.extra_hooks is not None:
@ -130,7 +165,7 @@ class ControlBase:
out += self.previous_controlnet.get_extra_hooks() out += self.previous_controlnet.get_extra_hooks()
return out return out
def copy_to(self, c): def copy_to(self, c: ControlBase):
c.cond_hint_original = self.cond_hint_original c.cond_hint_original = self.cond_hint_original
c.strength = self.strength c.strength = self.strength
c.timestep_percent_range = self.timestep_percent_range c.timestep_percent_range = self.timestep_percent_range
@ -284,6 +319,14 @@ class ControlNet(ControlBase):
self.copy_to(c) self.copy_to(c)
return c return c
def deepclone_multigpu(self, load_device, autoregister=False):
c = self.copy()
c.control_model = copy.deepcopy(c.control_model)
c.control_model_wrapped = comfy.model_patcher.ModelPatcher(c.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
if autoregister:
self.multigpu_clones[load_device] = c
return c
def get_models(self): def get_models(self):
out = super().get_models() out = super().get_models()
out.append(self.control_model_wrapped) out.append(self.control_model_wrapped)
@ -314,6 +357,10 @@ class QwenFunControlNet(ControlNet):
super().pre_run(model, percent_to_timestep_function) super().pre_run(model, percent_to_timestep_function)
self.set_extra_arg("base_model", model.diffusion_model) self.set_extra_arg("base_model", model.diffusion_model)
def cleanup(self):
self.extra_args.pop("base_model", None)
super().cleanup()
def copy(self): def copy(self):
c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
c.control_model = self.control_model c.control_model = self.control_model
@ -906,6 +953,14 @@ class T2IAdapter(ControlBase):
self.copy_to(c) self.copy_to(c)
return c return c
def deepclone_multigpu(self, load_device, autoregister=False):
c = self.copy()
c.t2i_model = copy.deepcopy(c.t2i_model)
c.device = load_device
if autoregister:
self.multigpu_clones[load_device] = c
return c
def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options
compression_ratio = 8 compression_ratio = 8
upscale_algorithm = 'nearest-exact' upscale_algorithm = 'nearest-exact'

20
comfy/ldm/hunyuan3dv2_1/hunyuandit.py
View File

@ -607,9 +607,13 @@ class HunYuanDiTPlain(nn.Module):
def forward(self, x, t, context, transformer_options = {}, **kwargs): def forward(self, x, t, context, transformer_options = {}, **kwargs):
x = x.movedim(-1, -2) x = x.movedim(-1, -2)
if context.shape[0] >= 2:
uncond_emb, cond_emb = context.chunk(2, dim = 0) swap_cfg_halves = context.shape[0] >= 2
context = torch.cat([cond_emb, uncond_emb], dim = 0)
if swap_cfg_halves:
first_half, second_half = context.chunk(2, dim = 0)
context = torch.cat([second_half, first_half], dim = 0)
main_condition = context main_condition = context
t = 1.0 - t t = 1.0 - t
@ -657,8 +661,8 @@ class HunYuanDiTPlain(nn.Module):
output = self.final_layer(combined) output = self.final_layer(combined)
output = output.movedim(-2, -1) * (-1.0) output = output.movedim(-2, -1) * (-1.0)
if output.shape[0] >= 2: if swap_cfg_halves:
cond_emb, uncond_emb = output.chunk(2, dim = 0) first_half, second_half = output.chunk(2, dim = 0)
return torch.cat([uncond_emb, cond_emb]) output = torch.cat([second_half, first_half], dim = 0)
else:
return output return output

8
comfy/ldm/lightricks/av_model.py
View File

@ -767,25 +767,25 @@ class LTXAVModel(LTXVModel):
# Cross-attention timesteps - compress these too # Cross-attention timesteps - compress these too
av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single(
timestep.max().expand_as(a_timestep_flat), a_timestep_flat,
{"resolution": None, "aspect_ratio": None}, {"resolution": None, "aspect_ratio": None},
batch_size=batch_size, batch_size=batch_size,
hidden_dtype=hidden_dtype, hidden_dtype=hidden_dtype,
) )
av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single(
a_timestep.max().expand_as(timestep_flat), timestep_flat,
{"resolution": None, "aspect_ratio": None}, {"resolution": None, "aspect_ratio": None},
batch_size=batch_size, batch_size=batch_size,
hidden_dtype=hidden_dtype, hidden_dtype=hidden_dtype,
) )
av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single(
a_timestep.max().expand_as(timestep_flat) * av_ca_factor, a_timestep_scaled.max().expand_as(timestep_flat) * av_ca_factor,
{"resolution": None, "aspect_ratio": None}, {"resolution": None, "aspect_ratio": None},
batch_size=batch_size, batch_size=batch_size,
hidden_dtype=hidden_dtype, hidden_dtype=hidden_dtype,
) )
av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single(
timestep.max().expand_as(a_timestep_flat) * av_ca_factor, timestep_scaled.max().expand_as(a_timestep_flat) * av_ca_factor,
{"resolution": None, "aspect_ratio": None}, {"resolution": None, "aspect_ratio": None},
batch_size=batch_size, batch_size=batch_size,
hidden_dtype=hidden_dtype, hidden_dtype=hidden_dtype,

1
comfy/ldm/lightricks/vae/causal_audio_autoencoder.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import torch import torch
from torch import nn from torch import nn
from torch.nn import functional as F from torch.nn import functional as F

1
comfy/ldm/lightricks/vae/causal_video_autoencoder.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import threading import threading
import torch import torch
from torch import nn from torch import nn

1
comfy/ldm/lumina/model.py
View File

@ -1,5 +1,4 @@
# Code from: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py # Code from: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py
from __future__ import annotations
from typing import List, Optional, Tuple from typing import List, Optional, Tuple

1
comfy/ldm/moge/geometry.py
View File

@ -1,6 +1,5 @@
"""Pure-torch + scipy geometry helpers for MoGe inference and mesh export.""" """Pure-torch + scipy geometry helpers for MoGe inference and mesh export."""
from __future__ import annotations
from typing import Optional, Tuple from typing import Optional, Tuple

1
comfy/ldm/moge/model.py
View File

@ -4,7 +4,6 @@ V1: DINOv2 backbone + multi-output head (points, mask).
V2: DINOv2 encoder + neck + per-output heads (points, mask, normal, optional metric-scale MLP). V2: DINOv2 encoder + neck + per-output heads (points, mask, normal, optional metric-scale MLP).
""" """
from __future__ import annotations
from numbers import Number from numbers import Number
from typing import Any, Dict, List, Optional, Tuple, Union from typing import Any, Dict, List, Optional, Tuple, Union

1
comfy/ldm/moge/modules.py
View File

@ -1,6 +1,5 @@
"""Building blocks for MoGe: residual conv stack, resamplers, MLP, DINOv2 encoder, v1 head.""" """Building blocks for MoGe: residual conv stack, resamplers, MLP, DINOv2 encoder, v1 head."""
from __future__ import annotations
from typing import List, Optional, Sequence, Tuple, Union from typing import List, Optional, Sequence, Tuple, Union

1
comfy/ldm/moge/panorama.py
View File

@ -6,7 +6,6 @@ equirect distance map via a multi-scale Poisson + gradient sparse solve.
Image sampling uses F.grid_sample (GPU); the sparse solve uses lsmr (CPU). Image sampling uses F.grid_sample (GPU); the sparse solve uses lsmr (CPU).
""" """
from __future__ import annotations
from typing import Callable, List, Optional, Tuple from typing import Callable, List, Optional, Tuple

1
comfy/lora.py
View File

@ -16,7 +16,6 @@
along with this program. If not, see <https://www.gnu.org/licenses/>. along with this program. If not, see <https://www.gnu.org/licenses/>.
""" """
from __future__ import annotations
import comfy.memory_management import comfy.memory_management
import comfy.utils import comfy.utils
import comfy.model_management import comfy.model_management

36
comfy/memory_management.py
View File

@ -1,6 +1,5 @@
import math import math
import ctypes import ctypes
import threading
import dataclasses import dataclasses
import torch import torch
from typing import NamedTuple from typing import NamedTuple
@ -10,7 +9,7 @@ from comfy.quant_ops import QuantizedTensor
class TensorFileSlice(NamedTuple): class TensorFileSlice(NamedTuple):
file_ref: object file_ref: object
thread_id: int lock: object
offset: int offset: int
size: int size: int
@ -43,7 +42,6 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N
file_obj = info.file_ref file_obj = info.file_ref
if (destination.device.type != "cpu" if (destination.device.type != "cpu"
or file_obj is None or file_obj is None
or threading.get_ident() != info.thread_id
or destination.numel() * destination.element_size() < info.size or destination.numel() * destination.element_size() < info.size
or tensor.numel() * tensor.element_size() != info.size or tensor.numel() * tensor.element_size() != info.size
or tensor.storage_offset() != 0 or tensor.storage_offset() != 0
@ -57,27 +55,29 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N
if hostbuf is not None: if hostbuf is not None:
stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0 stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0
device_ptr = destination2.data_ptr() if destination2 is not None else 0 device_ptr = destination2.data_ptr() if destination2 is not None else 0
hostbuf.read_file_slice(file_obj, info.offset, info.size, with info.lock:
offset=destination.data_ptr() - hostbuf.get_raw_address(), hostbuf.read_file_slice(file_obj, info.offset, info.size,
stream=stream_ptr, offset=destination.data_ptr() - hostbuf.get_raw_address(),
device_ptr=device_ptr, stream=stream_ptr,
device=None if destination2 is None else destination2.device.index) device_ptr=device_ptr,
device=None if destination2 is None else destination2.device.index)
return True return True
buf_type = ctypes.c_ubyte * info.size buf_type = ctypes.c_ubyte * info.size
view = memoryview(buf_type.from_address(destination.data_ptr())) view = memoryview(buf_type.from_address(destination.data_ptr()))
try: try:
file_obj.seek(info.offset) with info.lock:
done = 0 file_obj.seek(info.offset)
while done < info.size: done = 0
try: while done < info.size:
n = file_obj.readinto(view[done:]) try:
except OSError: n = file_obj.readinto(view[done:])
return False except OSError:
if n <= 0: return False
return False if n <= 0:
done += n return False
done += n
return True return True
finally: finally:
view.release() view.release()

153
comfy/model_management.py
View File

@ -15,6 +15,7 @@
You should have received a copy of the GNU General Public License You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>. along with this program. If not, see <https://www.gnu.org/licenses/>.
""" """
from __future__ import annotations
import psutil import psutil
import logging import logging
@ -27,13 +28,18 @@ import platform
import weakref import weakref
import gc import gc
import os import os
from contextlib import nullcontext from contextlib import contextmanager, nullcontext
import comfy.memory_management import comfy.memory_management
import comfy.utils import comfy.utils
import comfy.quant_ops import comfy.quant_ops
import comfy_aimdo.host_buffer import comfy_aimdo.host_buffer
import comfy_aimdo.vram_buffer import comfy_aimdo.vram_buffer
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from comfy.model_patcher import ModelPatcher
class VRAMState(Enum): class VRAMState(Enum):
DISABLED = 0 #No vram present: no need to move models to vram DISABLED = 0 #No vram present: no need to move models to vram
NO_VRAM = 1 #Very low vram: enable all the options to save vram NO_VRAM = 1 #Very low vram: enable all the options to save vram
@ -204,6 +210,107 @@ def get_torch_device():
else: else:
return torch.device(torch.cuda.current_device()) return torch.device(torch.cuda.current_device())
def get_all_torch_devices(exclude_current=False):
global cpu_state
devices = []
if cpu_state == CPUState.GPU:
# NVIDIA + AMD/ROCm both expose their GPUs through torch.cuda.*;
# without the AMD arm, single-GPU ROCm users get an empty list
# which silently turns unload_all_models() into a no-op.
if is_nvidia() or is_amd():
for i in range(torch.cuda.device_count()):
devices.append(torch.device("cuda", i))
elif is_intel_xpu():
for i in range(torch.xpu.device_count()):
devices.append(torch.device("xpu", i))
elif is_ascend_npu():
for i in range(torch.npu.device_count()):
devices.append(torch.device("npu", i))
elif is_mlu():
for i in range(torch.mlu.device_count()):
devices.append(torch.device("mlu", i))
else:
# Fallback for unhandled GPU backends (e.g. DirectML): at least
# report the current device so callers like unload_all_models()
# do not silently no-op.
devices.append(get_torch_device())
else:
devices.append(get_torch_device())
if exclude_current:
current = get_torch_device()
if current in devices:
devices.remove(current)
return devices
def get_gpu_device_options():
"""Return list of device option strings for node widgets.
Always includes "default" and "cpu". When multiple GPUs are present,
adds "gpu:0", "gpu:1", etc. (vendor-agnostic labels).
"""
options = ["default", "cpu"]
devices = get_all_torch_devices()
if len(devices) > 1:
for i in range(len(devices)):
options.append(f"gpu:{i}")
return options
def get_gpu_device_options_no_cpu():
"""Variant of get_gpu_device_options that omits "cpu".
Intended for components like the VAE selector where running on CPU
is impractical and should not be offered as a choice.
"""
return [o for o in get_gpu_device_options() if o != "cpu"]
def resolve_gpu_device_option(option: str):
"""Resolve a device option string to a torch.device.
Returns None for "default" (let the caller use its normal default).
Returns torch.device("cpu") for "cpu".
For "gpu:N", returns the Nth torch device. Returns None if the
index is out of range, the option string is malformed, or
unrecognized (callers are expected to log their own context-rich
message before falling back to the default device).
"""
if option is None or option == "default":
return None
if option == "cpu":
return torch.device("cpu")
if option.startswith("gpu:"):
try:
idx = int(option[4:])
except ValueError:
return None
devices = get_all_torch_devices()
if 0 <= idx < len(devices):
return devices[idx]
return None
@contextmanager
def cuda_device_context(device):
"""Context manager that sets torch.cuda.current_device to match *device*.
Used when running operations on a non-default CUDA device so that custom
CUDA kernels (e.g. comfy_kitchen fp8 quantization) pick up the correct
device index. The previous device is restored on exit.
No-op when *device* is not CUDA, has no explicit index, or already matches
the current device.
"""
prev = None
if device.type == "cuda" and device.index is not None:
prev = torch.cuda.current_device()
if prev != device.index:
torch.cuda.set_device(device)
else:
prev = None
try:
yield
finally:
if prev is not None:
torch.cuda.set_device(prev)
def get_total_memory(dev=None, torch_total_too=False): def get_total_memory(dev=None, torch_total_too=False):
global directml_enabled global directml_enabled
if dev is None: if dev is None:
@ -492,9 +599,13 @@ try:
logging.info("Device: {}".format(get_torch_device_name(get_torch_device()))) logging.info("Device: {}".format(get_torch_device_name(get_torch_device())))
except: except:
logging.warning("Could not pick default device.") logging.warning("Could not pick default device.")
try:
for device in get_all_torch_devices(exclude_current=True):
logging.info("Device: {}".format(get_torch_device_name(device)))
except:
pass
current_loaded_models: list[LoadedModel] = []
current_loaded_models = []
DIRTY_MMAPS = set() DIRTY_MMAPS = set()
@ -554,7 +665,7 @@ def ensure_pin_registerable(size, evict_active=False):
return shortfall <= REGISTERABLE_PIN_HYSTERESIS return shortfall <= REGISTERABLE_PIN_HYSTERESIS
class LoadedModel: class LoadedModel:
def __init__(self, model): def __init__(self, model: ModelPatcher):
self._set_model(model) self._set_model(model)
self.device = model.load_device self.device = model.load_device
self.real_model = None self.real_model = None
@ -562,7 +673,7 @@ class LoadedModel:
self.model_finalizer = None self.model_finalizer = None
self._patcher_finalizer = None self._patcher_finalizer = None
def _set_model(self, model): def _set_model(self, model: ModelPatcher):
self._model = weakref.ref(model) self._model = weakref.ref(model)
if model.parent is not None: if model.parent is not None:
self._parent_model = weakref.ref(model.parent) self._parent_model = weakref.ref(model.parent)
@ -573,6 +684,7 @@ class LoadedModel:
model = self._parent_model() model = self._parent_model()
if model is not None: if model is not None:
self._set_model(model) self._set_model(model)
self.device = model.load_device
@property @property
def model(self): def model(self):
@ -1217,7 +1329,7 @@ def get_aimdo_cast_buffer(offload_stream, device):
def get_pin_buffer(offload_stream): def get_pin_buffer(offload_stream):
pin_buffer = STREAM_PIN_BUFFERS.get(offload_stream, None) pin_buffer = STREAM_PIN_BUFFERS.get(offload_stream, None)
if pin_buffer is None: if pin_buffer is None:
pin_buffer = comfy_aimdo.host_buffer.HostBuffer(0, 0, pinned_hostbuf_size(8 * 1024**3)) pin_buffer = comfy_aimdo.host_buffer.HostBuffer(0, 0, pinned_hostbuf_size(8 * 1024**3), mark_cold=False)
STREAM_PIN_BUFFERS[offload_stream] = pin_buffer STREAM_PIN_BUFFERS[offload_stream] = pin_buffer
elif offload_stream is not None: elif offload_stream is not None:
event = getattr(pin_buffer, "_comfy_event", None) event = getattr(pin_buffer, "_comfy_event", None)
@ -1848,7 +1960,34 @@ def soft_empty_cache(force=False):
torch.cuda.ipc_collect() torch.cuda.ipc_collect()
def unload_all_models(): def unload_all_models():
free_memory(1e30, get_torch_device()) for device in get_all_torch_devices():
free_memory(1e30, device)
def unload_model_and_clones(model: ModelPatcher, unload_additional_models=True, all_devices=False):
'Unload only model and its clones - primarily for multigpu cloning purposes.'
initial_keep_loaded: list[LoadedModel] = current_loaded_models.copy()
additional_models = []
if unload_additional_models:
additional_models = model.get_nested_additional_models()
keep_loaded = []
for loaded_model in initial_keep_loaded:
if loaded_model.model is not None:
if model.clone_base_uuid == loaded_model.model.clone_base_uuid:
continue
# check additional models if they are a match
skip = False
for add_model in additional_models:
if add_model.clone_base_uuid == loaded_model.model.clone_base_uuid:
skip = True
break
if skip:
continue
keep_loaded.append(loaded_model)
if not all_devices:
free_memory(1e30, get_torch_device(), keep_loaded)
else:
for device in get_all_torch_devices():
free_memory(1e30, device, keep_loaded)
def debug_memory_summary(): def debug_memory_summary():
if is_amd() or is_nvidia(): if is_amd() or is_nvidia():

177
comfy/model_patcher.py
View File

@ -78,12 +78,15 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
def create_model_options_clone(orig_model_options: dict): def create_model_options_clone(orig_model_options: dict):
return comfy.patcher_extension.copy_nested_dicts(orig_model_options) return comfy.patcher_extension.copy_nested_dicts(orig_model_options)
def create_hook_patches_clone(orig_hook_patches): def create_hook_patches_clone(orig_hook_patches, copy_tuples=False):
new_hook_patches = {} new_hook_patches = {}
for hook_ref in orig_hook_patches: for hook_ref in orig_hook_patches:
new_hook_patches[hook_ref] = {} new_hook_patches[hook_ref] = {}
for k in orig_hook_patches[hook_ref]: for k in orig_hook_patches[hook_ref]:
new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:] new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:]
if copy_tuples:
for i in range(len(new_hook_patches[hook_ref][k])):
new_hook_patches[hook_ref][k][i] = tuple(new_hook_patches[hook_ref][k][i])
return new_hook_patches return new_hook_patches
def wipe_lowvram_weight(m): def wipe_lowvram_weight(m):
@ -329,7 +332,10 @@ class ModelPatcher:
self.is_clip = False self.is_clip = False
self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
self.cached_patcher_init: tuple[Callable, tuple] | None = None self.cached_patcher_init: tuple[Callable, tuple] | tuple[Callable, tuple, int] | None = None
self.is_multigpu_base_clone = False
self.clone_base_uuid = uuid.uuid4()
if not hasattr(self.model, 'model_loaded_weight_memory'): if not hasattr(self.model, 'model_loaded_weight_memory'):
self.model.model_loaded_weight_memory = 0 self.model.model_loaded_weight_memory = 0
@ -366,7 +372,8 @@ class ModelPatcher:
#than pays for CFG. So return everything both torch and Aimdo could give us #than pays for CFG. So return everything both torch and Aimdo could give us
aimdo_mem = 0 aimdo_mem = 0
if comfy.memory_management.aimdo_enabled: if comfy.memory_management.aimdo_enabled:
aimdo_mem = comfy_aimdo.model_vbar.vbars_analyze() aimdo_device = device.index if getattr(device, "type", None) == "cuda" else None
aimdo_mem = comfy_aimdo.model_vbar.vbars_analyze(aimdo_device)
return comfy.model_management.get_free_memory(device) + aimdo_mem return comfy.model_management.get_free_memory(device) + aimdo_mem
def get_clone_model_override(self): def get_clone_model_override(self):
@ -380,6 +387,8 @@ class ModelPatcher:
if self.cached_patcher_init is None: if self.cached_patcher_init is None:
raise RuntimeError("Cannot create non-dynamic delegate: cached_patcher_init is not initialized.") raise RuntimeError("Cannot create non-dynamic delegate: cached_patcher_init is not initialized.")
temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True) temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True)
if len(self.cached_patcher_init) > 2:
temp_model_patcher = temp_model_patcher[self.cached_patcher_init[2]]
model_override = temp_model_patcher.get_clone_model_override() model_override = temp_model_patcher.get_clone_model_override()
if model_override is None: if model_override is None:
model_override = self.get_clone_model_override() model_override = self.get_clone_model_override()
@ -438,19 +447,113 @@ class ModelPatcher:
n.hook_mode = self.hook_mode n.hook_mode = self.hook_mode
n.cached_patcher_init = self.cached_patcher_init n.cached_patcher_init = self.cached_patcher_init
n.is_multigpu_base_clone = self.is_multigpu_base_clone
n.clone_base_uuid = self.clone_base_uuid
for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE): for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE):
callback(self, n) callback(self, n)
return n return n
def deepclone_multigpu(self, new_load_device=None, models_cache: dict[uuid.UUID,ModelPatcher]=None):
logging.info(f"Creating deepclone of {self.model.__class__.__name__} for {new_load_device if new_load_device else self.load_device}.")
if self.cached_patcher_init is None:
raise RuntimeError(
f"Cannot create multigpu deepclone of {self.model.__class__.__name__}: "
"the loader that produced this model does not support multigpu "
"(cached_patcher_init is not initialized). Use a core loader "
"(CheckpointLoaderSimple, UNETLoader, CLIPLoader/DualCLIPLoader, VAELoader), "
"or have the custom loader register a cached_patcher_init factory."
)
comfy.model_management.unload_model_and_clones(self)
# Produce a freshly-loaded patcher from the loader factory so the multigpu
# clone owns its own untainted model weights (rather than relying on
# copy.deepcopy of an already-patched/already-loaded module).
temp_model_patcher: ModelPatcher | list[ModelPatcher] = self.cached_patcher_init[0](*self.cached_patcher_init[1])
if len(self.cached_patcher_init) > 2:
temp_model_patcher = temp_model_patcher[self.cached_patcher_init[2]]
# Override clone()'s normal "share self.model + share backup containers" with
# the pristine model from temp_model_patcher plus empty backup containers --
# the fresh model has no patches applied, so any deepcopy of self's stale
# backup/object_patches_backup/pinned would just propagate dead state that
# no longer corresponds to anything in n.model.
model_override = (temp_model_patcher.model, ({}, {}, {}, set()))
n = self.clone(model_override=model_override)
# clone() copies hook_backup by reference from self; reset since model is pristine.
n.hook_backup = {}
# set load device, if present
if new_load_device is not None:
n.load_device = new_load_device
# Ensure any per-device bookkeeping (e.g. ModelPatcherDynamic.dynamic_pins)
# has an entry for n.load_device on the freshly-loaded n.model. temp_model_patcher's
# __init__ only registered its own (default) load_device.
if hasattr(n, "register_load_device"):
n.register_load_device(n.load_device)
# multigpu clone should not have multigpu additional_models entry
n.remove_additional_models("multigpu")
# multigpu_clone all stored additional_models; make sure circular references are properly handled
if models_cache is None:
models_cache = {}
for key, model_list in n.additional_models.items():
for i in range(len(model_list)):
add_model = n.additional_models[key][i]
if add_model.clone_base_uuid not in models_cache:
models_cache[add_model.clone_base_uuid] = add_model.deepclone_multigpu(new_load_device=new_load_device, models_cache=models_cache)
n.additional_models[key][i] = models_cache[add_model.clone_base_uuid]
for callback in self.get_all_callbacks(CallbacksMP.ON_DEEPCLONE_MULTIGPU):
callback(self, n)
return n
def match_multigpu_clones(self):
multigpu_models = self.get_additional_models_with_key("multigpu")
if len(multigpu_models) > 0:
new_multigpu_models = []
for mm in multigpu_models:
# clone main model, but bring over relevant props from existing multigpu clone
n = self.clone()
n.load_device = mm.load_device
n.backup = mm.backup
n.object_patches_backup = mm.object_patches_backup
n.hook_backup = mm.hook_backup
n.model = mm.model
n.is_multigpu_base_clone = mm.is_multigpu_base_clone
n.remove_additional_models("multigpu")
orig_additional_models: dict[str, list[ModelPatcher]] = comfy.patcher_extension.copy_nested_dicts(n.additional_models)
n.additional_models = comfy.patcher_extension.copy_nested_dicts(mm.additional_models)
# figure out which additional models are not present in multigpu clone
models_cache = {}
for mm_add_model in mm.get_additional_models():
models_cache[mm_add_model.clone_base_uuid] = mm_add_model
remove_models_uuids = set(list(models_cache.keys()))
for key, model_list in orig_additional_models.items():
for orig_add_model in model_list:
if orig_add_model.clone_base_uuid not in models_cache:
models_cache[orig_add_model.clone_base_uuid] = orig_add_model.deepclone_multigpu(new_load_device=n.load_device, models_cache=models_cache)
existing_list = n.get_additional_models_with_key(key)
existing_list.append(models_cache[orig_add_model.clone_base_uuid])
n.set_additional_models(key, existing_list)
if orig_add_model.clone_base_uuid in remove_models_uuids:
remove_models_uuids.remove(orig_add_model.clone_base_uuid)
# remove duplicate additional models
for key, model_list in n.additional_models.items():
new_model_list = [x for x in model_list if x.clone_base_uuid not in remove_models_uuids]
n.set_additional_models(key, new_model_list)
for callback in self.get_all_callbacks(CallbacksMP.ON_MATCH_MULTIGPU_CLONES):
callback(self, n)
new_multigpu_models.append(n)
self.set_additional_models("multigpu", new_multigpu_models)
def is_clone(self, other): def is_clone(self, other):
if hasattr(other, 'model') and self.model is other.model: if hasattr(other, 'model') and self.model is other.model:
return True return True
return False return False
def clone_has_same_weights(self, clone: 'ModelPatcher'): def clone_has_same_weights(self, clone: ModelPatcher, allow_multigpu=False):
if not self.is_clone(clone): if allow_multigpu:
return False if self.clone_base_uuid != clone.clone_base_uuid:
return False
else:
if not self.is_clone(clone):
return False
if self.current_hooks != clone.current_hooks: if self.current_hooks != clone.current_hooks:
return False return False
@ -1232,7 +1335,7 @@ class ModelPatcher:
return self.additional_models.get(key, []) return self.additional_models.get(key, [])
def get_additional_models(self): def get_additional_models(self):
all_models = [] all_models: list[ModelPatcher] = []
for models in self.additional_models.values(): for models in self.additional_models.values():
all_models.extend(models) all_models.extend(models)
return all_models return all_models
@ -1286,9 +1389,18 @@ class ModelPatcher:
for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN): for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN):
callback(self) callback(self)
def prepare_state(self, timestep): def prepare_state(self, timestep, model_options):
ignore_multigpu = model_options.get("ignore_multigpu", False)
for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE): for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE):
callback(self, timestep) callback(self, timestep, model_options)
if not ignore_multigpu and "multigpu_clones" in model_options:
model_options["ignore_multigpu"] = True
try:
for p in model_options["multigpu_clones"].values():
p: ModelPatcher
p.prepare_state(timestep, model_options)
finally:
model_options.pop("ignore_multigpu", None)
def restore_hook_patches(self): def restore_hook_patches(self):
if self.hook_patches_backup is not None: if self.hook_patches_backup is not None:
@ -1301,12 +1413,18 @@ class ModelPatcher:
def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]): def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]):
curr_t = t[0] curr_t = t[0]
reset_current_hooks = False reset_current_hooks = False
multigpu_kf_changed_cache = None
transformer_options = model_options.get("transformer_options", {}) transformer_options = model_options.get("transformer_options", {})
for hook in hook_group.hooks: for hook in hook_group.hooks:
changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options) changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options)
# if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref; # if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref;
# this will cause the weights to be recalculated when sampling # this will cause the weights to be recalculated when sampling
if changed: if changed:
# cache changed for multigpu usage
if "multigpu_clones" in model_options:
if multigpu_kf_changed_cache is None:
multigpu_kf_changed_cache = []
multigpu_kf_changed_cache.append(hook)
# reset current_hooks if contains hook that changed # reset current_hooks if contains hook that changed
if self.current_hooks is not None: if self.current_hooks is not None:
for current_hook in self.current_hooks.hooks: for current_hook in self.current_hooks.hooks:
@ -1318,6 +1436,28 @@ class ModelPatcher:
self.cached_hook_patches.pop(cached_group) self.cached_hook_patches.pop(cached_group)
if reset_current_hooks: if reset_current_hooks:
self.patch_hooks(None) self.patch_hooks(None)
if "multigpu_clones" in model_options:
for p in model_options["multigpu_clones"].values():
p: ModelPatcher
p._handle_changed_hook_keyframes(multigpu_kf_changed_cache)
def _handle_changed_hook_keyframes(self, kf_changed_cache: list[comfy.hooks.Hook]):
'Used to handle multigpu behavior inside prepare_hook_patches_current_keyframe.'
if kf_changed_cache is None:
return
reset_current_hooks = False
# reset current_hooks if contains hook that changed
for hook in kf_changed_cache:
if self.current_hooks is not None:
for current_hook in self.current_hooks.hooks:
if current_hook == hook:
reset_current_hooks = True
break
for cached_group in list(self.cached_hook_patches.keys()):
if cached_group.contains(hook):
self.cached_hook_patches.pop(cached_group)
if reset_current_hooks:
self.patch_hooks(None)
def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None, def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None,
registered: comfy.hooks.HookGroup = None): registered: comfy.hooks.HookGroup = None):
@ -1566,16 +1706,27 @@ class ModelPatcherDynamic(ModelPatcher):
self.model.dynamic_vbars = {} self.model.dynamic_vbars = {}
if not hasattr(self.model, "dynamic_pins"): if not hasattr(self.model, "dynamic_pins"):
self.model.dynamic_pins = {} self.model.dynamic_pins = {}
if self.load_device not in self.model.dynamic_pins: self.register_load_device(self.load_device)
self.model.dynamic_pins[self.load_device] = { self.non_dynamic_delegate_model = None
assert load_device is not None
def register_load_device(self, device):
"""Ensure dynamic_pins has an entry for *device*.
Called from __init__ and also from any code that retargets an
already-constructed patcher to a new load_device (e.g. the
Select{Model,CLIP,VAE}Device selector nodes); without this entry
partially_unload_ram() raises KeyError when it tries to read the
per-device pin state.
"""
if device not in self.model.dynamic_pins:
self.model.dynamic_pins[device] = {
"weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]), "weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]),
"patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]), "patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]),
"hostbufs_initialized": False, "hostbufs_initialized": False,
"failed": False, "failed": False,
"active": False, "active": False,
} }
self.non_dynamic_delegate_model = None
assert load_device is not None
def is_dynamic(self): def is_dynamic(self):
return True return True

248
comfy/multigpu.py Normal file
View File

@ -0,0 +1,248 @@
from __future__ import annotations
import queue
import threading
import torch
import logging
from collections import namedtuple
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from comfy.model_patcher import ModelPatcher
import comfy.utils
import comfy.patcher_extension
import comfy.model_management
class MultiGPUThreadPool:
"""Persistent thread pool for multi-GPU work distribution.
Maintains one worker thread per extra GPU device. Each thread calls
torch.cuda.set_device() once at startup so that compiled kernel caches
(inductor/triton) stay warm across diffusion steps.
"""
def __init__(self, devices: list[torch.device]):
self._workers: list[threading.Thread] = []
self._work_queues: dict[torch.device, queue.Queue] = {}
self._result_queues: dict[torch.device, queue.Queue] = {}
for device in devices:
wq = queue.Queue()
rq = queue.Queue()
self._work_queues[device] = wq
self._result_queues[device] = rq
t = threading.Thread(target=self._worker_loop, args=(device, wq, rq), daemon=True)
t.start()
self._workers.append(t)
def _worker_loop(self, device: torch.device, work_q: queue.Queue, result_q: queue.Queue):
try:
torch.cuda.set_device(device)
except Exception as e:
logging.error(f"MultiGPUThreadPool: failed to set device {device}: {e}")
while True:
item = work_q.get()
if item is None:
return
result_q.put((None, e))
return
while True:
item = work_q.get()
if item is None:
break
fn, args, kwargs = item
try:
result = fn(*args, **kwargs)
result_q.put((result, None))
except Exception as e:
result_q.put((None, e))
def submit(self, device: torch.device, fn, *args, **kwargs):
self._work_queues[device].put((fn, args, kwargs))
def get_result(self, device: torch.device):
return self._result_queues[device].get()
@property
def devices(self) -> list[torch.device]:
return list(self._work_queues.keys())
def shutdown(self):
for wq in self._work_queues.values():
wq.put(None) # sentinel
for t in self._workers:
t.join(timeout=5.0)
class GPUOptions:
def __init__(self, device_index: int, relative_speed: float):
self.device_index = device_index
self.relative_speed = relative_speed
def clone(self):
return GPUOptions(self.device_index, self.relative_speed)
def create_dict(self):
return {
"relative_speed": self.relative_speed
}
class GPUOptionsGroup:
def __init__(self):
self.options: dict[int, GPUOptions] = {}
def add(self, info: GPUOptions):
self.options[info.device_index] = info
def clone(self):
c = GPUOptionsGroup()
for opt in self.options.values():
c.add(opt)
return c
def register(self, model: ModelPatcher):
opts_dict = {}
# get devices that are valid for this model
devices: list[torch.device] = [model.load_device]
for extra_model in model.get_additional_models_with_key("multigpu"):
extra_model: ModelPatcher
devices.append(extra_model.load_device)
# create dictionary with actual device mapped to its GPUOptions
device_opts_list: list[GPUOptions] = []
for device in devices:
device_opts = self.options.get(device.index, GPUOptions(device_index=device.index, relative_speed=1.0))
opts_dict[device] = device_opts.create_dict()
device_opts_list.append(device_opts)
# make relative_speed relative to 1.0
min_speed = min([x.relative_speed for x in device_opts_list])
for value in opts_dict.values():
value['relative_speed'] /= min_speed
model.model_options['multigpu_options'] = opts_dict
def create_multigpu_deepclones(model: ModelPatcher, max_gpus: int, gpu_options: GPUOptionsGroup=None, reuse_loaded=False):
'Prepare ModelPatcher to contain deepclones of its BaseModel and related properties.'
model = model.clone()
# check if multigpu is already prepared - get the load devices from them if possible to exclude
skip_devices = set()
multigpu_models = model.get_additional_models_with_key("multigpu")
if len(multigpu_models) > 0:
for mm in multigpu_models:
skip_devices.add(mm.load_device)
skip_devices = list(skip_devices)
# Exclude the primary model's actual device, not the global current device:
# after SelectModelDevice(gpu:N) the primary may not live on the process's
# current CUDA device, and excluding the wrong device picks bad extras.
all_devices = comfy.model_management.get_all_torch_devices(exclude_current=False)
full_extra_devices = [d for d in all_devices if d != model.load_device]
limit_extra_devices = full_extra_devices[:max_gpus-1]
extra_devices = limit_extra_devices.copy()
# exclude skipped devices
for skip in skip_devices:
if skip in extra_devices:
extra_devices.remove(skip)
# create new deepclones
if len(extra_devices) > 0:
for device in extra_devices:
device_patcher = None
if reuse_loaded:
# Only reuse a previously-loaded MultiGPU clone. A SelectModelDevice
# patcher on the same device shares clone_base_uuid but has
# is_multigpu_base_clone=False, which would later be filtered out by
# prepare_model_patcher_multigpu_clones() and silently shrink the
# work split back to one GPU.
loaded_models: list[ModelPatcher] = comfy.model_management.loaded_models()
for lm in loaded_models:
if lm.model is None:
continue
if lm.load_device != device:
continue
if lm.clone_base_uuid != model.clone_base_uuid:
continue
if not getattr(lm, "is_multigpu_base_clone", False):
continue
device_patcher = lm.clone()
logging.info(f"Reusing loaded multigpu deepclone of {device_patcher.model.__class__.__name__} for {device}")
break
if device_patcher is None:
device_patcher = model.deepclone_multigpu(new_load_device=device)
# Always flag the clone; whether reused or freshly deepcloned, it must
# advertise itself as a MultiGPU base clone so the cond scheduler picks
# it up in prepare_model_patcher_multigpu_clones().
device_patcher.is_multigpu_base_clone = True
multigpu_models = model.get_additional_models_with_key("multigpu")
multigpu_models.append(device_patcher)
model.set_additional_models("multigpu", multigpu_models)
model.match_multigpu_clones()
if gpu_options is None:
gpu_options = GPUOptionsGroup()
gpu_options.register(model)
else:
logging.info("No extra torch devices need initialization, skipping initializing MultiGPU Work Units.")
# only keep model clones that don't go 'past' the intended max_gpu count;
# this prunes any inherited multigpu clones whose load_device is no longer allowed
# when max_gpus is lowered between runs.
allowed_devices = set(limit_extra_devices)
allowed_devices.add(model.load_device)
multigpu_models = model.get_additional_models_with_key("multigpu")
new_multigpu_models = [m for m in multigpu_models if m.load_device in allowed_devices]
if len(new_multigpu_models) != len(multigpu_models):
model.set_additional_models("multigpu", new_multigpu_models)
model.match_multigpu_clones()
return model
LoadBalance = namedtuple('LoadBalance', ['work_per_device', 'idle_time'])
def load_balance_devices(model_options: dict[str], total_work: int, return_idle_time=False, work_normalized: int=None):
'Optimize work assigned to different devices, accounting for their relative speeds and splittable work.'
opts_dict = model_options['multigpu_options']
devices = list(model_options['multigpu_clones'].keys())
speed_per_device = []
work_per_device = []
# get sum of each device's relative_speed
total_speed = 0.0
for opts in opts_dict.values():
total_speed += opts['relative_speed']
# get relative work for each device;
# obtained by w = (W*r)/R
for device in devices:
relative_speed = opts_dict[device]['relative_speed']
relative_work = (total_work*relative_speed) / total_speed
speed_per_device.append(relative_speed)
work_per_device.append(relative_work)
# relative work must be expressed in whole numbers, but likely is a decimal;
# perform rounding while maintaining total sum equal to total work (sum of relative works)
work_per_device = round_preserved(work_per_device)
dict_work_per_device = {}
for device, relative_work in zip(devices, work_per_device):
dict_work_per_device[device] = relative_work
if not return_idle_time:
return LoadBalance(dict_work_per_device, None)
# divide relative work by relative speed to get estimated completion time of said work by each device;
# time here is relative and does not correspond to real-world units
completion_time = [w/r for w,r in zip(work_per_device, speed_per_device)]
# calculate relative time spent by the devices waiting on each other after their work is completed
idle_time = abs(min(completion_time) - max(completion_time))
# if need to compare work idle time, need to normalize to a common total work
if work_normalized:
idle_time *= (work_normalized/total_work)
return LoadBalance(dict_work_per_device, idle_time)
def round_preserved(values: list[float]):
'Round all values in a list, preserving the combined sum of values.'
# get floor of values; casting to int does it too
floored = [int(x) for x in values]
total_floored = sum(floored)
# get remainder to distribute
remainder = round(sum(values)) - total_floored
# pair values with fractional portions
fractional = [(i, x-floored[i]) for i, x in enumerate(values)]
# sort by fractional part in descending order
fractional.sort(key=lambda x: x[1], reverse=True)
# distribute the remainder
for i in range(remainder):
index = fractional[i][0]
floored[index] += 1
return floored

3
comfy/patcher_extension.py
View File

@ -1,8 +1,9 @@
from __future__ import annotations
from typing import Callable from typing import Callable
class CallbacksMP: class CallbacksMP:
ON_CLONE = "on_clone" ON_CLONE = "on_clone"
ON_DEEPCLONE_MULTIGPU = "on_deepclone_multigpu"
ON_MATCH_MULTIGPU_CLONES = "on_match_multigpu_clones"
ON_LOAD = "on_load_after" ON_LOAD = "on_load_after"
ON_DETACH = "on_detach_after" ON_DETACH = "on_detach_after"
ON_CLEANUP = "on_cleanup" ON_CLEANUP = "on_cleanup"

65
comfy/sampler_helpers.py
View File

@ -1,16 +1,18 @@
from __future__ import annotations from __future__ import annotations
import torch
import uuid import uuid
import math import math
import collections import collections
import comfy.model_management import comfy.model_management
import comfy.conds import comfy.conds
import comfy.model_patcher
import comfy.utils import comfy.utils
import comfy.hooks import comfy.hooks
import comfy.patcher_extension import comfy.patcher_extension
from typing import TYPE_CHECKING from typing import TYPE_CHECKING
if TYPE_CHECKING: if TYPE_CHECKING:
from comfy.model_patcher import ModelPatcher
from comfy.model_base import BaseModel from comfy.model_base import BaseModel
from comfy.model_patcher import ModelPatcher
from comfy.controlnet import ControlBase from comfy.controlnet import ControlBase
def prepare_mask(noise_mask, shape, device): def prepare_mask(noise_mask, shape, device):
@ -119,6 +121,47 @@ def cleanup_additional_models(models):
if hasattr(m, 'cleanup'): if hasattr(m, 'cleanup'):
m.cleanup() m.cleanup()
def preprocess_multigpu_conds(conds: dict[str, list[dict[str]]], model: ModelPatcher, model_options: dict[str]):
'''If multigpu acceleration required, creates deepclones of ControlNets and GLIGEN per device.'''
multigpu_models: list[ModelPatcher] = model.get_additional_models_with_key("multigpu")
if len(multigpu_models) == 0:
return
extra_devices = [x.load_device for x in multigpu_models]
# handle controlnets
controlnets: set[ControlBase] = set()
for k in conds:
for kk in conds[k]:
if 'control' in kk:
controlnets.add(kk['control'])
if len(controlnets) > 0:
# first, unload all controlnet clones
for cnet in list(controlnets):
cnet_models = cnet.get_models()
for cm in cnet_models:
comfy.model_management.unload_model_and_clones(cm, unload_additional_models=True)
# next, make sure each controlnet has a deepclone for all relevant devices
for cnet in controlnets:
curr_cnet = cnet
while curr_cnet is not None:
for device in extra_devices:
if device not in curr_cnet.multigpu_clones:
curr_cnet.deepclone_multigpu(device, autoregister=True)
curr_cnet = curr_cnet.previous_controlnet
# since all device clones are now present, recreate the linked list for cloned cnets per device
for cnet in controlnets:
curr_cnet = cnet
while curr_cnet is not None:
prev_cnet = curr_cnet.previous_controlnet
for device in extra_devices:
device_cnet = curr_cnet.get_instance_for_device(device)
prev_device_cnet = None
if prev_cnet is not None:
prev_device_cnet = prev_cnet.get_instance_for_device(device)
device_cnet.set_previous_controlnet(prev_device_cnet)
curr_cnet = prev_cnet
# potentially handle gligen - since not widely used, ignored for now
def estimate_memory(model, noise_shape, conds): def estimate_memory(model, noise_shape, conds):
cond_shapes = collections.defaultdict(list) cond_shapes = collections.defaultdict(list)
cond_shapes_min = {} cond_shapes_min = {}
@ -143,7 +186,8 @@ def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None
return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload) return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload)
def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False): def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
real_model: BaseModel = None model.match_multigpu_clones()
preprocess_multigpu_conds(conds, model, model_options)
models, inference_memory = get_additional_models(conds, model.model_dtype()) models, inference_memory = get_additional_models(conds, model.model_dtype())
models += get_additional_models_from_model_options(model_options) models += get_additional_models_from_model_options(model_options)
models += model.get_nested_additional_models() # TODO: does this require inference_memory update? models += model.get_nested_additional_models() # TODO: does this require inference_memory update?
@ -155,7 +199,7 @@ def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=Non
memory_required += inference_memory memory_required += inference_memory
minimum_memory_required += inference_memory minimum_memory_required += inference_memory
comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
real_model = model.model real_model: BaseModel = model.model
return real_model, conds, models return real_model, conds, models
@ -201,3 +245,18 @@ def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name], comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name],
copy_dict1=False) copy_dict1=False)
return to_load_options return to_load_options
def prepare_model_patcher_multigpu_clones(model_patcher: ModelPatcher, loaded_models: list[ModelPatcher], model_options: dict):
'''
In case multigpu acceleration is enabled, prep ModelPatchers for each device.
'''
multigpu_patchers: list[ModelPatcher] = [x for x in loaded_models if x.is_multigpu_base_clone]
if len(multigpu_patchers) > 0:
multigpu_dict: dict[torch.device, ModelPatcher] = {}
multigpu_dict[model_patcher.load_device] = model_patcher
for x in multigpu_patchers:
x.hook_patches = comfy.model_patcher.create_hook_patches_clone(model_patcher.hook_patches, copy_tuples=True)
x.hook_mode = model_patcher.hook_mode # match main model's hook_mode
multigpu_dict[x.load_device] = x
model_options["multigpu_clones"] = multigpu_dict
return multigpu_patchers

310
comfy/samplers.py
View File

@ -1,7 +1,9 @@
from __future__ import annotations from __future__ import annotations
import comfy.model_management
from .k_diffusion import sampling as k_diffusion_sampling from .k_diffusion import sampling as k_diffusion_sampling
from .extra_samplers import uni_pc from .extra_samplers import uni_pc
from typing import TYPE_CHECKING, Callable, NamedTuple from typing import TYPE_CHECKING, Callable, NamedTuple, Any
if TYPE_CHECKING: if TYPE_CHECKING:
from comfy.model_patcher import ModelPatcher from comfy.model_patcher import ModelPatcher
from comfy.model_base import BaseModel from comfy.model_base import BaseModel
@ -16,6 +18,7 @@ import comfy.model_patcher
import comfy.patcher_extension import comfy.patcher_extension
import comfy.hooks import comfy.hooks
import comfy.context_windows import comfy.context_windows
import comfy.multigpu
import comfy.utils import comfy.utils
import scipy.stats import scipy.stats
import numpy import numpy
@ -141,7 +144,7 @@ def can_concat_cond(c1, c2):
return cond_equal_size(c1.conditioning, c2.conditioning) return cond_equal_size(c1.conditioning, c2.conditioning)
def cond_cat(c_list): def cond_cat(c_list, device=None):
temp = {} temp = {}
for x in c_list: for x in c_list:
for k in x: for k in x:
@ -153,6 +156,8 @@ def cond_cat(c_list):
for k in temp: for k in temp:
conds = temp[k] conds = temp[k]
out[k] = conds[0].concat(conds[1:]) out[k] = conds[0].concat(conds[1:])
if device is not None and hasattr(out[k], 'to'):
out[k] = out[k].to(device)
return out return out
@ -212,7 +217,12 @@ def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torc
) )
return executor.execute(model, conds, x_in, timestep, model_options) return executor.execute(model, conds, x_in, timestep, model_options)
def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options): def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
# NOTE: keep in sync with _calc_cond_batch_multigpu below. Shared logic
# (hooked_to_run accumulation, memory-fit batching, per-chunk output
# aggregation) is duplicated there with per-device scheduling layered on top.
if 'multigpu_clones' in model_options:
return _calc_cond_batch_multigpu(model, conds, x_in, timestep, model_options)
out_conds = [] out_conds = []
out_counts = [] out_counts = []
# separate conds by matching hooks # separate conds by matching hooks
@ -244,7 +254,7 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
if has_default_conds: if has_default_conds:
finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options) finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
model.current_patcher.prepare_state(timestep) model.current_patcher.prepare_state(timestep, model_options)
# run every hooked_to_run separately # run every hooked_to_run separately
for hooks, to_run in hooked_to_run.items(): for hooks, to_run in hooked_to_run.items():
@ -344,6 +354,239 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
return out_conds return out_conds
def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
# NOTE: keep in sync with _calc_cond_batch above. Same conds-by-hooks
# accumulation, memory-fit batching, and output aggregation, but adds a
# per-device scheduler, per-device patcher/control lookup, tensor .to(device)
# placement, and MultiGPUThreadPool dispatch around the inner loop.
out_conds = []
out_counts = []
# separate conds by matching hooks
hooked_to_run: dict[comfy.hooks.HookGroup,list[tuple[tuple,int]]] = {}
default_conds = []
has_default_conds = False
output_device = x_in.device
for i in range(len(conds)):
out_conds.append(torch.zeros_like(x_in))
out_counts.append(torch.ones_like(x_in) * 1e-37)
cond = conds[i]
default_c = []
if cond is not None:
for x in cond:
if 'default' in x:
default_c.append(x)
has_default_conds = True
continue
p = get_area_and_mult(x, x_in, timestep)
if p is None:
continue
if p.hooks is not None:
model.current_patcher.prepare_hook_patches_current_keyframe(timestep, p.hooks, model_options)
hooked_to_run.setdefault(p.hooks, list())
hooked_to_run[p.hooks] += [(p, i)]
default_conds.append(default_c)
if has_default_conds:
finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
model.current_patcher.prepare_state(timestep, model_options)
devices = list(model_options['multigpu_clones'].keys())
device_batched_hooked_to_run: dict[torch.device, list[tuple[comfy.hooks.HookGroup, tuple]]] = {}
# Track conds currently scheduled per device; single source of truth for capacity checks.
device_load: dict[torch.device, int] = {d: 0 for d in devices}
total_conds = sum(len(to_run) for to_run in hooked_to_run.values())
conds_per_device = max(1, math.ceil(total_conds / len(devices)))
def next_available_device(start: int) -> tuple[int, torch.device]:
"""Return (index, device) for the next device with remaining capacity, starting at `start`.
Scans at most len(devices) positions, so this always terminates. Raises if no device
has remaining capacity, which would indicate a bug in conds_per_device accounting.
"""
for offset in range(len(devices)):
i = (start + offset) % len(devices)
if device_load[devices[i]] < conds_per_device:
return i, devices[i]
raise RuntimeError(
f"MultiGPU scheduler: all {len(devices)} devices at capacity "
f"({conds_per_device}) but conds remain to schedule"
)
# run every hooked_to_run separately
index_device = 0
for hooks, to_run in hooked_to_run.items():
while len(to_run) > 0:
index_device, current_device = next_available_device(index_device)
remaining_capacity = conds_per_device - device_load[current_device]
first = to_run[0]
first_shape = first[0][0].shape
# collect candidate indices that can be concatenated with `first`, up to remaining capacity
to_batch_temp = []
for x in range(len(to_run)):
if can_concat_cond(to_run[x][0], first[0]) and len(to_batch_temp) < remaining_capacity:
to_batch_temp += [x]
to_batch_temp.reverse()
to_batch = to_batch_temp[:1]
free_memory = comfy.model_management.get_free_memory(current_device)
for i in range(1, len(to_batch_temp) + 1):
batch_amount = to_batch_temp[:len(to_batch_temp)//i]
input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
cond_shapes = collections.defaultdict(list)
for tt in batch_amount:
for k, v in to_run[tt][0].conditioning.items():
cond_shapes[k].append(v.size())
if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory:
to_batch = batch_amount
break
conds_to_batch = [to_run.pop(x) for x in to_batch]
device_load[current_device] += len(conds_to_batch)
device_batched_hooked_to_run.setdefault(current_device, []).append((hooks, conds_to_batch))
if device_load[current_device] >= conds_per_device:
index_device += 1
class thread_result(NamedTuple):
output: Any
mult: Any
area: Any
batch_chunks: int
cond_or_uncond: Any
error: Exception = None
def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]):
try:
# TODO: non-NVIDIA support -- guard with `if device.type == "cuda":` once
# we extend multigpu QA beyond CUDA. Unconditional call crashes on
# XPU/NPU/MPS/CPU/DirectML backends.
torch.cuda.set_device(device)
model_current: BaseModel = model_options["multigpu_clones"][device].model
# run every hooked_to_run separately
with torch.no_grad():
for hooks, to_batch in batch_tuple:
input_x = []
mult = []
c = []
cond_or_uncond = []
uuids = []
area = []
control: ControlBase = None
patches = None
for x in to_batch:
o = x
p = o[0]
input_x.append(p.input_x)
mult.append(p.mult)
c.append(p.conditioning)
area.append(p.area)
cond_or_uncond.append(o[1])
uuids.append(p.uuid)
control = p.control
patches = p.patches
batch_chunks = len(cond_or_uncond)
input_x = torch.cat(input_x).to(device)
c = cond_cat(c, device=device)
timestep_ = torch.cat([timestep.to(device)] * batch_chunks)
transformer_options = model_current.current_patcher.apply_hooks(hooks=hooks)
if 'transformer_options' in model_options:
transformer_options = comfy.patcher_extension.merge_nested_dicts(transformer_options,
model_options['transformer_options'],
copy_dict1=False)
if patches is not None:
transformer_options["patches"] = comfy.patcher_extension.merge_nested_dicts(
transformer_options.get("patches", {}),
patches
)
transformer_options["cond_or_uncond"] = cond_or_uncond[:]
transformer_options["uuids"] = uuids[:]
transformer_options["sigmas"] = timestep.to(device)
transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device)
transformer_options["multigpu_thread_device"] = device
cast_transformer_options(transformer_options, device=device)
c['transformer_options'] = transformer_options
if control is not None:
device_control = control.get_instance_for_device(device)
c['control'] = device_control.get_control(input_x, timestep_, c, len(cond_or_uncond), transformer_options)
if 'model_function_wrapper' in model_options:
output = model_options['model_function_wrapper'](model_current.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).to(output_device).chunk(batch_chunks)
else:
output = model_current.apply_model(input_x, timestep_, **c).to(output_device).chunk(batch_chunks)
# TODO: non-NVIDIA support -- the `.to(output_device)` copies
# above are async on CUDA, so the main thread's aggregation
# could race with in-flight transfers. CUDA-only QA has not
# surfaced this in practice, but before extending multigpu
# beyond NVIDIA add a `torch.cuda.synchronize(output_device)`
# here (guarded by `output_device.type == "cuda"`).
results.append(thread_result(output, mult, area, batch_chunks, cond_or_uncond))
except Exception as e:
results.append(thread_result(None, None, None, None, None, error=e))
raise
def _handle_batch_pooled(device, batch_tuple):
worker_results = []
_handle_batch(device, batch_tuple, worker_results)
return worker_results
results: list[thread_result] = []
thread_pool: comfy.multigpu.MultiGPUThreadPool = model_options.get("multigpu_thread_pool")
# Submit all GPU work to pool threads
pool_devices = []
for device, batch_tuple in device_batched_hooked_to_run.items():
if thread_pool is not None:
thread_pool.submit(device, _handle_batch_pooled, device, batch_tuple)
pool_devices.append(device)
else:
# Fallback: no pool, run everything on main thread
_handle_batch(device, batch_tuple, results)
# Collect results from pool workers
for device in pool_devices:
worker_results, error = thread_pool.get_result(device)
if error is not None:
raise error
results.extend(worker_results)
for output, mult, area, batch_chunks, cond_or_uncond, error in results:
if error is not None:
raise error
for o in range(batch_chunks):
cond_index = cond_or_uncond[o]
a = area[o]
if a is None:
out_conds[cond_index] += output[o] * mult[o]
out_counts[cond_index] += mult[o]
else:
out_c = out_conds[cond_index]
out_cts = out_counts[cond_index]
dims = len(a) // 2
for i in range(dims):
out_c = out_c.narrow(i + 2, a[i + dims], a[i])
out_cts = out_cts.narrow(i + 2, a[i + dims], a[i])
out_c += output[o] * mult[o]
out_cts += mult[o]
for i in range(len(out_conds)):
out_conds[i] /= out_counts[i]
return out_conds
def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove
logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.") logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.")
return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options)) return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options))
@ -642,12 +885,21 @@ def calculate_start_end_timesteps(model, conds):
def pre_run_control(model, conds): def pre_run_control(model, conds):
s = model.model_sampling s = model.model_sampling
# Per-device model lookup so multigpu control clones get the matching
# diffusion_model (e.g. QwenFunControlNet stashes it into extra_args).
device_models: dict = {}
patcher = getattr(model, "current_patcher", None)
if patcher is not None:
for p in patcher.get_additional_models_with_key("multigpu"):
device_models[p.load_device] = p.model
for t in range(len(conds)): for t in range(len(conds)):
x = conds[t] x = conds[t]
percent_to_timestep_function = lambda a: s.percent_to_sigma(a) percent_to_timestep_function = lambda a: s.percent_to_sigma(a)
if 'control' in x: if 'control' in x:
x['control'].pre_run(model, percent_to_timestep_function) x['control'].pre_run(model, percent_to_timestep_function)
for device, device_cnet in x['control'].multigpu_clones.items():
device_cnet.pre_run(device_models.get(device, model), percent_to_timestep_function)
def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func): def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
cond_cnets = [] cond_cnets = []
@ -890,7 +1142,9 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
to_load_options = model_options.get("to_load_options", None) to_load_options = model_options.get("to_load_options", None)
if to_load_options is None: if to_load_options is None:
return return
cast_transformer_options(to_load_options, device, dtype)
def cast_transformer_options(transformer_options: dict[str], device=None, dtype=None):
casts = [] casts = []
if device is not None: if device is not None:
casts.append(device) casts.append(device)
@ -899,18 +1153,17 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
# if nothing to apply, do nothing # if nothing to apply, do nothing
if len(casts) == 0: if len(casts) == 0:
return return
# try to call .to on patches # try to call .to on patches
if "patches" in to_load_options: if "patches" in transformer_options:
patches = to_load_options["patches"] patches = transformer_options["patches"]
for name in patches: for name in patches:
patch_list = patches[name] patch_list = patches[name]
for i in range(len(patch_list)): for i in range(len(patch_list)):
if hasattr(patch_list[i], "to"): if hasattr(patch_list[i], "to"):
for cast in casts: for cast in casts:
patch_list[i] = patch_list[i].to(cast) patch_list[i] = patch_list[i].to(cast)
if "patches_replace" in to_load_options: if "patches_replace" in transformer_options:
patches = to_load_options["patches_replace"] patches = transformer_options["patches_replace"]
for name in patches: for name in patches:
patch_list = patches[name] patch_list = patches[name]
for k in patch_list: for k in patch_list:
@ -920,8 +1173,8 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
# try to call .to on any wrappers/callbacks # try to call .to on any wrappers/callbacks
wrappers_and_callbacks = ["wrappers", "callbacks"] wrappers_and_callbacks = ["wrappers", "callbacks"]
for wc_name in wrappers_and_callbacks: for wc_name in wrappers_and_callbacks:
if wc_name in to_load_options: if wc_name in transformer_options:
wc: dict[str, list] = to_load_options[wc_name] wc: dict[str, list] = transformer_options[wc_name]
for wc_dict in wc.values(): for wc_dict in wc.values():
for wc_list in wc_dict.values(): for wc_list in wc_dict.values():
for i in range(len(wc_list)): for i in range(len(wc_list)):
@ -929,7 +1182,6 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
for cast in casts: for cast in casts:
wc_list[i] = wc_list[i].to(cast) wc_list[i] = wc_list[i].to(cast)
class CFGGuider: class CFGGuider:
def __init__(self, model_patcher: ModelPatcher): def __init__(self, model_patcher: ModelPatcher):
self.model_patcher = model_patcher self.model_patcher = model_patcher
@ -984,16 +1236,32 @@ class CFGGuider:
self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options) self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
device = self.model_patcher.load_device device = self.model_patcher.load_device
noise = noise.to(device=device, dtype=torch.float32) multigpu_patchers = comfy.sampler_helpers.prepare_model_patcher_multigpu_clones(self.model_patcher, self.loaded_models, self.model_options)
latent_image = latent_image.to(device=device, dtype=torch.float32)
sigmas = sigmas.to(device)
cast_to_load_options(self.model_options, device=device, dtype=self.model_patcher.model_dtype())
try: # Create persistent thread pool for all GPU devices (main + extras)
self.model_patcher.pre_run() if multigpu_patchers:
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) extra_devices = [p.load_device for p in multigpu_patchers]
finally: all_devices = [device] + extra_devices
self.model_patcher.cleanup() self.model_options["multigpu_thread_pool"] = comfy.multigpu.MultiGPUThreadPool(all_devices)
with comfy.model_management.cuda_device_context(device):
try:
noise = noise.to(device=device, dtype=torch.float32)
latent_image = latent_image.to(device=device, dtype=torch.float32)
sigmas = sigmas.to(device)
cast_to_load_options(self.model_options, device=device, dtype=self.model_patcher.model_dtype())
self.model_patcher.pre_run()
for multigpu_patcher in multigpu_patchers:
multigpu_patcher.pre_run()
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
finally:
thread_pool = self.model_options.pop("multigpu_thread_pool", None)
if thread_pool is not None:
thread_pool.shutdown()
self.model_patcher.cleanup()
for multigpu_patcher in multigpu_patchers:
multigpu_patcher.cleanup()
comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models) comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models)
del self.inner_model del self.inner_model

382
comfy/sd.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import json import json
import torch import torch
from enum import Enum from enum import Enum
@ -336,41 +335,43 @@ class CLIP:
self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.cond_stage_model.set_clip_options({"projected_pooled": False})
self.load_model(tokens) self.load_model(tokens)
self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) device = self.patcher.load_device
self.cond_stage_model.set_clip_options({"execution_device": device})
all_hooks.reset() all_hooks.reset()
self.patcher.patch_hooks(None) self.patcher.patch_hooks(None)
if show_pbar: if show_pbar:
pbar = ProgressBar(len(scheduled_keyframes)) pbar = ProgressBar(len(scheduled_keyframes))
for scheduled_opts in scheduled_keyframes: with model_management.cuda_device_context(device):
t_range = scheduled_opts[0] for scheduled_opts in scheduled_keyframes:
# don't bother encoding any conds outside of start_percent and end_percent bounds t_range = scheduled_opts[0]
if "start_percent" in add_dict: # don't bother encoding any conds outside of start_percent and end_percent bounds
if t_range[1] < add_dict["start_percent"]: if "start_percent" in add_dict:
continue if t_range[1] < add_dict["start_percent"]:
if "end_percent" in add_dict: continue
if t_range[0] > add_dict["end_percent"]: if "end_percent" in add_dict:
continue if t_range[0] > add_dict["end_percent"]:
hooks_keyframes = scheduled_opts[1] continue
for hook, keyframe in hooks_keyframes: hooks_keyframes = scheduled_opts[1]
hook.hook_keyframe._current_keyframe = keyframe for hook, keyframe in hooks_keyframes:
# apply appropriate hooks with values that match new hook_keyframe hook.hook_keyframe._current_keyframe = keyframe
self.patcher.patch_hooks(all_hooks) # apply appropriate hooks with values that match new hook_keyframe
# perform encoding as normal self.patcher.patch_hooks(all_hooks)
o = self.cond_stage_model.encode_token_weights(tokens) # perform encoding as normal
cond, pooled = o[:2] o = self.cond_stage_model.encode_token_weights(tokens)
pooled_dict = {"pooled_output": pooled} cond, pooled = o[:2]
# add clip_start_percent and clip_end_percent in pooled pooled_dict = {"pooled_output": pooled}
pooled_dict["clip_start_percent"] = t_range[0] # add clip_start_percent and clip_end_percent in pooled
pooled_dict["clip_end_percent"] = t_range[1] pooled_dict["clip_start_percent"] = t_range[0]
# add/update any keys with the provided add_dict pooled_dict["clip_end_percent"] = t_range[1]
pooled_dict.update(add_dict) # add/update any keys with the provided add_dict
# add hooks stored on clip pooled_dict.update(add_dict)
self.add_hooks_to_dict(pooled_dict) # add hooks stored on clip
all_cond_pooled.append([cond, pooled_dict]) self.add_hooks_to_dict(pooled_dict)
if show_pbar: all_cond_pooled.append([cond, pooled_dict])
pbar.update(1) if show_pbar:
model_management.throw_exception_if_processing_interrupted() pbar.update(1)
model_management.throw_exception_if_processing_interrupted()
all_hooks.reset() all_hooks.reset()
return all_cond_pooled return all_cond_pooled
@ -384,8 +385,12 @@ class CLIP:
self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.cond_stage_model.set_clip_options({"projected_pooled": False})
self.load_model(tokens) self.load_model(tokens)
self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) device = self.patcher.load_device
o = self.cond_stage_model.encode_token_weights(tokens) self.cond_stage_model.set_clip_options({"execution_device": device})
with model_management.cuda_device_context(device):
o = self.cond_stage_model.encode_token_weights(tokens)
cond, pooled = o[:2] cond, pooled = o[:2]
if return_dict: if return_dict:
out = {"cond": cond, "pooled_output": pooled} out = {"cond": cond, "pooled_output": pooled}
@ -447,9 +452,12 @@ class CLIP:
self.cond_stage_model.reset_clip_options() self.cond_stage_model.reset_clip_options()
self.load_model(tokens) self.load_model(tokens)
device = self.patcher.load_device
self.cond_stage_model.set_clip_options({"layer": None}) self.cond_stage_model.set_clip_options({"layer": None})
self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) self.cond_stage_model.set_clip_options({"execution_device": device})
return self.cond_stage_model.generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed, presence_penalty=presence_penalty)
with model_management.cuda_device_context(device):
return self.cond_stage_model.generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed, presence_penalty=presence_penalty)
def decode(self, token_ids, skip_special_tokens=True): def decode(self, token_ids, skip_special_tokens=True):
return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens) return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens)
@ -1027,50 +1035,52 @@ class VAE:
do_tile = False do_tile = False
if self.latent_dim == 2 and samples_in.ndim == 5: if self.latent_dim == 2 and samples_in.ndim == 5:
samples_in = samples_in[:, :, 0] samples_in = samples_in[:, :, 0]
try:
memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
free_memory = self.patcher.get_free_memory(self.device)
batch_number = int(free_memory / memory_used)
batch_number = max(1, batch_number)
# Pre-allocate output for VAEs that support direct buffer writes with model_management.cuda_device_context(self.device):
preallocated = False try:
if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype()) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
preallocated = True free_memory = self.patcher.get_free_memory(self.device)
batch_number = int(free_memory / memory_used)
batch_number = max(1, batch_number)
for x in range(0, samples_in.shape[0], batch_number): # Pre-allocate output for VAEs that support direct buffer writes
samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype) preallocated = False
if preallocated: if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options) pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype())
else: preallocated = True
out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True)
if pixel_samples is None:
pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
pixel_samples[x:x+batch_number].copy_(out)
del out
self.process_output(pixel_samples[x:x+batch_number])
except Exception as e:
model_management.raise_non_oom(e)
logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
#NOTE: We don't know what tensors were allocated to stack variables at the time of the
#exception and the exception itself refs them all until we get out of this except block.
#So we just set a flag for tiler fallback so that tensor gc can happen once the
#exception is fully off the books.
do_tile = True
if do_tile: for x in range(0, samples_in.shape[0], batch_number):
comfy.model_management.soft_empty_cache() samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype)
dims = samples_in.ndim - 2 if preallocated:
if dims == 1 or self.extra_1d_channel is not None: self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options)
pixel_samples = self.decode_tiled_1d(samples_in) else:
elif dims == 2: out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True)
pixel_samples = self.decode_tiled_(samples_in) if pixel_samples is None:
elif dims == 3: pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
tile = 256 // self.spacial_compression_decode() pixel_samples[x:x+batch_number].copy_(out)
overlap = tile // 4 del out
pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) self.process_output(pixel_samples[x:x+batch_number])
except Exception as e:
model_management.raise_non_oom(e)
logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
#NOTE: We don't know what tensors were allocated to stack variables at the time of the
#exception and the exception itself refs them all until we get out of this except block.
#So we just set a flag for tiler fallback so that tensor gc can happen once the
#exception is fully off the books.
do_tile = True
if do_tile:
comfy.model_management.soft_empty_cache()
dims = samples_in.ndim - 2
if dims == 1 or self.extra_1d_channel is not None:
pixel_samples = self.decode_tiled_1d(samples_in)
elif dims == 2:
pixel_samples = self.decode_tiled_(samples_in)
elif dims == 3:
tile = 256 // self.spacial_compression_decode()
overlap = tile // 4
pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1) pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1)
return pixel_samples return pixel_samples
@ -1088,20 +1098,21 @@ class VAE:
if overlap is not None: if overlap is not None:
args["overlap"] = overlap args["overlap"] = overlap
if dims == 1 or self.extra_1d_channel is not None: with model_management.cuda_device_context(self.device):
args.pop("tile_y") if dims == 1 or self.extra_1d_channel is not None:
output = self.decode_tiled_1d(samples, **args) args.pop("tile_y")
elif dims == 2: output = self.decode_tiled_1d(samples, **args)
output = self.decode_tiled_(samples, **args) elif dims == 2:
elif dims == 3: output = self.decode_tiled_(samples, **args)
if overlap_t is None: elif dims == 3:
args["overlap"] = (1, overlap, overlap) if overlap_t is None:
else: args["overlap"] = (1, overlap, overlap)
args["overlap"] = (max(1, overlap_t), overlap, overlap) else:
if tile_t is not None: args["overlap"] = (max(1, overlap_t), overlap, overlap)
args["tile_t"] = max(2, tile_t) if tile_t is not None:
args["tile_t"] = max(2, tile_t)
output = self.decode_tiled_3d(samples, **args) output = self.decode_tiled_3d(samples, **args)
return output.movedim(1, -1) return output.movedim(1, -1)
def encode(self, pixel_samples): def encode(self, pixel_samples):
@ -1114,44 +1125,46 @@ class VAE:
pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0) pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
else: else:
pixel_samples = pixel_samples.unsqueeze(2) pixel_samples = pixel_samples.unsqueeze(2)
try:
memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) with model_management.cuda_device_context(self.device):
model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) try:
free_memory = self.patcher.get_free_memory(self.device) memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
batch_number = int(free_memory / max(1, memory_used)) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
batch_number = max(1, batch_number) free_memory = self.patcher.get_free_memory(self.device)
samples = None batch_number = int(free_memory / max(1, memory_used))
for x in range(0, pixel_samples.shape[0], batch_number): batch_number = max(1, batch_number)
pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype) samples = None
if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): for x in range(0, pixel_samples.shape[0], batch_number):
out = self.first_stage_model.encode(pixels_in, device=self.device) pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype)
if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
out = self.first_stage_model.encode(pixels_in, device=self.device)
else:
pixels_in = pixels_in.to(self.device)
out = self.first_stage_model.encode(pixels_in)
out = out.to(self.output_device).to(dtype=self.vae_output_dtype())
if samples is None:
samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
samples[x:x + batch_number] = out
except Exception as e:
model_management.raise_non_oom(e)
logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
#NOTE: We don't know what tensors were allocated to stack variables at the time of the
#exception and the exception itself refs them all until we get out of this except block.
#So we just set a flag for tiler fallback so that tensor gc can happen once the
#exception is fully off the books.
do_tile = True
if do_tile:
comfy.model_management.soft_empty_cache()
if self.latent_dim == 3:
tile = 256
overlap = tile // 4
samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
elif self.latent_dim == 1 or self.extra_1d_channel is not None:
samples = self.encode_tiled_1d(pixel_samples)
else: else:
pixels_in = pixels_in.to(self.device) samples = self.encode_tiled_(pixel_samples)
out = self.first_stage_model.encode(pixels_in)
out = out.to(self.output_device).to(dtype=self.vae_output_dtype())
if samples is None:
samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
samples[x:x + batch_number] = out
except Exception as e:
model_management.raise_non_oom(e)
logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
#NOTE: We don't know what tensors were allocated to stack variables at the time of the
#exception and the exception itself refs them all until we get out of this except block.
#So we just set a flag for tiler fallback so that tensor gc can happen once the
#exception is fully off the books.
do_tile = True
if do_tile:
comfy.model_management.soft_empty_cache()
if self.latent_dim == 3:
tile = 256
overlap = tile // 4
samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
elif self.latent_dim == 1 or self.extra_1d_channel is not None:
samples = self.encode_tiled_1d(pixel_samples)
else:
samples = self.encode_tiled_(pixel_samples)
return samples return samples
@ -1177,26 +1190,27 @@ class VAE:
if overlap is not None: if overlap is not None:
args["overlap"] = overlap args["overlap"] = overlap
if dims == 1: with model_management.cuda_device_context(self.device):
args.pop("tile_y") if dims == 1:
samples = self.encode_tiled_1d(pixel_samples, **args) args.pop("tile_y")
elif dims == 2: samples = self.encode_tiled_1d(pixel_samples, **args)
samples = self.encode_tiled_(pixel_samples, **args) elif dims == 2:
elif dims == 3: samples = self.encode_tiled_(pixel_samples, **args)
if tile_t is not None: elif dims == 3:
tile_t_latent = max(2, self.downscale_ratio[0](tile_t)) if tile_t is not None:
else: tile_t_latent = max(2, self.downscale_ratio[0](tile_t))
tile_t_latent = 9999 else:
args["tile_t"] = self.upscale_ratio[0](tile_t_latent) tile_t_latent = 9999
args["tile_t"] = self.upscale_ratio[0](tile_t_latent)
if overlap_t is None: if overlap_t is None:
args["overlap"] = (1, overlap, overlap) args["overlap"] = (1, overlap, overlap)
else: else:
args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), overlap, overlap) args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), overlap, overlap)
maximum = pixel_samples.shape[2] maximum = pixel_samples.shape[2]
maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum)) maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum))
samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args) samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args)
return samples return samples
@ -1720,12 +1734,52 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata, disable_dynamic=disable_dynamic) out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata, disable_dynamic=disable_dynamic)
if out is None: if out is None:
raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd))) raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
if output_model and out[0] is not None: if out[0] is not None:
out[0].cached_patcher_init = (load_checkpoint_guess_config_model_only, (ckpt_path, embedding_directory, model_options, te_model_options)) out[0].cached_patcher_init = (load_checkpoint_guess_config, (ckpt_path, False, False, False, embedding_directory, output_model, model_options, te_model_options), 0)
if output_clip and out[1] is not None: # Register reload factories for the CLIP and VAE produced by the same checkpoint so
out[1].patcher.cached_patcher_init = (load_checkpoint_guess_config_clip_only, (ckpt_path, embedding_directory, model_options, te_model_options)) # ModelPatcher.deepclone_multigpu can spawn per-device copies (Select{CLIP,VAE}Device,
# MultiGPU work-units, etc.) without falling back to copy.deepcopy of an
# already-loaded module.
if out[1] is not None and getattr(out[1], "patcher", None) is not None:
out[1].patcher.cached_patcher_init = (load_checkpoint_clip_patcher, (ckpt_path, embedding_directory, model_options, te_model_options))
if out[2] is not None and getattr(out[2], "patcher", None) is not None:
out[2].patcher.cached_patcher_init = (load_checkpoint_vae_patcher, (ckpt_path, embedding_directory, model_options, te_model_options))
return out return out
def load_checkpoint_clip_patcher(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
"""Reload only the CLIP patcher from a checkpoint. Used as the cached_patcher_init
factory for the CLIP returned by load_checkpoint_guess_config."""
_, clip, _, _ = load_checkpoint_guess_config(
ckpt_path,
output_vae=False,
output_clip=True,
output_clipvision=False,
embedding_directory=embedding_directory,
output_model=False,
model_options=model_options,
te_model_options=te_model_options,
disable_dynamic=disable_dynamic,
)
return clip.patcher
def load_checkpoint_vae_patcher(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
"""Reload only the VAE patcher from a checkpoint. Used as the cached_patcher_init
factory for the VAE returned by load_checkpoint_guess_config."""
_, _, vae, _ = load_checkpoint_guess_config(
ckpt_path,
output_vae=True,
output_clip=False,
output_clipvision=False,
embedding_directory=embedding_directory,
output_model=False,
model_options=model_options,
te_model_options=te_model_options,
disable_dynamic=disable_dynamic,
)
return vae.patcher
def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False): def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
model, *_ = load_checkpoint_guess_config(ckpt_path, False, False, False, model, *_ = load_checkpoint_guess_config(ckpt_path, False, False, False,
embedding_directory=embedding_directory, embedding_directory=embedding_directory,
@ -1752,7 +1806,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd) diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix) parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix)
weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix) weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
load_device = model_management.get_torch_device() load_device = model_options.get("load_device", model_management.get_torch_device())
custom_operations = model_options.get("custom_operations", None) custom_operations = model_options.get("custom_operations", None)
if custom_operations is None: if custom_operations is None:
@ -1792,13 +1846,15 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype) inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device) model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device)
ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher
model_patcher = ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device()) offload_device = model_options.get("offload_device", model_management.unet_offload_device())
model_patcher = ModelPatcher(model, load_device=load_device, offload_device=offload_device)
model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic()) model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic())
if output_vae: if output_vae:
vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True) vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True)
vae_sd = model_config.process_vae_state_dict(vae_sd) vae_sd = model_config.process_vae_state_dict(vae_sd)
vae = VAE(sd=vae_sd, metadata=metadata) vae_device = model_options.get("load_device", None)
vae = VAE(sd=vae_sd, metadata=metadata, device=vae_device)
if output_clip: if output_clip:
if te_model_options.get("custom_operations", None) is None: if te_model_options.get("custom_operations", None) is None:
@ -1882,7 +1938,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
parameters = comfy.utils.calculate_parameters(sd) parameters = comfy.utils.calculate_parameters(sd)
weight_dtype = comfy.utils.weight_dtype(sd) weight_dtype = comfy.utils.weight_dtype(sd)
load_device = model_management.get_torch_device() load_device = model_options.get("load_device", model_management.get_torch_device())
model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata) model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata)
if model_config is not None: if model_config is not None:
@ -1907,7 +1963,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
else: else:
logging.warning("{} {}".format(diffusers_keys[k], k)) logging.warning("{} {}".format(diffusers_keys[k], k))
offload_device = model_management.unet_offload_device() offload_device = model_options.get("offload_device", model_management.unet_offload_device())
unet_weight_dtype = list(model_config.supported_inference_dtypes) unet_weight_dtype = list(model_config.supported_inference_dtypes)
if model_config.quant_config is not None: if model_config.quant_config is not None:
weight_dtype = None weight_dtype = None
@ -1949,6 +2005,26 @@ def load_diffusion_model(unet_path, model_options={}, disable_dynamic=False):
model.cached_patcher_init = (load_diffusion_model, (unet_path, model_options)) model.cached_patcher_init = (load_diffusion_model, (unet_path, model_options))
return model return model
def load_vae_patcher(vae_path, metadata=None, device=None, disable_dynamic=False):
"""Reload a disk-backed VAE from ``vae_path`` and return its patcher.
Used as the ``cached_patcher_init`` factory on ``VAE.patcher`` so
:meth:`comfy.model_patcher.ModelPatcher.deepclone_multigpu` can produce a
fresh, untainted VAE patcher (no inherited per-device load state, no
in-place quantization fallout) for multigpu work-units and the
SelectVAEDevice node. The optional ``device`` matches the source loader's
VAE initialization path; the deepclone's ``load_device`` still controls
where the cloned patcher is targeted.
"""
if metadata is None:
sd, metadata = comfy.utils.load_torch_file(vae_path, return_metadata=True)
else:
sd = comfy.utils.load_torch_file(vae_path)
vae = VAE(sd=sd, metadata=metadata, device=device)
vae.throw_exception_if_invalid()
return vae.patcher
def load_unet(unet_path, dtype=None): def load_unet(unet_path, dtype=None):
logging.warning("The load_unet function has been deprecated and will be removed please switch to: load_diffusion_model") logging.warning("The load_unet function has been deprecated and will be removed please switch to: load_diffusion_model")
return load_diffusion_model(unet_path, model_options={"dtype": dtype}) return load_diffusion_model(unet_path, model_options={"dtype": dtype})

3
comfy/utils.py
View File

@ -86,6 +86,7 @@ def load_safetensors(ckpt):
import comfy_aimdo.model_mmap import comfy_aimdo.model_mmap
f = open(ckpt, "rb", buffering=0) f = open(ckpt, "rb", buffering=0)
file_lock = threading.Lock()
model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt) model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt)
file_size = os.path.getsize(ckpt) file_size = os.path.getsize(ckpt)
mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get())) mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get()))
@ -111,7 +112,7 @@ def load_safetensors(ckpt):
storage = tensor.untyped_storage() storage = tensor.untyped_storage()
setattr(storage, setattr(storage,
"_comfy_tensor_file_slice", "_comfy_tensor_file_slice",
comfy.memory_management.TensorFileSlice(f, threading.get_ident(), data_base_offset + start, end - start)) comfy.memory_management.TensorFileSlice(f, file_lock, data_base_offset + start, end - start))
setattr(storage, "_comfy_tensor_mmap_refs", (model_mmap, mv)) setattr(storage, "_comfy_tensor_mmap_refs", (model_mmap, mv))
sd[name] = tensor sd[name] = tensor

2
comfy_api/latest/__init__.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
from abc import ABC, abstractmethod from abc import ABC, abstractmethod
from typing import TYPE_CHECKING from typing import TYPE_CHECKING
from comfy_api.internal import ComfyAPIBase from comfy_api.internal import ComfyAPIBase

1
comfy_api/latest/_input_impl/video_types.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from av.container import InputContainer from av.container import InputContainer
from av.subtitles.stream import SubtitleStream from av.subtitles.stream import SubtitleStream
from fractions import Fraction from fractions import Fraction

1
comfy_api/latest/_util/video_types.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from dataclasses import dataclass from dataclasses import dataclass
from enum import Enum from enum import Enum
from fractions import Fraction from fractions import Fraction

1
comfy_api_nodes/apis/__init__.py generated
View File

@ -3,7 +3,6 @@
# timestamp: 2025-07-30T08:54:00+00:00 # timestamp: 2025-07-30T08:54:00+00:00
# pylint: disable # pylint: disable
from __future__ import annotations
from datetime import date, datetime from datetime import date, datetime
from enum import Enum from enum import Enum

2
comfy_api_nodes/apis/bfl.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
from enum import Enum from enum import Enum
from typing import Any, Dict, Optional from typing import Any, Dict, Optional

2
comfy_api_nodes/apis/stability.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
from enum import Enum from enum import Enum
from typing import Optional from typing import Optional

1
comfy_execution/graph.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from typing import Type, Literal from typing import Type, Literal
import nodes import nodes

2
comfy_execution/progress.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
from typing import TypedDict, Dict, Optional, Tuple from typing import TypedDict, Dict, Optional, Tuple
from typing_extensions import override from typing_extensions import override
from PIL import Image from PIL import Image

1
comfy_execution/validation.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from comfy_api.latest import IO from comfy_api.latest import IO

1
comfy_extras/mediapipe/face_geometry.py
View File

@ -2,7 +2,6 @@
+ weighted Procrustes solver. Computes the 4x4 facial transformation matrix. + weighted Procrustes solver. Computes the 4x4 facial transformation matrix.
""" """
from __future__ import annotations
import math import math
import numpy as np import numpy as np

1
comfy_extras/mediapipe/face_landmarker.py
View File

@ -1,7 +1,6 @@
"""Pure-PyTorch port of MediaPipe's face_landmarker_v2_with_blendshapes.task: """Pure-PyTorch port of MediaPipe's face_landmarker_v2_with_blendshapes.task:
BlazeFace detector → FaceMesh v2 → ARKit-52 blendshapes.""" BlazeFace detector → FaceMesh v2 → ARKit-52 blendshapes."""
from __future__ import annotations
import math import math
from functools import lru_cache from functools import lru_cache

2
comfy_extras/nodes_audio.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import av import av
import torchaudio import torchaudio
import torch import torch

1
comfy_extras/nodes_context_windows.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from comfy_api.latest import ComfyExtension, io from comfy_api.latest import ComfyExtension, io
import comfy.context_windows import comfy.context_windows
import nodes import nodes

2
comfy_extras/nodes_curve.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import numpy as np import numpy as np
from comfy_api.latest import ComfyExtension, io from comfy_api.latest import ComfyExtension, io

410
comfy_extras/nodes_images.py
View File

@ -1,17 +1,23 @@
from __future__ import annotations
import nodes import nodes
import folder_paths import folder_paths
import av
import json import json
import os import os
import re import re
import math import math
import numpy as np
import struct
import torch import torch
import zlib
import comfy.utils import comfy.utils
from fractions import Fraction
from server import PromptServer from server import PromptServer
from comfy_api.latest import ComfyExtension, IO, UI from comfy_api.latest import ComfyExtension, IO, UI
from comfy.cli_args import args
from typing_extensions import override from typing_extensions import override
SVG = IO.SVG.Type # TODO: temporary solution for backward compatibility, will be removed later. SVG = IO.SVG.Type # TODO: temporary solution for backward compatibility, will be removed later.
@ -835,6 +841,405 @@ class ImageMergeTileList(IO.ComfyNode):
return IO.NodeOutput(merged_image) return IO.NodeOutput(merged_image)
# ---------------------------------------------------------------------------
# Format specifications
# ---------------------------------------------------------------------------
# Maps (file_format, bit_depth, has_alpha) -> (numpy dtype scale, av pixel format,
# stream pix_fmt). Keeps the encode path declarative instead of branchy.
_FORMAT_SPECS = {
("png", "8-bit", False): {"scale": 255.0, "dtype": np.uint8, "frame_fmt": "rgb24", "stream_fmt": "rgb24"},
("png", "8-bit", True): {"scale": 255.0, "dtype": np.uint8, "frame_fmt": "rgba", "stream_fmt": "rgba"},
("png", "16-bit", False): {"scale": 65535.0, "dtype": np.uint16, "frame_fmt": "rgb48le", "stream_fmt": "rgb48be"},
("png", "16-bit", True): {"scale": 65535.0, "dtype": np.uint16, "frame_fmt": "rgba64le", "stream_fmt": "rgba64be"},
("exr", "32-bit float", False): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrpf32le", "stream_fmt": "gbrpf32le"},
("exr", "32-bit float", True): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrapf32le", "stream_fmt": "gbrapf32le"},
}
# ---------------------------------------------------------------------------
# Color transforms
# ---------------------------------------------------------------------------
def srgb_to_linear(t: torch.Tensor) -> torch.Tensor:
"""Inverse sRGB EOTF (IEC 61966-2-1). Operates on RGB channels only;
alpha (if present as the 4th channel) is passed through unchanged."""
if t.shape[-1] == 4:
rgb, alpha = t[..., :3], t[..., 3:]
return torch.cat([srgb_to_linear(rgb), alpha], dim=-1)
# Piecewise: linear toe below 0.04045, gamma curve above.
low = t / 12.92
high = ((t.clamp(min=0.0) + 0.055) / 1.055) ** 2.4
return torch.where(t <= 0.04045, low, high)
# HLG OETF constants from BT.2100 Table 5.
_HLG_A = 0.17883277
_HLG_B = 0.28466892
_HLG_C = 0.55991072928 # = 0.5 - a*ln(4*a)
def hlg_to_linear(t: torch.Tensor) -> torch.Tensor:
"""Inverse HLG OETF (BT.2100). Maps a non-linear HLG signal in [0, 1] to
*scene*-linear light in [0, 1]. Per BT.2100 Note 5a, this is the correct
transform when converting HLG to a linear scene-light representation
(rather than display-light, which would also involve the HLG OOTF).
Operates on RGB channels only; alpha is passed through unchanged."""
if t.shape[-1] == 4:
rgb, alpha = t[..., :3], t[..., 3:]
return torch.cat([hlg_to_linear(rgb), alpha], dim=-1)
# Piecewise: sqrt branch below 0.5, log branch above.
# Clamp inside the log branch so negative / out-of-range values don't blow up;
# values above 1.0 are allowed and extrapolate naturally.
low = (t ** 2) / 3.0
high = (torch.exp((t.clamp(min=_HLG_C) - _HLG_C) / _HLG_A) + _HLG_B) / 12.0
return torch.where(t <= 0.5, low, high)
# ---------------------------------------------------------------------------
# Metadata injection
# ---------------------------------------------------------------------------
_PNG_SIGNATURE = b"\x89PNG\r\n\x1a\n"
def _png_chunk(chunk_type: bytes, data: bytes) -> bytes:
"""Build a single PNG chunk: length | type | data | CRC32(type+data)."""
crc = zlib.crc32(chunk_type + data) & 0xFFFFFFFF
return struct.pack(">I", len(data)) + chunk_type + data + struct.pack(">I", crc)
def _png_text_chunk(keyword: str, text: str) -> bytes:
"""tEXt chunk: latin-1 keyword + NUL + latin-1 text."""
payload = keyword.encode("latin-1") + b"\x00" + text.encode("latin-1", errors="replace")
return _png_chunk(b"tEXt", payload)
def inject_png_metadata(png_bytes: bytes, prompt: dict | None, extra_pnginfo: dict | None) -> bytes:
"""Insert ComfyUI prompt/workflow as tEXt chunks right after IHDR."""
if not png_bytes.startswith(_PNG_SIGNATURE):
return png_bytes
chunks: list[bytes] = []
if prompt is not None:
chunks.append(_png_text_chunk("prompt", json.dumps(prompt)))
if extra_pnginfo:
for key, value in extra_pnginfo.items():
chunks.append(_png_text_chunk(key, json.dumps(value)))
if not chunks:
return png_bytes
# IHDR is always the first chunk; insert ours immediately after it.
ihdr_length = struct.unpack(">I", png_bytes[8:12])[0]
ihdr_end = 8 + 8 + ihdr_length + 4 # signature + (len+type) + data + crc
return png_bytes[:ihdr_end] + b"".join(chunks) + png_bytes[ihdr_end:]
# Standard chromaticities (CIE 1931 xy) for the colorspaces this node writes.
# Each tuple is (Rx, Ry, Gx, Gy, Bx, By, Wx, Wy). All share D65 white point.
_CHROMATICITIES = {
# ITU-R BT.709 / sRGB primaries
"Rec.709": (0.6400, 0.3300, 0.3000, 0.6000, 0.1500, 0.0600, 0.3127, 0.3290),
# ITU-R BT.2020 (UHDTV / wide-gamut HDR) primaries
"Rec.2020": (0.7080, 0.2920, 0.1700, 0.7970, 0.1310, 0.0460, 0.3127, 0.3290),
}
def _pack_chromaticities(primaries: tuple) -> bytes:
"""Serialize 8 chromaticity floats into the EXR `chromaticities` payload."""
return struct.pack("<8f", *primaries)
def _exr_attribute(name: str, attr_type: str, value: bytes) -> bytes:
"""Serialize one EXR header attribute: name\\0 type\\0 size:int32 value."""
return (
name.encode("utf-8") + b"\x00"
+ attr_type.encode("utf-8") + b"\x00"
+ struct.pack("<i", len(value))
+ value
)
def inject_exr_metadata(
exr_bytes: bytes,
prompt: dict | None,
extra_pnginfo: dict | None,
colorspace: str | None = None,
) -> bytes:
"""Insert ComfyUI metadata and color-space info into an EXR header.
Color: EXR pixels are linear by convention. The standard way to describe
their RGB→XYZ relationship is the `chromaticities` attribute. We pick the
primaries that match what the user told us their input was:
colorspace="sRGB" → Rec. 709 / sRGB primaries (D65)
colorspace="HDR" → Rec. 2020 / BT.2100 primaries (D65)
Pixels are always converted to linear scene light upstream (sRGB EOTF
inverse for sRGB; HLG OETF inverse for HDR), so the file content is
scene-linear in the indicated gamut. OpenEXR has no standard transfer-
function attribute (the OpenEXR TSC has discussed adding one but it
doesn't exist), so we don't invent one — `chromaticities` plus the EXR
linear-by-convention rule fully specifies the color.
Prompt/workflow: written as plain `string` attributes using the same keys
(`prompt`, `workflow`, ...) that Comfy uses for PNG tEXt chunks, so the
same readers can pull them out symmetrically.
Implementation note: the chunk-offset table that follows the header stores
*absolute* byte offsets into the file. Inserting N bytes into the header
means every offset must be incremented by N or the file becomes unreadable.
"""
if len(exr_bytes) < 8 or exr_bytes[:4] != b"\x76\x2f\x31\x01":
return exr_bytes
new_blob = b""
if prompt is not None:
new_blob += _exr_attribute("prompt", "string", json.dumps(prompt).encode("utf-8"))
if extra_pnginfo:
for key, value in extra_pnginfo.items():
new_blob += _exr_attribute(key, "string", json.dumps(value).encode("utf-8"))
if colorspace is not None:
# Map each colorspace option to the RGB primaries the linear pixels
# are now in. "sRGB" and "linear" both produce Rec. 709 linear; "HDR"
# (HLG-encoded Rec. 2020 input) produces Rec. 2020 linear.
primaries_name = {
"sRGB": "Rec.709",
"linear": "Rec.709",
"HDR": "Rec.2020",
}.get(colorspace, "Rec.709")
new_blob += _exr_attribute(
"chromaticities",
"chromaticities",
_pack_chromaticities(_CHROMATICITIES[primaries_name]),
)
if not new_blob:
return exr_bytes
# Walk header attributes to find the terminating null byte, and pick up
# dataWindow + compression so we know how many chunks the offset table has.
pos = 8 # past magic (4) + version (4)
data_window = None
compression = 0
while pos < len(exr_bytes) and exr_bytes[pos] != 0:
name_end = exr_bytes.index(b"\x00", pos)
attr_name = exr_bytes[pos:name_end].decode("latin-1", errors="replace")
type_end = exr_bytes.index(b"\x00", name_end + 1)
attr_type = exr_bytes[name_end + 1:type_end].decode("latin-1", errors="replace")
size = struct.unpack("<i", exr_bytes[type_end + 1:type_end + 5])[0]
value_start = type_end + 5
value = exr_bytes[value_start:value_start + size]
if attr_name == "dataWindow" and attr_type == "box2i":
data_window = struct.unpack("<iiii", value) # xMin, yMin, xMax, yMax
elif attr_name == "compression" and attr_type == "compression":
compression = value[0]
pos = value_start + size
if data_window is None:
return exr_bytes # required attribute missing — don't risk corrupting
# Scanlines per chunk by compression, from the OpenEXR spec.
scanlines_per_block = {
0: 1, # NO_COMPRESSION
1: 1, # RLE
2: 1, # ZIPS
3: 16, # ZIP
4: 32, # PIZ
5: 16, # PXR24
6: 32, # B44
7: 32, # B44A
8: 256, # DWAA
9: 256, # DWAB
}.get(compression, 1)
_, y_min, _, y_max = data_window
height = y_max - y_min + 1
num_chunks = (height + scanlines_per_block - 1) // scanlines_per_block
header_end = pos # position of the terminating null byte
table_start = header_end + 1
pixel_start = table_start + num_chunks * 8
delta = len(new_blob)
old_offsets = struct.unpack(f"<{num_chunks}Q", exr_bytes[table_start:pixel_start])
new_table = struct.pack(f"<{num_chunks}Q", *(o + delta for o in old_offsets))
return (
exr_bytes[:header_end] # header attributes
+ new_blob # our new attributes
+ exr_bytes[header_end:table_start] # terminating null byte
+ new_table # shifted offset table
+ exr_bytes[pixel_start:] # pixel data, untouched
)
# ---------------------------------------------------------------------------
# Encoding
# ---------------------------------------------------------------------------
def _encode_image(
img_tensor: torch.Tensor,
file_format: str,
bit_depth: str,
colorspace: str,
) -> bytes:
"""Encode a single HxWxC tensor to PNG or EXR bytes in memory.
For EXR the input is interpreted according to `colorspace` and converted
to scene-linear (EXR's convention) before writing:
"sRGB" → input is sRGB-encoded Rec. 709; apply inverse sRGB EOTF.
"HDR" → input is HLG-encoded Rec. 2020 (BT.2100); apply inverse HLG
OETF to get scene-linear, per BT.2100 Note 5a.
"linear" → input is already scene-linear (Rec. 709 primaries); write
through unchanged. Use this for renderer/compositor output.
For PNG, colorspace selection does not modify pixels — PNG is delivered
sRGB-encoded and there is no PNG path for wide-gamut HDR in this node.
"""
height, width, num_channels = img_tensor.shape
has_alpha = num_channels == 4
spec = _FORMAT_SPECS[(file_format, bit_depth, has_alpha)]
if spec["dtype"] == np.float32:
# EXR path: preserve full range, no clamp.
if colorspace == "sRGB":
img_tensor = srgb_to_linear(img_tensor)
elif colorspace == "HDR":
img_tensor = hlg_to_linear(img_tensor)
img_np = img_tensor.cpu().numpy().astype(np.float32)
else:
# PNG path: quantize to integer range.
scaled = (img_tensor * spec["scale"]).clamp(0, spec["scale"])
img_np = scaled.to(torch.int32).cpu().numpy().astype(spec["dtype"])
# Encode directly via CodecContext. PyAV's `image2` muxer does NOT write to
# BytesIO (it expects a real file path), so we bypass the container entirely.
# For single-frame PNG/EXR the raw codec output IS the file.
codec = av.CodecContext.create(file_format, "w")
codec.width = width
codec.height = height
codec.pix_fmt = spec["stream_fmt"]
codec.time_base = Fraction(1, 1)
frame = av.VideoFrame.from_ndarray(img_np, format=spec["frame_fmt"])
if spec["frame_fmt"] != spec["stream_fmt"]:
frame = frame.reformat(format=spec["stream_fmt"])
frame.pts = 0
frame.time_base = codec.time_base
packets = list(codec.encode(frame)) + list(codec.encode(None)) # flush with None
return b"".join(bytes(p) for p in packets)
# ---------------------------------------------------------------------------
# Node
# ---------------------------------------------------------------------------
class SaveImageAdvanced(IO.ComfyNode):
@classmethod
def define_schema(cls):
return IO.Schema(
node_id="SaveImageAdvanced",
search_aliases=["save", "save image", "export image", "output image", "write image"],
display_name="Save Image (Advanced)",
description="Saves the input images to your ComfyUI output directory.",
category="image",
essentials_category="Basics",
inputs=[
IO.Image.Input("images", tooltip="The images to save."),
IO.String.Input(
"filename_prefix",
default="ComfyUI",
tooltip=(
"The prefix for the file to save. May include formatting tokens "
"such as %date:yyyy-MM-dd% or %Empty Latent Image.width%."
),
),
IO.DynamicCombo.Input(
"format",
options=[
IO.DynamicCombo.Option("png", [
IO.Combo.Input("bit_depth", options=["8-bit", "16-bit"],
default="8-bit", advanced=True),
IO.Combo.Input("input_color_space", options=["sRGB"],
default="sRGB", advanced=True),
]),
IO.DynamicCombo.Option("exr", [
IO.Combo.Input("bit_depth", options=["32-bit float"],
default="32-bit float", advanced=True),
IO.Combo.Input(
"input_color_space",
options=["sRGB", "HDR", "linear"],
default="sRGB",
advanced=True,
tooltip=(
"Colorspace of the input tensor. The EXR is "
"always written as scene-linear in the matching "
"gamut.\n"
" 'sRGB' — input is sRGB-encoded Rec.709; "
"the inverse sRGB EOTF is applied.\n"
" 'HDR' — input is HLG-encoded Rec.2020 "
"(BT.2100); the inverse HLG OETF is applied "
"to get scene-linear light.\n"
" 'linear' — input is already scene-linear "
"(Rec.709 primaries); written through unchanged. "
"Use this for renderer/compositor output."
),
),
]),
],
tooltip="The file format in which to save the image.",
),
],
hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo],
is_output_node=True,
)
@classmethod
def execute(cls, images, filename_prefix: str, format: dict) -> IO.NodeOutput:
file_format = format["format"]
bit_depth = format["bit_depth"]
colorspace = format.get("input_color_space", "sRGB")
output_dir = folder_paths.get_output_directory()
full_output_folder, filename, counter, subfolder, filename_prefix = (
folder_paths.get_save_image_path(
filename_prefix, output_dir, images[0].shape[1], images[0].shape[0]
)
)
prompt = cls.hidden.prompt
extra_pnginfo = cls.hidden.extra_pnginfo
write_metadata = not args.disable_metadata
results = []
for batch_number, image in enumerate(images):
encoded = _encode_image(image, file_format, bit_depth, colorspace)
if write_metadata:
if file_format == "png":
encoded = inject_png_metadata(encoded, prompt, extra_pnginfo)
elif file_format == "exr":
encoded = inject_exr_metadata(encoded, prompt, extra_pnginfo, colorspace)
name = filename.replace("%batch_num%", str(batch_number))
file = f"{name}_{counter:05}.{file_format}"
with open(os.path.join(full_output_folder, file), "wb") as f:
f.write(encoded)
results.append({"filename": file, "subfolder": subfolder, "type": "output"})
counter += 1
return IO.NodeOutput(ui={"images": results})
class ImagesExtension(ComfyExtension): class ImagesExtension(ComfyExtension):
@override @override
async def get_node_list(self) -> list[type[IO.ComfyNode]]: async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@ -847,6 +1252,7 @@ class ImagesExtension(ComfyExtension):
ImageAddNoise, ImageAddNoise,
SaveAnimatedWEBP, SaveAnimatedWEBP,
SaveAnimatedPNG, SaveAnimatedPNG,
SaveImageAdvanced,
SaveSVGNode, SaveSVGNode,
ImageStitch, ImageStitch,
ResizeAndPadImage, ResizeAndPadImage,

1
comfy_extras/nodes_logic.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from typing import TypedDict from typing import TypedDict
from typing_extensions import override from typing_extensions import override
from comfy_api.latest import ComfyExtension, io from comfy_api.latest import ComfyExtension, io

1
comfy_extras/nodes_math.py
View File

@ -4,7 +4,6 @@ Provides a ComfyMathExpression node that evaluates math expressions
against dynamically-grown numeric inputs. against dynamically-grown numeric inputs.
""" """
from __future__ import annotations
import math import math
import string import string

1
comfy_extras/nodes_mediapipe.py
View File

@ -10,7 +10,6 @@ Custom IO types:
MediaPipeFaceLandmarker also emits the core BOUNDING_BOX type — pair with DrawBBoxes. MediaPipeFaceLandmarker also emits the core BOUNDING_BOX type — pair with DrawBBoxes.
""" """
from __future__ import annotations
import numpy as np import numpy as np
import torch import torch

1
comfy_extras/nodes_moge.py
View File

@ -1,6 +1,5 @@
"""ComfyUI nodes for the native MoGe (Monocular Geometry Estimation) integration.""" """ComfyUI nodes for the native MoGe (Monocular Geometry Estimation) integration."""
from __future__ import annotations
import torch import torch

408
comfy_extras/nodes_multigpu.py Normal file
View File

@ -0,0 +1,408 @@
from __future__ import annotations
import copy
import logging
from inspect import cleandoc
from typing import TYPE_CHECKING
from typing_extensions import override
from comfy_api.latest import ComfyExtension, io
if TYPE_CHECKING:
from comfy.model_patcher import ModelPatcher
from comfy.sd import CLIP, VAE
import torch
import comfy.model_management
import comfy.multigpu
class MultiGPUCFGSplitNode(io.ComfyNode):
"""
Prepares model to have sampling accelerated via splitting work units.
Should be placed after nodes that modify the model object itself, such as compile or attention-switch nodes.
Other than those exceptions, this node can be placed in any order.
"""
@classmethod
def define_schema(cls):
return io.Schema(
node_id="MultiGPU_WorkUnits",
display_name="MultiGPU CFG Split",
category="advanced/multigpu",
description=cleandoc(cls.__doc__),
inputs=[
io.Model.Input("model"),
io.Int.Input("max_gpus", default=2, min=1, step=1),
],
outputs=[
io.Model.Output(),
],
)
@classmethod
def execute(cls, model: ModelPatcher, max_gpus: int) -> io.NodeOutput:
model = comfy.multigpu.create_multigpu_deepclones(model, max_gpus, reuse_loaded=True)
return io.NodeOutput(model)
def _force_supported_compute_dtype(patcher: ModelPatcher, device: torch.device):
"""Cast compute dtype to one the device supports; no-op if already supported."""
weight_dtype = patcher.model_dtype()
cast_dtype = comfy.model_management.unet_manual_cast(weight_dtype, device)
if cast_dtype is None:
return
logging.info(f"Select Model Device: using {cast_dtype} compute dtype on {device} (model weight dtype was {weight_dtype}).")
patcher.set_model_compute_dtype(cast_dtype)
def _remember_base_devices(patcher: ModelPatcher):
"""Stash the original load/offload device on the underlying model.
Stored on patcher.model (which is shared with the input patcher), so
later "default" selections can recover the loader's original routing.
Only the first Select on a given chain writes these attrs; subsequent
deepclones inherit them onto their freshly-loaded model below.
"""
if not hasattr(patcher.model, "_select_base_load_device"):
patcher.model._select_base_load_device = patcher.load_device
patcher.model._select_base_offload_device = patcher.offload_device
def _propagate_base_devices(src_model, dst_model):
"""Carry the loader-original device attrs onto the freshly-deepcloned model."""
if hasattr(src_model, "_select_base_load_device") and not hasattr(dst_model, "_select_base_load_device"):
dst_model._select_base_load_device = src_model._select_base_load_device
dst_model._select_base_offload_device = src_model._select_base_offload_device
def _retarget_patcher(patcher: ModelPatcher, target_load_device, target_offload_device):
"""Return a patcher whose actual model weights live on *target_load_device*.
If *patcher* is already on *target_load_device* we just retarget the
(already-cloned) patcher's metadata in place. Otherwise we call
:meth:`ModelPatcher.deepclone_multigpu` to spawn a fresh model from
the loader's ``cached_patcher_init`` factory -- the only safe way to
move weights that may already be partially loaded onto another device.
NOTE: reusing the input patcher's model when the requested device
matches its current load_device is a deliberate fast path. Anything
that has already mutated the original model (e.g. a prior KSampler
invocation on the same model) will be observed here. This is by
design and documented on the SelectXDeviceNode docstrings -- placing
Select X Device after a node that consumes the same model is not
recommended.
"""
if patcher.load_device == target_load_device:
# Fast path: weights already on the desired device, just update offload.
patcher.offload_device = target_offload_device
return patcher
src_model = patcher.model
patcher = patcher.deepclone_multigpu(new_load_device=target_load_device)
patcher.offload_device = target_offload_device
_propagate_base_devices(src_model, patcher.model)
if hasattr(patcher, "register_load_device"):
patcher.register_load_device(patcher.load_device)
return patcher
def _apply_patcher_device(patcher: ModelPatcher, resolved, base_offload_override=None):
"""Resolve the requested device and produce a patcher routed there.
For "default" we restore the loader's original load/offload pair.
For CPU we pin both load and offload to CPU (and, on a dynamic
patcher, downgrade to a plain ModelPatcher so the dynamic-only
code paths are bypassed).
For an explicit GPU we keep the loader's original offload but
target the requested load device; if that differs from the current
load device the patcher is deepcloned onto the new device.
"""
_remember_base_devices(patcher)
base_load = patcher.model._select_base_load_device
base_offload = base_offload_override if base_offload_override is not None else patcher.model._select_base_offload_device
if resolved is None:
# "default" -> route back to the loader's original devices.
return _retarget_patcher(patcher, base_load, base_offload)
if resolved.type == "cpu":
if patcher.is_dynamic():
# clone(disable_dynamic=True) requires cached_patcher_init; let the
# exception surface to the caller (Select*DeviceNode.execute), which
# will translate it into a passthrough+log so unsupported loaders
# don't hard-fail the workflow.
patcher = patcher.clone(disable_dynamic=True)
patcher.load_device = resolved
patcher.offload_device = resolved
return patcher
return _retarget_patcher(patcher, resolved, base_offload)
def _prune_multigpu_collision(model: ModelPatcher, primary_device):
"""Drop any multigpu clone whose load_device matches *primary_device*.
Without pruning, MultiGPU CFG Split would have stacked a clone on
the same device the primary now occupies (i.e. the workflow places
MultiGPU CFG Split before Select Model Device). Keeps the clone set
consistent with the new primary placement.
"""
multigpu_models = model.get_additional_models_with_key("multigpu")
if not multigpu_models:
return
filtered = [m for m in multigpu_models if m.load_device != primary_device]
if len(filtered) != len(multigpu_models):
logging.info(f"Select Model Device: pruning MultiGPU clone on {primary_device} that now collides with the primary model.")
model.set_additional_models("multigpu", filtered)
if hasattr(model, "match_multigpu_clones"):
model.match_multigpu_clones()
class SelectModelDeviceNode(io.ComfyNode):
"""
Place the diffusion model on a specific device (default / cpu / gpu:N).
- "default" restores the device assigned by the loader (even after a
prior Select Model Device call).
- "cpu" pins both the load and offload device to CPU.
- "gpu:N" pins the load device to the Nth available GPU; the offload
device is restored to the loader's original choice.
When the requested device differs from the device the input model is
already on, a fresh model is spawned via the loader's reload factory
(cached_patcher_init) so the new patcher owns independent weights on
the new device. Loaders that don't support multigpu (no factory) will
cause the node to pass through unchanged with a warning.
If the workflow already has MultiGPU CFG Split applied and the chosen
GPU collides with one of the existing multigpu clones, that clone is
dropped so two patchers don't end up bound to the same device.
When the selected device does not exist on the current machine
(e.g. a workflow built on a 2-GPU box opened on a 1-GPU box),
the node passes the model through unchanged and logs a message
instead of failing.
NOTE: Placing Select Model Device *after* a node that has already
consumed the same model (e.g. a KSampler that ran on this model on
the original device) is not recommended -- any state the prior
consumer mutated on the original model will be observed when the
selected device matches the original (fast path). Place Select Model
Device before any consumer of the model.
"""
@classmethod
def define_schema(cls):
return io.Schema(
node_id="SelectModelDevice",
display_name="Select Model Device",
category="advanced/multigpu",
description=cleandoc(cls.__doc__),
inputs=[
io.Model.Input("model"),
io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options()),
],
outputs=[
io.Model.Output(),
],
)
@classmethod
def validate_inputs(cls, device="default"):
# Allow unknown gpu:N values so portable workflows do not error
# at validation time; runtime fallback will handle them.
return True
@classmethod
def execute(cls, model: ModelPatcher, device: str = "default") -> io.NodeOutput:
model = model.clone()
resolved = comfy.model_management.resolve_gpu_device_option(device)
if resolved is None and device not in (None, "default"):
logging.info(f"Select Model Device: requested device '{device}' not available, passing through unchanged.")
return io.NodeOutput(model)
try:
model = _apply_patcher_device(model, resolved)
except RuntimeError as e:
logging.warning(f"Select Model Device: cannot retarget model, passing through unchanged. ({e})")
return io.NodeOutput(model)
if resolved is not None:
_force_supported_compute_dtype(model, resolved)
_prune_multigpu_collision(model, model.load_device)
return io.NodeOutput(model)
class SelectCLIPDeviceNode(io.ComfyNode):
"""
Place the CLIP text encoder on a specific device (default / cpu / gpu:N).
- "default" restores the device assigned by the loader.
- "cpu" pins both the load and offload device to CPU.
- "gpu:N" pins the load device to the Nth available GPU.
When the selected device does not exist on the current machine
(e.g. a workflow built on a 2-GPU box opened on a 1-GPU box),
the node passes the CLIP through unchanged and logs a message
instead of failing.
"""
@classmethod
def define_schema(cls):
return io.Schema(
node_id="SelectCLIPDevice",
display_name="Select CLIP Device",
category="advanced/multigpu",
description=cleandoc(cls.__doc__),
inputs=[
io.Clip.Input("clip"),
io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options()),
],
outputs=[
io.Clip.Output(),
],
)
@classmethod
def validate_inputs(cls, device="default"):
return True
@classmethod
def execute(cls, clip: CLIP, device: str = "default") -> io.NodeOutput:
clip = clip.clone()
resolved = comfy.model_management.resolve_gpu_device_option(device)
if resolved is None and device not in (None, "default"):
logging.info(f"Select CLIP Device: requested device '{device}' not available, passing through unchanged.")
return io.NodeOutput(clip)
try:
clip.patcher = _apply_patcher_device(clip.patcher, resolved)
except RuntimeError as e:
logging.warning(f"Select CLIP Device: cannot retarget CLIP, passing through unchanged. ({e})")
return io.NodeOutput(clip)
class SelectVAEDeviceNode(io.ComfyNode):
"""
Place the VAE on a specific device (default / gpu:N).
- "default" restores the device assigned by the loader.
- "gpu:N" pins the load device to the Nth available GPU; the offload
device is set to the standard VAE offload device.
CPU is intentionally not exposed in the UI for the VAE; if a workflow
supplies "cpu" anyway (e.g. opened from another machine), the request
is dropped with a log message and the VAE is passed through unchanged.
When the selected device does not exist on the current machine
(e.g. a workflow built on a 2-GPU box opened on a 1-GPU box),
the node passes the VAE through unchanged and logs a message
instead of failing.
"""
@classmethod
def define_schema(cls):
return io.Schema(
node_id="SelectVAEDevice",
display_name="Select VAE Device",
category="advanced/multigpu",
description=cleandoc(cls.__doc__),
inputs=[
io.Vae.Input("vae"),
io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options_no_cpu()),
],
outputs=[
io.Vae.Output(),
],
)
@classmethod
def validate_inputs(cls, device="default"):
return True
@classmethod
def execute(cls, vae: VAE, device: str = "default") -> io.NodeOutput:
# VAE has no .clone(); shallow-copy the wrapper and clone the patcher
# so we can retarget load/offload device without affecting the input VAE.
vae = copy.copy(vae)
vae.patcher = vae.patcher.clone()
resolved = comfy.model_management.resolve_gpu_device_option(device)
if resolved is None and device not in (None, "default"):
logging.info(f"Select VAE Device: requested device '{device}' not available, passing through unchanged.")
return io.NodeOutput(vae)
if resolved is not None and resolved.type == "cpu":
logging.info("Select VAE Device: CPU is not a supported choice, passing through unchanged.")
return io.NodeOutput(vae)
if not hasattr(vae, "_select_base_device"):
vae._select_base_device = vae.device
try:
vae.patcher = _apply_patcher_device(
vae.patcher, resolved,
base_offload_override=comfy.model_management.vae_offload_device(),
)
except RuntimeError as e:
logging.warning(f"Select VAE Device: cannot retarget VAE, passing through unchanged. ({e})")
return io.NodeOutput(vae)
# Keep VAE wrapper in sync with whatever model the patcher now owns;
# deepclone_multigpu may have produced a fresh first_stage_model.
vae.first_stage_model = vae.patcher.model
vae.device = vae._select_base_device if resolved is None else resolved
return io.NodeOutput(vae)
class MultiGPUOptionsNode(io.ComfyNode):
"""
Select the relative speed of GPUs in the special case they have significantly different performance from one another.
NOTE (not registered yet, see MultiGPUExtension.get_node_list below):
The output GPUOptionsGroup is plumbed through create_multigpu_deepclones() and stored on
model.model_options['multigpu_options'] via GPUOptionsGroup.register(), but the cond
scheduler in comfy/samplers.py (calc_cond_batch_outer_multigpu) does NOT yet consult
relative_speed when distributing conds across devices; it uses a uniform conds_per_device
round-robin via next_available_device(). Before re-enabling this node, wire its
relative_speed into the scheduler (e.g. via comfy.multigpu.load_balance_devices(),
which already implements the proportional split) so the input actually affects work
distribution.
"""
@classmethod
def define_schema(cls):
return io.Schema(
node_id="MultiGPU_Options",
display_name="MultiGPU Options",
category="advanced/multigpu",
description=cleandoc(cls.__doc__),
inputs=[
io.Int.Input("device_index", default=0, min=0, max=64),
io.Float.Input("relative_speed", default=1.0, min=0.0, step=0.01),
io.Custom("GPU_OPTIONS").Input("gpu_options", optional=True),
],
outputs=[
io.Custom("GPU_OPTIONS").Output(),
],
)
@classmethod
def execute(cls, device_index: int, relative_speed: float, gpu_options: comfy.multigpu.GPUOptionsGroup = None) -> io.NodeOutput:
if not gpu_options:
gpu_options = comfy.multigpu.GPUOptionsGroup()
else:
gpu_options = gpu_options.clone()
opt = comfy.multigpu.GPUOptions(device_index=device_index, relative_speed=relative_speed)
gpu_options.add(opt)
return io.NodeOutput(gpu_options)
class MultiGPUExtension(ComfyExtension):
@override
async def get_node_list(self) -> list[type[io.ComfyNode]]:
return [
MultiGPUCFGSplitNode,
SelectModelDeviceNode,
SelectCLIPDeviceNode,
SelectVAEDeviceNode,
# MultiGPUOptionsNode,
]
async def comfy_entrypoint() -> MultiGPUExtension:
return MultiGPUExtension()

1
comfy_extras/nodes_number_convert.py
View File

@ -4,7 +4,6 @@ Provides a single node that converts INT, FLOAT, STRING, and BOOL
inputs into FLOAT and INT outputs. inputs into FLOAT and INT outputs.
""" """
from __future__ import annotations
import math import math

2
comfy_extras/nodes_painter.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import hashlib import hashlib
import os import os

1
comfy_extras/nodes_resolution.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import math import math
from enum import Enum from enum import Enum
from typing_extensions import override from typing_extensions import override

1
comfy_extras/nodes_toolkit.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
from typing_extensions import override from typing_extensions import override
from comfy_api.latest import ComfyExtension, io from comfy_api.latest import ComfyExtension, io

2
comfy_extras/nodes_video.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import os import os
import av import av
import torch import torch

2
folder_paths.py
View File

@ -1,5 +1,3 @@
from __future__ import annotations
import os import os
import time import time
import mimetypes import mimetypes

10
main.py
View File

@ -218,7 +218,7 @@ import comfy.model_patcher
if args.enable_dynamic_vram or (enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl()): if args.enable_dynamic_vram or (enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl()):
if (not args.enable_dynamic_vram) and (comfy.model_management.torch_version_numeric < (2, 8)): if (not args.enable_dynamic_vram) and (comfy.model_management.torch_version_numeric < (2, 8)):
logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index): elif comfy_aimdo.control.init_devices(d.index for d in comfy.model_management.get_all_torch_devices()):
if args.verbose == 'DEBUG': if args.verbose == 'DEBUG':
comfy_aimdo.control.set_log_debug() comfy_aimdo.control.set_log_debug()
elif args.verbose == 'CRITICAL': elif args.verbose == 'CRITICAL':
@ -286,8 +286,8 @@ def prompt_worker(q, server_instance):
cache_ram = 0 cache_ram = 0
cache_ram_inactive = 0 cache_ram_inactive = 0
if not args.cache_classic and not args.cache_none and args.cache_lru <= 0: if not args.cache_classic and not args.cache_none and args.cache_lru <= 0:
cache_ram = min(32.0, max(4.0, comfy.model_management.total_ram * 0.25 / 1024.0)) cache_ram = min(10.0, max(2.0, comfy.model_management.total_ram * 0.10 / 1024.0))
cache_ram_inactive = min(96.0, max(12.0, comfy.model_management.total_ram * 0.75 / 1024.0)) cache_ram_inactive = min(96.0, comfy.model_management.total_ram / 1024.0)
if len(args.cache_ram) > 0: if len(args.cache_ram) > 0:
cache_ram = args.cache_ram[0] cache_ram = args.cache_ram[0]
if len(args.cache_ram) > 1: if len(args.cache_ram) > 1:
@ -344,9 +344,9 @@ def prompt_worker(q, server_instance):
# Log Time in a more readable way after 10 minutes # Log Time in a more readable way after 10 minutes
if execution_time > 600: if execution_time > 600:
execution_time = time.strftime("%H:%M:%S", time.gmtime(execution_time)) execution_time = time.strftime("%H:%M:%S", time.gmtime(execution_time))
logging.info(f"Prompt executed in {execution_time}") logging.info(f"Prompt executed in {execution_time}", extra={'color': 'green'})
else: else:
logging.info("Prompt executed in {:.2f} seconds".format(execution_time)) logging.info("Prompt executed in {:.2f} seconds".format(execution_time), extra={'color': 'green'})
if not asset_seeder.is_disabled(): if not asset_seeder.is_disabled():
paths = _collect_output_absolute_paths(e.history_result) paths = _collect_output_absolute_paths(e.history_result)

11
nodes.py
View File

@ -1,4 +1,3 @@
from __future__ import annotations
import torch import torch
@ -795,6 +794,7 @@ class VAELoader:
#TODO: scale factor? #TODO: scale factor?
def load_vae(self, vae_name): def load_vae(self, vae_name):
metadata = None metadata = None
vae_path = None
if vae_name == "pixel_space": if vae_name == "pixel_space":
sd = {} sd = {}
sd["pixel_space_vae"] = torch.tensor(1.0) sd["pixel_space_vae"] = torch.tensor(1.0)
@ -813,6 +813,14 @@ class VAELoader:
metadata["tae_latent_channels"] = 128 metadata["tae_latent_channels"] = 128
vae = comfy.sd.VAE(sd=sd, metadata=metadata) vae = comfy.sd.VAE(sd=sd, metadata=metadata)
vae.throw_exception_if_invalid() vae.throw_exception_if_invalid()
# Register a reload factory on the patcher so multigpu deepclones
# (Select VAE Device, future MultiGPU VAE work-units) can produce
# per-device clones from the same loader context. Only set when we
# actually have a single backing file -- pixel_space and the
# image TAESDs (composed from separate encoder/decoder files via
# load_taesd) are not addressable by a single vae_path.
if vae_path is not None:
vae.patcher.cached_patcher_init = (comfy.sd.load_vae_patcher, (vae_path, metadata, None))
return (vae,) return (vae,)
class ControlNetLoader: class ControlNetLoader:
@ -2389,6 +2397,7 @@ async def init_builtin_extra_nodes():
"nodes_lt_audio.py", "nodes_lt_audio.py",
"nodes_lt.py", "nodes_lt.py",
"nodes_hooks.py", "nodes_hooks.py",
"nodes_multigpu.py",
"nodes_load_3d.py", "nodes_load_3d.py",
"nodes_cosmos.py", "nodes_cosmos.py",
"nodes_video.py", "nodes_video.py",

4
requirements.txt
View File

@ -1,6 +1,6 @@
comfyui-frontend-package==1.44.19 comfyui-frontend-package==1.44.19
comfyui-workflow-templates==0.9.82 comfyui-workflow-templates==0.9.82
comfyui-embedded-docs==0.5.0 comfyui-embedded-docs==0.5.1
torch torch
torchsde torchsde
torchvision torchvision
@ -23,7 +23,7 @@ SQLAlchemy>=2.0.0
filelock filelock
av>=14.2.0 av>=14.2.0
comfy-kitchen>=0.2.8 comfy-kitchen>=0.2.8
comfy-aimdo==0.4.3 comfy-aimdo==0.4.5
requests requests
simpleeval>=1.0.0 simpleeval>=1.0.0
blake3 blake3

39
server.py
View File

@ -646,18 +646,37 @@ class PromptServer():
@routes.get("/system_stats") @routes.get("/system_stats")
async def system_stats(request): async def system_stats(request):
device = comfy.model_management.get_torch_device() primary_device = comfy.model_management.get_torch_device()
device_name = comfy.model_management.get_torch_device_name(device)
cpu_device = comfy.model_management.torch.device("cpu") cpu_device = comfy.model_management.torch.device("cpu")
ram_total = comfy.model_management.get_total_memory(cpu_device) ram_total = comfy.model_management.get_total_memory(cpu_device)
ram_free = comfy.model_management.get_free_memory(cpu_device) ram_free = comfy.model_management.get_free_memory(cpu_device)
vram_total, torch_vram_total = comfy.model_management.get_total_memory(device, torch_total_too=True)
vram_free, torch_vram_free = comfy.model_management.get_free_memory(device, torch_free_too=True)
required_frontend_version = FrontendManager.get_required_frontend_version() required_frontend_version = FrontendManager.get_required_frontend_version()
installed_templates_version = FrontendManager.get_installed_templates_version() installed_templates_version = FrontendManager.get_installed_templates_version()
required_templates_version = FrontendManager.get_required_templates_version() required_templates_version = FrontendManager.get_required_templates_version()
comfy_package_versions = FrontendManager.get_comfy_package_versions() comfy_package_versions = FrontendManager.get_comfy_package_versions()
# Report every torch device visible to multigpu, with the primary
# device first so existing clients that read devices[0] keep working.
torch_devices = comfy.model_management.get_all_torch_devices()
if primary_device in torch_devices:
torch_devices = [primary_device] + [d for d in torch_devices if d != primary_device]
else:
torch_devices = [primary_device] + list(torch_devices)
device_entries = []
for d in torch_devices:
vram_total, torch_vram_total = comfy.model_management.get_total_memory(d, torch_total_too=True)
vram_free, torch_vram_free = comfy.model_management.get_free_memory(d, torch_free_too=True)
device_entries.append({
"name": comfy.model_management.get_torch_device_name(d),
"type": d.type,
"index": d.index,
"vram_total": vram_total,
"vram_free": vram_free,
"torch_vram_total": torch_vram_total,
"torch_vram_free": torch_vram_free,
})
system_stats = { system_stats = {
"system": { "system": {
"os": sys.platform, "os": sys.platform,
@ -673,17 +692,7 @@ class PromptServer():
"embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded", "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded",
"argv": sys.argv "argv": sys.argv
}, },
"devices": [ "devices": device_entries
{
"name": device_name,
"type": device.type,
"index": device.index,
"vram_total": vram_total,
"vram_free": vram_free,
"torch_vram_total": torch_vram_total,
"torch_vram_free": torch_vram_free,
}
]
} }
return web.json_response(system_stats) return web.json_response(system_stats)

2
tests-unit/assets_test/conftest.py
View File

@ -236,6 +236,8 @@ def seeded_asset(request: pytest.FixtureRequest, http: requests.Session, api_bas
r = http.post(api_base + "/api/assets", files=files, data=form_data, timeout=120) r = http.post(api_base + "/api/assets", files=files, data=form_data, timeout=120)
body = r.json() body = r.json()
assert r.status_code == 201, body assert r.status_code == 201, body
from helpers import assert_hash_fields_consistent
assert_hash_fields_consistent(body)
return body return body

23
tests-unit/assets_test/helpers.py
View File

@ -26,3 +26,26 @@ def trigger_sync_seed_assets(session: requests.Session, base_url: str) -> None:
def get_asset_filename(asset_hash: str, extension: str) -> str: def get_asset_filename(asset_hash: str, extension: str) -> str:
return asset_hash.removeprefix("blake3:") + extension return asset_hash.removeprefix("blake3:") + extension
def assert_hash_fields_consistent(body: dict, expected_hash: str | None = None) -> None:
"""Assert hash and asset_hash invariants on an Asset response.
Both must be present or both absent (so a regression that drops only one
is caught). When present, they must equal each other and, if expected_hash
is provided, must equal that value.
"""
hash_present = "hash" in body
asset_hash_present = "asset_hash" in body
assert hash_present == asset_hash_present, (
f"hash and asset_hash must both be present or both absent: "
f"hash present={hash_present}, asset_hash present={asset_hash_present}"
)
if hash_present:
h = body["hash"]
ah = body["asset_hash"]
assert h == ah, f"hash and asset_hash must match: hash={h!r}, asset_hash={ah!r}"
if expected_hash is not None:
assert h == expected_hash, (
f"hash must equal expected: got {h!r}, expected {expected_hash!r}"
)

4
tests-unit/assets_test/test_assets_missing_sync.py
View File

@ -40,7 +40,9 @@ def test_seed_asset_removed_when_file_is_deleted(
# there should be exactly one with that name # there should be exactly one with that name
matches = [a for a in body1.get("assets", []) if a.get("name") == name] matches = [a for a in body1.get("assets", []) if a.get("name") == name]
assert matches assert matches
assert matches[0].get("asset_hash") is None # Seed assets have no hash; exclude_none drops both keys from the response
assert "asset_hash" not in matches[0]
assert "hash" not in matches[0]
asset_info_id = matches[0]["id"] asset_info_id = matches[0]["id"]
# Remove the underlying file and sync again # Remove the underlying file and sync again

9
tests-unit/assets_test/test_crud.py
View File

@ -21,6 +21,8 @@ def test_create_from_hash_success(
b1 = r1.json() b1 = r1.json()
assert r1.status_code == 201, b1 assert r1.status_code == 201, b1
assert b1["asset_hash"] == h assert b1["asset_hash"] == h
assert b1["hash"] == h
assert b1["hash"] == b1["asset_hash"]
assert b1["created_new"] is False assert b1["created_new"] is False
aid = b1["id"] aid = b1["id"]
@ -39,6 +41,7 @@ def test_get_and_delete_asset(http: requests.Session, api_base: str, seeded_asse
detail = rg.json() detail = rg.json()
assert rg.status_code == 200, detail assert rg.status_code == 200, detail
assert detail["id"] == aid assert detail["id"] == aid
assert detail["hash"] == detail["asset_hash"]
assert "user_metadata" in detail assert "user_metadata" in detail
assert "filename" in detail["user_metadata"] assert "filename" in detail["user_metadata"]
@ -97,6 +100,7 @@ def test_delete_upon_reference_count(
copy = r2.json() copy = r2.json()
assert r2.status_code == 201, copy assert r2.status_code == 201, copy
assert copy["asset_hash"] == src_hash assert copy["asset_hash"] == src_hash
assert copy["hash"] == src_hash
assert copy["created_new"] is False assert copy["created_new"] is False
# Soft-delete original reference (default) -> asset identity must remain # Soft-delete original reference (default) -> asset identity must remain
@ -139,6 +143,7 @@ def test_update_asset_fields(http: requests.Session, api_base: str, seeded_asset
body = ru.json() body = ru.json()
assert ru.status_code == 200, body assert ru.status_code == 200, body
assert body["name"] == payload["name"] assert body["name"] == payload["name"]
assert body["hash"] == body["asset_hash"]
assert body["tags"] == original_tags # tags unchanged assert body["tags"] == original_tags # tags unchanged
assert body["user_metadata"]["purpose"] == "updated" assert body["user_metadata"]["purpose"] == "updated"
# filename should still be present and normalized by server # filename should still be present and normalized by server
@ -289,7 +294,9 @@ def test_metadata_filename_is_set_for_seed_asset_without_hash(
assert r1.status_code == 200, body assert r1.status_code == 200, body
matches = [a for a in body.get("assets", []) if a.get("name") == name] matches = [a for a in body.get("assets", []) if a.get("name") == name]
assert matches, "Seed asset should be visible after sync" assert matches, "Seed asset should be visible after sync"
assert matches[0].get("asset_hash") is None # still a seed # Seed assets have no hash; exclude_none drops both keys from the response
assert "asset_hash" not in matches[0]
assert "hash" not in matches[0]
aid = matches[0]["id"] aid = matches[0]["id"]
r2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) r2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120)

5
tests-unit/assets_test/test_list_filter.py
View File

@ -3,6 +3,7 @@ import uuid
import pytest import pytest
import requests import requests
from helpers import assert_hash_fields_consistent
def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asset_factory, make_asset_bytes):
@ -26,6 +27,10 @@ def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asse
got1 = [a["name"] for a in b1["assets"]] got1 = [a["name"] for a in b1["assets"]]
assert got1 == sorted(names)[:2] assert got1 == sorted(names)[:2]
assert b1["has_more"] is True assert b1["has_more"] is True
# Populated assets in list responses must carry both `hash` and `asset_hash` consistently
for asset in b1["assets"]:
assert_hash_fields_consistent(asset)
assert "hash" in asset, "populated asset must emit hash on list endpoint"
r2 = http.get( r2 = http.get(
api_base + "/api/assets", api_base + "/api/assets",

Some files were not shown because too many files have changed in this diff Show More