wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-14 00:00:57 +08:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
flyingshutter 2023-06-11 18:05:31 +02:00
commit 2759982b37
16 changed files with 421 additions and 249 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
.gitignore vendored
View File

@ -9,3 +9,8 @@ custom_nodes/
!custom_nodes/example_node.py.example
extra_model_paths.yaml
/.vs
.idea/
venv/
web/extensions/*
!web/extensions/logging.js.example
!web/extensions/core/

65
README.md
View File

@ -38,28 +38,28 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
## Shortcuts
| Keybind | Explanation |
| - | - |
| Ctrl + Enter | Queue up current graph for generation |
| Ctrl + Shift + Enter | Queue up current graph as first for generation |
| Ctrl + S | Save workflow |
| Ctrl + O | Load workflow |
| Ctrl + A | Select all nodes |
| Ctrl + M | Mute/unmute selected nodes |
| Delete/Backspace | Delete selected nodes |
| Ctrl + Delete/Backspace | Delete the current graph |
| Space | Move the canvas around when held and moving the cursor |
| Ctrl/Shift + Click | Add clicked node to selection |
| Ctrl + C/Ctrl + V | Copy and paste selected nodes (without maintaining connections to outputs of unselected nodes) |
| Ctrl + C/Ctrl + Shift + V| Copy and paste selected nodes (maintaining connections from outputs of unselected nodes to inputs of pasted nodes) |
| Shift + Drag | Move multiple selected nodes at the same time |
| Ctrl + D | Load default graph |
| Q | Toggle visibility of the queue |
| H | Toggle visibility of history |
| R | Refresh graph |
| Double-Click LMB | Open node quick search palette |
| Keybind | Explanation |
|---------------------------|--------------------------------------------------------------------------------------------------------------------|
| Ctrl + Enter | Queue up current graph for generation |
| Ctrl + Shift + Enter | Queue up current graph as first for generation |
| Ctrl + S | Save workflow |
| Ctrl + O | Load workflow |
| Ctrl + A | Select all nodes |
| Ctrl + M | Mute/unmute selected nodes |
| Delete/Backspace | Delete selected nodes |
| Ctrl + Delete/Backspace | Delete the current graph |
| Space | Move the canvas around when held and moving the cursor |
| Ctrl/Shift + Click | Add clicked node to selection |
| Ctrl + C/Ctrl + V | Copy and paste selected nodes (without maintaining connections to outputs of unselected nodes) |
| Ctrl + C/Ctrl + Shift + V | Copy and paste selected nodes (maintaining connections from outputs of unselected nodes to inputs of pasted nodes) |
| Shift + Drag | Move multiple selected nodes at the same time |
| Ctrl + D | Load default graph |
| Q | Toggle visibility of the queue |
| H | Toggle visibility of history |
| R | Refresh graph |
| Double-Click LMB | Open node quick search palette |
Ctrl can also be replaced with Cmd instead for MacOS users
Ctrl can also be replaced with Cmd instead for macOS users
# Installing
@ -119,13 +119,26 @@ After this you should have everything installed and can proceed to running Comfy
### Others:
[Intel Arc](https://github.com/comfyanonymous/ComfyUI/discussions/476)
#### [Intel Arc](https://github.com/comfyanonymous/ComfyUI/discussions/476)
Mac/MPS: There is basic support in the code but until someone makes some install instruction you are on your own.
#### Apple Mac silicon
You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS version.
1. Install pytorch. For instructions, read the [Accelerated PyTorch training on Mac](https://developer.apple.com/metal/pytorch/) Apple Developer guide.
1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux.
1. Install the ComfyUI [dependencies](#dependencies). If you have another Stable Diffusion UI [you might be able to reuse the dependencies](#i-already-have-another-ui-for-stable-diffusion-installed-do-i-really-have-to-install-all-of-these-dependencies).
1. Launch ComfyUI by running `python main.py`.
> **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux).
#### DirectML (AMD Cards on Windows)
```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
### I already have another UI for Stable Diffusion installed do I really have to install all of these dependencies?
You don't. If you have another UI installed and working with it's own python venv you can use that venv to run ComfyUI. You can open up your favorite terminal and activate it:
You don't. If you have another UI installed and working with its own python venv you can use that venv to run ComfyUI. You can open up your favorite terminal and activate it:
```source path_to_other_sd_gui/venv/bin/activate```
@ -135,7 +148,7 @@ With Powershell: ```"path_to_other_sd_gui\venv\Scripts\Activate.ps1"```
With cmd.exe: ```"path_to_other_sd_gui\venv\Scripts\activate.bat"```
And then you can use that terminal to run Comfyui without installing any dependencies. Note that the venv folder might be called something else depending on the SD UI.
And then you can use that terminal to run ComfyUI without installing any dependencies. Note that the venv folder might be called something else depending on the SD UI.
# Running
@ -159,6 +172,8 @@ You can use () to change emphasis of a word or phrase like: (good code:1.2) or (
You can use {day|night}, for wildcard/dynamic prompts. With this syntax "{wild|card|test}" will be randomly replaced by either "wild", "card" or "test" by the frontend every time you queue the prompt. To use {} characters in your actual prompt escape them like: \\{ or \\}.
Dynamic prompts also support C-style comments, like `// comment` or `/* comment */`.
To use a textual inversion concepts/embeddings in a text prompt put them in the models/embeddings directory and use them in the CLIPTextEncode node like this (you can omit the .pt extension):
```embedding:embedding_filename.pt```

28
comfy/diffusers_load.py
View File

@ -4,7 +4,7 @@ import yaml
import folder_paths
from comfy.ldm.util import instantiate_from_config
from comfy.sd import ModelPatcher, load_model_weights, CLIP, VAE
from comfy.sd import ModelPatcher, load_model_weights, CLIP, VAE, load_checkpoint
import os.path as osp
import re
import torch
@ -84,28 +84,4 @@ def load_diffusers(model_path, fp16=True, output_vae=True, output_clip=True, emb
# Put together new checkpoint
sd = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
clip = None
vae = None
class WeightsLoader(torch.nn.Module):
pass
w = WeightsLoader()
load_state_dict_to = []
if output_vae:
vae = VAE(scale_factor=scale_factor, config=vae_config)
w.first_stage_model = vae.first_stage_model
load_state_dict_to = [w]
if output_clip:
clip = CLIP(config=clip_config, embedding_directory=embedding_directory)
w.cond_stage_model = clip.cond_stage_model
load_state_dict_to = [w]
model = instantiate_from_config(config["model"])
model = load_model_weights(model, sd, verbose=False, load_state_dict_to=load_state_dict_to)
if fp16:
model = model.half()
return ModelPatcher(model), clip, vae
return load_checkpoint(embedding_directory=embedding_directory, state_dict=sd, config=config)

97
comfy/model_base.py Normal file
View File

@ -0,0 +1,97 @@
import torch
from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel
from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation
from comfy.ldm.modules.diffusionmodules.util import make_beta_schedule
import numpy as np
class BaseModel(torch.nn.Module):
def __init__(self, unet_config, v_prediction=False):
super().__init__()
self.register_schedule(given_betas=None, beta_schedule="linear", timesteps=1000, linear_start=0.00085, linear_end=0.012, cosine_s=8e-3)
self.diffusion_model = UNetModel(**unet_config)
self.v_prediction = v_prediction
if self.v_prediction:
self.parameterization = "v"
else:
self.parameterization = "eps"
if "adm_in_channels" in unet_config:
self.adm_channels = unet_config["adm_in_channels"]
else:
self.adm_channels = 0
print("v_prediction", v_prediction)
print("adm", self.adm_channels)
def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
if given_betas is not None:
betas = given_betas
else:
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
alphas = 1. - betas
alphas_cumprod = np.cumprod(alphas, axis=0)
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
timesteps, = betas.shape
self.num_timesteps = int(timesteps)
self.linear_start = linear_start
self.linear_end = linear_end
self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32))
self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32))
self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32))
def apply_model(self, x, t, c_concat=None, c_crossattn=None, c_adm=None, control=None, transformer_options={}):
if c_concat is not None:
xc = torch.cat([x] + c_concat, dim=1)
else:
xc = x
context = torch.cat(c_crossattn, 1)
return self.diffusion_model(xc, t, context=context, y=c_adm, control=control, transformer_options=transformer_options)
def get_dtype(self):
return self.diffusion_model.dtype
def is_adm(self):
return self.adm_channels > 0
class SD21UNCLIP(BaseModel):
def __init__(self, unet_config, noise_aug_config, v_prediction=True):
super().__init__(unet_config, v_prediction)
self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**noise_aug_config)
def encode_adm(self, **kwargs):
unclip_conditioning = kwargs.get("unclip_conditioning", None)
device = kwargs["device"]
if unclip_conditioning is not None:
adm_inputs = []
weights = []
noise_aug = []
for unclip_cond in unclip_conditioning:
adm_cond = unclip_cond["clip_vision_output"].image_embeds
weight = unclip_cond["strength"]
noise_augment = unclip_cond["noise_augmentation"]
noise_level = round((self.noise_augmentor.max_noise_level - 1) * noise_augment)
c_adm, noise_level_emb = self.noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device))
adm_out = torch.cat((c_adm, noise_level_emb), 1) * weight
weights.append(weight)
noise_aug.append(noise_augment)
adm_inputs.append(adm_out)
if len(noise_aug) > 1:
adm_out = torch.stack(adm_inputs).sum(0)
#TODO: add a way to control this
noise_augment = 0.05
noise_level = round((self.noise_augmentor.max_noise_level - 1) * noise_augment)
c_adm, noise_level_emb = self.noise_augmentor(adm_out[:, :self.noise_augmentor.time_embed.dim], noise_level=torch.tensor([noise_level], device=device))
adm_out = torch.cat((c_adm, noise_level_emb), 1)
else:
adm_out = torch.zeros((1, self.adm_channels))
return adm_out
class SDInpaint(BaseModel):
def __init__(self, unet_config, v_prediction=False):
super().__init__(unet_config, v_prediction)
self.concat_keys = ("mask", "masked_image")

44
comfy/samplers.py
View File

@ -248,7 +248,7 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, con
c['transformer_options'] = transformer_options
output = model_function(input_x, timestep_, cond=c).chunk(batch_chunks)
output = model_function(input_x, timestep_, **c).chunk(batch_chunks)
del input_x
model_management.throw_exception_if_processing_interrupted()
@ -460,36 +460,18 @@ def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
uncond[temp[1]] = [o[0], n]
def encode_adm(noise_augmentor, conds, batch_size, device):
def encode_adm(model, conds, batch_size, device):
for t in range(len(conds)):
x = conds[t]
adm_out = None
if 'adm' in x[1]:
adm_inputs = []
weights = []
noise_aug = []
adm_in = x[1]["adm"]
for adm_c in adm_in:
adm_cond = adm_c[0].image_embeds
weight = adm_c[1]
noise_augment = adm_c[2]
noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
c_adm, noise_level_emb = noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device))
adm_out = torch.cat((c_adm, noise_level_emb), 1) * weight
weights.append(weight)
noise_aug.append(noise_augment)
adm_inputs.append(adm_out)
if len(noise_aug) > 1:
adm_out = torch.stack(adm_inputs).sum(0)
#TODO: add a way to control this
noise_augment = 0.05
noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
c_adm, noise_level_emb = noise_augmentor(adm_out[:, :noise_augmentor.time_embed.dim], noise_level=torch.tensor([noise_level], device=device))
adm_out = torch.cat((c_adm, noise_level_emb), 1)
adm_out = x[1]["adm"]
else:
adm_out = torch.zeros((1, noise_augmentor.time_embed.dim * 2), device=device)
x[1] = x[1].copy()
x[1]["adm_encoded"] = torch.cat([adm_out] * batch_size)
params = x[1].copy()
adm_out = model.encode_adm(device=device, **params)
if adm_out is not None:
x[1] = x[1].copy()
x[1]["adm_encoded"] = torch.cat([adm_out] * batch_size).to(device)
return conds
@ -591,14 +573,14 @@ class KSampler:
apply_empty_x_to_equal_area(positive, negative, 'control', lambda cond_cnets, x: cond_cnets[x])
apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
if self.model.model.diffusion_model.dtype == torch.float16:
if self.model.get_dtype() == torch.float16:
precision_scope = torch.autocast
else:
precision_scope = contextlib.nullcontext
if hasattr(self.model, 'noise_augmentor'): #unclip
positive = encode_adm(self.model.noise_augmentor, positive, noise.shape[0], self.device)
negative = encode_adm(self.model.noise_augmentor, negative, noise.shape[0], self.device)
if self.model.is_adm():
positive = encode_adm(self.model, positive, noise.shape[0], self.device)
negative = encode_adm(self.model, negative, noise.shape[0], self.device)
extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": self.model_options}

93
comfy/sd.py
View File

@ -15,8 +15,15 @@ from . import utils
from . import clip_vision
from . import gligen
from . import diffusers_convert
from . import model_base
def load_model_weights(model, sd, verbose=False, load_state_dict_to=[]):
replace_prefix = {"model.diffusion_model.": "diffusion_model."}
for rp in replace_prefix:
replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), sd.keys())))
for x in replace:
sd[x[1]] = sd.pop(x[0])
m, u = model.load_state_dict(sd, strict=False)
k = list(sd.keys())
@ -31,17 +38,6 @@ def load_model_weights(model, sd, verbose=False, load_state_dict_to=[]):
if ids.dtype == torch.float32:
sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()
keys_to_replace = {
"cond_stage_model.model.positional_embedding": "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight",
"cond_stage_model.model.token_embedding.weight": "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight",
"cond_stage_model.model.ln_final.weight": "cond_stage_model.transformer.text_model.final_layer_norm.weight",
"cond_stage_model.model.ln_final.bias": "cond_stage_model.transformer.text_model.final_layer_norm.bias",
}
for x in keys_to_replace:
if x in sd:
sd[keys_to_replace[x]] = sd.pop(x)
sd = utils.transformers_convert(sd, "cond_stage_model.model", "cond_stage_model.transformer.text_model", 24)
for x in load_state_dict_to:
@ -193,7 +189,7 @@ def model_lora_keys(model, key_map={}):
counter = 0
for b in range(12):
tk = "model.diffusion_model.input_blocks.{}.1".format(b)
tk = "diffusion_model.input_blocks.{}.1".format(b)
up_counter = 0
for c in LORA_UNET_MAP_ATTENTIONS:
k = "{}.{}.weight".format(tk, c)
@ -204,13 +200,13 @@ def model_lora_keys(model, key_map={}):
if up_counter >= 4:
counter += 1
for c in LORA_UNET_MAP_ATTENTIONS:
k = "model.diffusion_model.middle_block.1.{}.weight".format(c)
k = "diffusion_model.middle_block.1.{}.weight".format(c)
if k in sdk:
lora_key = "lora_unet_mid_block_attentions_0_{}".format(LORA_UNET_MAP_ATTENTIONS[c])
key_map[lora_key] = k
counter = 3
for b in range(12):
tk = "model.diffusion_model.output_blocks.{}.1".format(b)
tk = "diffusion_model.output_blocks.{}.1".format(b)
up_counter = 0
for c in LORA_UNET_MAP_ATTENTIONS:
k = "{}.{}.weight".format(tk, c)
@ -234,7 +230,7 @@ def model_lora_keys(model, key_map={}):
ds_counter = 0
counter = 0
for b in range(12):
tk = "model.diffusion_model.input_blocks.{}.0".format(b)
tk = "diffusion_model.input_blocks.{}.0".format(b)
key_in = False
for c in LORA_UNET_MAP_RESNET:
k = "{}.{}.weight".format(tk, c)
@ -253,7 +249,7 @@ def model_lora_keys(model, key_map={}):
counter = 0
for b in range(3):
tk = "model.diffusion_model.middle_block.{}".format(b)
tk = "diffusion_model.middle_block.{}".format(b)
key_in = False
for c in LORA_UNET_MAP_RESNET:
k = "{}.{}.weight".format(tk, c)
@ -267,7 +263,7 @@ def model_lora_keys(model, key_map={}):
counter = 0
us_counter = 0
for b in range(12):
tk = "model.diffusion_model.output_blocks.{}.0".format(b)
tk = "diffusion_model.output_blocks.{}.0".format(b)
key_in = False
for c in LORA_UNET_MAP_RESNET:
k = "{}.{}.weight".format(tk, c)
@ -343,7 +339,7 @@ class ModelPatcher:
patch_list[i] = patch_list[i].to(device)
def model_dtype(self):
return self.model.diffusion_model.dtype
return self.model.get_dtype()
def add_patches(self, patches, strength=1.0):
p = {}
@ -775,7 +771,7 @@ def load_controlnet(ckpt_path, model=None):
for x in controlnet_data:
c_m = "control_model."
if x.startswith(c_m):
sd_key = "model.diffusion_model.{}".format(x[len(c_m):])
sd_key = "diffusion_model.{}".format(x[len(c_m):])
if sd_key in model_sd:
cd = controlnet_data[x]
cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
@ -942,9 +938,10 @@ def load_gligen(ckpt_path):
model = model.half()
return model
def load_checkpoint(config_path, ckpt_path, output_vae=True, output_clip=True, embedding_directory=None):
with open(config_path, 'r') as stream:
config = yaml.safe_load(stream)
def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
if config is None:
with open(config_path, 'r') as stream:
config = yaml.safe_load(stream)
model_config_params = config['model']['params']
clip_config = model_config_params['cond_stage_config']
scale_factor = model_config_params['scale_factor']
@ -953,8 +950,19 @@ def load_checkpoint(config_path, ckpt_path, output_vae=True, output_clip=True, e
fp16 = False
if "unet_config" in model_config_params:
if "params" in model_config_params["unet_config"]:
if "use_fp16" in model_config_params["unet_config"]["params"]:
fp16 = model_config_params["unet_config"]["params"]["use_fp16"]
unet_config = model_config_params["unet_config"]["params"]
if "use_fp16" in unet_config:
fp16 = unet_config["use_fp16"]
noise_aug_config = None
if "noise_aug_config" in model_config_params:
noise_aug_config = model_config_params["noise_aug_config"]
v_prediction = False
if "parameterization" in model_config_params:
if model_config_params["parameterization"] == "v":
v_prediction = True
clip = None
vae = None
@ -974,9 +982,16 @@ def load_checkpoint(config_path, ckpt_path, output_vae=True, output_clip=True, e
w.cond_stage_model = clip.cond_stage_model
load_state_dict_to = [w]
model = instantiate_from_config(config["model"])
sd = utils.load_torch_file(ckpt_path)
model = load_model_weights(model, sd, verbose=False, load_state_dict_to=load_state_dict_to)
if config['model']["target"].endswith("LatentInpaintDiffusion"):
model = model_base.SDInpaint(unet_config, v_prediction=v_prediction)
elif config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"):
model = model_base.SD21UNCLIP(unet_config, noise_aug_config["params"], v_prediction=v_prediction)
else:
model = model_base.BaseModel(unet_config, v_prediction=v_prediction)
if state_dict is None:
state_dict = utils.load_torch_file(ckpt_path)
model = load_model_weights(model, state_dict, verbose=False, load_state_dict_to=load_state_dict_to)
if fp16:
model = model.half()
@ -1073,47 +1088,59 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
"legacy": False
}
if len(sd['model.diffusion_model.input_blocks.1.1.proj_in.weight'].shape) == 2:
if len(sd['model.diffusion_model.input_blocks.4.1.proj_in.weight'].shape) == 2:
unet_config['use_linear_in_transformer'] = True
unet_config["use_fp16"] = fp16
unet_config["model_channels"] = sd['model.diffusion_model.input_blocks.0.0.weight'].shape[0]
unet_config["in_channels"] = sd['model.diffusion_model.input_blocks.0.0.weight'].shape[1]
unet_config["context_dim"] = sd['model.diffusion_model.input_blocks.1.1.transformer_blocks.0.attn2.to_k.weight'].shape[1]
unet_config["context_dim"] = sd['model.diffusion_model.input_blocks.4.1.transformer_blocks.0.attn2.to_k.weight'].shape[1]
sd_config["unet_config"] = {"target": "comfy.ldm.modules.diffusionmodules.openaimodel.UNetModel", "params": unet_config}
model_config = {"target": "comfy.ldm.models.diffusion.ddpm.LatentDiffusion", "params": sd_config}
unclip_model = False
inpaint_model = False
if noise_aug_config is not None: #SD2.x unclip model
sd_config["noise_aug_config"] = noise_aug_config
sd_config["image_size"] = 96
sd_config["embedding_dropout"] = 0.25
sd_config["conditioning_key"] = 'crossattn-adm'
unclip_model = True
model_config["target"] = "comfy.ldm.models.diffusion.ddpm.ImageEmbeddingConditionedLatentDiffusion"
elif unet_config["in_channels"] > 4: #inpainting model
sd_config["conditioning_key"] = "hybrid"
sd_config["finetune_keys"] = None
model_config["target"] = "comfy.ldm.models.diffusion.ddpm.LatentInpaintDiffusion"
inpaint_model = True
else:
sd_config["conditioning_key"] = "crossattn"
if unet_config["context_dim"] == 1024:
unet_config["num_head_channels"] = 64 #SD2.x
else:
if unet_config["context_dim"] == 768:
unet_config["num_heads"] = 8 #SD1.x
else:
unet_config["num_head_channels"] = 64 #SD2.x
unclip = 'model.diffusion_model.label_emb.0.0.weight'
if unclip in sd_keys:
unet_config["num_classes"] = "sequential"
unet_config["adm_in_channels"] = sd[unclip].shape[1]
v_prediction = False
if unet_config["context_dim"] == 1024 and unet_config["in_channels"] == 4: #only SD2.x non inpainting models are v prediction
k = "model.diffusion_model.output_blocks.11.1.transformer_blocks.0.norm1.bias"
out = sd[k]
if torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out.
v_prediction = True
sd_config["parameterization"] = 'v'
model = instantiate_from_config(model_config)
if inpaint_model:
model = model_base.SDInpaint(unet_config, v_prediction=v_prediction)
elif unclip_model:
model = model_base.SD21UNCLIP(unet_config, noise_aug_config["params"], v_prediction=v_prediction)
else:
model = model_base.BaseModel(unet_config, v_prediction=v_prediction)
model = load_model_weights(model, sd, verbose=False, load_state_dict_to=load_state_dict_to)
if fp16:

2
comfy/sd1_clip.py
View File

@ -82,6 +82,8 @@ class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):
next_new_token += 1
else:
print("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored", y.shape[0], current_embeds.weight.shape[1])
while len(tokens_temp) < len(x):
tokens_temp += [self.empty_tokens[0][-1]]
out_tokens += [tokens_temp]
if len(embedding_weights) > 0:

12
comfy/utils.py
View File

@ -24,6 +24,18 @@ def load_torch_file(ckpt, safe_load=False):
return sd
def transformers_convert(sd, prefix_from, prefix_to, number):
keys_to_replace = {
"{}.positional_embedding": "{}.embeddings.position_embedding.weight",
"{}.token_embedding.weight": "{}.embeddings.token_embedding.weight",
"{}.ln_final.weight": "{}.final_layer_norm.weight",
"{}.ln_final.bias": "{}.final_layer_norm.bias",
}
for k in keys_to_replace:
x = k.format(prefix_from)
if x in sd:
sd[keys_to_replace[k].format(prefix_to)] = sd.pop(x)
resblock_to_replace = {
"ln_1": "layer_norm1",
"ln_2": "layer_norm2",

20
main.py
View File

@ -37,21 +37,25 @@ def prompt_worker(q, server):
e.execute(item[2], item[1], item[3], item[4])
q.task_done(item_id, e.outputs_ui)
async def run(server, address='', port=8188, verbose=True, call_on_start=None):
await asyncio.gather(server.start(address, port, verbose, call_on_start), server.publish_loop())
def hijack_progress(server):
def hook(value, total, preview_image_bytes):
server.send_sync("progress", { "value": value, "max": total}, server.client_id)
server.send_sync("progress", {"value": value, "max": total}, server.client_id)
if preview_image_bytes is not None:
server.send_sync(BinaryEventTypes.PREVIEW_IMAGE, preview_image_bytes, server.client_id)
comfy.utils.set_progress_bar_global_hook(hook)
def cleanup_temp():
temp_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "temp")
if os.path.exists(temp_dir):
shutil.rmtree(temp_dir, ignore_errors=True)
def load_extra_path_config(yaml_path):
with open(yaml_path, 'r') as stream:
config = yaml.safe_load(stream)
@ -72,6 +76,7 @@ def load_extra_path_config(yaml_path):
print("Adding extra search path", x, full_path)
folder_paths.add_model_folder_path(x, full_path)
if __name__ == "__main__":
cleanup_temp()
@ -92,7 +97,7 @@ if __name__ == "__main__":
server.add_routes()
hijack_progress(server)
threading.Thread(target=prompt_worker, daemon=True, args=(q,server,)).start()
threading.Thread(target=prompt_worker, daemon=True, args=(q, server,)).start()
if args.output_directory:
output_dir = os.path.abspath(args.output_directory)
@ -106,15 +111,12 @@ if __name__ == "__main__":
if args.auto_launch:
def startup_server(address, port):
import webbrowser
webbrowser.open("http://{}:{}".format(address, port))
webbrowser.open(f"http://{address}:{port}")
call_on_start = startup_server
if os.name == "nt":
try:
loop.run_until_complete(run(server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start))
except KeyboardInterrupt:
pass
else:
try:
loop.run_until_complete(run(server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start))
except KeyboardInterrupt:
print("\nStopped server")
cleanup_temp()

8
nodes.py
View File

@ -623,11 +623,11 @@ class unCLIPConditioning:
c = []
for t in conditioning:
o = t[1].copy()
x = (clip_vision_output, strength, noise_augmentation)
if "adm" in o:
o["adm"] = o["adm"][:] + [x]
x = {"clip_vision_output": clip_vision_output, "strength": strength, "noise_augmentation": noise_augmentation}
if "unclip_conditioning" in o:
o["unclip_conditioning"] = o["unclip_conditioning"][:] + [x]
else:
o["adm"] = [x]
o["unclip_conditioning"] = [x]
n = [t[0], o]
c.append(n)
return (c, )

226
web/extensions/core/contextMenuFilter.js
View File

@ -1,132 +1,138 @@
import { app } from "/scripts/app.js";
import {app} from "/scripts/app.js";
// Adds filtering to combo context menus
const id = "Comfy.ContextMenuFilter";
app.registerExtension({
name: id,
const ext = {
name: "Comfy.ContextMenuFilter",
init() {
const ctxMenu = LiteGraph.ContextMenu;
LiteGraph.ContextMenu = function (values, options) {
const ctx = ctxMenu.call(this, values, options);
// If we are a dark menu (only used for combo boxes) then add a filter input
if (options?.className === "dark" && values?.length > 10) {
const filter = document.createElement("input");
Object.assign(filter.style, {
width: "calc(100% - 10px)",
border: "0",
boxSizing: "border-box",
background: "#333",
border: "1px solid #999",
margin: "0 0 5px 5px",
color: "#fff",
});
filter.classList.add("comfy-context-menu-filter");
filter.placeholder = "Filter list";
this.root.prepend(filter);
let selectedIndex = 0;
let items = this.root.querySelectorAll(".litemenu-entry");
let itemCount = items.length;
let selectedItem;
const items = Array.from(this.root.querySelectorAll(".litemenu-entry"));
let displayedItems = [...items];
let itemCount = displayedItems.length;
// Apply highlighting to the selected item
function updateSelected() {
if (selectedItem) {
selectedItem.style.setProperty("background-color", "");
selectedItem.style.setProperty("color", "");
}
selectedItem = items[selectedIndex];
if (selectedItem) {
selectedItem.style.setProperty("background-color", "#ccc", "important");
selectedItem.style.setProperty("color", "#000", "important");
}
}
// We must request an animation frame for the current node of the active canvas to update.
requestAnimationFrame(() => {
const currentNode = LGraphCanvas.active_canvas.current_node;
const clickedComboValue = currentNode.widgets
.filter(w => w.type === "combo" && w.options.values.length === values.length)
.find(w => w.options.values.every((v, i) => v === values[i]))
.value;
const positionList = () => {
const rect = this.root.getBoundingClientRect();
// If the top is off screen then shift the element with scaling applied
if (rect.top < 0) {
const scale = 1 - this.root.getBoundingClientRect().height / this.root.clientHeight;
const shift = (this.root.clientHeight * scale) / 2;
this.root.style.top = -shift + "px";
}
}
updateSelected();
// Arrow up/down to select items
filter.addEventListener("keydown", (e) => {
if (e.key === "ArrowUp") {
if (selectedIndex === 0) {
selectedIndex = itemCount - 1;
} else {
selectedIndex--;
}
updateSelected();
e.preventDefault();
} else if (e.key === "ArrowDown") {
if (selectedIndex === itemCount - 1) {
selectedIndex = 0;
} else {
selectedIndex++;
}
updateSelected();
e.preventDefault();
} else if ((selectedItem && e.key === "Enter") || e.keyCode === 13 || e.keyCode === 10) {
selectedItem.click();
} else if(e.key === "Escape") {
this.close();
}
});
filter.addEventListener("input", () => {
// Hide all items that dont match our filter
const term = filter.value.toLocaleLowerCase();
items = this.root.querySelectorAll(".litemenu-entry");
// When filtering recompute which items are visible for arrow up/down
// Try and maintain selection
let visibleItems = [];
for (const item of items) {
const visible = !term || item.textContent.toLocaleLowerCase().includes(term);
if (visible) {
item.style.display = "block";
if (item === selectedItem) {
selectedIndex = visibleItems.length;
}
visibleItems.push(item);
} else {
item.style.display = "none";
if (item === selectedItem) {
selectedIndex = 0;
}
}
}
items = visibleItems;
let selectedIndex = values.findIndex(v => v === clickedComboValue);
let selectedItem = displayedItems?.[selectedIndex];
updateSelected();
// If we have an event then we can try and position the list under the source
if (options.event) {
let top = options.event.clientY - 10;
const bodyRect = document.body.getBoundingClientRect();
const rootRect = this.root.getBoundingClientRect();
if (bodyRect.height && top > bodyRect.height - rootRect.height - 10) {
top = Math.max(0, bodyRect.height - rootRect.height - 10);
}
this.root.style.top = top + "px";
positionList();
// Apply highlighting to the selected item
function updateSelected() {
selectedItem?.style.setProperty("background-color", "");
selectedItem?.style.setProperty("color", "");
selectedItem = displayedItems[selectedIndex];
selectedItem?.style.setProperty("background-color", "#ccc", "important");
selectedItem?.style.setProperty("color", "#000", "important");
}
});
requestAnimationFrame(() => {
// Focus the filter box when opening
filter.focus();
const positionList = () => {
const rect = this.root.getBoundingClientRect();
positionList();
});
// If the top is off-screen then shift the element with scaling applied
if (rect.top < 0) {
const scale = 1 - this.root.getBoundingClientRect().height / this.root.clientHeight;
const shift = (this.root.clientHeight * scale) / 2;
this.root.style.top = -shift + "px";
}
}
// Arrow up/down to select items
filter.addEventListener("keydown", (event) => {
switch (event.key) {
case "ArrowUp":
event.preventDefault();
if (selectedIndex === 0) {
selectedIndex = itemCount - 1;
} else {
selectedIndex--;
}
updateSelected();
break;
case "ArrowRight":
event.preventDefault();
selectedIndex = itemCount - 1;
updateSelected();
break;
case "ArrowDown":
event.preventDefault();
if (selectedIndex === itemCount - 1) {
selectedIndex = 0;
} else {
selectedIndex++;
}
updateSelected();
break;
case "ArrowLeft":
event.preventDefault();
selectedIndex = 0;
updateSelected();
break;
case "Enter":
selectedItem?.click();
break;
case "Escape":
this.close();
break;
}
});
filter.addEventListener("input", () => {
// Hide all items that don't match our filter
const term = filter.value.toLocaleLowerCase();
// When filtering, recompute which items are visible for arrow up/down and maintain selection.
displayedItems = items.filter(item => {
const isVisible = !term || item.textContent.toLocaleLowerCase().includes(term);
item.style.display = isVisible ? "block" : "none";
return isVisible;
});
selectedIndex = 0;
if (displayedItems.includes(selectedItem)) {
selectedIndex = displayedItems.findIndex(d => d === selectedItem);
}
itemCount = displayedItems.length;
updateSelected();
// If we have an event then we can try and position the list under the source
if (options.event) {
let top = options.event.clientY - 10;
const bodyRect = document.body.getBoundingClientRect();
const rootRect = this.root.getBoundingClientRect();
if (bodyRect.height && top > bodyRect.height - rootRect.height - 10) {
top = Math.max(0, bodyRect.height - rootRect.height - 10);
}
this.root.style.top = top + "px";
positionList();
}
});
requestAnimationFrame(() => {
// Focus the filter box when opening
filter.focus();
positionList();
});
})
}
return ctx;
@ -134,4 +140,6 @@ app.registerExtension({
LiteGraph.ContextMenu.prototype = ctxMenu.prototype;
},
});
}
app.registerExtension(ext);

9
web/extensions/core/dynamicPrompts.js
View File

@ -3,6 +3,13 @@ import { app } from "../../scripts/app.js";
// Allows for simple dynamic prompt replacement
// Inputs in the format {a|b} will have a random value of a or b chosen when the prompt is queued.
/*
* Strips C-style line and block comments from a string
*/
function stripComments(str) {
return str.replace(/\/\*[\s\S]*?\*\/|\/\/.*/g,'');
}
app.registerExtension({
name: "Comfy.DynamicPrompts",
nodeCreated(node) {
@ -15,7 +22,7 @@ app.registerExtension({
for (const widget of widgets) {
// Override the serialization of the value to resolve dynamic prompts for all widgets supporting it in this node
widget.serializeValue = (workflowNode, widgetIndex) => {
let prompt = widget.value;
let prompt = stripComments(widget.value);
while (prompt.replace("\\{", "").includes("{") && prompt.replace("\\}", "").includes("}")) {
const startIndex = prompt.replace("\\{", "00").indexOf("{");
const endIndex = prompt.replace("\\}", "00").indexOf("}");

17
web/extensions/core/widgetInputs.js
View File

@ -200,8 +200,23 @@ app.registerExtension({
applyToGraph() {
if (!this.outputs[0].links?.length) return;
function get_links(node) {
let links = [];
for (const l of node.outputs[0].links) {
const linkInfo = app.graph.links[l];
const n = node.graph.getNodeById(linkInfo.target_id);
if (n.type == "Reroute") {
links = links.concat(get_links(n));
} else {
links.push(l);
}
}
return links;
}
let links = get_links(this);
// For each output link copy our value over the original widget value
for (const l of this.outputs[0].links) {
for (const l of links) {
const linkInfo = app.graph.links[l];
const node = this.graph.getNodeById(linkInfo.target_id);
const input = node.inputs[linkInfo.target_slot];

25
web/scripts/app.js
View File

@ -125,10 +125,14 @@ export class ComfyApp {
if(ComfyApp.clipspace.imgs && node.imgs) {
if(node.images && ComfyApp.clipspace.images) {
if(ComfyApp.clipspace['img_paste_mode'] == 'selected') {
app.nodeOutputs[node.id + ""].images = node.images = [ComfyApp.clipspace.images[ComfyApp.clipspace['selectedIndex']]];
node.images = [ComfyApp.clipspace.images[ComfyApp.clipspace['selectedIndex']]];
}
else
app.nodeOutputs[node.id + ""].images = node.images = ComfyApp.clipspace.images;
else {
node.images = ComfyApp.clipspace.images;
}
if(app.nodeOutputs[node.id + ""])
app.nodeOutputs[node.id + ""].images = node.images;
}
if(ComfyApp.clipspace.imgs) {
@ -161,7 +165,16 @@ export class ComfyApp {
if(ComfyApp.clipspace.widgets) {
ComfyApp.clipspace.widgets.forEach(({ type, name, value }) => {
const prop = Object.values(node.widgets).find(obj => obj.type === type && obj.name === name);
if (prop && prop.type != 'button') {
if (prop && prop.type != 'image') {
if(typeof prop.value == "string" && value.filename) {
prop.value = (value.subfolder?value.subfolder+'/':'') + value.filename + (value.type?` [${value.type}]`:'');
}
else {
prop.value = value;
prop.callback(value);
}
}
else if (prop && prop.type != 'button') {
prop.value = value;
prop.callback(value);
}
@ -365,6 +378,10 @@ export class ComfyApp {
}
node.prototype.setSizeForImage = function () {
if (this.inputHeight) {
this.setSize(this.size);
return;
}
const minHeight = getImageTop(this) + 220;
if (this.size[1] < minHeight) {
this.setSize([this.size[0], minHeight]);

4
web/scripts/widgets.js
View File

@ -115,12 +115,12 @@ function addMultilineWidget(node, name, opts, app) {
// See how large each text input can be
freeSpace -= widgetHeight;
freeSpace /= multi.length;
freeSpace /= multi.length + (!!node.imgs?.length);
if (freeSpace < MIN_SIZE) {
// There isnt enough space for all the widgets, increase the size of the node
freeSpace = MIN_SIZE;
node.size[1] = y + widgetHeight + freeSpace * multi.length;
node.size[1] = y + widgetHeight + freeSpace * (multi.length + (!!node.imgs?.length));
node.graph.setDirtyCanvas(true);
}

15
web/style.css
View File

@ -50,7 +50,7 @@ body {
padding: 30px 30px 10px 30px;
background-color: var(--comfy-menu-bg); /* Modal background */
color: var(--error-text);
box-shadow: 0px 0px 20px #888888;
box-shadow: 0 0 20px #888888;
border-radius: 10px;
top: 50%;
left: 50%;
@ -84,7 +84,7 @@ body {
font-size: 15px;
position: absolute;
top: 50%;
right: 0%;
right: 0;
text-align: center;
z-index: 100;
width: 170px;
@ -252,7 +252,7 @@ button.comfy-queue-btn {
bottom: 0 !important;
left: auto !important;
right: 0 !important;
border-radius: 0px;
border-radius: 0;
}
.comfy-menu span.drag-handle {
visibility:hidden
@ -291,7 +291,7 @@ button.comfy-queue-btn {
.litegraph .dialog {
z-index: 1;
font-family: Arial;
font-family: Arial, sans-serif;
}
.litegraph .litemenu-entry.has_submenu {
@ -330,6 +330,13 @@ button.comfy-queue-btn {
color: var(--input-text) !important;
}
.comfy-context-menu-filter {
box-sizing: border-box;
border: 1px solid #999;
margin: 0 0 5px 5px;
width: calc(100% - 10px);
}
/* Search box */
.litegraph.litesearchbox {