wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-07-21 23:41:28 +08:00
Code Issues Actions 10 Packages Projects Releases Wiki Activity

Merge branch 'master' of github.com:comfyanonymous/ComfyUI

Browse Source
This commit is contained in:
doctorpangloss
2024-08-29 16:37:30 -07:00
parent ccdbd957ef ea3f39bd69
commit fd503d8a96
12 changed files with 332 additions and 206 deletions
Download Patch File Download Diff File Expand all files Collapse all files
comfy/ldm/common_dit.py
+13
View File
@@ -1,4 +1,5 @@
import torch
import comfy.ops
def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
@@ -6,3 +7,15 @@ def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)
try:
rms_norm_torch = torch.nn.functional.rms_norm
except:
rms_norm_torch = None
def rms_norm(x, weight, eps=1e-6):
if rms_norm_torch is not None:
return rms_norm_torch(x, weight.shape, weight=comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
else:
rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps)
return (x * rrms) * comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device)
comfy/ldm/flux/controlnet.py
+140
View File
@@ -0,0 +1,140 @@
#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
import torch
import math
from torch import Tensor, nn
from einops import rearrange, repeat
from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
MLPEmbedder, SingleStreamBlock,
timestep_embedding)
from .model import Flux
import comfy.ldm.common_dit
class ControlNetFlux(Flux):
def __init__(self, latent_input=False, image_model=None, dtype=None, device=None, operations=None, **kwargs):
super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
self.main_model_double = 19
self.main_model_single = 38
# add ControlNet blocks
self.controlnet_blocks = nn.ModuleList([])
for _ in range(self.params.depth):
controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
self.controlnet_blocks.append(controlnet_block)
self.controlnet_single_blocks = nn.ModuleList([])
for _ in range(self.params.depth_single_blocks):
self.controlnet_single_blocks.append(operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device))
self.gradient_checkpointing = False
self.latent_input = latent_input
self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
if not self.latent_input:
self.input_hint_block = nn.Sequential(
operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
)
def forward_orig(
self,
img: Tensor,
img_ids: Tensor,
controlnet_cond: Tensor,
txt: Tensor,
txt_ids: Tensor,
timesteps: Tensor,
y: Tensor,
guidance: Tensor = None,
) -> Tensor:
if img.ndim != 3 or txt.ndim != 3:
raise ValueError("Input img and txt tensors must have 3 dimensions.")
# running on sequences img
img = self.img_in(img)
if not self.latent_input:
controlnet_cond = self.input_hint_block(controlnet_cond)
controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
controlnet_cond = self.pos_embed_input(controlnet_cond)
img = img + controlnet_cond
vec = self.time_in(timestep_embedding(timesteps, 256))
if self.params.guidance_embed:
vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
vec = vec + self.vector_in(y)
txt = self.txt_in(txt)
ids = torch.cat((txt_ids, img_ids), dim=1)
pe = self.pe_embedder(ids)
controlnet_double = ()
for i in range(len(self.double_blocks)):
img, txt = self.double_blocks[i](img=img, txt=txt, vec=vec, pe=pe)
controlnet_double = controlnet_double + (self.controlnet_blocks[i](img),)
img = torch.cat((txt, img), 1)
controlnet_single = ()
for i in range(len(self.single_blocks)):
img = self.single_blocks[i](img, vec=vec, pe=pe)
controlnet_single = controlnet_single + (self.controlnet_single_blocks[i](img[:, txt.shape[1] :, ...]),)
repeat = math.ceil(self.main_model_double / len(controlnet_double))
if self.latent_input:
out_input = ()
for x in controlnet_double:
out_input += (x,) * repeat
else:
out_input = (controlnet_double * repeat)
out = {"input": out_input[:self.main_model_double]}
if len(controlnet_single) > 0:
repeat = math.ceil(self.main_model_single / len(controlnet_single))
out_output = ()
if self.latent_input:
for x in controlnet_single:
out_output += (x,) * repeat
else:
out_output = (controlnet_single * repeat)
out["output"] = out_output[:self.main_model_single]
return out
def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
patch_size = 2
if self.latent_input:
hint = comfy.ldm.common_dit.pad_to_patch_size(hint, (patch_size, patch_size))
hint = rearrange(hint, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
else:
hint = hint * 2.0 - 1.0
bs, c, h, w = x.shape
x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
h_len = ((h + (patch_size // 2)) // patch_size)
w_len = ((w + (patch_size // 2)) // patch_size)
img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance)
comfy/ldm/flux/controlnet_xlabs.py
-104
View File
@@ -1,104 +0,0 @@
#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
import torch
from torch import Tensor, nn
from einops import rearrange, repeat
from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
MLPEmbedder, SingleStreamBlock,
timestep_embedding)
from .model import Flux
import comfy.ldm.common_dit
class ControlNetFlux(Flux):
def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
# add ControlNet blocks
self.controlnet_blocks = nn.ModuleList([])
for _ in range(self.params.depth):
controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
# controlnet_block = zero_module(controlnet_block)
self.controlnet_blocks.append(controlnet_block)
self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
self.gradient_checkpointing = False
self.input_hint_block = nn.Sequential(
operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
nn.SiLU(),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
)
def forward_orig(
self,
img: Tensor,
img_ids: Tensor,
controlnet_cond: Tensor,
txt: Tensor,
txt_ids: Tensor,
timesteps: Tensor,
y: Tensor,
guidance: Tensor = None,
) -> Tensor:
if img.ndim != 3 or txt.ndim != 3:
raise ValueError("Input img and txt tensors must have 3 dimensions.")
# running on sequences img
img = self.img_in(img)
controlnet_cond = self.input_hint_block(controlnet_cond)
controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
controlnet_cond = self.pos_embed_input(controlnet_cond)
img = img + controlnet_cond
vec = self.time_in(timestep_embedding(timesteps, 256))
if self.params.guidance_embed:
vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
vec = vec + self.vector_in(y)
txt = self.txt_in(txt)
ids = torch.cat((txt_ids, img_ids), dim=1)
pe = self.pe_embedder(ids)
block_res_samples = ()
for block in self.double_blocks:
img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
block_res_samples = block_res_samples + (img,)
controlnet_block_res_samples = ()
for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
block_res_sample = controlnet_block(block_res_sample)
controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
return {"input": (controlnet_block_res_samples * 10)[:19]}
def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
hint = hint * 2.0 - 1.0
bs, c, h, w = x.shape
patch_size = 2
x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
h_len = ((h + (patch_size // 2)) // patch_size)
w_len = ((w + (patch_size // 2)) // patch_size)
img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance)
comfy/ldm/flux/layers.py
+2 -5
View File
@@ -5,7 +5,7 @@ import torch
from torch import Tensor, nn
from .math import attention, rope
from ... import ops
from ..common_dit import rms_norm
class EmbedND(nn.Module):
@@ -63,10 +63,7 @@ class RMSNorm(torch.nn.Module):
self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
def forward(self, x: Tensor):
x_dtype = x.dtype
x = x.float()
rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
return (x * rrms).to(dtype=x_dtype) * ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
return rms_norm(x, self.scale, 1e-6)
class QKNorm(torch.nn.Module):
comfy/ldm/modules/diffusionmodules/mmdit.py
+2 -22
View File
@@ -356,29 +356,9 @@ class RMSNorm(torch.nn.Module):
else:
self.register_parameter("weight", None)
def _norm(self, x):
"""
Apply the RMSNorm normalization to the input tensor.
Args:
x (torch.Tensor): The input tensor.
Returns:
torch.Tensor: The normalized tensor.
"""
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x):
"""
Forward pass through the RMSNorm layer.
Args:
x (torch.Tensor): The input tensor.
Returns:
torch.Tensor: The output tensor after applying RMSNorm.
"""
x = self._norm(x)
if self.learnable_scale:
return x * self.weight.to(device=x.device, dtype=x.dtype)
else:
return x
return comfy.ldm.common_dit.rms_norm(x, self.weight, self.eps)
class SwiGLUFeedForward(nn.Module):