298977887/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-04-19 19:03:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

InstantX canny controlnet.

Browse Source
This commit is contained in:
comfyanonymous 2024-08-28 18:56:33 -04:00
parent 34eda0f853
commit b33cd61070
3 changed files with 63 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
comfy/controlnet.py
View File

@ -34,7 +34,7 @@ import comfy.t2i_adapter.adapter
import comfy.ldm.cascade.controlnet import comfy.ldm.cascade.controlnet
import comfy.cldm.mmdit import comfy.cldm.mmdit
import comfy.ldm.hydit.controlnet import comfy.ldm.hydit.controlnet
import comfy.ldm.flux.controlnet_xlabs import comfy.ldm.flux.controlnet
def broadcast_image_to(tensor, target_batch_size, batched_number): def broadcast_image_to(tensor, target_batch_size, batched_number):
@ -433,12 +433,25 @@ def load_controlnet_hunyuandit(controlnet_data):
def load_controlnet_flux_xlabs(sd): def load_controlnet_flux_xlabs(sd):
model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd) model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd)
control_model = comfy.ldm.flux.controlnet_xlabs.ControlNetFlux(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config) control_model = comfy.ldm.flux.controlnet.ControlNetFlux(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
control_model = controlnet_load_state_dict(control_model, sd) control_model = controlnet_load_state_dict(control_model, sd)
extra_conds = ['y', 'guidance'] extra_conds = ['y', 'guidance']
control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds) control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
return control return control
def load_controlnet_flux_instantx(sd):
new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(new_sd)
for k in sd:
new_sd[k] = sd[k]
control_model = comfy.ldm.flux.controlnet.ControlNetFlux(latent_input=True, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
control_model = controlnet_load_state_dict(control_model, new_sd)
latent_format = comfy.latent_formats.Flux()
extra_conds = ['y', 'guidance']
control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
return control
def load_controlnet(ckpt_path, model=None): def load_controlnet(ckpt_path, model=None):
controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True) controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
@ -504,8 +517,10 @@ def load_controlnet(ckpt_path, model=None):
elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
if "double_blocks.0.img_attn.norm.key_norm.scale" in controlnet_data: if "double_blocks.0.img_attn.norm.key_norm.scale" in controlnet_data:
return load_controlnet_flux_xlabs(controlnet_data) return load_controlnet_flux_xlabs(controlnet_data)
else: elif "pos_embed_input.proj.weight" in controlnet_data:
return load_controlnet_mmdit(controlnet_data) return load_controlnet_mmdit(controlnet_data)
elif "controlnet_x_embedder.weight" in controlnet_data:
return load_controlnet_flux_instantx(controlnet_data)
pth_key = 'control_model.zero_convs.0.0.weight' pth_key = 'control_model.zero_convs.0.0.weight'
pth = False pth = False

67
comfy/ldm/flux/controlnet_xlabs.py → comfy/ldm/flux/controlnet.py
View File

@ -1,6 +1,7 @@
#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py #Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
import torch import torch
import math
from torch import Tensor, nn from torch import Tensor, nn
from einops import rearrange, repeat from einops import rearrange, repeat
@ -13,34 +14,38 @@ import comfy.ldm.common_dit
class ControlNetFlux(Flux): class ControlNetFlux(Flux):
def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs): 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) 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 # add ControlNet blocks
self.controlnet_blocks = nn.ModuleList([]) self.controlnet_blocks = nn.ModuleList([])
for _ in range(self.params.depth): for _ in range(self.params.depth):
controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device) controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
# controlnet_block = zero_module(controlnet_block) # controlnet_block = zero_module(controlnet_block)
self.controlnet_blocks.append(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.gradient_checkpointing = False
self.input_hint_block = nn.Sequential( self.latent_input = latent_input
operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device), self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
nn.SiLU(), if not self.latent_input:
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device), self.input_hint_block = nn.Sequential(
nn.SiLU(), operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device), nn.SiLU(),
nn.SiLU(), operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device), nn.SiLU(),
nn.SiLU(), operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device), nn.SiLU(),
nn.SiLU(), operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device), nn.SiLU(),
nn.SiLU(), operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device), nn.SiLU(),
nn.SiLU(), operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
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( def forward_orig(
self, self,
@ -58,8 +63,10 @@ class ControlNetFlux(Flux):
# running on sequences img # running on sequences img
img = self.img_in(img) img = self.img_in(img)
controlnet_cond = self.input_hint_block(controlnet_cond) if not self.latent_input:
controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2) 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) controlnet_cond = self.pos_embed_input(controlnet_cond)
img = img + controlnet_cond img = img + controlnet_cond
vec = self.time_in(timestep_embedding(timesteps, 256)) vec = self.time_in(timestep_embedding(timesteps, 256))
@ -82,13 +89,25 @@ class ControlNetFlux(Flux):
block_res_sample = controlnet_block(block_res_sample) block_res_sample = controlnet_block(block_res_sample)
controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,) controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
return {"input": (controlnet_block_res_samples * 10)[:19]}
repeat = math.ceil(self.main_model_double / len(controlnet_block_res_samples))
if self.latent_input:
out_input = ()
for x in controlnet_block_res_samples:
out_input += (x,) * repeat
else:
out_input = (controlnet_block_res_samples * repeat)
return {"input": out_input[:self.main_model_double]}
def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs): def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
hint = hint * 2.0 - 1.0 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 bs, c, h, w = x.shape
patch_size = 2
x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size)) 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) img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)

2
comfy/utils.py
View File

@ -528,6 +528,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"), ("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"),
("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift), ("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift), ("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
("pos_embed_input.bias", "controlnet_x_embedder.bias"),
("pos_embed_input.weight", "controlnet_x_embedder.weight"),
} }
for k in MAP_BASIC: for k in MAP_BASIC: