flux controlnet fix (control_modes batch & others) #9507

yiyixuxu · 2024-09-23T20:29:57Z

this PR:

fix control_mode shape when num_images_per_prompt > 1 (continue the PR from @christopher-beckham Fix flux controlnet mode to take into account batch size #9406)
fix a bug when multi-controlnet (follow up on from Several fixes to Flux ControlNet pipelines #9472)
fix a bug when use multi-controlnet with regular controllers that may not have single blocks

slow tests here

import torch
torch.cuda.set_device(2) 

# num_images_per_prompt = 1
num_images_per_prompt = 2
branch = "test" 
# branch = "main"

import gc
def flush():
    gc.collect()
    torch.cuda.empty_cache()

# test1: single controlnet (canny)
import torch
from diffusers.utils import load_image
from diffusers.pipelines.flux.pipeline_flux_controlnet import FluxControlNetPipeline
from diffusers.models.controlnet_flux import FluxControlNetModel

base_model = 'black-forest-labs/FLUX.1-dev'
controlnet_model = 'InstantX/FLUX.1-dev-Controlnet-Canny'
controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.bfloat16)
pipe.to("cuda")

control_image = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny/resolve/main/canny.jpg")
prompt = "A girl in city, 25 years old, cool, futuristic"
generator = torch.Generator(device="cuda").manual_seed(42)
images_out = pipe(
    prompt, 
    control_image=control_image,
    controlnet_conditioning_scale=0.6,
    num_inference_steps=28, 
    num_images_per_prompt=num_images_per_prompt,
    guidance_scale=3.5,
    generator=generator,
).images
print(images_out)
for i, image in enumerate(images_out):
    image.save(f"yiyi_test_7_{branch}_num_images_per_prompt_{num_images_per_prompt}_test1_out_{i}.png")


del pipe
flush()

# test2: single controlnet (union)
import torch
from diffusers.utils import load_image
from diffusers import FluxControlNetPipeline, FluxControlNetModel

base_model = 'black-forest-labs/FLUX.1-dev'
controlnet_model = 'InstantX/FLUX.1-dev-Controlnet-Union'

controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.bfloat16)
pipe.to("cuda")

control_image_canny = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alpha/resolve/main/images/canny.jpg")
controlnet_conditioning_scale = 0.5
control_mode = 0

width, height = control_image_canny.size

prompt = 'A bohemian-style female travel blogger with sun-kissed skin and messy beach waves.'
generator = torch.Generator(device="cuda").manual_seed(42)

images_out = pipe(
    prompt, 
    control_image=control_image_canny,
    control_mode=control_mode,
    width=width,
    height=height,
    controlnet_conditioning_scale=controlnet_conditioning_scale,
    num_inference_steps=24, 
    num_images_per_prompt=num_images_per_prompt,
    generator=generator,
    guidance_scale=3.5,
).images
for i, image in enumerate(images_out):
    image.save(f"yiyi_test_7_{branch}_num_images_per_prompt_{num_images_per_prompt}_test2_out_{i}.png")

del pipe
flush()



# test3: multiple controlnet (regular controlnets)
# note that we only have 1 regular controlnet now, so testing with 2 canny (this has no real use case but want to make sure it works regardless)
import torch
from diffusers.utils import load_image
from diffusers.pipelines.flux.pipeline_flux_controlnet import FluxControlNetPipeline
from diffusers.models.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel

base_model = 'black-forest-labs/FLUX.1-dev'
controlnet_canny = 'InstantX/FLUX.1-dev-Controlnet-Canny'
controlnet = FluxControlNetModel.from_pretrained(controlnet_canny, torch_dtype=torch.bfloat16)
multi_controlnet = FluxMultiControlNetModel([controlnet] * 2)
pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=multi_controlnet, torch_dtype=torch.bfloat16)
pipe.to("cuda")

control_image = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny/resolve/main/canny.jpg")
prompt = "A girl in city, 25 years old, cool, futuristic"
generator = torch.Generator(device="cuda").manual_seed(42)
# currently does not work on main branch
try:    
    images_out = pipe(
        prompt,
        control_image=[control_image, control_image],
        controlnet_conditioning_scale=[0.6, 0.6],
        num_inference_steps=28,
        guidance_scale=3.5,
        num_images_per_prompt=num_images_per_prompt,
        generator=generator,
    ).images
    for i, image in enumerate(images_out):
        image.save(f"yiyi_test_7_{branch}_num_images_per_prompt_{num_images_per_prompt}_test3_out_{i}.png")
except Exception as e:
    print(e)

del pipe
flush()

# test4: multi controlnet with union
import torch
from diffusers.utils import load_image
from diffusers import FluxControlNetPipeline, FluxControlNetModel, FluxMultiControlNetModel

base_model = 'black-forest-labs/FLUX.1-dev'
controlnet_model_union = 'InstantX/FLUX.1-dev-Controlnet-Union'

controlnet_union = FluxControlNetModel.from_pretrained(controlnet_model_union, torch_dtype=torch.bfloat16)
controlnet = FluxMultiControlNetModel([controlnet_union]) # we always recommend loading via FluxMultiControlNetModel

pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.bfloat16)
pipe.to("cuda")

prompt = 'A bohemian-style female travel blogger with sun-kissed skin and messy beach waves.'
control_image_depth = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union/resolve/main/images/depth.jpg")
control_mode_depth = 2

control_image_canny = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union/resolve/main/images/canny.jpg")
control_mode_canny = 0

width, height = control_image_canny.size
generator = torch.Generator(device="cuda").manual_seed(42)
out_images = pipe(
    prompt,
    control_image=[control_image_depth, control_image_canny],
    control_mode=[control_mode_depth, control_mode_canny],
    width=width,
    height=height,
    controlnet_conditioning_scale=[0.2, 0.4],
    num_inference_steps=24, 
    num_images_per_prompt=num_images_per_prompt,
    guidance_scale=3.5,
    generator=generator,
).images


for i, out in enumerate(out_images):
    out.save(f'yiyi_test_7_{branch}_num_images_per_prompt_{num_images_per_prompt}_test4_out_{i}.png')

…p control mode handling for multi case

…rolnet_modes' into flux_controlnet_modes_yiyi

yiyixuxu · 2024-09-23T20:33:43Z

src/diffusers/models/controlnet_flux.py

-                            control_single_block_samples, single_block_samples
-                        )
-                    ]
+                    if block_samples is not None and control_block_samples is not None:


flux controlnet output can contain None depends on the controlnet has single layers or not,
e.g. union has it https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union/blob/main/config.json
but canny does not https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny/blob/main/config.json#L16

yiyixuxu · 2024-09-23T20:34:19Z

src/diffusers/pipelines/flux/pipeline_flux_controlnet.py

                control_mode = torch.tensor(control_mode).to(device, dtype=torch.long)
-                control_mode = control_mode.reshape([-1, 1])
+                control_mode = control_mode.view(-1, 1).expand(control_image.shape[0], 1)


see more context on this PR #9406

yiyixuxu · 2024-09-23T20:35:45Z

src/diffusers/pipelines/flux/pipeline_flux_controlnet.py

-                guidance = (
-                    torch.tensor([guidance_scale], device=device) if self.controlnet.config.guidance_embeds else None
-                )
+                if isinstance(self.controlnet, FluxMultiControlNetModel):


this PR #9472 makes it possible to use a schnell base with dev controlnet, however it breaks the case when it is a FluxMultiControlNetModel because FluxMultiControlNetModel does not config

I think it is safer to use this :
in controlnet_flux.py, in the foward method check this:

if isinstance(self.time_text_embed, CombinedTimestepTextProjEmbeddings): guidance = None

HuggingFaceDocBuilderDev · 2024-09-23T20:36:15Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

asomoza

LGTM I also tested it just in case.

christopher-beckham · 2024-09-25T20:22:25Z

One suggestion I have, it doesn't appear that any checking is actually done to see the lengths of all the arguments in zip are the same length. This could lead to a "silent bug" where e.g. the user may accidentally mis-specify something like conditioning_scale=[0.6] (i.e. a length-1 list instead of 2) and zip doesn't give any warnings if the length of each of its arguments are not the same.

diffusers/src/diffusers/models/controlnet_flux.py

Lines 479 to 484 in 560449d

    
           # Regular Multi-ControlNets 
        
           # load all ControlNets into memories 
        
           else: 
        
               for i, (image, mode, scale, controlnet) in enumerate( 
        
                   zip(controlnet_cond, controlnet_mode, conditioning_scale, self.nets) 
        
               ):

It might be useful to add a check in for that before the zip call. I wrote this below:

lengths = [
    len(controlnet_cond), len(controlnet_mode), len(conditioning_scale), 
        len(self.nets)
]
if len(set(lengths)) != 1:
    raise ValueError(
        "`controlnet_cond`, `controlnet_mode`, `conditioning_scale`, `self.nets` " + 
        f"  must all have the same length, we got (respectively): {lengths}"
    )

Otherwise it lgtm. Thanks!

Edit: actually ideally it would also be good to check if they're lists too... Type checking in Python is annoying, isn't it?

yiyixuxu · 2024-09-26T05:09:43Z

@christopher-beckham nice suggestion! thanks!
I think we should add a input check for all our controlnet (will do that in a follow up PR)

christopher-beckham · 2024-09-26T15:14:56Z

@yiyixuxu thanks! Just fyi I have something already in the works for the input checking, and can PR that today if you want. It's actually a lot easier than I thought it was, just some logic which needs to be pulled in from the corresponding controlnet pipeline for SDXL. Thanks.

* flux controlnet mode to take into account batch size * incorporate yiyixuxu's suggestions (cleaner logic) as well as clean up control mode handling for multi case * fix * fix use_guidance when controlnet is a multi and does not have config --------- Co-authored-by: Christopher Beckham <[email protected]> Co-authored-by: Sayak Paul <[email protected]>

christopher-beckham and others added 6 commits September 10, 2024 15:20

flux controlnet mode to take into account batch size

370f382

incorporate yiyixuxu's suggestions (cleaner logic) as well as clean u…

c333d89

…p control mode handling for multi case

fix

7c95f0b

Merge branch 'main' into christopher-beckham/fix_flux_controlnet_modes

61e0950

Merge remote-tracking branch 'chris/christopher-beckham/fix_flux_cont…

db8178a

…rolnet_modes' into flux_controlnet_modes_yiyi

fix use_guidance when controlnet is a multi and does not have config

560449d

yiyixuxu requested a review from asomoza September 23, 2024 20:31

yiyixuxu commented Sep 23, 2024

View reviewed changes

yiyixuxu mentioned this pull request Sep 23, 2024

Fix flux controlnet mode to take into account batch size #9406

Closed

6 tasks

asomoza approved these changes Sep 25, 2024

View reviewed changes

yiyixuxu mentioned this pull request Sep 26, 2024

[FluxMultiControlNetModel] object has no attribute 'config #9534

Closed

Merge branch 'main' into flux_controlnet_modes_yiyi

53bee55

yiyixuxu merged commit 9cd3755 into main Sep 26, 2024
17 of 18 checks passed

yiyixuxu deleted the flux_controlnet_modes_yiyi branch September 26, 2024 05:09

christopher-beckham mentioned this pull request Sep 28, 2024

Image dimension checking for ControlNet FLUX #9550

Open

6 tasks

rshah240 mentioned this pull request Oct 4, 2024

config attribute not foud error for FluxImagetoImage Pipeline for multi controlnet solved #9586

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

flux controlnet fix (control_modes batch & others) #9507

flux controlnet fix (control_modes batch & others) #9507

yiyixuxu commented Sep 23, 2024 •

edited

Loading

yiyixuxu Sep 23, 2024

yiyixuxu Sep 23, 2024

yiyixuxu Sep 23, 2024

xziayro Sep 25, 2024 •

edited

Loading

HuggingFaceDocBuilderDev commented Sep 23, 2024

asomoza left a comment •

edited

Loading

christopher-beckham commented Sep 25, 2024 •

edited

Loading

yiyixuxu commented Sep 26, 2024

christopher-beckham commented Sep 26, 2024

flux controlnet fix (control_modes batch & others) #9507

flux controlnet fix (control_modes batch & others) #9507

Conversation

yiyixuxu commented Sep 23, 2024 • edited Loading

yiyixuxu Sep 23, 2024

Choose a reason for hiding this comment

yiyixuxu Sep 23, 2024

Choose a reason for hiding this comment

yiyixuxu Sep 23, 2024

Choose a reason for hiding this comment

xziayro Sep 25, 2024 • edited Loading

Choose a reason for hiding this comment

HuggingFaceDocBuilderDev commented Sep 23, 2024

asomoza left a comment • edited Loading

Choose a reason for hiding this comment

christopher-beckham commented Sep 25, 2024 • edited Loading

yiyixuxu commented Sep 26, 2024

christopher-beckham commented Sep 26, 2024

yiyixuxu commented Sep 23, 2024 •

edited

Loading

xziayro Sep 25, 2024 •

edited

Loading

asomoza left a comment •

edited

Loading

christopher-beckham commented Sep 25, 2024 •

edited

Loading