[WIP] [Core] Add support for ControlNet LoRA

[shauray8]

What does this PR do?

Adds support for ControlNet LoRA. continuing on @sayakpaul design #4899

Fixes#5800

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

The current design

Load things

fromdiffusersimport (
StableDiffusionXLControlNetPipeline, ControlNetModel, UNet2DConditionModel
)
importtorchpipe_id="stabilityai/stable-diffusion-xl-base-1.0"lora_id="stabilityai/control-lora"lora_filename="control-LoRAs-rank128/control-lora-canny-rank128.safetensors"unet=UNet2DConditionModel.from_pretrained(pipe_id, subfolder="unet", torch_dtype=torch.float16).to("cuda")
controlnet=ControlNetModel.from_unet(unet).to(device="cuda", dtype=torch.float16)
controlnet.load_lora_weights(lora_id, weight_name=lora_filename, controlnet_config=controlnet.config)

Infer

fromdiffusersimportAutoencoderKLfromdiffusers.utilsimportload_image, make_image_gridfromPILimportImageimportnumpyasnpimportcv2prompt="aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"negative_prompt="low quality, bad quality, sketches"image=load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png")
controlnet_conditioning_scale=0.5# recommended for good generalizationvae=AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe=StableDiffusionXLControlNetPipeline.from_pretrained(
pipe_id,
unet=unet,
controlnet=controlnet,
vae=vae,
torch_dtype=torch.float16,
).to("cuda")
image=np.array(image)
image=cv2.Canny(image, 100, 200)
image=image[:, :, None]
image=np.concatenate([image, image, image], axis=2)
image=Image.fromarray(image)
images=pipe(
prompt, negative_prompt=negative_prompt, image=image,
controlnet_conditioning_scale=controlnet_conditioning_scale,
num_images_per_prompt=4
).imagesfinal_image= [image] +imagesgrid=make_image_grid(final_image, 1, 5)
grid

Who can review?

@sayakpaul can you provide me with access to the notebook you mentioned

[sayakpaul]

Here you go: https://colab.research.google.com/drive/1S-NDshYL7N4S1ugF9Y86d3SRY-Fe0Qe-?usp=sharing

[George0726]

Hi folks, are there any other updates on the PR?

[andypotato]

+1 for this

[sayakpaul]

We don't have the bandwidth for this at the moment and hence #4899

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[AbnerCSZ]

@sayakpaul@shauray8@George0726@andypotato Requesting assistance from everyone, thank you very much.
controlnet = ControlNetModel.from_unet(unet).to(device="cuda", dtype=torch.float16)
controlnet.load_lora_weights(lora_id, weight_name=lora_filename, controlnet_config=controlnet.config)

The following error occurred during runtime:
AttributeError: 'ControlNetModel' object has no attribute 'load_lora_weights'
Is this due to a version issue? May I know which version you are using? My running version is as follows:
diffusers 0.27.0.dev0
transformers 4.35.2

[sayakpaul]

This feature hasn't been shipped yet.

[AbnerCSZ]

This feature hasn't been shipped yet.

@sayakpaul
Is there a test version that has been implemented that can be used? thanks！

[bghira]

@sayakpaul maybe this can come back sometime

[a-r-r-o-w]

Thanks for reminding about this @bghira! I'm going to take a look at this soon considering we now are working on adding support for Flux Control LoRA

[sayakpaul]

Just flagging some differences so that we don't conflate ControlNet LoRA (from SAI) and what BFL has released (Flux Control LoRA):

1. SAI's version is like so:
   a. Initialize a ControlNet from the weights of the denoiser.
   b. Add LoRA layers.
   c. Train the LoRA layers.
   d. Their pipeline is actually a ControlNet pipeline i.e., during denoising we leverage two things: the ControlNet (with LoRA added) module as well as the denoiser.

2. OTOH, BFL is doing something different (it's more akin to how regular img2img is done but with differences)
   a. Expand the input dimension of the denoiser ([3072, 64] -> [3072, 128]). First half is pre-trained weights and the rest is initialized to zeros.
   b. Add LoRA layers to the attention layers (and others) including the input embedding layer (img_in) and train them.
   c. During inference, compute the condition of a given image (depth condition, for example), encode it with the VAE's encoder, and then do regular img2img. There is NO separate ControlNet module involved.

Hope that helps clear out the differences a bit.