FluxInpaintPipeline overrides pixels outside the mask

[Clement-Lelievre]

Describe the bug

When inpainting (with diffusers==0.31.0 and torch==2.4.1) using FluxInpaintPipeline, I get some pixels outside the mask (and pretty far away from the mask border) that are overrided whereas the mask at their indices was black.
With either flux schnell or dev.

Is this expected?

I could paste the input image on the output image using the mask to get rid of this, but I'd like to check first if this is a bug or not.

Here is my input image and mask:

Here is the kind of diff I get:

Reproduction

I am using a small, square (224*224) image (see above), and a non-rectangular mask that has values only in {0, 255}.

from diffusers import FluxInpaintPipeline

pipe = FluxInpaintPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16, device='cuda')
prompt = "skull"
image= # above, convert it to RGB
mask = # above,  convert it to RGB
image = pipe(prompt=prompt, image=source, mask_image=mask, height=224, width=224, num_inference_steps=2, guidance_scale=3.5, strength=0.76).images[0]  # the latter params are probably not necessary to reproduce the issue
image.save("flux_inpainting.png")

# now visualize the pixel diff between the output you get and the input image, you should see non-zero pixels outside the mask, particularly on the right part

Logs

No response

System Info

• 🤗 Diffusers version: 0.31.0.dev0 or 0.31.0
• Platform: Linux-5.15.0-1048-aws-x86_64-with-glibc2.31
• Running on Google Colab?: No
• Python version: 3.12.0
• PyTorch version (GPU?): 2.4.1+cu121 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.25.0
• Transformers version: 4.44.2
• Accelerate version: 0.34.2
• PEFT version: 0.12.0
• Bitsandbytes version: not installed
• Safetensors version: 0.4.5
• xFormers version: not installed
• Accelerator: NVIDIA A10G, 23028 MiB

Who can help?

@asomoza @sayakpaul

[sayakpaul]

PyTorch 3.2.0? That sounds like invalid.

[Clement-Lelievre]

PyTorch 3.2.0? That sounds like invalid.

torch==2.4.1

[yiyixuxu]

cc @asomoza

[christopher5106]

I believe it's due to VAE encode/decode. It's not a bijective method. If confirmed, a blend with the decoded image at the end will solve the difference.

[SahilCarterr]

That can be done but it would require changing Pipelines including all ControlNets and Inpainting pipelines because they all have the same difference.

I believe it's due to VAE encode/decode. It's not a bijective method. If confirmed, a blend with the decoded image at the end will solve the difference.

[crapthings]

I believe it's due to VAE encode/decode. It's not a bijective method. If confirmed, a blend with the decoded image at the end will solve the difference.

@christopher5106
can this be apply to controlnet inpainting?
is this method same as apply_overlay?

what is different?

https://huggingface.co/docs/diffusers/api/image_processor#diffusers.image_processor.VaeImageProcessor.apply_overlay

https://huggingface.co/docs/diffusers/api/image_processor#diffusers.image_processor.VaeImageProcessor.postprocess

[christopher5106]

Right, you can use apply_overlay on the final PIL image in both inpaint and controlnet inpaint pipelines to correct the small VAE differences. But I would apply the mask directly on the image tensor before postprocess just after decode and probably not on PIL object

[christopher5106]

            ...
            image = self.vae.decode(latents, return_dict=False)[0]
            image = (1 - init_mask) * init_image + init_mask * image
            image = self.image_processor.postprocess(image, output_type=output_type)
            ...

[crapthings]

            ...
            image = self.vae.decode(latents, return_dict=False)[0]
            image = (1 - init_mask) * init_image + init_mask * image
            image = self.image_processor.postprocess(image, output_type=output_type)
            ...

What are the advantages of doing it this way? Compared to directly applying overlay? Also, can this be applied to control networks?

i'm trying to apply this into controlnet by only copy this line

image = self.vae.decode(latents, return_dict=False)[0]

to pag controlnet pipeline, but i got a grayish result

https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd.py#L1317

[putdanil]

I have issues too, parts of the original image (black parts) show up on the inpainted image.

Removing this line "solves" this, but seems to damage the diffusion process
latents = (1 - init_mask) * init_latents_proper + init_mask * latents

[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.