Update pipeline_stable_diffusion_inpaint_legacy.py

[Randolph-zeng]

This PR is related to the issue https://github.com/huggingface/diffusers/issues/1581. Please refer to it as why for the proposed PR, if more extensive local tests needed please let me know : )

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[Randolph-zeng]

btw, It seems tha there are some errors regarding compatability and formatting, if someone can confirm that this is indeed an undesired off by one error I am happy to follow the contributing procedure to run local test to fix those CI errors

[averad]

Any thoughts about adding the changes to the Onnx Pipeline as well?

pipeline_onnx_stable_diffusion_inpaint_legacy.py - Line 374

# add noise to latents using the timesteps
noise = generator.randn(*init_latents.shape).astype(latents_dtype)
t_minus_one = timesteps - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
if t_minus_one > 0:
	init_latents = self.scheduler.add_noise(
		torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(t_minus_one)
	)
else:
	init_latents = self.scheduler.add_noise(
		torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps)
	)
init_latents = init_latents.numpy()

Not 100% sure I incorporated the changes correctly

[Randolph-zeng]

oh yeah, thanks for the catch! I will incorporate that and fix the errors by doing some local test

[averad]

Code change is similar to PR #1585 - Linking so both are reviewed at the same time.

[averad]

@patrickvonplaten analog to https://github.com/CompVis/stable-diffusion/pull/533 ?

[patrickvonplaten]

I'm not sure this change is correct here @Randolph-zeng - can you explain a bit more why you think the timestep should be changed? Also, the inpaint stable diffusion legacy pipeline really isn't an official pipeline because there is no reference paper. I'm not sure it should follow the RePaint example . @anton-l wdyt?

[Randolph-zeng]

Hi @patrickvonplaten , I totally agree with you that SD-Legacy is not a standard pipeline following any paper and is different from RePaint. It is just that every diffusion models are trained with fixed noise schedule that associates with the timestamps. If you are using a sampler that samples in 20 steps instead of 1000 steps, then mixing parts of images that are 50 steps away from each other might result in a very different input than the models were trained with, and thus deteriorate the model performance.

[github-actions[bot]]

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.