Add StableDiffusion repaint pipeline by nathanielherman · Pull Request #1341 · huggingface/diffusers

nathanielherman · 2022-11-19T07:25:52Z

No description provided.

nathanielherman · 2022-11-19T07:27:18Z


        timesteps = np.array(timesteps) * (self.config.num_train_timesteps // self.num_inference_steps)
        self.timesteps = torch.from_numpy(timesteps).to(device)
+        self.timesteps += self.config.steps_offset


repaint scheduler wasn't doing this but other schedulers do, I assume this step is supposed to be here? (it doesn't seem to affect output much)

nathanielherman · 2022-11-19T07:28:05Z

+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        # TODO: steps_offset is usually 1, so this effectively cuts the first step out when strength=1.0, is that desired? (for inpaint/img2img)


is this a bug in inpaint_legacy or intended? (ie inpaint_legacy will remove the first step when steps_offset is set to a default of 1)

HuggingFaceDocBuilderDev · 2022-11-19T07:29:53Z

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

patrickvonplaten

I'm actually ok to leave it here given that the code uses a lot of "Copied from statements" - @anton-l what do you think?

nathanielherman · 2022-11-28T22:14:50Z

Bump on this PR! Also @patrickvonplaten wdym by "leave it here"?

anton-l

Hi @nathanielherman, the PR looks great already! We've had to make some changes to the Stable Diffusion pipelines last week to accommodate for SD 2.0, so we'll need to do some tweaking here as well, hope that's ok :)

Most of the updates will be copied over when you run python utils/check_copies.py --fix_and_overwrite thanks to the # Copied from comments!

Also, could you add one integration test similar to https://github.com/huggingface/diffusers/blob/main/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint_legacy.py#L352 so that we have the same reference? 🙏

Randolph-zeng · 2022-12-08T09:28:48Z

+        prompt: Union[str, List[str]],
+        init_image: Union[torch.FloatTensor, PIL.Image.Image],
+        mask_image: Union[torch.FloatTensor, PIL.Image.Image],
+        num_inference_steps: Optional[int] = 50,


Sorry to jump in, but strength argument is missing in here

ah true, I updated to remove it entirely since I don't actually use it anymore (repaint just initializes the latents to random noise)

Randolph-zeng · 2022-12-08T09:38:21Z

+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            if t >= t_last:


and sorry to jump in again, but this line actually causes a bug:
in original repaint pipeline, the t >= last is actually a condition that wraps the main denoise logic, if you instead put such condition check in here, the first time that t >= last is satisfied, the latents size will be doubled but skipped the unet forward that puts its shape back, thus causing error in the next round.
An easy way to reproduce this is to set jump_n_sample to 2 or anything larger than 1

Just wanna quickly suggest a candidate version of this for loop to avoid the shape doubling bug mentioned above:

for i, t in enumerate(self.progress_bar(timesteps)): if t < t_last: # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, init_latents_orig, mask, generator).prev_sample # call the callback, if provided if callback is not None and i % callback_steps == 0: callback(i, t, latents) else: # compute the reverse: x_t-1 -> x_t latents = self.scheduler.undo_step(latents, t_last, generator) t_last = t

Ah I see, I think I can just put the if t >= t_last 2 lines earlier before the latent_model_input = to achieve the same effect.

anton-l · 2022-12-08T14:01:41Z

Thanks for catching the issues @Randolph-zeng!

@nathanielherman let me know if you don't have bandwidth this week, I'd be happy to help getting the PR ready for merging :)

…sion_repaint.py Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

nathanielherman · 2022-12-08T19:34:34Z

Hey! I did most of the updates and will look at adding an integration test. On that note, AFAICT from here https://github.com/huggingface/diffusers/blob/main/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint_legacy.py#L352, the code is actually loading the non-legacy pipeline rather than the legacy one? I'm a bit confused how that's not breaking it though, since it's initializing it with a non-inpainting model ("CompVis/stable-diffusion-v1-4")

Randolph-zeng · 2022-12-09T02:24:47Z

Just curious, am I the only one that experienced the DDIM degradation here ? When I use the code in this PR I noticed that the DDIM almost failed completely in producing any meaningful impaint image that corresponds to the prompt.
@nathanielherman Are you troubled by this same issue #1602 or does it work fine with you ? Thanks a lot if you can share some insight : )

nathanielherman · 2022-12-09T22:26:12Z

@Randolph-zeng hmm I'm confused by the linked issue, do you only get the issue for CFG outside of 6-7, or for any CFG? I only really use the default CFG of 7.5 but for that I get pretty reasonable outputs. (Though I wouldn't say repaint is obviously better results than default inpaint_legacy.)

nathanielherman · 2022-12-09T22:29:37Z

@anton-l bump on my question for the unit test, I just want to make sure I'm understanding correctly before I add my own unit test

anton-l · 2022-12-12T12:55:39Z

@nathanielherman regarding your question

the code is actually loading the non-legacy pipeline rather than the legacy one

The tests there carried over from the time when the legacy inpainting pipeline wasn't yet Legacy :) The pipeline loader substitutes the appropriate class for now:

diffusers/src/diffusers/pipeline_utils.py

Lines 539 to 544 in 31444f5

    
           if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse( 
        
               version.parse(config_dict["_diffusers_version"]).base_version 
        
           ) <= version.parse("0.5.1"): 
        
               from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy 
        
               pipeline_class = StableDiffusionInpaintPipelineLegacy

So you can safely assume that those tests are actually using StableDiffusionInpaintPipelineLegacy (we should probably update them, thanks for bringing it up!)

nathanielherman · 2022-12-12T22:37:02Z

red_cat_sitting_on_a_park_bench_repaint.npy.zip

nathanielherman · 2022-12-12T22:38:43Z

@anton-l makes sense! I've added the test and attached the npy file as a comment on this PR — IIUC from the docs, someone would need to upload that npy file and then I can update the test to download it from the url?

anton-l · 2022-12-15T12:01:07Z

@nathanielherman uploaded it to the repo: https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/red_cat_sitting_on_a_park_bench_repaint.npy
But also feel free to set up a personal repository on the hub to link the files, we can adapt later! :)

Great progress on the PR, let me know if it's ready for the final review!

nathanielherman · 2022-12-16T21:44:57Z

@anton-l perfect, it should be good for final review now!

patrickvonplaten · 2023-01-03T11:37:51Z

Gently ping @anton-l for a final review

patrickvonplaten · 2023-01-03T11:38:55Z

+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:


Can we remove all those deprecation messages? We should not add new models with deprecation messages :-)

patrickvonplaten · 2023-01-03T11:39:00Z

+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:


patrickvonplaten · 2023-01-03T11:39:08Z

+        ) < version.parse("0.9.0.dev0")
+        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+            deprecation_message = (


patrickvonplaten · 2023-01-03T11:45:52Z

Thanks for the nice PR @nathanielherman!

Three things from my side:

1. Could we also add fast tests here similar to:
  
  diffusers/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py
  
  Line 43 in f17fae6
  
  class StableDiffusionInpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
1. Could we also add docs as explained here: https://github.com/huggingface/diffusers/tree/main/docs
1. Could we remove all the deprecation messages?

Thanks!

anton-l · 2023-01-03T13:47:51Z

+@slow
+@require_torch_gpu
+class StableDiffusionRepaintPipelineIntegrationTests(unittest.TestCase):


Now that we're trying to move all of the slow integration tests to nightly runs (reference PR: #1664), this cab be moved as well:

Suggested change

@slow

@require_torch_gpu

class StableDiffusionRepaintPipelineIntegrationTests(unittest.TestCase):

@nightly

@require_torch_gpu

class StableDiffusionRepaintPipelineNightlyTests(unittest.TestCase):

Then the tests can be launched locally with RUN_NIGHTLY=1 pytest <your usual path and args>

anton-l · 2023-01-03T13:53:41Z

+
+from diffusers import RePaintScheduler, StableDiffusionRepaintPipeline
+from diffusers.utils import load_image, slow, torch_device
+from diffusers.utils.testing_utils import load_numpy, require_torch_gpu


As Patrick mentioned above, most of the models are now getting covered by common tests from PipelineTesterMixin that check API compatibility, common functionality, etc.
What we need here is just a test class similar to RepaintPipelineFastTests:

diffusers/tests/pipelines/repaint/test_repaint.py

Line 31 in f17fae6

class RepaintPipelineFastTests(PipelineTesterMixin, unittest.TestCase):

with pipeline_class = StableDiffusionRepaintPipeline and slightly adapted get_dummy_components() and get_dummy_inputs() which you can probably borrow without many changes from StableDiffusionInpaintPipelineFastTests:

diffusers/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py

Line 46 in f17fae6

def get_dummy_components(self):

These added tests will probably uncover some missing pieces in the pipeline, so feel free to ping us if something is tough to fix! :)

github-actions · 2023-02-26T15:04:15Z

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.

Markus-Pobitzer · 2023-03-09T09:34:07Z

Good Morning

Thanks for the great work. I am wondering why this pull request has been closed and how one can help.

vlordier · 2023-04-09T20:28:46Z

@anton-l is there something left we can do to merge this pipeline ?

anton-l · 2023-04-12T08:59:28Z

@vlordier the TODO is mostly just to update the tests as per the comments above (and fix any API issues uncovered by the common tests), and resolve the merge conflicts.

github-actions · 2023-05-06T15:03:23Z

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.

Repaint pipeline

5a478cf

nathanielherman commented Nov 19, 2022

View reviewed changes

patrickvonplaten reviewed Nov 21, 2022

View reviewed changes

anton-l reviewed Nov 29, 2022

View reviewed changes

Comment thread src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_repaint.py

Comment thread src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_repaint.py Outdated

Comment thread src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_repaint.py Outdated

anton-l reviewed Nov 29, 2022

View reviewed changes

Comment thread src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_repaint.py

Randolph-zeng suggested changes Dec 8, 2022

View reviewed changes

nathanielherman and others added 8 commits December 8, 2022 10:46

Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffu…

17fd219

…sion_repaint.py Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffu…

a696c14

…sion_repaint.py Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffu…

80737f4

…sion_repaint.py Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffu…

e9890e9

…sion_repaint.py Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

.

a01b16a

Merge remote-tracking branch 'origin/main' into repaint

b4dc538

fix bug + rm strength

269bcb1

run check_copies.py

9cb5d44

nathanielherman added 2 commits December 8, 2022 19:51

rename init_image to image

4cdec74

add test file

41833a5

fixes

ce924ec

add integration test

7996688

nathanielherman added 2 commits December 16, 2022 21:31

update image url

7f728b0

run make style and make quality

33e37eb

nathanielherman force-pushed the repaint branch from d75f51d to 33e37eb Compare December 16, 2022 21:42

Merge branch 'main' into repaint

1fabaf9

nathanielherman added 2 commits December 21, 2022 20:42

fix unit test + style warning

3f0ffc6

make fix-copies

3984383

patrickvonplaten reviewed Jan 3, 2023

View reviewed changes

anton-l reviewed Jan 3, 2023

View reviewed changes

github-actions Bot added the stale Issues that haven't received updates label Feb 26, 2023

github-actions Bot closed this Mar 7, 2023

anton-l removed the stale Issues that haven't received updates label Mar 9, 2023

anton-l reopened this Mar 9, 2023

github-actions Bot added the stale Issues that haven't received updates label Apr 2, 2023

huggingface deleted a comment from github-actions Bot Apr 4, 2023

github-actions Bot closed this May 14, 2023

Conversation

nathanielherman commented Nov 19, 2022

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

HuggingFaceDocBuilderDev commented Nov 19, 2022

Uh oh!

patrickvonplaten left a comment

Choose a reason for hiding this comment

Uh oh!

nathanielherman commented Nov 28, 2022

Uh oh!

anton-l left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anton-l commented Dec 8, 2022

Uh oh!

nathanielherman commented Dec 8, 2022

Uh oh!

Randolph-zeng commented Dec 9, 2022

Uh oh!

nathanielherman commented Dec 9, 2022

Uh oh!

nathanielherman commented Dec 9, 2022

Uh oh!

anton-l commented Dec 12, 2022

Uh oh!

nathanielherman commented Dec 12, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nathanielherman commented Dec 12, 2022

Uh oh!

anton-l commented Dec 15, 2022

Uh oh!

nathanielherman commented Dec 16, 2022

Uh oh!

patrickvonplaten commented Jan 3, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

patrickvonplaten commented Jan 3, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anton-l Jan 3, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Feb 26, 2023

Uh oh!

Markus-Pobitzer commented Mar 9, 2023

Uh oh!

vlordier commented Apr 9, 2023

Uh oh!

anton-l commented Apr 12, 2023

Uh oh!

github-actions Bot commented May 6, 2023

Uh oh!

nathanielherman commented Dec 12, 2022 •

edited

Loading

anton-l Jan 3, 2023 •

edited

Loading