Fix wrong repeat of negative prompt

[hadaev8]

With current codebase this snippet would return error

num_images = 2
prompt = ["a photograph of an astronaut riding a horse"] * num_images

images = pipe(prompt, negative_prompt=prompt).images

Number of positive and negative prompts already checked in line elif batch_size != len(negative_prompt):
So no reason to repeat it by batch size.

[HuggingFaceDocBuilderDev]

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

[hadaev8]

Im not sure whats wrong with code quality.

[Thomas-MMJ]

Im not sure whats wrong with code quality.

Have you tried running

python -m black source_file_name.py
and seeing what changes?

[shirayu]

Anther formatting way is to run

make style

[pcuenca]

Hi @hadaev8, thanks a lot! However I think this was already in progress in #1002 (issue #779), which was just missing some tests. Can you please take a look there and see if there's anything missing?

[hadaev8]

@pcuenca Yes, I checked and change is same, but I fixed all (at least what I found) pipes.
Is it possible to merge our PRs?

[hadaev8]

@Thomas-MMJ
It added whitespace.

            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])

Whitespace was removed by vs code and I considered its fine.
Why whitespace should be here? Or is the problem on the side of code quality check?
Im confused.

[Thomas-MMJ]

Whitespace was removed by vs code and I considered its fine.
Why whitespace should be here? Or is the problem on the side of code quality check?
Im confused.

It just happens to be the style standard used for this code base. There are a variety of essentially arbitrary choices for consistency in style and this happens to be one.

[patrickvonplaten]

@pcuenca could you take a look here as it's very similar to your open PR here: #1002

[pcuenca]

Looks good to me! It's the same as the other PR but includes all the pipelines.

The only thing remaining would be to create some tests, as outlined by @patil-suraj here: #1002 (review).

@hadaev8 will you have time to work on it? Otherwise happy to take it :)

[patil-suraj]

Thanks for fixing this! I have same comments as @pcuenca, good for merge once those are resolved.

[hadaev8]

@pcuenca
I would do it, should you point me to an example of test?

[pcuenca]

@hadaev8 For now I'd use something like maybe https://github.com/huggingface/diffusers/blob/main/tests/test_pipelines.py#L265 to exercise a small pipeline with different inputs, and verify that no exceptions are raised and the number of images returned is correct. These tests would include all the cases:

• Single items for positive and negative prompts.
• Single items + num_prompts_per_image
• Lists for both.
• Mismatched cardinality between positive and negative (verify an error is returned).
• Lists + num_prompts_per_image

A better test would be to actually run a full stable diffusion pipeline and verify some slice of the results, but this is somewhat more involved and could be done later in my opinion.

[pcuenca]

I made the final changes here, I hope you don't mind @hadaev8, we want to include this very nice fix in our next release :) I also opened #1117 to create the tests so this PR should be ready to merge now. Thanks a lot @hadaev8!

[pcuenca]

Running the fast and slow tests locally.

[hadaev8]

@pcuenca
No, i dont mind.

So about this reshape, if repeat doesnt suit, where should be another function to avoid reshaping. Maybe expand or something.

How do you think, does it make sense for another pr or leave it as it is?

[pcuenca]

I had to open #1120 to fix a typo I made in my suggestion and pass the tests locally. Already merged, so we can close this one.

Thanks again @hadaev8, this is very very helpful!

[anton-l]

Continued in #1120