Fix positive / negative prompts #1002
Conversation
I tested a few different combinations: - Lists are supplied for both. - Lists + num_images_per_prompt. - Single positive, single negative, num_images_per_prompt > 1 - Same without negative prompts. It works, but it'd be nice to come up with a unit test.
|
The documentation is not available anymore as the PR was closed or merged. |
There was a problem hiding this comment.
Looks good to me, just posted one question in the comments. Also for tests ,maybe we could test each combination with dummy model
- single prompt
- list of prompts
- single prompt with
num_images_per_prompt - list of prompts with
num_images_per_prompt
and the same including the negative prompt. This will add a lot of tests but I think it should be fine.
| @@ -283,7 +283,7 @@ def __call__( | |||
|
|
|||
| # duplicate unconditional embeddings for each generation per prompt, using mps friendly method | |||
| seq_len = uncond_embeddings.shape[1] | |||
| uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1) | |||
There was a problem hiding this comment.
The change looks good to me, but I'm surprised that this doesn't work, what was the issue with it ?
There was a problem hiding this comment.
If we have three prompts and three negative prompts, batch_size would be three and we would be creating uncond_embeddings with shape (9, 77, 768) in this step. But text_embeddings is still (3, 77, 768), and latents_shape in line 299 would be wrong.
|
I think we should merge this soon as more people are finding the problem (it used to be a silent error, but it now fails). See for example #1035. @patrickvonplaten @anton-l can one of you please review on Monday? Meanwhile I'll prepare a couple of tests. |
|
Superseded by #1035. |
I tested a few different combinations:
It seems to work, but it'd be nice to come up with a unit test somehow. Not easy to do because the code to be tested is inside the pipeline function.
Fixes #779.