[docs/nits] Fix return values based on return_dict and minor doc updates
#7105
Conversation
| @@ -71,17 +71,14 @@ def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: | |||
| for batch_idx in range(batch_size): | |||
There was a problem hiding this comment.
Maybe this tensor2vid function could be moved into the utils folder because it has common use across multiple pipelines and there is yet to be a case where the implementation would be different? @ DN6
|
@DN6 can you take a look here? |
| outputs.append(batch_output) | ||
|
|
||
| if output_type == "np": | ||
| outputs = np.stack(outputs) | ||
|
|
||
| elif output_type == "pt": | ||
| elif output_type == "pt" or output_type == "latent": |
There was a problem hiding this comment.
how would the latent type output be used for animate diff?
can we pass them as input to this or another pipeline?
There was a problem hiding this comment.
We already support latents as input to __call__ right? Animate vid2vid, pia, controlnet, etc. can use latents generated by other pipelines.
The main focus of this PR is to handle return_dict correctly because all pipelines here would return a class object when return_dict=False and output_type=latent , but the expected is a tuple. Since pt and latent are essentially the same output type, like the other pipelines, the extra if-statement doesn't make sense in post processing.
There was a problem hiding this comment.
I'm AFK for the next few hours so please hold on if this gets approved for merge 😔 The return_dict parameter is not properly implemented in the community pipelines of animatediff as well. This is causing some issues on a downstream repo, so making the behaviour consistent and fixing across all pipes here would be awesome.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
@a-r-r-o-w let's resolve the merge conflicts and ship it! |
Thanks, I believe this is ready 🤗 |
What does this PR do?
I noticed that the
return_dictparameter was not correctly honored whenoutput_type="latent"and have made corrections to the relevant pipelines based on what I believe is the expected behaviour.Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@DN6 @yiyixuxu