Create VAE feature extractor class

[patrickvonplaten]

We currently do not have a unified API that makes sure that for pipelines that accept image inputs that image input and output format always stay the same. This PR shows that nicely: #1882 (comment)

We should make sure that:

• a) All pipelines that accept images, can treat images of type PIL, numpy and torch
• b) if images are passed in numpy or torch, then the input format (image scale) should match 1-to-1 the output format
• c) Pipelines should be able to return images in PT format besides PIL and numpy so that one can run multiple image-to-image generations on GPU
• d) there is a lot of boiler plate code around "preparing images" and "preparing masks" => we should unify this code in a feature extractor as it's usually pretty much always the same
• e) Test that pipelines give the same results for all image inputs

This change will require to open a more involved PR, but it's time to tackle this! It would greatly help users that use img-2-img to make movies etc...

[patrickvonplaten]

@williamberman @yiyixuxu @pcuenca let me know if such a PR interests you, otherwise I can try to tackle it in 1-2 weeks :-)

[williamberman]

@patrickvonplaten I believe last time we talked about this, the idea was to use an image processing mixin that we would pull the common image processing code into. Is that still the case?

[patrickvonplaten]

Well actually I don't think we have to use a Mixin class, think we could just create a "weight" and "config" less VAEFeatureExtractor class that is instantiated in all the respective pipelines under:

self.vae_extractor = VAEExtractor(image_size=..., ...)

[williamberman]

ok cool, will put on TODO list but someone else feel free to pick up if have room :)

[patrickvonplaten]

@yiyixuxu feel free to take over :-)

[patrickvonplaten]

Being worked on in: #2617

[github-actions]

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.

[pcuenca]

I think this is done.