Generate 'Latent Noise' in 'Transparent' areas during Img2Img.

[DrakenZA]

Currently, you can do a 'sort of' uncrop with SD, via doing img2img , and you fill the 'empty' space you are trying to uncrop into, with RGB noise or a dupe of the image flipped around.

Im pretty sure this would work a lot better if instead, you send the pipeline an image that has transparent areas, and those get filled with latent noise, which can then resolve into an uncrop.

I think that is how Dall-E 'handles it'. If you send Dall-E an image with transparent areas, those areas are instantly the 'masked areas' that need to get filled in, and they also get latent noise.