[wip] sdxl inpainting

[williamberman]

No description provided.

[HuggingFaceDocBuilderDev]

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

[bghira]

was this intended on going into the text to image script? it seems like a substantial change and makes no effort for backwards compatibility between these two very different training approaches.

are you planning on including offset noise and v_prediction sample prediction?

was there a specific reason that the final validation output was removed?

it also looks like the epoch count is hardcoded, and when resuming from ckpt, information on how to resume is discarded.

the wds train_dataset variable population looks like it discarded all of the embed/latent precompute logic, which is likely something you'll want to fit back in.

[williamberman]

@bghira this is just a checkpoint we're working on that we pushed the wip script to github to colab on a bit. If we decide to push a script to diffusers source, it won't be as adhoc as this :)