search

HL-hanlin / Ctrl-Adapter

Official implementation of Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model
https://ctrl-adapter.github.io/
Apache License 2.0
342 stars 14 forks source link

Inpainting #17

Open dengtianbi opened 1 month ago

dengtianbi commented 1 month ago

Hi, I noticed that there are inpainting examples in the project, but it seems they are not included in the code, and I haven't seen any related models available for download. Could you please advise on how to perform video inpainting?

HL-hanlin commented 1 month ago

Hi there,

Thanks for your interest in this application! We haven't updated support for inpainting ControlNet code yet (will be added in a couple of weeks).

If you are in urgent need of that model, you can try to implement by some simple modification of our codebase:

During training, instead of giving a sequence of depth/canny frames as input, now give the random block masked frames as input. All other training hyper-parameters can basically remain the same. During inference, you can give a sequence of block masked frames, and the model should be able to recover the object in the masked area according to the guidance of text prompts.