Would it be feasible to use PiD for image restoration or inpainting tasks?

[chenyu20230702]

Thank you for your great work. Would it be feasible to use PiD for image restoration or inpainting tasks? Also, when do you plan to open-source the training code?

[yifanlu0227]

Hi,

Image restoration and inpainting can be formulated as conditional generation tasks. In principle, the setup would be similar to prior latent diffusion-based restoration methods.

A practical approach would be to design a conditioning mechanism that injects the low-quality image, mask, or other task-specific guidance into the pixel diffusion backbone, but not necessarily encoding the input through a VAE :)

[yifanlu0227]

Re training code:

We are still going through the company's approval process, including more checkpoints, training code. The training code will also be released later, but we do not have an accurate estimate at the moment.