Bake texture in latent space or in image space

[wchengad]

hi, thanks again for the great effort. Have you tried baking the texture in image space in diffusion steps? Why do you choose to bake the texture directly in latent space, rather than decode to image space, bake then encode back? It is more intuitive to me, like computing LPIPS loss in image space in SnycDiffusion.

Also, I have tried to texture in latent space while getting fuzzy results directly, could you help me analyze the difference?

Direct baking

[LIU-Yuxin]

Could you explain in details about your experiment of "directly texture in latent space"? Do you mean encoding a RGB texture to latent space, apply it to the surface of the object, render a latent camera view, then decode the camera view to RGB?