Open xiaoxiaodadada opened 4 months ago
Hi, thanks for your interest! That part of code is quite different from the code base shared now, so I didn't provide it here. I recalled we combines Patch-wise training with Latent Diffusion during another project and find it could also produce reasonably well results and hence put that in paper. The whole modification is not very complicated. Just change the conv_in layer in Stable Diffusion to let it accepts image+positions (4+2=6 channels) and then use the stochastic patch technique in this paper to extract patches and train only on patches. Hope this helps your implementation.
Thank you very much for your reply, I will modify the code according to it, it helped me a lot.
Can you share the Latent Patch Diffusion Model (LPDM) modification code that you mentioned in your paper that combines this method with the Latent Diffusion framework? If not, it doesn't matter. Thank you and I look forward to your reply.