Open Jarvisss opened 3 weeks ago
Hi, here is a question, I want to bake latent texture from middle steps. For example, run 10 steps without UV projection, and start the UV projection from step 11.
The following code you provide seems lose the noise distribution in the original latent, due to the sparsity of gradient and the voronoi solve operation. https://github.com/LIU-Yuxin/SyncMVD/blob/8750811f571276a4c5d137c2cfd7c2e39d1ba62d/src/syncmvd/step.py#L75C1-L78C62
if texture is None: sample_views = [view for view in sample] sample_views, texture, _ = uvp.bake_texture(views=sample_views, main_views=main_views, exp=exp) sample_views = torch.stack(sample_views, axis=0)[:,:-1,...]
Can I use the baked original texture and add noise to get the noisy texture instead? $x_t=\sqrt{\alpha_t}x_0+\sqrt{1-\alpha_t}\epsilon$
Any advice can be helpful, thanks!
Yes, you can output the original latent views and use it for baking, then add noise. The bake function is not intended to be used for noisy latents.
Hi, here is a question, I want to bake latent texture from middle steps. For example, run 10 steps without UV projection, and start the UV projection from step 11.
The following code you provide seems lose the noise distribution in the original latent, due to the sparsity of gradient and the voronoi solve operation. https://github.com/LIU-Yuxin/SyncMVD/blob/8750811f571276a4c5d137c2cfd7c2e39d1ba62d/src/syncmvd/step.py#L75C1-L78C62
Can I use the baked original texture and add noise to get the noisy texture instead? $x_t=\sqrt{\alpha_t}x_0+\sqrt{1-\alpha_t}\epsilon$
Any advice can be helpful, thanks!