xxlong0
commented
11 months ago Hello. You point is great! We also consider the depth representation in the project, incorporating depth domain might be helpful. However, a three-domain multi-view diffusion model will be very memory-consuming for training, so we prefer normal domain instead of depth domain. This is because normal is rotation-invariant and translation-invariant for multi-views in a global coordinate system, but depth does not.
Great work!
I'm trying to experiment with depth maps. Usually for 3D reconstruction it's great to have a normals + depth pair (like in the humanNORM paper), but adding a 3rd domain with rendered depthmaps in Wonder3D might be too intensive.
In your opinion, if you try to only create the best geometry possible, and ignore textures, is RGB+Normals better than RGB+Depth? Or is stable diffusion very RGB dependant?