Meet problems when exporting meshes with refined textures

[snowflakewang]

Hello, thank you for your brilliant work in HR multi-view diffusion models and Wonder3D-style 3D generation!

I run your code on my case and it works well during diffusion inference stages and NeuS optimization stages. But when I visualize the refined mesh _refine_ironmanrgba.obj using MeshLab, I meet some color artifacts as shown in the Figure 1.

But the video _refine_ironmanrgba.mp4 does not have such flaws as the Figure 2.

I wonder whether I missed some details.

Thanks a lot! :)

[qingtingJ]

where can I get the texture? I only find obj without texture in my result folder. thx

[snowflakewang]

@qingtingJ Hello, I cannot get texture UV, either. The mesh above only has vertex colors.

[altava-sgp]

I wonder this too. @pengHTYX Could you please explain about this ?

[pengHTYX]

It is because the differentiable renderer does not involve all vertices. One possible solution is to increase the rendering resolution and decrease the learning rate when refining texture.

Alternatively, you can directly unproject the multi-view color images onto the mesh. To do this, first project the vertices onto the multi-view image plane using the camera pose. Then, sample the pixel values using torch.grid_sample(). Finally, fuse the multi-view colors, weighting them by the cosine distance between the normal direction and the camera rays.

I hope these suggestions are helpful for you guys.

[altava-sgp]

@pengHTYX I will try soon. Thanks for your kind advice. 👍