justimyhxu / GRM

Large Gaussian Reconstruction Model for Efficient 3D Reconstruction and Generation
https://justimyhxu.github.io/projects/grm/
503 stars 27 forks source link

Why do most objects that the code creates look like cartoons? #6

Open Bardo-Konrad opened 3 months ago

a1600012888 commented 3 months ago

Because its mostly trained on ObjVerse. This dataset are somehow biased towards carton-like objects.

justimyhxu commented 3 months ago

I guess it is mainly caused by the multi-view diffusion models, which are always trained on objaverse. On the other hand, GRM is also only trained on objaverse and can't handle very complex lighting and scenes.