PixWorld unifies 3D scene generation and reconstruction in pixel space
PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space
Hugging FaceTrending 21
Summary
PixWorld is a unified diffusion model that jointly addresses 3D reconstruction and generation by supervising diffusion directly on rendered images. It introduces geometry perception loss aligned with a pretrained 3D foundation model, outperforming latent-space methods while matching state-of-the-art reconstruction.