junshutang/Make-It-3D

Employs a two-stage pipeline combining neural radiance fields with 2D diffusion priors: the coarse stage optimizes geometry from frontal views while hallucinating novel views, then the refine stage converts to textured point clouds for enhanced realism. Integrates Stable Diffusion 2.0, DPT depth estimation, SAM segmentation, and BLIP2 captioning, with support for text-conditioned 3D generation and texture editing via contextual loss optimization.

1,886 stars. No commits in the last 6 months.