deepseek-ai/DreamCraft3D

Employs a multi-stage optimization pipeline combining NeRF/NeuS geometry sculpting with view-dependent diffusion guidance, followed by texture refinement via Bootstrapped Score Distillation using DreamBooth personalization. Integrates Zero123 for multi-view priors, Omnidata for depth/normal prediction, and builds on the threestudio framework with support for both text and image-conditioned generation. The approach alternates between refining the 3D scene representation and training scene-specific diffusion models, creating a self-reinforcing loop that prioritizes geometric consistency before texture enhancement.

3,008 stars. No commits in the last 6 months.