DreamCraft3D is DeepSeek AI’s official implementation of its hierarchical 3D generation method, presented at ICLR 2024. Released under the MIT license on GitHub, where it has about 3,040 stars, it generates textured 3D assets from a text prompt or a single image. The project is a research codebase, last updated in 2025, and is notable for the quality and view-consistency of the objects it produces compared with earlier score-distillation approaches.
DreamCraft3D works in stages. First it generates a coherent reference image and an initial 3D representation, then it refines geometry using a bootstrapped diffusion prior: the model renders its current 3D object from many viewpoints, fine-tunes a view-conditioned diffusion model on those renderings, and uses that personalised prior to improve the object. A final texture-boosting stage applies Bootstrapped Score Distillation to recover fine surface detail. This coarse-to-fine, self-improving loop is what gives DreamCraft3D outputs their unusually consistent geometry and rich textures.
The bootstrapping idea is what sets DreamCraft3D apart: instead of relying only on a fixed 2D diffusion model to judge its output, it periodically renders its current 3D object from many angles and fine-tunes a personalised diffusion prior on those renderings, so the critic improves alongside the object. The repository provides the full training pipeline, configuration files and examples for generating assets from both text prompts and single images. As research code it expects a capable GPU and patience (each asset takes a substantial optimisation run), but it remains a reference point for quality-focused 3D generation.
DreamCraft3D is free and open source under the MIT license. There is no hosted service; generation runs on your own GPUs and takes meaningful compute time per asset.
3D generation researchers and technical artists prototyping AI asset pipelines for games, film pre-visualisation and virtual worlds.