Direct3D-V3 is a work by teams from Nanjing University, DreamTech, HKU, and Oxford (SIGGRAPH Asia 2026 / ACM TOG). It uses a diffusion Transformer without a latent space to directly generate raw SDF voxel grids, refining them progressively across three resolution levels: 512, 1024, and 1536. Inference code and weights were released on 11 19; the input consists of a reference image (it is recommended to first use GPT Image 2.5, Lotus-2, or Qwen-Image-2.1 to create a normal map or a texture-free white model) and a coarse mesh aligned in orientation. The output is a GLB file, with decimation to 200 ten thousand triangles by default. The released version uses four layers of DINOv3 features for image conditioning and switches to the Muon optimizer for training.
Why it matters By modeling SDF voxels directly without VAE compression, the authors claim that the surface details in the released version are significantly finer than those in the paper's evaluation version; it is under the MIT license and available for commercial use.
Hugging Face Likes 4 · Hugging Face Upvotes 4 · Hugging Face Downloads 191
License: Commercial use allowed · ComfyUI: Not available · VRAM: ≈18 GB estimated
Reasons
- The model card provides three usage methods: Python calls, command-line batch processing, and a Gradio demo, along with accompanying sample images and coarse meshes
- The code is MIT licensed; DINOv3 used for image encoding and BiRefNet used for background segmentation each have their own licenses
Usage
- Linux + CUDA 13.0: Clone DreamTechAI/Direct3D-V3, set up a Python 3.12 environment according to the model card, install FlashAttention 2 and two voxelization extensions, then run python infer.py --image assets/input_imgs/6_normal.png --coarse-mesh assets/coarse_meshes/6.obj --resolution 1024 --output generated.glb, or run python app.py to open the web demo
- The coarse mesh generation model has not been released yet; use Hunyuan3D-2.1 or UltraShape to generate an OBJ / GLB first, keeping the same orientation as the reference image
Risks
- This is only the refinement stage: coarse mesh generation, texture generation, and training code are still on the to-do list, so you need to borrow other models to generate the coarse mesh first
- The model card does not specify VRAM requirements, only suggesting lowering max_tokens when running out of VRAM (which will lose detail); DINOv3 weights require requesting access on Hugging Face; ComfyUI support is not yet available