Direct3D-S2

Direct3D-S2 turns a reference image into a source mesh for artists who can handle the rest of the asset. Its sparse-volume approach targets detailed geometry; textures, rigs and animations still need their own pipeline. The documented v1.1 setup favors 1024 resolution, while the authors describe 512 as a lower-quality intermediate and advise against using it as the final output. It is a sensible candidate for baking and retopology when you have the GPU memory and finishing budget. Judge the extra detail at your game's camera distance: geometry nobody sees still costs time to clean up. The paper's training-efficiency claims say nothing about the exported mesh's frame rate.

Commercial use & license

The project and model card state MIT; retain the required notices. Dependencies and datasets have separate terms. PyMeshFix lists GPL-3.0, while its README says commercial MeshFix use requires a licensing contract. Resolve that mismatch before adopting the mesh-repair code for production.

Adult content rules

The MIT licence contains no adult-content prohibition, but it does not grant rights to reference images, likenesses or third-party assets. Check the exact checkpoints and dependencies you deploy. Hosted services add separate rules; no blanket adult-content permission is established.

Read the policy ↗

What you need

The repository documents Ubuntu 22.04, CUDA 12.1, PyTorch 2.5.1 and TorchSparse. Allow at least 10 GB VRAM for 512; the recommended 1024 route needs around 24 GB.

Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →