Direct3D-S2

https://github.com/Deathdadev/Direct3D-S2-Pinokioupdated 6/18/2025, 12:07:42 PMindexed 1/23/2026, 7:45:55 PMOwner@death

[NVIDIA ONLY] Direct3D-S2 is a scalable 3D shape generation framework leveraging sparse volumetric representations for high-resolution outputs. It features Spatial Sparse Attention (SSA), a novel mechanism that accelerates Diffusion Transformer computations on sparse data, achieving up to 9.6× speedup in training. The unified Sparse VAE architecture maintains a consistent sparse volumetric format across input, latent, and output stages, significantly improving efficiency and stability.