LibreDirect3DS2v11

Direct3D-S2 v1.1, DreamTech's image-to-3D model ("Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention"), repackaged for LibreYOLO. One picture of an object in, the whole object out as a mesh, generated as a signed distance field on sparse voxels at up to 1024^3.

LibreYOLO runs it in plain PyTorch, without torchsparse, flash-attn, Triton or custom CUDA kernels, so the same model runs on NVIDIA GPUs, Apple silicon (MPS) and CPU.

from libreyolo import LibreGen

model = LibreGen("direct3d-s2")      # downloads this repo
model("chair.png").save("chair.obj")

libreyolo generate source=chair.png model=direct3d-s2

This is a mirror, not a new model

The Direct3D-S2 weights are DreamTech's. Nothing was retrained, converted or re-encoded; every file under direct3d-s2-v-1-1/ is byte-identical to upstream and keeps upstream's layout.

• Upstream repository: wushuang98/Direct3D-S2
• Upstream revision: 8b04a8eddb7a56a0f4e89fe5f5b840c7d5610c00 (folder direct3d-s2-v-1-1)
• Code: DreamTechAI/Direct3D-S2
• Paper: arXiv:2505.17412

File	Networks
direct3d-s2-v-1-1/model_dense.ckpt	dense flow transformer and dense VAE
direct3d-s2-v-1-1/model_sparse_512.ckpt	sparse flow transformer and SDF VAE, 512^3
direct3d-s2-v-1-1/model_sparse_1024.ckpt	sparse flow transformer and SDF VAE, 1024^3
direct3d-s2-v-1-1/model_refiner.ckpt	512^3 SDF refiner
direct3d-s2-v-1-1/model_refiner_1024.ckpt	1024^3 sign refiner
direct3d-s2-v-1-1/config.yaml	network and sampler configuration

LibreYOLO reads the DINOv2 image encoder from facebook/dinov2-with-registers-large (Apache-2.0) and the BiRefNet background remover from LibreYOLO/LibreTRELLISl (MIT). SHA-256 of every file is listed in NOTICE.

License

MIT, as published by DreamTech. See LICENSE and NOTICE.