Direct3D-V3 Paper
SIGGRAPH Asia 2026 / ACM TOG

Direct3D-V3

Latent-Free Diffusion Transformers
in the Native Voxel Domain

High-resolution 3D generation directly on signed distance fields, at up to 15363 voxel resolution.

Armored rover

Shuang Wu1,2Jiachen Qian3,2Youtian Lin1Feihu Zhang2Philip Torr4Yao Yao1,†Xun Cao1,†

1 Nanjing University2 DreamTech3 The University of Hong Kong4 University of Oxford† Corresponding authors

Overview

Diffusion Transformers
in Voxel Space.

Direct3D-V3 learns to generate geometry directly on raw signed distance fields. The patch-level DiT models coarse shape semantics, while the voxel-level DiT restores fine-grained geometric detail.

Image-conditioned meshes generated by Direct3D-V3.

Interactive results

Geometry explorer.

Compare six front-view renders against the RGB and normal inputs. Enlarge a preview to inspect surface detail, or load the full-resolution mesh to orbit and zoom.

14 cases

Serpent guardian

Direct3D-V3 and Hi3DV1.0 use the same coarse model.

Static previews · Orthographic front view

Method

Framework.

Direct3D-V3 removes the VAE and performs training and sampling directly in signed distance field (SDF) voxel space, scaling to a resolution of 15363.

Generation framework

Surface Patch Tokenization

Only a small fraction of a high-resolution SDF grid lies near the surface. To exploit this sparsity, we partition the grid into non-overlapping patches and retain those that intersect a coarse mesh. Each retained patch contributes one token containing its raw SDF values. We train the DiT directly on this sparse token sequence.

Dual-granularity DiT

The patch-level DiT combines surface patch tokens with features from the input image. Each patch feature then modulates its voxel tokens at every block of the voxel-level DiT, guiding the prediction of individual SDF values.

Dual-granularity DiT architecture

Patch-level and Voxel-level DiT

The patch-level DiT provides global semantic structure over sparse surface patches, while the voxel-level DiT restores local SDF variation within each patch. This dual-granularity design avoids applying expensive attention to all voxel tokens directly, yet still preserves the fine geometric detail needed for high-resolution shape generation.

Citation

Cite this work.

ACM Transactions on Graphics · Vol. 45, No. 6 · 2026

Paper