Alibaba's TaoFlowForge: Image-to-3D That Keeps Thin Parts, Chamfer 0.0655 vs 0.1553

jiqizhixin · x · 2026-11-20

Alibaba's Taobao3D team presents TaoFlowForge in an arXiv preprint, tackling a production pain point of image-to-3D: dense, irregular triangulation makes meshes hard to edit, while simplification erases thin structures.

Approach: The model decouples vertex placement from connectivity prediction. Cascaded flow matching first builds coarse structure, then progressively refines vertices; a final stage estimates vertex connectivity and predicts normals. Geometry-aware losses encourage structurally important vertices to survive.

Results: Under single-image conditioning on Toys4K, it reports a Chamfer distance of 0.0655, versus 0.1553 for LATO.2 and 0.3264 for EdgeRunner, while producing artist-editable meshes.

Original post →

More from Multimodal

Multimodal channel →