Digit3D: 3D Multimodal MNIST Benchmark

Transforming 2D MNIST into rich 3D geometries (watertight meshes, sparse SDFs, and 6D normal-oriented point clouds) with continuous flow matching generative pipelines.

Khoi DO
Conquer3D Research & Open-Source Geometry Engine

Sparse Voxel VAE 3D Reconstruction

Continuous 3D surface extraction from sparse voxel grids. SimpleSparseVAE compresses sparse Signed Distance Fields into a 16-channel latent grid, decoded into continuous watertight triangle meshes via differentiable marching cubes.

0 Sample #000
Ground Truth
Reconstructed
1 Sample #002
Ground Truth
Reconstructed
2 Sample #004
Ground Truth
Reconstructed
3 Sample #006
Ground Truth
Reconstructed
4 Sample #008
Ground Truth
Reconstructed
5 Sample #010
Ground Truth
Reconstructed
6 Sample #012
Ground Truth
Reconstructed
7 Sample #014
Ground Truth
Reconstructed
8 Sample #016
Ground Truth
Reconstructed
9 Sample #018
Ground Truth
Reconstructed

Interactive 3D Mesh Inspection • Left: Ground Truth • Right: Sparse VAE Reconstruction • Shaded by Surface Normals