MeshFlow

MeshFlow is a research method from Meta AI and HKUST for generating artist-like 3D meshes with explicit vertices and connectivity. Instead of autoregressively predicting discrete face tokens, it compresses a continuous vertex-based mesh representation (positions, normals and contrastively learned edge embeddings) into compact MeshVAE latents, then generates all latent tokens in parallel with a flow-based diffusion transformer. It was presented at CVPR 2026 as a Highlight.\n\nThe approach avoids coordinate quantization and scales linearly with mesh size. The project page reports inference of roughly one second, and a YouTube creator summary cites a claim of being about 18x faster than other methods. MeshVAE reconstructs meshes with 512 latent codes, using TokenMerge/TokenSplit to downsample and restore vertex tokens. The paper supports unconditional, point-cloud-conditioned and image-conditioned generation.\n\nThe released model expects a surface point cloud as input for coarse shape guidance. The authors state that single-image conditioning is still being developed. Code has reportedly been released, but no repository, weights or license details appeared in the provided material, so those are left unspecified.