facebookresearch/meshflow

Repository for the CVPR 2026 paper MeshFlow Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer by Weiyu Li, Antoine Toisoul, Tom Monnier, Roman Shapovalov, Rakesh Ranjan, Ping Tan and Andrea Vedaldi.

What it solves

MeshFlow is designed to generate high-quality, artistic 3D meshes quickly. It addresses the challenge of creating detailed 3D geometry that looks professionally crafted, reducing the generation time to approximately one second per mesh.

How it works

The system uses a two-part architecture combining a Variational Autoencoder (VAE) and a Diffusion Transformer (DiT):

  • MeshVAE: Encodes the topology of a mesh into continuous latent spaces and decodes them back into vertices, normals, and adjacency information.
  • MeshFlowDiT: A flow-matching Diffusion Transformer that operates on these latents. It can be conditioned on input geometry (using voxel RoPE) and an optional reference image (using a DINOv3 visual encoder).
  • Pipeline: The end-to-end process involves surface sampling, flow matching in the latent space, and final decoding via the VAE.

Who it’s for

This tool is for 3D artists, researchers in computer vision, and developers building applications that require rapid, high-fidelity 3D mesh generation from geometry or image references.

Highlights

  • High Speed: Generates artistic meshes in about one second.
  • Flexible Conditioning: Supports both geometry-based conditioning and reference-image conditioning.
  • ** keluarga DINOv3**: Integrates with DINOv3 for high-quality visual tokens when using images.
  • Resolution Control: Includes a mechanism to roughly control the number of vertices in the generated mesh.

Related

  • Project
    guochengqian/Magic123

    A system for generating high-quality 3D objects from a single 2D image by combining 2D and 3D diffusion priors in a coarse-to-fine pipeline.

  • Project
    haidog-yaqub/MeanFlow

    A PyTorch implementation of Mean Flows and Improved Mean Flows for high-quality, one-step image generation.

  • Project
    nv-tlabs/XCube

    XCube is a generative model for high-resolution sparse 3D voxel grids that uses hierarchical latent diffusion and VDB data structures to create detailed 3D objects and large-scale scenes.

  • Dispatch
    Stable Diffusion 3 Medium Integration with Diffusers

    Hugging Face has integrated Stable Diffusion 3 Medium (2B parameters) into the Diffusers library, introducing a Multimodal Diffusion Transformer (MMDiT) and rectified flow-matching for improved text-to-image synthesis.

  • Dispatch
    OpenAI Point-E: Generating 3D Point Clouds from Text Prompts

    OpenAI has introduced Point-E, a system that generates 3D point clouds from complex text prompts in 1-2 minutes on a single GPU, prioritizing sampling speed over absolute sample quality.