MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer

MeshFlow is a state-of-the-art 3D generative model introduced as a CVPR 2026 Highlight by researchers from Meta AI and HKUST (Weiyu Li, Antoine Toisoul, Tom Monnier, Roman Shapovalov, Rakesh Ranjan, Ping Tan, and Andrea Vedaldi).

MeshFlow addresses the fundamental limitations of prior Auto-Regressive (AR) 3D mesh generators—such as slow token-by-token sequence generation and quantization artifacts from discretizing vertex coordinates—by establishing a continuous latent space representation and leveraging parallel flow matching diffusion transformers.

🔬 Architecture & Methodology

MeshFlow operates via a two-stage architecture:

1. MeshVAE (Variational Autoencoder):
   • Learns a compact continuous latent space for complex 3D triangular meshes.
   • Encodes continuous 3D vertex positions, surface normals, and face connectivity without requiring coordinate discretization or quantization.
   • Optimized with graph contrastive losses and topological regularization to ensure clean manifold mesh reconstruction.

1. Flow-Based Diffusion Transformer:
   • Built on Rectified Flow formulation over continuous MeshVAE latents.
   • Replaces slow sequential AR token prediction with parallel generation of all mesh vertices and faces simultaneously.
   • Conditioned on text prompts, reference images, or sparse point clouds via cross-attention mechanisms.

Input (Text / Image / Point Cloud) ──► Latent Condition Encoder ──► Flow-based DiT ──► Continuous Mesh Latent ──► MeshVAE Decoder ──► 3D Triangular Mesh (.obj / .ply)

⚡ Key Features & Benchmarks

• 18× – 26× Faster Inference: Generates artist-grade 3D meshes in ~1 second per asset on modern GPUs, drastically outperforming autoregressive baseline models like MeshGPT.
• Zero Coordinate Quantization: Operates in continuous space, eliminating vertex snapping and staircase artifacts common in discrete coordinate models.
• Explicit Mesh Output: Outputs native 3D triangular meshes with clean connectivity and explicit vertices, ready for game engines (Unreal/Unity) and VFX suites (Blender/Maya).
• Flexible Conditioning: Supports text-to-3D, image-to-3D, and point-cloud-to-3D generation pipelines.

Benchmark Performance

🚀 Quickstart & Usage

# Clone official MeshFlow repository
git clone https://github.com/facebookresearch/meshflow.git
cd meshflow

# Create Conda environment & install dependencies
conda create -n meshflow python=3.10 -y
conda activate meshflow
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

# Inference: Text-to-3D Mesh Generation
python sample.py \
  --config configs/meshflow_t23d.yaml \
  --prompt "An artistic wooden chair with ornate carving" \
  --output_dir ./outputs/

🔗 Official Links & Resources

• Official Project Page
• arXiv Paper (arXiv:2606.04621)
• Paper PDF Download
• GitHub Repository (facebookresearch/meshflow)
• Hugging Face Space Demo