MeshFlow generates artist-like meshes in ~1 second with MeshVAE + flow-matching DiT, using input geometry and an optional reference image.

Before running the code, download the MeshFlow checkpoint bundle and place it under ckpt/meshflow/:

ckpt/meshflow/
├── config.yaml
└── model.pth

You can also prepare the directory manually:

mkdir -p ckpt/meshflow
# download config.yaml and model.pth into ckpt/meshflow/

If you use reference-image conditioning (inference_dit.py --ref_image, pipeline.run(image=...), or the Gradio image upload), you also need to configure DINOv3. Mesh / point-cloud-only inference does not load DINOv3.

git clone https://github.com/facebookresearch/dinov3.git \
  ~/.cache/torch/hub/facebookresearch_dinov3_main

visual_condition:
  hub_model: dinov3_vitl16
  hub_dir: /root/.cache/torch/hub/facebookresearch_dinov3_main
  hub_weights: /path/to/dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth
  pretrained: true
  image_size: 512

First, clone this repository and install the dependencies:

git clone https://github.com/facebookresearch/meshflow.git
cd meshflow
pip install -r requirements.txt

Download the MeshFlow checkpoint into ckpt/meshflow/ as described above.

Now, try the model with a few lines of code:

from meshflow.pipelines import MeshFlowPipeline

pipeline = MeshFlowPipeline.from_pretrained(
    "ckpt/meshflow",
    device="cuda",
    dtype="fp16",
)

mesh = pipeline.run(
    mesh="path/to/input.ply",       # mesh / point cloud for RoPE geometry condition
    image=None,                     # optional reference image (.png / .jpg / .webp)
    steps=28,
    guidance_scale=2.5,             # only effective when `image` is provided (CFG on visual cond)
    seed=42,
)
mesh.to_trimesh().export("output.glb")

Online demo: facebook/meshflow on Hugging Face Spaces

You can also launch the Gradio demo locally:

python gradio_app.py --gpu 0 --dtype fp16

Omit --model_path to use local ckpt/meshflow/ if present, otherwise download config.yaml and model.pth from facebook/meshflow into ~/.cache/meshflow/. Or pass an explicit bundle path:

python gradio_app.py --model_path ckpt/meshflow --num_verts 4096

Upload a mesh or point cloud for RoPE surface sampling, optionally add a reference image, and generate a new mesh in the browser. torch.compile is enabled by default on CUDA (--no-compile to disable). When the model config sets denoiser_model.use_proj_cond_on_temb: true, use the num_verts slider to send a DiT control signal that roughly controls generated mesh vertex count.

More results and method details are on the project page.

python inference_vae.py \
  --model_path ckpt/meshflow \
  --input <mesh_file_or_dir> \
  --output outputs/meshflow_vae/run1

Outputs: inputs_meshes/ (.ply), vae_recon/ (.ply).

Optional: --dtype bf16|fp16|fp32 (default: fp16).

python inference_dit.py \
  --model_path ckpt/meshflow \
  --input <mesh_file_or_dir> \
  --ref_image <image_file_or_dir> \
  --output outputs/meshflow_dit/run1 \
  --steps 28 \
  --compile \
  --guidance_scale 2.5  # only when --ref_image is provided

--ref_image is optional — if omitted, a zero visual condition is used. When using a reference image, configure DINOv3 as described in Pretrained models.

Outputs: input_meshes/, input_images/ (when --ref_image is set), surface_pc/ (.ply), rope_cond/ (.ply), generated_meshes/ (.glb).

Chamfer and Hausdorff distances between GT and reconstructed meshes:

python evaluate.py \
  --gt_path outputs/meshflow_vae/run1/inputs_meshes \
  --pred_path outputs/meshflow_vae/run1/vae_recon \
  --output_path outputs/meshflow_vae/run1/eval_results.txt

• Input meshes should respect the configured vertex budget (mesh_model.num_latents, 4096 by default). --num_verts is the DiT control numerator (proj_cond_on_temb = num_verts / num_latents), only when denoiser_model.use_proj_cond_on_temb is enabled.
• Optional RMBG matting is in meshflow/pipelines/utils.py; enable with MeshFlowPipeline(use_rmbg=True).

@inproceedings{li2026meshflow,
  title={MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer},
  author={Li, Weiyu and Toisoul, Antoine and Monnier, Tom and Shapovalov, Roman and Ranjan, Rakesh and Tan, Ping and Vedaldi, Andrea},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026},
  note={Highlight}
}

See the LICENSE file for details about the license under which this code is made available.