TriFlow — Pretrained Weights
Pretrained checkpoints for TriFlow: Generating Artist-Like 3D Mesh Topology via Nearest-Vertex Vector Fields (ECCV 2026).
TriFlow generates compact 3D meshes with artist-like triangle topology from input geometry conditions such as signed distance fields. Mesh topology is represented as a Nearest-Vertex Vector Field (NVF) over the surface; a latent flow-matching model synthesises this field, and a constrained Quadric Error Metric (QEM) simplification extracts the final mesh.
Files
| File | Size | Stage |
|---|---|---|
sdf_vae.safetensors | 152 MB | Stage 1 — SDF VAE |
nvv_vae.safetensors | 412 MB | Stage 2 — NVF VAE |
flow_model.safetensors | 755 MB | Stage 3 — Latent Flow Matching |
Usage
These weights are used by the TriFlow codebase — see the repository for setup, data preparation, training and inference instructions.
hf download lihcxr/TriFlow \
flow_model.safetensors sdf_vae.safetensors nvv_vae.safetensors \
--local-dir checkpoints
inference.py loads them from checkpoints/ by default.
License
Released under the Automotive Development Public Non-Commercial License v1.0
(ADPNCL) — see LICENSE. This license permits non-commercial use only;
please read it in full before use.
Note that TriFlow's mesh-processing pipeline additionally depends on MeshLib, which is not distributed under an open-source license and carries its own terms.
Citation
@inproceedings{li2026triflow,
title = {TriFlow: Generating Artist-Like 3D Mesh Topology via Nearest-Vertex Vector Fields},
author = {Li, Haoxuan and Erko{\c{c}}, Ziya and Sirigatti, Daniele and Rosov, Vladislav and Li, Lei and Dai, Angela and Nie{\ss}ner, Matthias},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026},
}
This work was funded by AUDI AG.