meshtailor

MeshTailor reproduction checkpoint (paper protocol, 100k train)

Best checkpoint of our reproduction of MeshTailor: Cutting Seams via Generative Mesh Traversal (arXiv:2603.27309). Companion code repository: github.com/Xinghan-Wang/meshtailor.

• File: best_paper100k.pt (~1.14 GB), a full PyTorch training checkpoint (model weights + optimizer state + config, saved by meshtailor/train.py).
• Trained on 100k garments from GarmentCodeDataset with paper-style "maximal chain" labels (paper App. B.1) and the paper sequence protocol.
• 10k test results: macro edge recall 0.846, precision 0.931, chart count 0.88×GT, PartUV overall distortion within 4% of GT (paper reports parity).

Usage

import torch
# The checkpoint contains non-tensor objects (optimizer state/config), so it must
# be loaded with weights_only=False. Only load checkpoints from trusted sources.
ckpt = torch.load("best_paper100k.pt", weights_only=False)
# ckpt["model"] holds the state dict; load with meshtailor.models.model.MeshTailor
# (see the companion GitHub repository for the full pipeline).

Recommended inference config: temperature=0.1, no penalties (p0 protocol), --bf16 on modern GPUs.

External encoder

MeshTailor uses the frozen point-cloud encoder from NeuralCarver/Michelangelo. Clone the upstream repository as Michelangelo/ at the root of the companion MeshTailor repository, then download the two required weight directories:

git clone https://github.com/NeuralCarver/Michelangelo.git Michelangelo
git -C Michelangelo checkout 6d83b0b
hf download Maikou/Michelangelo \
    checkpoints/aligned_shape_latents/shapevae-256.ckpt \
    --local-dir Michelangelo
hf download Maikou/Michelangelo \
    --include "checkpoints/clip/clip-vit-large-patch14/*" \
    --local-dir Michelangelo

Michelangelo is not included in this model repository. Its code remains under the upstream GPL-3.0 license, and its pretrained weights retain their upstream terms. See the companion repository README for the complete setup and data preprocessing instructions.

License

MIT (code); dataset rights belong to GarmentCodeDataset's authors.