Cloth-splatters/unigarment-native-coarse-deformation

UniGarment native coarse — deformation

Category-specific dense-correspondence checkpoints trained from scratch on the 2026-08-22 DexGarmentLab lift/deformation collection. This is the current best completed deformation recipe by dense correspondence metrics. It is not a cloth state estimator or dynamics model.

Checkpoints

Select the checkpoint matching the garment category:

tops/best.pt
dress/best.pt
trousers/best.pt

Each directory also contains its exact training configuration, validation/test metrics, and a held-out deformation sweep.

Macro averages over the three held-out test categories:

Metric	Result
Dense mean error	2.28 cm
Dense within 2 cm	56.61%
Exact surface-ID top-1	12.32%
Semantic mean error	5.01 cm
Semantic within 5 cm	64.78%
Left/right swap rate	1.95%

These are seed-0 results. Category-level counts and phase/bin breakdowns are in each test_metrics.json.

Download

From the pinned UniGarmentManip repository:

pip install huggingface_hub
python scripts/download_unigarment_deformation.py --category all

For one category:

python scripts/download_unigarment_deformation.py --category tops --output checkpoints/unigarment-deformation

The helper downloads from Cloth-splatters/unigarment-native-coarse-deformation and verifies every file against SHA256SUMS.

Loading

Use the repository loader so architecture and checkpoint conventions remain consistent:

from dexgarment_adaptation.model import build_descriptor_model, load_adapted_checkpoint

model = build_descriptor_model("unigarment", dgl_root, feature_dim=512)
load_adapted_checkpoint(model, "checkpoints/unigarment-deformation/tops/best.pt", "cuda:0")
model = model.to("cuda:0").eval()

dgl_root must contain the released DexGarmentLab GAM PointNet++ support code expected by UniGarmentManip.

Training provenance

• UniGarmentManip commit: 0701bc475877da68bc07f727a1995fd78a3de511
• Code fingerprint: 665d3f5e2509bc26ea46fb467d53056a0fcea09998e9d0955e027c05e715843a
• Recipe: unigarment_native_coarse
• Backbone: UniGarment PointNet++ descriptor, 512 features
• Initialization: random
• Loss: published UniGarment coarse correspondence loss
• Optimizer: Adam, LR 1e-3, weight decay 1e-5
• Batch size: 16
• Steps: 40,000 per category
• Cross-phase exact-pair probability: 0.75
• Seed: 0

Training code:

• scripts/train_dexgarment_flat.py — trainer (supports flat and deformation caches)
• dexgarment_adaptation/ — model, data, losses, and evaluation
• experiments/marlowe_matrix.json — reviewed recipe registry
• scripts/marlowe/run_manifest_entry.py — reproducible manifest-to-command launcher