Cloth-splatters/dexgarment-dense-descriptors-20260826

DexGarment dense descriptor model matrix (2026-08-26)

This repository contains all 48 trained seed-0 checkpoints from the final DexGarment dense-correspondence comparison. Models cover two observation regimes (flat and deformation), eight recipes, and three garment categories (tops, dress, and trousers).

These are point-cloud descriptor models for garment correspondence. They are not cloth state estimators or dynamics models.

Layout

models/<regime>/<recipe>/<category>/seed_0/
  best.pt
  config.json
  validation_metrics.json
  test_metrics.json
comparison/
  summary.json
  runs.csv
experiment/
  marlowe_matrix.json
SHA256SUMS

The eight trained recipes are:

• gam_scratch_shared
• unigarment_scratch_shared
• gam_pretrained_adapt
• unigarment_native_coarse
• gam_shared_c2f
• gam_pretrained_c2f
• unigarment_shared_c2f
• unigarment_native_c2f

The separately released, unadapted GAM baseline is not duplicated here. Its source checkpoints remain part of DexGarmentLab/Model-HALO rather than outputs of this training matrix.

Recommended checkpoints

Macro averages are over tops, dress, and trousers.

Use case	Recipe	Dense error	Dense <=2 cm	Exact top-1
Flat correspondence	unigarment_shared_c2f	0.613 cm	95.365%	44.375%
Deformation correspondence	unigarment_native_c2f	2.208 cm	59.890%	14.305%

For deformation semantic metrics, unigarment_scratch_shared is strongest in this comparison (4.531 cm semantic error and 69.576% within 5 cm).

Loading

Use the pinned UniGarmentManip model builder and checkpoint loader:

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, checkpoint_path, "cuda:0")
model = model.to("cuda:0").eval()

Select gam or unigarment from the recipe name. The exact architecture and training settings for every checkpoint are recorded in its adjacent config.json.

Provenance and scope

• UniGarmentManip experiment commit: 0701bc475877da68bc07f727a1995fd78a3de511
• Code fingerprint: 665d3f5e2509bc26ea46fb467d53056a0fcea09998e9d0955e027c05e715843a
• Dataset: 1,800 garment-disjoint episodes; 28,800 deformation observations
• Categories: tops, dress, trousers
• Seeds published here: seed 0 only
• Completion: 48/48 trained cells; no failed or non-finite runs

The full category-level results are in comparison/summary.json and comparison/runs.csv. Verify downloaded artifacts with SHA256SUMS.