Cloth-splatters/dexgarmentlab-lift-20260822-state-est-gps-regression-tf2

Superseded (2026-09-16) by dexgarmentlab-lift-20260822-state-est-gps-regression-tf2-noscale, the D-GNN of the ClothAtlas paper. This is the 10-epoch arm with template-scale randomization. Current release: the ClothAtlas collection.

DexGarmentLab Lift 2026-08-22 — GPS TF2 deterministic regression baseline

The paper's "is generation needed?" ablation (Section V-B). A GPSStateEstModel with exactly the architecture, data, augmentation and optimiser of the yaw-mixture DDPM estimator (dexgarmentlab-lift-20260822-state-est-gps-tf2, launcher arm 20) but trained with a direct regression loss instead of diffusion: the scheduler Regression_StateEstGPS feeds [0 ‖ rest] at timestep 0 and minimises a masked MSE on the clean vertex positions. At inference the pipeline collapses to one deterministic forward pass (init_noise_sigma 0, a single timestep), so --steps 1 --num-samples 1 is the complete evaluation and extra steps or samples change nothing. model/ contains the best EMA checkpoint.

• —Dataset: dexgarmentlab-lift-correspondence-20260822, file dexgarmentlab_lift_full_state_20260822.h5 (random-garment-v0 split, revision 771714a2)
• —Dataset: 1,800 episodes, 221 garments, 798,510 frames, garment-disjoint train/validation/test splits
• —Architecture: GPSStateEstModel, temporal context 2 (pcd_prev=None single-frame fallback), coord_scale 1
• —Training: bf16, EMA, heavy point-cloud augmentation, joint random scale + yaw, template scale mismatch (log-uniform 0.5-2) and template yaw mismatch with probability 0.5; ~198k steps
• —Source branch: mesh-hypothesis-pf, config configs/train_state_est_gps_dexgarment_v2_tf2_lift20260822_yawmix_regression.yaml

Measured (garment-disjoint lift test split, 22 zero-shot garments)

Exact-template reconstruction over 7,537 test frames (169 trajectories, stride 10): median vertex error 22.9 mm (flat 7.7 / lift 20.1 / bimanual 29.4), surface Chamfer 13.1 mm — against 19.0 mm for the DDPM estimator and 17.3 mm for the flow estimator, with the whole gap in the manipulated phases (flat frames nearly tie). The recipe's robustness is kept: 23.2-23.4 mm across template scale factors 0.5-2, 22.6-23.2 mm across all template yaw offsets. Canonical-mesh retrieval from a flat frame against the 221-mesh library: 18/22 rank 1 (top-5 20/22; the sampled estimators 22/22), 9/22 rank 1 when all nine probe frames are accumulated (vs 17-18/22), 26/132 rank 1 on lifted frames (vs 49/41). Use it as the deterministic reference, not as a deployment checkpoint.

• —Source experiment: dexgarment_regression_v2_tf2_lift20260822_yawmix_2026-09-04_09-47-01_1035269 (checkpoint-best, EMA)
• —Slurm job: 1035269_34 (Training exited with code 0)
• —Best validation loss: 0.00018204110257023784
• —Weight SHA-256: 77ddc33620d1a08a0e5810e609f61824299ad200aa5b2cc6a51b4a28c19a4aab

Load through UniClothDiff's src.hub.resolve_checkpoint (dexgarmentlab-lift-20260822-state-est-gps-regression-tf2). The complete training recipe is config.yml; the Diffusers-compatible model is under model/.