Modelpublic

Cloth-splatters/dexgarmentlab-lift-20260822-state-est-gps-flow-tf2-noscale

sourceHugging Facemitupdated 13d agoView on Hugging Face

0likes

Model Card

DexGarmentLab Lift — GPS TF2 flow matching, no template-scale randomization

Released 2026-09-14. This is the 20-epoch no-scale flow model, trained from scratch in job 1291773_31. model/ contains the validation-selected EMA weights; config.yml is the complete saved training configuration.

"No scale" means no template-size mismatch augmentation: template_mismatch_range: [1.0, 1.0]. Joint scale augmentation of the cloud, target and template remains enabled ([0.8, 1.25]), as do joint random yaw, template yaw mismatch (180 degrees, probability 0.5), and cloud corruption. The model uses coordinate scale 3 and no scale normalization.

—Architecture: GPSStateEstModel, sequential cross-attention, two context frames, hidden dimension 128, 8 layers / 8 heads, 2048-node padding cap.
—Scheduler: FlowMatch_StateEstGPS, 1000 training timesteps, shift 1.
—Training: 20 epochs / 398,700 optimizer updates, batch size 32, seed 259, bf16, EMA decay 0.9999, learning rate 2e-4, cosine schedule, 2000 warm-up steps.
—Dataset: dexgarmentlab_lift_full_state_20260822.h5, original random-garment-v0 split, dataset revision 771714a2; 176/23/22 train/validation/test garments.
—Experiment: dexgarment_flow_v2_cs3_tf2_lift20260822_yawmix_noscale_ep20_1291773_31 (training completed successfully).
—Best validation sampling loss: 6.283105492252189e-05.
—Weights SHA-256: f3faa4b11dfb36774ecb5e4f5f68afcbcdcb122888fa63a351d6741fe0f9c56b.

Completed reconstruction evaluation

Exact-template reconstruction on all 7,537 test frames, 169 trajectories, 22 held-out garments, frame stride 10. Four samples, seed 0, 30 flow steps; sample selection by observation-to-surface energy. Warm starts use strength 0.8 and 3 steps. Values below are medians over frames, in millimeters.

Model / inference	Vertex error	Surface Chamfer	Occluded surface Chamfer
No scale, independent	14.78	11.73	12.05
No scale, warm start	13.70	11.49	11.79
Full randomization, independent	15.79	11.72	12.04
Full randomization, warm start	14.88	11.53	11.81

The reference is the full-randomization 20-epoch flow release `03e61682`. The no-scale model improves vertex correspondence error; independent surface Chamfer is essentially unchanged. These exact-template results do not measure robustness to template-size mismatch. Full evaluation summaries and arguments are included under evaluation/.

Loading in UniClothDiff

python

from src.hub import resolve_checkpoint

checkpoint = resolve_checkpoint(
    "Cloth-splatters/dexgarmentlab-lift-20260822-state-est-gps-flow-tf2-noscale"
)

Use UniClothDiff's ClothStateEstGPSPipeline with the saved config.yml so the flow scheduler and dataset conventions are restored. The custom model is Diffusers-compatible and lives under the model/ subdirectory. The evaluation above uses 30 steps and four samples; a dedicated few-step sweep for this checkpoint is not included in this release.