CoolFace
Modelpublic

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

sourceHugging Facemitupdated 11d agoView on Hugging Face
0likes
Model Card
Superseded (2026-09-16). The ClothAtlas paper uses flow matching only; its estimator is dexgarmentlab-lift-20260822-state-est-gps-flow-tf2-noscale. This DDPM estimator was the August 2026 standard and is kept as the control it served as. Current release: the ClothAtlas collection.

DexGarmentLab Lift 2026-08-22 — GPS TF2 state estimator

GPSStateEstModel fine-tuned on the full-state 2026-08-22 lift collection. It reconstructs the variable-topology cloth mesh from visible point clouds and uses two-frame temporal conditioning (num_context_frames: 2). model/ contains the best EMA checkpoint.

  • —Dataset: dexgarmentlab-lift-correspondence-20260822, file dexgarmentlab_lift_full_state_20260822.h5
  • —Dataset SHA-256: 8eca09186491af5842ec7f09f28a3da885d25f6ba1eab6e74ead89126a977d40
  • —Dataset: 1,800 episodes, 221 garments, 798,510 frames, garment-disjoint train/validation/test splits
  • —Architecture: DDPM GPSStateEstModel, temporal context 2, 4.33M parameters
  • —Training: seed 259, bf16, batch 32, cosine LR, EMA, heavy point-cloud augmentation
  • —Best validation loss: 8.689093247579871e-05
  • —Slurm job: 451439 (COMPLETED, 0:0)
  • —Source branch: mesh-hypothesis-pf
  • —Source revision at training/publication audit: f01f8091d052d77ef0b1e9edcea3e3b8683841b5
  • —Weight SHA-256: 7f982c47f1f63943de0caa082f48518df87820509c2ed1495b87400718dadaf5

Load through UniClothDiff's src.hub.resolve_checkpoint after checking out the source revision recorded in this card. The complete training recipe is config.yml; the Diffusers-compatible model is under model/.

yaw (2026-09-01) — this revision

The lift best recipe (v2 EMA + heavy augmentation + two-frame temporal + template-size mismatch) plus orientation augmentation: joint random yaw about gravity (random_z_rotation: true) and template-alone yaw mismatch (template_yaw_max_deg: 180), so template yaw stops being ground truth and in-plane orientation must be read from the observation. Measured on the real take-2 pants with no alignment at all: flat orientation error 41–55° → 2–11°, symmetric Chamfer 12.55 → 8.73 mm (best take-2 arm of the campaign); the PCA pre-alignment step is unnecessary for this revision. Cost: unaugmented validation loss 1.74e-4 m² vs 1.33e-4 (arm 16, job 93991016, tag `noyaw`) for the non-yaw recipe, and near-symmetric flat silhouettes can resolve to the wrong 180° facing (the template-facing prior is deleted; use the orientation tie-break, or wait for the mixture-prior `templateyaw_p 0.5` arms).

  • —Source experiment: dexgarment_ddpm_v2_tf2_lift20260822_yaw_2026-08-31_21-57-30_976630 (checkpoint-best, EMA)
  • —Slurm job: 976630_18 (COMPLETED, 0:0)
  • —Best validation loss: 0.0001741877611560995
  • —Weight SHA-256: 45d10b0a18d3cf3f4621c8f3c81a5f3d4351c23136f8863719d0427ca709b10e
  • —Previous release (job 451439, 2026-08-28): revision 869f885256039b85ae68703bb62dfbe9cd23e026

yaw-mixture (2026-09-03) — this revision

The yaw recipe of the 2026-09-01 release with a mixture prior on the template yaw (template_yaw_p: 0.5: half the training samples keep the exact template yaw, the other half get a uniform random one), i.e. the complete training randomization of the paper: joint random scale + yaw of cloud, target and rest mesh; template-alone scale mismatch (log-uniform 0.5-2) and yaw mismatch with probability 0.5; point jitter, dropout, wedge dropout and outliers. This is the checkpoint behind every simulation result in the paper.

Measured on the garment-disjoint lift test split (22 garments, 221-mesh library): flat-frame canonical-mesh retrieval 21/22 rank 1 and 22/22 top-5, median rank 2 over 198 observations; energy-selected vertex error 7.8 mm on flat garments and 19.2 mm over all 7,537 test frames; error flat within 1 mm across template scale factors 0.5-2 and all template yaw offsets. Reconstruction saturates at 3-5 denoising steps (18.5 mm at 5 steps vs 19.4 at 30) — use 5. Real take-2 pants, no alignment: 8.36 mm symmetric Chamfer, a tie with the 2026-09-01 yaw release; the 180° facing ambiguity of near-symmetric flat silhouettes remains.

  • —Source experiment: dexgarment_ddpm_v2_tf2_lift20260822_yawmix_2026-09-01_08-35-23_982082 (checkpoint-best, EMA)
  • —Slurm job: 982082_20 (COMPLETED, 0:0)
  • —Best validation loss: 0.00015062040602383228
  • —Weight SHA-256: 80dcd7624b0bf22d1aeec7f9585e32dac3e529673d4d74590ee2c3bfee59e4ff
  • —Previous release (yaw recipe, 2026-09-01): revision 9a68e30a3b6bd920f2aca7f5ee7f1dc5cdee2485