kelthuzadyl/instant-fold-assets

INSTANT-FOLD Assets

Assets for the public INSTANT-FOLD workflows. Most users only need the eval bundle; use the repository CLI to download and verify files rather than extracting archives manually.

Evaluation And Representation

instant-fold assets download --bundle eval
instant-fold eval-sim demo \
  --demo-dir data/eval_demo_trajectories/seen/way1/L \
  --way 1 \
  --execution-mode L \
  --runs 5

Change --demo-dir, --way, and --execution-mode together to evaluate a different released context.

Exact-init replay additionally requires benchmark-eval and an explicit replay manifest:

instant-fold assets download --bundle benchmark-eval
instant-fold eval-sim demo \
  --demo-dir data/eval_demo_trajectories/seen/way1/L \
  --way 1 \
  --execution-mode L \
  --replay-init-root benchmark_replay/replay_manifest.json \
  --runs 5

For compact paper-table reproduction:

instant-fold assets download --bundle benchmark-repro
instant-fold benchmark-eval same-init-table \
  --rollouts-root benchmark_rollouts \
  --reference-feature-cache benchmark_repro/reference_features_38400.npz \
  --reference-metric-rows benchmark_repro/reference_metric_rows_38400.csv

The representation example uses the separately downloadable encoder bundle.

Data Generation

instant-fold assets download --bundle data-gen
bash scripts/generate_data.sh \
  --ways 1 \
  --only-modes L \
  --cloth-list data/smoke_mesh_ids.txt \
  --traj-total-per-mesh 1 \
  --num-workers 1 \
  --mesh-path data/meshes/tops \
  --target-camera-height 1.13

Training

For a small workflow check:

instant-fold assets download --bundle train-smoke

For the full eight-context policy-training corpus:

instant-fold assets download --bundle train-full

train-full contains 34,560 trajectories across the exact eight policy-training strategies used by the released paper model:

way1, way2, way3, way4, way5, way6, way9, way11

The extracted data are normalized under data/cloth_8ctx_release/, and data/train_full/data_dirs.txt is shared by public encoder training and policy training. The released encoder checkpoint used a broader 22-mode corpus, so this bundle does not claim exact encoder-checkpoint reproduction. Ways 7, 8, and 10 are held-out folding strategies and are not part of the bundle.

Bundles

Bundle	Purpose
encoder	Released LoRA-adapted DINOv3 encoder
eval	Released policy, demos, evaluation meshes, and mesh IDs
benchmark-eval	Exact simulator initialization replay
benchmark-repro	Compact same-init paper-table reproduction
data-gen	Released 360-mesh data-generation bank
train-smoke	Small training workflow check
train-full	Exact eight-context policy-training corpus

This repository contains assets with different terms. The public source code is MIT-licensed, while the bundled PyFlex runtime, Cloth3D-derived data, and DINOv3-derived checkpoints have separate notices or access terms. See the source repository and the upstream DINOv3 terms before using or redistributing each asset. Workflows using DINOv3 require access to the upstream weights.