FoldTrace
Checkpoints for FoldTrace: an autoregressive diffusion model that generates rigid-foldable single-branch origami crease patterns, one vertex unit at a time, so that folding moves the end vertex along a target 3D curve. Project page.
| File | Condition |
|---|---|
foldtrace_abs.pt |
the target curve in the driving-crease frame (32 points by arc length) |
foldtrace_rel.pt |
the curve's shape and size; the model chooses where it starts |
Both: joint discrete + continuous diffusion heads, d_model 256, 3 encoder / 6 decoder layers, 10.8M parameters,
300k steps on 300k single-branch crops of the CreasePatternDataset
(Hochuli and Stankovic). On 500 held-out curves, the best of 16 foldtrace_abs designs is within 5% of the curve's
reach for 84.8% of curves (median error 2.1%), against 51.8% for the nearest of 1M training designs.
git clone https://github.com/Teresa-l23/FoldTrace && cd FoldTrace
pip install -r requirements.txt
python scripts/sample.py --target wideS --ckpt abs --k 16
Each file is a dict with model (state dict), model_config (constructor arguments of
foldtrace.model.SingleBranchGenerator) and step; it loads with torch.load(..., weights_only=True).