particulate-hinge final models (private)
Continued training of PARTICULATE (arXiv 2512.11798; official weights rayli/Particulate), same architecture (16 parts).
Delivered model (2026-10-15): R7-WiSE α=.5, selected by the pre-registered rule (protocol-r6 amendments 2/3/5). It is
0.5·official + 0.5·R7, where R7 = F's recipe + own v2 + data-source token + ≥16 training points per part. The source token
is folded into two biases, so each file below is an official-format checkpoint for the unmodified infer.py; pick the
file by the kind of input.
| File | What |
|---|---|
R7-WiSE050-lw/model.pt |
default: Articraft/Lightwheel-style appliances, small parts (source token 2; as evaluated on Lightwheel) |
R7-WiSE050-pn/model.pt |
PartNet-Mobility-style furniture (source token 0; as evaluated on PN74) |
R7-WiSE050-own/model.pt |
assets from our Articraft generator (source token 3; as evaluated on own val/test) |
F/model.pt |
earlier model: official weights + 20k steps on PartNet-Mobility .45 / GRScenes-object .20 / Articraft-10K .25 / own generated assets .10 |
F-WiSE050/model.pt |
0.5·official + 0.5·F (alpha fixed in advance); better motion parameters on PartNet than R7-WiSE |
sha256 and provenance: MODELS.json. Usage, evaluation and results: the project repository (docs/manual.md,
docs/research/2026-10-13-scaling-regression/RESULTS.md).
cd particulate
python infer.py --input_mesh mesh.glb --ckpt_path R7-WiSE050-lw/model.pt --model_config configs/particulate-B.yaml --up_dir Y
Not for redistribution: training data includes PartNet-Mobility (non-commercial) and GRScenes (CC BY-NC-SA).