PRISM v2 — pretrained geodesic models

Pretrained weights for PRISM (ICML 2026), From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data.

• Paper: arXiv:2606.02268
• Code: github.com/AidenZhao/PRISM

PRISM predicts geodesic distances between arbitrary pairs of points on the underlying surface from a point cloud (xyz + normals). PRISM v2 uses an anchor-attention transformer instead of the paper's Point Transformer V3 backbone: about half the parameters and lower error at every size.

Models

File	Params	MRE (%) ↓	L1 (%) ↓
prism_S.pt	15.3M	2.21	2.57
prism_B.pt	61.2M	1.60	1.87
prism_L.pt	250.9M	1.36	1.64

ShapeNetCore.v2 test split (2,000 shapes, 55 categories), exact MMP ground truth, shapes normalized to the unit sphere. For comparison, the paper's PTv3 models reach 6.5 / 4.4 / 3.9 % MRE with 31M / 124M / 494M parameters.

Usage

git clone https://github.com/AidenZhao/PRISM.git && cd PRISM && pip install -e .

from huggingface_hub import hf_hub_download
from prism.evaluate import load_model

ckpt = hf_hub_download("AidenZhaoy/PRISM", "prism_B.pt")
model, _ = load_model(ckpt, "cuda")
D = model.distance_matrix(points, normals)   # (N, N); points (N,3) normalized to the unit sphere

Or from the command line, on a mesh or a point cloud (distances returned in input units):

python -m prism.predict --ckpt prism_B.pt --input shape.obj --n_points 4096 --out geodesic.npz

Each checkpoint stores model (state dict), args.model_cfg (architecture), epoch and val_mre.

Citation

@inproceedings{prism2026,
  title     = {From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data},
  author    = {Zhao, Yuming and Hou, Junhui and Zhang, Qijian and Qin, Jia and He, Ying},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}