llama2thedog/BIM-JEPA-FM

BIM-JEPA

A self-supervised point-cloud encoder for 3D Building Information Modeling (BIM) geometry, pre-trained with LeJEPA on over 2.1 million individual BIM elements.

• Paper: Toward generalizable foundation models for 3D BIM geometry using a joint embedding predictive architecture — Automation in Construction 191:107169 (2026)
• Code: jackswl/bim-jepa-fm

Model

Encoder	12-layer transformer, 384 dim, 6 heads (21.3M params)
Tokenizer	PointNet, 64 groups × 32 points
Input	4096-point cloud (xyz)
Objective	2 global + 8 local views; invariance to the global-view centroid + SIGReg

There is no EMA teacher and no predictor. The encoder is trained directly, with a sketched isotropic-Gaussian regulariser (SIGReg: 1024 directions, 17 quadrature knots) preventing collapse.

Training

Data	IFC-884K, IFCNet, BIMGEOM, BIMCompNet (16 building categories)
Schedule	100 epochs, 411,500 steps
Optimiser	AdamW, wd 0.05, lr 5e-4 → 5e-6 cosine, 10-epoch warmup
Precision	bf16-mixed, 4×GPU DDP, batch 128/GPU

Test splits of the downstream datasets are excluded from pre-training.

Usage

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("llama2thedog/BIM-JEPA-FM", "last.ckpt")

last.ckpt is a PyTorch Lightning checkpoint. In the code release, point any downstream config at it:

model:
  pretrained_ckpt_path: /path/to/last.ckpt

Loading renames student.* to encoder.* and drops the LeJEPA-only projector.* and sigreg.* weights, which are not used downstream.

Citation

@article{shi2026toward,
  title={Toward generalizable foundation models for 3D BIM geometry using a joint embedding predictive architecture},
  author={Shi, Jack Wei Lun and Solihin, Wawan and Weng, Yufeng and Liang, Houhao and Zhao, Yimin and Poh, Leong Hien and Yeoh, Justin K.W.},
  journal={Automation in Construction},
  volume={191},
  pages={107169},
  year={2026},
  publisher={Elsevier}
}

Acknowledgements

We thank the authors of LeJEPA, Point-JEPA, SpaRSE-BIM/IFCNet, BIMGEOM, BIMCompNet and BIMNet for releasing their code, data and models.