BrickFlow: Connectivity-Guided Brick Reconstruction

Peter Kulits and Cordelia Schmid

NeurIPS 2026

The weights of BrickFlow's denoisers. See kulits/BrickFlow for the dataset. The brickflow package is on GitHub and has more instructions.

Weights

file model size
weights/pt-s.pt PT-S: pretrained on train/pt (width 384, depth 12) 127 MiB
weights/pt-b.pt PT-B: pretrained on train/pt (width 768, depth 12) 503 MiB
weights/pt-l.pt PT-L: pretrained on train/pt (width 1024, depth 24) 1.72 GiB
weights/ic-s.pt IC-S: image-conditioned, finetuned from PT-S on train/ic 206 MiB
weights/ic-b.pt IC-B: image-conditioned, finetuned from PT-B on train/ic 811 MiB
weights/ic-l.pt IC-L: image-conditioned, finetuned from PT-L on train/ic 2.48 GiB

Each file is torch.save(dict(model=state_dict)) of a brickflow.model.Denoiser, the post-hoc EMA of the training snapshots. IC models read DINOv3 features (facebook/dinov3-vitl16-pretrain-lvd1689m) and use the PT model of the same size to finish sampling.

Download

brickflow fetches the weights by name into brickflow/ in BrickNet's data directory ($BRICKNET_DATA, else the platform user-data directory), at the same paths as here:

pip install brick-flow
python -m brickflow fetch                                   # assets, val/procbrick, and weights/pt-l
python -m brickflow fetch weights/pt-s weights/ic-s
python -m brickflow sample weights/pt-l.pt val/procbrick --samples 100 --guidance --out pt-l

License

The weights are under the MIT license.

Citation

@InProceedings{kulits2026brickflow,
  author    = {Kulits, Peter and Schmid, Cordelia},
  title     = {BrickFlow: Connectivity-Guided Brick Reconstruction},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Dataset used to train kulits/BrickFlow