JKAlice/neural-octree

NeuOctree — Neural Explicit Shape Representation

Paper · Code and running instructions

Files

• weights/feedforward.pth: the paper's d9-union feedforward checkpoint (36,888,057 parameters; four recurrent refinement stages). Original checkpoint bytes, without optimizer state. MD5: 2db9de355b2e04fba84ca167479d1392.
• weights/geometry/<tier>KiB/<model>/inference_fp16.pt: 28 packed inference checkpoints for seven objects at the 81, 122, 244, and 823 KiB budgets. The budgets describe representation payloads, not PyTorch container sizes.
• configs/: matching model configurations and measured geometry capacities.
• filelists/objaverse/train.txt, val.txt: the original ordered training and validation manifests, with objaverse/<uid>/<view> entries. They contain 31,761 training objects and 1,000 validation objects, with eight views per object and no object overlap.
• filelists/objaverse/train_uids.txt, val_uids.txt: the same splits as one UID per line.
• filelists/objaverse/object_paths.json: official Objaverse source paths for these 32,761 objects.
• filelists/objaverse/eval30_uids.txt: the 30 validation objects used in the paper's Objaverse evaluation. This UID list alone does not specify the paper's fixed evaluation camera protocol.
• SHA256SUMS: file integrity checksums.

Download and prepare data

After installing the code, run from its root:

pip install -U huggingface_hub
hf download JKAlice/neural-octree --include 'weights/feedforward.pth' 'filelists/*' --local-dir assets
python scripts/prepare_objaverse.py --filelists assets/filelists/objaverse \
  --output data/objaverse --blender /path/to/blender
neuoct-evaluate-feedforward --config data/objaverse/config.yaml \
  --checkpoint assets/weights/feedforward.pth --output outputs/feedforward \
  --hit-threshold 0.4 --orient-normals

The preparation script downloads meshes from the official allenai/objaverse dataset, renders eight 512×512 views per object, constructs 200,000 colored points, and generates depth-9 mesh occupancy. It preserves the published object split. Re-run the same command to resume; use --limit 1 with a separate output directory for a small check. The evaluation command scores the validation manifest, rather than the separate fixed-camera paper benchmark.

Only checkpoints and filelists are hosted here. Meshes, textures, rendered images, and processed point clouds are not redistributed. Original Objaverse assets retain their individual source licenses; this filelist does not grant additional rights to them. No new license is assigned to the project weights by this release.

Citation

@article{wang2026neural,
  title={Neural Explicit Shape Representation},
  author={Wang, Jun-Hao and Lu, Changan and Chen, Baoquan and Wang, Peng-Shuai},
  journal={arXiv preprint arXiv:2610.38441},
  year={2026}
}