---
license: mit
tags:
  - nerf
  - neural-fields
  - permutation-symmetry
  - model-merging
---

# Trained NeRF pairs for "Progressive Permutation Alignment with Temperature-Controlled Spectral Sharpening" (ACCV 2026)

Code: https://github.com/yuzu0308/progressive-permutation-alignment

16 vanilla NeRFs (nerfstudio 1.1.5 `vanilla-nerf`, 1M iterations) trained on the 8 scenes of NeRF Synthetic
with two random seeds each. These are the model pairs used for the interpolation (linear mode connectivity)
experiments of the paper.

| File | Content |
|---|---|
| `<scene>/model_a.pth` | state dict of the model trained with seed 42 (only the `field_coarse.*` / `field_fine.*` entries) |
| `<scene>/model_b.pth` | same for seed 100 |
| `<scene>/ref_perms_cpu.pt` | permutations found by the original research code (`pa`: Ours w/o Ref., `pa_wm`: Ours), used by the regression test |

Scenes: lego, chair, ficus, hotdog, mic, ship, materials, drums.

```bash
pip install huggingface_hub
python scripts/download_weights.py            # from the code repository
```

```python
from ppa.nerf import load_state_dict
a = load_state_dict("weights/lego/model_a.pth")
```

The LPIPS network weights that nerfstudio stores in its checkpoints are not included; they are not needed
for alignment or rendering.

## Citation

```bibtex
@inproceedings{horioka2026progressive,
  title     = {Progressive Permutation Alignment with Temperature-Controlled Spectral Sharpening},
  author    = {Horioka, Yuta and Chiba, Naoya and Uranishi, Yuki},
  booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}
```