---
language: en
license: cc-by-nc-sa-4.0
library_name: pytorch
pipeline_tag: image-to-image
tags:
- diffusion
- image-restoration
- remote-sensing
- underwater
---
# KAUSTics-Net: selected result checkpoints

Pixel-space conditional velocity-diffusion model for through-water UAV restoration. This model has no latent autoencoder. Use the accompanying KAUSTics-Net GitHub code; these files are not Diffusers pipeline checkpoints.

## Contents

| Directory | Seed checkpoints | Initialization / selected epoch |
|---|---|---|
| `kaustics/` | `seed1.pt`, `seed2.pt`, `seed3.pt` | Scratch KAUSTics training; epochs975/925/925 |
| `sea_undistort/` | `seed1.pt`, `seed2.pt`, `seed3.pt` | Corresponding KAUSTics seed, sampler-aware fine-tuning; epochs95/100/100 |

Only the six selected model weights are included. No optimizer states, intermediate epochs, pilots, ablations or competitor weights. `selected.json` records SHA256 checksums, original archival checkpoint checksums and unchanged tensor digests. The lightweight release exports CPU tensors plus primitive metadata; load with `torch.load(..., map_location='cpu', weights_only=True)`.

## Inference

Use `infer.py --checkpoint checkpoints/kaustics/seed1.pt --input image.png --output predictions/example` in the code release. Input/output256×256 RGB; input normalization[-1,1];1000-step cosine schedule;25-step DDIM at strength0.95 and classifier-free guidance2.0. Preserve image filenames for deterministic per-image noise.

KAUSTics training uses source-disjoint26/3/3 UAV source IDs. Sea-Undistort fine-tuning uses840 all-effects training scenes and240 all-effects validation scenes; all three120-scene test variants are scored against original GT. See the code release’s `docs/PROTOCOL.md` for loss weights, selection, stopping rules, benchmark exposure and portability limits.

These weights target the paper’s dataset/task. Restoration quality and geometric accuracy on unrelated real UAV scenes are not guaranteed. Sea-Undistort adaptation shows a fidelity/perceptual trade-off; it is not a claim of universally improved image quality.

## License

The model weights are licensed under [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/).

## Citation

If you use this release in your research, please cite our [arXiv preprint](https://arxiv.org/abs/2611.03120).

```bibtex
@article{scilla2026kausticsnet,
  title   = {{KAUSTics-Net}: Diffusion-Based Through-Water {UAV} Restoration with a Semi-Synthetic Dataset},
  author  = {Scilla, Dario and Lopez, Omar A. and Johansen, Kasper and Nieuwenhuis, Brian O. and McCabe, Matthew F.},
  journal = {arXiv preprint arXiv:2611.03120},
  year    = {2026},
  doi     = {10.48550/arXiv.2611.03120},
  url     = {https://arxiv.org/abs/2611.03120}
}
```