See the Invisible with SWIR Models
Official checkpoint collection for the SWIR denoising method and its reported baselines and ablations. Download and verify every published artifact with:
pip install "huggingface-hub>=0.25"
python scripts/download_artifacts.py --kind models
| File | Experiment | Architecture |
|---|---|---|
two_stage_shared_task.bin |
neutral legacy shared-task checkpoint | two-stage residual U-Net |
pg.bin |
P-G noise model | U-Net |
eld.bin |
ELD noise model | U-Net |
sfrn.bin |
SFRN noise model | U-Net |
ablation_single_stage.bin |
single-stage ablation | U-Net |
fpnv2.bin |
no-FPN evaluation ablation | two-stage residual U-Net |
The exact publication allowlist, SHA-256 hashes, embedded metadata,
configurations, and legacy script identifiers are included in
checkpoint_inventory.json.
PG, ELD, SFRN, single-stage, and no-FPN artifacts have unique legacy namespaces.
The full-method and shot-noise trainers accidentally shared one output namespace;
that surviving weight is labeled two_stage_shared_task, not uniquely as either
experiment. See docs/checkpoint_provenance.md before citing or repackaging it.
Model weights are released under the repository's MIT license. Training data are separately licensed; SID images are not redistributed.
@inproceedings{jiang2025see,
title={See the Invisible with SWIR: Learning to Enhance via Synthetic Noise Modeling},
author={Jiang, Haiyang and Wang, Hongjun and Zheng, Yinqiang},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
pages={7449--7458},
year={2025}
}