SNaP -- pretrained checkpoints
Weights for SNaP, a one-step posterior sampler for linear inverse problems. The code, configs and download script live in the GitHub repository:
These files are meant to be fetched by the repository's downloader, which verifies every file against its sha256:
python scripts/download_checkpoints.py --all
Each checkpoint is only valid together with its frozen config in
configs/pretrained/<model>.yaml in the code repository.
| file | problem | one-step PSNR / SSIM / LPIPS | size |
|---|---|---|---|
celeba_inpaint.pt |
CelebA 128 - random inpainting, 70% pixels missing, sigma=0.01 | 31.98 / 0.928 / 0.018 | 35 MB |
celeba_box.pt |
CelebA 128 - centered 40x40 box inpainting, sigma=0.05 | 30.73 / 0.934 / 0.021 | 35 MB |
celeba_deblur.pt |
CelebA 128 - Gaussian deblurring, sigma_b=1.0, k=61, sigma=0.05 | 32.92 / 0.919 / 0.019 | 70 MB |
celeba_sr.pt |
CelebA 128 - 2x super-resolution, sigma=0.05 | 31.27 / 0.903 / 0.023 | 35 MB |
afhq_inpaint.pt |
AFHQ-Cat 256 - random inpainting, 70% pixels missing, sigma=0.01 | 30.49 / 0.854 / 0.066 | 422 MB |
afhq_box.pt |
AFHQ-Cat 256 - centered 80x80 box inpainting, sigma=0.05 | 26.47 / 0.892 / 0.054 | 422 MB |
afhq_deblur.pt |
AFHQ-Cat 256 - Gaussian deblurring, sigma_b=3.0, k=61, sigma=0.05 | 26.25 / 0.674 / 0.160 | 422 MB |
afhq_sr.pt |
AFHQ-Cat 256 - 4x super-resolution, sigma=0.05 | 26.07 / 0.704 / 0.131 | 422 MB |
brain_r4_20db.pt |
fastMRI brain - multi-coil CS-MRI, R=4, 20 dB | 32.07 / 0.889 | 156 MB |
brain_r4_30db.pt |
fastMRI brain - multi-coil CS-MRI, R=4, 30 dB | 32.89 / 0.899 | 156 MB |
brain_r8_20db.pt |
fastMRI brain - multi-coil CS-MRI, R=8, 20 dB | 28.11 / 0.819 | 156 MB |
brain_r8_30db.pt |
fastMRI brain - multi-coil CS-MRI, R=8, 30 dB | 29.12 / 0.845 | 156 MB |
Metrics: 100 held-out test images, one step (k=1), a single posterior draw. MRI uses the
magnitude, per-slice dynamic-range convention and has no LPIPS. SHA256SUMS lists the
checksum of every file.
License
The weights are trained on third-party datasets and inherit their terms: CelebA (non-commercial research only), AFHQ (CC BY-NC 4.0) and fastMRI (its data-use agreement). Use them for non-commercial research. The code is MIT-licensed separately.