Samdany/photofix-models

PhotoFix models

Trained weights for PhotoFix, an automatic photo enhancement pipeline. See the project README for a demo video, benchmarks and the write-up.

File	Model	Params	Trained on	Held-out result
analyzer.pt	Dual-view EfficientNet-B0 defect classifier (7 defects)	8.7M	DIV2K + synthetic defects	mean AUC 0.978; noise/blur F1 ≥ 0.99
lut.pt	Image-Adaptive 3D LUT (3 basis LUTs, 33³)	0.57M	MIT-Adobe FiveK, Expert C (1,897 pairs)	22.41 dB vs expert (no edit: 21.01); ΔE on finished photos 2.45
restorer.pt	NAFNet, width 24 (blind denoise/deblur/JPEG cleanup)	3.9M	DIV2K + realistic synthetic damage	28.80 dB (input 26.51); noise +2.9 dB over NL-means

All three were trained on a single Apple M1 Pro (PyTorch MPS) in about 35–70 minutes each.

Usage

git clone https://github.com/sam666-deb/PhotoFix && cd PhotoFix
hf download Samdany/photofix-models analyzer.pt lut.pt restorer.pt --local-dir checkpoints
pip install -r requirements.txt && uvicorn server.main:app

from photofix.pipeline import Pipeline
from photofix.imageio import load_path

result = Pipeline.load("checkpoints").run(load_path("photo.jpg"), style="natural")  # or "pro"
print(result.steps)  # e.g. ['Neural denoise & deblur', 'Exposure (brighten, gamma 0.54)', ...]

Intended use and limitations

• These weights are for photo enhancement research and demos. They're trained on synthetic defects and one retoucher's style, so they don't represent every taste or camera.
• Deliberately dark or bright photos (silhouettes, night skies) can be over-corrected. PhotoFix's guardrail mitigates this but doesn't solve it.
• Restoration was validated on synthetic noise and blur; real sensor noise may behave differently.

Licenses

• lut.pt is trained on MIT-Adobe FiveK, which is licensed for research use only (terms). Don't use it commercially.
• analyzer.pt and restorer.pt are trained on DIV2K (academic research use). The analyzer starts from torchvision's ImageNet EfficientNet-B0 weights.
• Architectures re-implemented from Zeng et al. (TPAMI 2020, 3D LUT) and Chen et al. (ECCV 2022, NAFNet).