resenh weights

These are the weights for resenh, a library for robust low-light image restoration and enhancement. An adaptive classical pipeline (Retinex, CLAHE, edge-aware denoising) and a Retinexformer network run side by side, and a quality controller blends the two photo by photo.

Try it in your browser: resenh.sharadh.blog · Install: pip install "resenh[dl]" · Report: technical report (PDF)

import resenh

res = resenh.enhance("night.jpg")   # downloads and verifies the weights below on first use
res.save("night_restored.png")
print(res.summary())

Files

File	What it is	Used by
resenh_light_general.pth	Retinexformer with classical-prior inputs, fine-tuned on LOL + LSRW (de-duplicated)	package default (mode="hybrid")
resenh_light_lolv1.pth, resenh_light_lolv2_real.pth, resenh_light_lolv2_syn.pth	the same architecture fine-tuned per benchmark	model= argument, evaluation
LOL_v1.pth, LOL_v2_real.pth, LOL_v2_synthetic.pth	official Retinexformer checkpoints (Cai et al., ICCV 2023, MIT)	mode="dl", fallback
web/resenh-web-hybrid.onnx, web/resenh-web-dl.onnx	ONNX exports (opset 17, dynamic size) for ONNX Runtime Web	the in-browser demo

The package pins every file's SHA-256 (src/resenh/models/weights.py) and refuses a mismatch.

Results

Plain RGB PSNR on the saved 8-bit image. There is no ground-truth brightness correction, no per-image knobs, and no checkpoint picked on the test set. Our harness reproduces the official Retinexformer checkpoints exactly. Full tables with 95% confidence intervals: eval_P1_T46.md.

PSNR (dB) ↑	LOL-v1 (15)	LOL-v2-Real (100)	LOL-v2-Syn (100)	LSRW test (50)
Classical branch	16.75	17.44	16.83	16.15
Retinexformer, network only	25.15	22.79	25.67	17.19
resenh hybrid	25.13	22.82	25.35	19.26

LOL-v2-Real test	PSNR ↑	SSIM ↑	LPIPS ↓	NIQE ↓
Classical branch	17.44	0.724	0.417	4.83
Retinexformer, network only	22.79	0.840	0.171	3.59
resenh hybrid	22.82	0.840	0.171	3.59

• LOL hybrid rows: the official per-benchmark backbone with the benchmark preset.
• LSRW hybrid row: resenh_light_general.pth, which is in-domain for LSRW cameras (it was trained on LSRW train). It was chosen on LSRW validation before the test set was scored.
• Fine-tuned per-benchmark weights: trained with the GT-mean loss. They improve SSIM and LPIPS (LOL-v2-Real: 0.852 / 0.139) but score lower in plain PSNR (20.82 dB). See the ablation.
• Latency: a 1 MP photo takes 0.49 s for the network and 1.9 s for the full hybrid on an RTX 4070 laptop GPU, and about 6 s for the hybrid in a laptop browser with WebGPU.

Training

• Architecture: Retinexformer, 1.6 M parameters, with a zero-initialised stem extension for the classical priors (illumination map, Retinex output).
• Loss: Charbonnier + 0.2·(1−SSIM) + 0.02·VGG-19 perceptual, wrapped in the GT-mean loss.
• Recipe: progressive patches (128→256→384), EMA 0.999, gamma and noise augmentation, and scene-grouped validation splits.

Team

Team V-Max (T46): Sharadh Naidu, Shivani, S. V. Swamy, Sumukha K. MIT licence. Retinexformer weights © their authors (MIT).