We have released the paper on SD-Retinex, the low-light image enhancement method at the core of Beacon, on Zenodo, together with the reference implementation, evaluation code and per-image metric tables on GitHub.

SD-Retinex is a Retinex-based method designed around the derivative of the sigmoid function. By estimating illumination with a pole-free real-axis kernel and lifting brightness with a bounded Naka–Rushton-type tone map, it achieves stable enhancement without training data and without a GPU.

On the public LOL benchmark (15 evaluation pairs), it recorded the best PSNR (16.47 dB) and the lowest color error ΔE00 (15.90) among the training-free methods tested (the repository’s SSR, MSR and LIME-like implementations and XCR), processing a 600×400 image in about 18 ms on a single CPU thread (about 54 FPS). It also exceeds the pretrained deep models Zero-DCE and SCI in PSNR and color error.

See the technology page for details.