LUMEN: New AI enhances low-light images using depth-guided flash and transformers

LUMEN tackles the fundamental challenge of low-light image enhancement: existing methods apply uniform processing regardless of depth, ignoring that light attenuates and sensor noise increases with distance. The multi-stage framework first uses an encoder-decoder to estimate depth from the input image. A soft clustering module then partitions pixels into depth-aware regions. For each region, a virtual flash is simulated based on the estimated depth, mimicking the effect of a real flash but without hardware. The core innovation is the efficient attention-based fusion block that combines original image features, depth representations, and simulated flash features to restore global context while preserving fine details. A composite loss function—including reconstruction, perceptual, structural, color, edge, and depth consistency terms—ensures visually natural results.

Extensive experiments on the LOL-v1 and LOL-v2 benchmarks show LUMEN achieves state-of-the-art performance both quantitatively (PSNR, SSIM) and qualitatively. The 6-page paper includes 2 figures and 1 table demonstrating that the depth-guided approach effectively reduces halos and artifacts common in other flash-based methods. Accepted for the IEEE ICIP 2026 conference, the work was authored by Bibhabasu Debnath, Sahana Ray, and Sanjay Ghosh. The code and additional materials have been made available on arXiv (ID: 2605.17893). This approach has practical implications for mobile photography, surveillance, and autonomous driving where reliable low-light vision is critical.