Low-light image enhancement is more than turning up the brightness. Real low-light images typically combine uneven illumination with amplified noise, color distortion, and lost local detail. The paper from arXiv (2610.22958) addresses this by focusing on how to condition enhancement models on illumination in a stable way.

The authors propose a method built on state-space models (SSMs), which are increasingly used for image processing. To keep illumination conditioning stable, they add what they call bottleneck prompt sidecars — small auxiliary modules that inject illumination information into the model without disturbing its main processing path. This is designed to handle spatially varying illumination, where different parts of an image need different levels of correction.

Because this is a single source, there are no competing claims to compare. The abstract does not yet report quantitative results, but the contribution is a structural fix for a known instability in SSM-based low-light enhancement. If the approach holds up, it could make such models more reliable for real-world images where lighting is rarely uniform.