Beijing Normal Univ's DeflareMambav2 removes lens flares with adaptive priors

Lens flares, caused by complex optical aberrations in lenses, severely degrade image quality—especially in nighttime photography. Traditional restoration methods process the entire image uniformly, which fails to balance the conflicting needs of preserving bright light sources, removing flare artifacts, and recovering dark background details. This spatial uniformity problem has limited the effectiveness of existing deep learning approaches.

To overcome this, researchers at Beijing Normal University developed DeflareMambav2, a prior-guided model built on the Mamba architecture (a state space model alternative to Transformers). The key innovation is a Flare Prior Network (FPN) that estimates per-pixel flare probabilities, enabling adaptive processing. A novel radial serialization strategy samples pixels along circular paths emanating from bright sources, naturally encoding the spatial structure of flares while preserving long-range dependencies via SSMs. The backbone then applies a dual-level adaptive scheme: it explicitly protects light-source regions from over-processing, and applies curriculum-based restoration to contaminated areas with pixel-level intensity calibration. Extensive experiments show DeflareMambav2 achieves state-of-the-art results on standard benchmarks with a reduced parameter count, making it practical for real-world deployment.