Lens Flare Removal and Reconstruction

arXiv cs.GR details a lens-flare removal and reconstruction pipeline aimed at a problem that matters for both capture and rendering workflows: flare can wreck downstream 3D scene reconstruction because it belongs to the camera, not the scene. The authors target large, image-filling flares rather than the small halos most prior methods handle well.

To tackle removal, they build a new large-flare dataset by combining public real-world data with procedural generation, then fine-tune a diffusion model to suppress complex flare patterns. That makes the work relevant to anyone building preprocessing for photogrammetry, SLAM, or image-based reconstruction, where bad input frames can poison the whole pipeline.

The more unusual part is reconstruction. The paper introduces an explicit flare representation that exploits symmetry around the camera principal point, then jointly optimizes that model with a Gaussian splatting scene representation. In practice, that lets the system decompose a scene into flare and non-flare components, re-render the flare, and transfer it to novel images or even new 3D scenes.

The authors also say the pipeline holds up even when automatic light-source localization is imperfect, and they evaluate both removal quality and flare/scene decomposition on an established benchmark plus a new large-reflective-flare set. For game teams, the takeaway is a potentially useful bridge between cleanup tooling and stylized, controllable flare effects.