← Back to Computer Vision
cs.CV

Relight any photo with precise control over virtual lights

Miguel Farinha, Ronald Clark

May 18, 2026

PIXLRelight solves the problem of relighting single images with fine-grained control over light direction and color—a task where existing methods either limit control to text prompts and environment maps, accumulate errors through inverse-then-forward rendering, or require slow per-image optimization. The approach uses physically based rendering (PBR) intrinsics (albedo, diffuse shading, non-diffuse residuals) as conditioning signals for a transformer-based neural renderer. At training, it learns from multi-illumination photograph sets; at inference, it extracts the same intrinsics from a quick path-traced render of a coarse 3D reconstruction under user-specified lights, then applies target illumination to the source photo while preserving fine detail. Runs in under 100ms per image. Code and trained models are released.
Published as PIXLRelight: Controllable Relighting via Intrinsic Conditioning arXiv:2605.18735
Read the original paper →