SyncLight: Controllable and Consistent Multi-View Relighting

SyncLight is a generative framework that enables consistent, parametric relighting across multiple uncalibrated views. Built on a multi-view diffusion model with Latent Bridge Matching, our method allows users to specify lighting changes (such as adjusting light intensity and color) on a single reference view, which then propagates coherently across all synchronized viewpoints in a single inference step. Unlike single-view methods that process each camera independently and produce geometrically inconsistent results, SyncLight explicitly models cross-view interactions through modified self-attention, learning to maintain rigorous spatial and photometric consistency from training on stereo pairs. Remarkably, despite being trained only on image pairs (N=2), the architecture generalizes zero-shot to arbitrary view counts at inference, making it suitable for multi-camera systems in virtual production, stereoscopic cinema, and sports broadcasting. We train on a large-scale hybrid dataset combining synthetic environments (Infinigen, BlenderKit) with high-fidelity real-world OLAT captures, enabling the model to bridge the sim-to-real gap while achieving 10-50× speedup over iterative diffusion baselines.