NVIDIA and USC's HorizonRelight Fixes Flicker in Long Video Relighting

Diffusion video models are mostly trained on short clips, so relighting a longer video means processing it in chunks, and NVIDIA and USC researchers say the lighting can visibly jump between chunks. Their method, HorizonRelight, targets that seam. According to the paper's abstract, it trains the model to continue from temporally masked propagated context, which the paper calls masked target-domain self-conditioning, so each sliding window picks up where the last left off. It also starts from a relit anchor frame produced by a controllable generative model, which sets the initial target lighting state.

The abstract reports markedly improved temporal consistency on long in-the-wild videos, with chunk-boundary artifacts largely reduced. These are the authors' reported results, no numbers or baselines appear in the abstract, and the demo video that accompanied NVIDIA's announcement was not reviewed.

“Most diffusion-based video models are trained on short clips.”

“When they’re used to relight longer videos in chunks, the lighting can visibly jump between them.”