NVIDIA and USC Tackle Video Relighting Flicker

A new research effort called HorizonRelight aims to fix the visible lighting jumps that appear when diffusion models relight long videos in short chunks.

Diffusion-based video models are typically trained on brief clips, which creates a problem when they're applied to longer footage: since the tools can only process short segments at a time, relighting a full-length video means stitching together multiple chunks. That stitching often produces a visible seam, with lighting conditions shifting noticeably from one segment to the next. NVIDIA's research team, working with USC, says it has a new approach to that problem called HorizonRelight. According to a post from NVIDIA AI, the method works by carrying forward context from the target lighting domain across a sliding window as the video progresses, rather than treating each chunk as an isolated relighting task. The idea is to give consecutive segments shared information about the desired lighting outcome, so transitions between chunks stay consistent instead of resetting with each new clip. This kind of continuity issue is a known limitation for diffusion video tools generally, since most are architected around short-form generation rather than long-form editing. NVIDIA has not published full technical details, benchmarks, or a release timeline for HorizonRelight beyond the initial announcement. The project appears to be an early research showcase rather than a shipped product, but it points to ongoing work at NVIDIA and its academic partners to solve one of the more persistent quality problems in AI-driven video relighting.