BodyReLux: Temporally Consistent Full-Body Video Relighting
BodyReLux is a subject-specific diffusion model for full-body video relighting, and the main technical win is temporal consistency. Instead of relighting single frames, it targets entire performance clips so lighting changes stay stable across motion, which is exactly where many generative video methods fall apart.
The paper’s dataset strategy is also notable: it combines traditional static one-light-at-a-time capture with a dynamic capture method that rapidly interleaves two smoothly varying lighting sequences. Because the lighting stays above the flicker-fusion threshold, the interleaving doesn’t visibly strobe. The model is initialized from a pretrained text-to-video system, then uses per-light token conditioning plus masked attention for sequences of lighting, aiming for both control and realism.
For game devs, this is most relevant to cinematic pipelines, virtual production, and character-facing tools rather than runtime gameplay. If the approach holds up outside the paper, it...
“temporally consistent full-body human performances”
- what
- BodyReLux is a subject-specific video diffusion framework for temporally consistent full-body human relighting.
- who
- Authors include Li Ma, Mingming He, Xueming Yu, David M. George, Ahmet Levent Taşel, Paul Debevec, and Julien Philip.
- when
- Submitted to arXiv on 20 May 2026 (arXiv:2605.21766).
- impact
- Could help cinematic, performance-capture, and virtual-production workflows by making relighting edits less flickery and more controllable.
Promising relighting quality with clear production relevance
Follow relighting updates
See relevant stories in your personalized news feed.
Discussion