Kling AI team and partners unveil JointLight, a SIGGRAPH Asia 2026 video relighting method
jiqizhixin · x · 2026-11-06
JointLight, from UCAS/CAS, Cardiff University and Kuaishou's Kling AI team (SIGGRAPH Asia 2026 conference paper), relights video while preserving identity and motion. Key ideas: use estimated meshes and target HDR lighting to build coarse irradiance maps, then have a diffusion transformer jointly generate relit RGB and refined irradiance rather than trusting imperfect geometry; run separate denoising noise schedules for RGB and irradiance with training-time domain matching; fine-tune on synthetic paired videos and test transfer on real footage. Results are reported on the SynthDynamic benchmark.
More from Multimodal
- Meta's Muse app adds real-time voice: talk to your custom avatar like FaceTime — Scobleizer · 2026-11-06
- AI-generated portrait 'Shadow Play' shows photorealistic skin and lighting detail — aziz4ai · 2026-11-06
- Land Rover Defender 110 concept brand campaign made entirely with Firefly and AI video — ENSAYME · 2026-11-06
- Matthew Berman demos a 4B AI model that generates 3D game assets — Matthew Berman · 2026-11-06
- Qwen-Image-2.1 Uncensored Omni: 1M-image LoRA in 9 quantized formats — metalfans · 2026-11-06
- World Labs launches Frameboy: real-time Atlas world model streaming at 15 fps — drfeifei · 2026-11-06