UbiEnlight: User-Perception-Aware Real-Time Low-Light Video Enhancement on Mobile Devices and Smartglasses
Minfan Wang, Sicong Liu, Teng Li, Zimu Zhou, Sixun He, Bin Guo, Junzhao Du, Zhiwen Yu · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies · 2026
Low-light video capture is now common on smartphones and wearables, yet dynamic illumination coupled with mobile/werable user motion causes noise, blur, and flicker that make videos hard to perceive and use , especially for users with night blindness and near-eye displays. Prior work enhances frames with post-hoc temporal constraints or optimizes videos directly; however, low-light and motion degradations are spatiotemporally coupled , so post-hoc consistency after per-frame restoration often leaves flicker and motion artifacts. We present UbiEnlight, the first diffusion-transformer-based on-device system for real-time, temporally stable low-light video enhancement. UbiEnlight builds on two insights: (1) illumination and structure separate more robustly in the frequency domain, enabling a spectrum-guided diffusion transformer that injects Fourier amplitude/phase priors to correct illumination while anchoring structure without optical flow; and (2) user perceptual tolerance varies by context, motivating a perception-aware controller that adapts sampling depth, attention reuse, and frame caching at runtime. We implement UbiEnlight on smartphones and smartglasses, and evaluate it against state-of-the-art baselines. UbiEnlight improves temporal stability by up to 54.76% under dynamic/extreme conditions while sustaining real-time on-device performance.