A new paper from vivo, LiveMoments, has been accepted at ICLR 2026, marking the first dedicated approach to restoring image quality in Live Photo cover frames after user reselection. The core innovation is a reference-guided diffusion model that uses the original high-quality cover frame as a guide, combined with a motion alignment module to handle the temporal inconsistencies inherent in Live Photos. This addresses a common pain point: when users select a different frame as the cover, the resulting image often suffers from motion blur, noise, or compression artifacts. The method is notable for its practical focus—it targets a real-world mobile photography feature used by millions—while still pushing the technical boundary of diffusion models in video-to-image tasks. For developers and researchers, this signals a shift toward applying advanced generative models to specific, user-facing problems rather than generic image generation. The commercial angle is clear: improving user experience in camera apps can directly impact device satisfaction and brand loyalty. The paper's acceptance at a top-tier venue like ICLR also validates the research direction. For overseas engineers, this is a signal to watch for similar applied diffusion model work emerging from Chinese tech companies, especially in the mobile imaging space.
LiveMoments, accepted at ICLR 2026, is the first dedicated solution for restoring quality in Live Photo cover frames after user reselection. It leverages the original high-quality cover as a reference and introduces a motion-aligned diffusion model to address degradation. This work highlights a growing trend of applying generative AI to niche but high-impact mobile imaging tasks.