LiveMoments: Reference‑Guided Diffusion Boosts Live Photo Cover Frame Quality (ICLR 2026)

Background

Live Photo combines a high‑resolution still cover and a short video. Users often reselect a frame from the video as a new cover, but the selected frames suffer from low resolution and heavy compression, resulting in poor visual quality.

New Task Definition

The authors define “Reselected Key Photo Restoration” – using the original high‑quality cover as a reference to guide the enhancement of the low‑quality reselected frame. This differs from traditional reference‑based super‑resolution, which relies on external databases, and from video‑wide super‑resolution, which processes entire sequences.

Technical Approach

LiveMoments builds on a diffusion model and introduces a reference‑guided restoration framework. An attention‑based feature fusion injects details from the original cover into the diffusion process, preserving structure while improving clarity.

To address motion misalignment between the original and reselected frames, a motion‑alignment module operates both in latent space (motion‑guided attention) and pixel space (block‑matching retrieval), aligning the reference correctly before fusion.

Experiments

The team constructed two real‑world Live Photo datasets captured with vivo X200 Pro and iPhone 15 Pro, plus a synthetic dataset. Evaluation metrics were adapted for the task. Table 1 compares LiveMoments with existing reference‑SR and single‑frame SR methods, showing superior scores on both datasets.

Qualitative results (Figure 4) demonstrate that LiveMoments restores fine details and achieves “cover‑level” sharpness, outperforming competing methods.

Conclusion and Outlook

By exploiting the native high‑resolution cover as a reference and integrating a dual‑branch diffusion architecture with motion alignment, LiveMoments reliably restores reselected frames even in complex dynamic scenes. The work opens avenues for higher‑quality cover selection in everyday photography, content creation, and further research on reference‑guided diffusion.