ICLR 2026: How LiveMoments Restores Live Photo Cover Frames Without Blur

Task Definition

Reselected Key Photo Restoration uses the high‑resolution original cover frame of a Live Photo as a reference to restore a newly selected cover frame, aiming for visual quality comparable to the original cover.

Method Overview (LiveMoments)

LiveMoments adopts a reference‑guided diffusion framework with two branches:

Generative branch : a diffusion model provides a strong generative prior for high‑fidelity texture synthesis.

Motion‑alignment module : aligns the original and reselection frames both in latent space (motion‑guided attention) and in pixel space (block‑matching) to handle temporal offset and motion misalignment.

An attention‑based feature‑fusion mechanism injects fine details from the original cover into the diffusion process, preserving structural consistency while enhancing sharpness.

Baseline Design for Detail Transfer

The diffusion model extracts fine‑grained features; a reference‑guided attention fusion incorporates texture from the original cover, enabling stable restoration of the reselection frame.

Motion Alignment Design

Because the original and reselection frames often exhibit noticeable time shift and motion displacement, the motion‑alignment module operates in two stages:

In latent space, motion‑guided attention enhances consistency of fused features.

In pixel space, block‑matching retrieves the correct reference region, reducing artifacts caused by motion.

Dataset and Evaluation

Three datasets were built:

Two real‑world Live Photo collections captured with vivo X200 Pro and iPhone 15 Pro.

One synthetic dataset.

Evaluation metrics were adapted from no‑reference image‑quality measures to suit the reselection task.

Experimental Results

LiveMoments achieved the highest scores on both real datasets, surpassing state‑of‑the‑art reference‑super‑resolution (RefSR) and single‑frame super‑resolution methods. Visual comparisons show clearer edges and fewer artifacts.

Conclusion

By exploiting the naturally available high‑resolution original cover as a reference, the dual‑branch diffusion model with unified motion alignment reliably transfers structure and texture to reselection frames, even in complex dynamic scenes. The work introduces a new reference‑guided image‑restoration paradigm for computational photography on mobile devices.