PrismRestore compacts multi-task image fixes into one model

PrismRestore: A Compact Restoration Toolbox for Agentic Image Restoration

Computer Vision and Pattern Recognition

Summary

Handling images with many kinds of damage usually means storing many separate fix-it tools. The authors propose PrismRestore, which creates a small set of fix-it actions from one all-in-one model. It groups parts of the model based on how they react to tasks, then picks which parts to adjust for each kind of fix. A policy learns to combine these actions for the best image restoration quality, all with one model that runs faster than other multi-tool approaches.

What this means in practice

  • For photo editing software developers: Integrate PrismRestore to enhance image repair features with a single efficient model supporting multiple damage types without extra restoration networks.$Commercial implications: Enables photo editing tools to sell faster and more compact multi-effect image repair powered by a single unified model.
  • For mobile app developers: Use PrismRestore to deliver versatile image restoration with improved speed and less storage for multi-degradation fixes in mobile photo apps.

Authors

Zhaocheng Yu, Kui Jiang, Jing Cao, Yiang Chen, Xiaocheng Feng, Junjun Jiang, Xianming Liu

Abstract

Agentic image restoration composes operations to handle mixed degradations, yet its toolbox often requires storing multiple restoration networks. We propose PrismRestore, which constructs a compact bank of callable restoration actions from a single frozen all-in-one model. Our approach builds on task-dependent yet overlapping channel responses in the shared model. An offline task-prior search uses activation and gradient statistics to group channels with similar response profiles, then searches task-specific modulation coefficients over non-representative channels while leaving representative channels unmodulated. The resulting prior bank enables explicit action selection and composition without additional restoration backbones or task-specific adapter modules. A state-conditioned policy learns to compose these actions using Group Relative Policy Optimization. Its objective separately normalizes fidelity and perceptual feedback within groups of trajectories sampled for the same input, and combines terminal evaluation with return-to-go derived from stepwise quality changes. Experiments on synthetic composite-degradation and real-world multi-task benchmarks show improvements over the evaluated all-in-one baselines and restoration quality competitive with multi-model agents. PrismRestore achieves these results with a single restoration backbone and faster inference than the evaluated agentic baselines.