JK Joonkyu Park

ICLR, 2026.

Input-Aware Corruption for Image Restoration

Single-Step General Image Restoration

IAC-IR explores an input-aware approach to image restoration, modeling degradations in a way that better reflects the visual evidence preserved in each input image.

All Projects ←
Input-Aware Corruption for Image Restoration
PROJECT IAC-IR (with Samsung MX)

Overview

Real-world image restoration requires more than applying a fixed degradation assumption to every image. The visual evidence available in an input can vary substantially across regions, degradation types, and severity levels.

IAC-IR investigates input-aware corruption modeling for image restoration, enabling the restoration process to better preserve reliable image content while recovering visually plausible details in degraded regions.

01 — INPUT Adapt restoration behavior to the characteristics of each degraded image.
02 — CORRUPTION Model degradation in a manner that reflects image-dependent uncertainty.
03 — RESTORATION Balance perceptual quality with faithful preservation of input content.

Interactive Comparison

The examples below compare Bicubic upsampling and IAC-IR. Drag the divider horizontally to inspect the restored details.

Key Idea

IAC-IR focuses on restoration settings where different regions of an image can require different levels of generative correction. Instead of treating every region identically, the framework uses input evidence to guide where restoration should preserve, enhance, or plausibly reconstruct details.

Results

Qualitative comparisons illustrate the behavior of IAC-IR under realistic degradation conditions. In particular, the method aims to preserve reliable image content while recovering visually plausible details in regions that require stronger restoration.

  • Paper: Coming soon
  • Code: Coming soon