Learning to Corrupt for Better Restoration, ECCV, 2026.

Recalling that diffusion models map noisy images to clean ones, the choice of noise addition~(\ie, corruption) critically influences the denoising process. However, prior diffusion-based image restoration~(IR) methods often employ na"ive corruption strategies—such as randomly sampling noise, fixing the timestep, or even omitting corruption altogether—which may not accurately reflect the actual degradation of each image. To handle this, we propose Input-Aware Corruption for IR~(IAC-IR) framework, which maps each low-quality~(LQ) input into an optimal noisy sample that lies on the pretrained diffusion trajectory. Instead of choosing corruption heuristically, we predict the timestep and noise using supervision derived from the properties of the pretrained diffusion model. Specifically, the predicted timestep aligns the corrupted sample with the Gaussian noise corrupted distribution, while the predicted noise preserves the recoverable content of the input. Moreover, we use these input-aware corruption factors to improve conventional score-based distillation. Rather than relying on random corruption, which often produce unreliable target scores and weak gradients, we perform distillation with input-aware corruption, yielding more reliable score estimates and more stable distillation. By modeling input-aware corruption and integrating it into distillation, our method better leverages the pretrained diffusion prior, achieving superior perceptual quality in image restoration.