Research map / Image generation / Image restoration
Learning to Corrupt for Better Restoration
Joonkyu Park, Wooseok Lee, Jaeha Kim, Sehoon Kim, Bokyeung Lee, Kyoung Mu Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| predict input-aware timestep instead of fixing it (ablation) | IAC-IR · RealSR | LIQE: 4.571 (t=400) -> 4.668 | Improved |
| larger fixed timestep (ablation) | IAC-IR · RealSR | PSNR vs LIQE: PSNR 23.064 -> 20.492; LIQE 3.816 -> 4.571 | Mixed |
| use ground-truth timestep tGT instead of predicted (ablation) | IAC-IR · RealSR | PSNR: 22.509 -> 22.698 | Improved |
| add explicit noise supervision L_eps (ablation) | IAC-IR · RealSR | MANIQA: 0.483 -> 0.589 | Improved |
| replace L_eps with KL divergence noise loss (ablation) | IAC-IR · RealSR | MANIQA: 0.589 -> 0.523 | Hurt |
| add timestep loss Lt (ablation) | IAC-IR · RealSR | MANIQA: 0.457 -> 0.483 | Improved |
| use both Lt and L_eps together (ablation) | IAC-IR · RealSR | MANIQA: 0.457 -> 0.589 | Improved |
| distillation with predicted timestep and noise instead of random (ablation) | Ldistill · RealSR | MANIQA: 0.479 -> 0.589 | Improved |
| use Gaussianity measure G() instead of KL (ablation) | IAC-IR · RealSR | PSNR: 21.190 -> 22.509 | Improved |
| increase iterations from 1 to 5 (ablation) | IAC-IR · RealSR | PSNR / MANIQA: PSNR 22.509 -> 18.798; MANIQA 0.589 -> 0.635 | Mixed |
| IAC-IR vs prior one-step diffusion IR on perceptual metrics | IAC-IR · RealSR | LIQE: 4.068 (OSEDiff) -> 4.668 | Improved |
| IAC-IR vs OSEDiff on distortion | IAC-IR · RealSR | PSNR: 23.589 -> 22.509 | Hurt |
Closest papers by meaning
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- Blind Image Restoration via Fast Diffusion Inversion (NeurIPS 2024 · similarity 0.88)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 2025 · similarity 0.88)
- IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models (ICML 2025 · similarity 0.87)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
