AI Paper Atlas

Research map / Image generation / Image restoration

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

I-Hsiang Chen, Isma Hadji, Enrique Sanchez, Adrian Bulat, Sy-Yen Kuo, Radu Timofte, Georgios Tzimiropoulos, Brais Martinez

CVPR 2026 · Image generation › Image restoration · approach: Efficient restoration architectures · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace CLIP IQA with DepictQA free-form IQA (ablation)SD1.5 restorer · UDC unknownPSNR: 20.67 -> 22.34Improved
upgrade backbone from SD1.5 to SD3.5 (ablation)RAR · UDC unknownPSNR: 22.34 -> 24.41Improved
upgrade backbone from SD1.5 to SD3.5 (ablation)RAR · Composite Group CPSNR: 17.67 -> 17.88No clear effect
add noise-free direct degraded-to-clean mapping (ablation)RAR SD3.5 · UDC unknownPSNR: 24.41 -> 24.70Improved
feed IQA in latent space instead of pixel space (ablation)RAR SD3.5 · Composite Group CPSNR: 17.89 -> 18.72Improved
condition on IQA embeddings instead of decoded text (ablation)RAR SD3.5 · UDC unknownPSNR: 24.90 -> 28.49Improved
condition on IQA embeddings instead of decoded text (ablation)RAR SD3.5 · Composite Group CLPIPS: 0.1576 -> 0.1730Hurt
add iterative conditioning during training (ablation)SD1.5 diffusion · UDC unknownPSNR: 25.57 -> 23.05Hurt
add iterative conditioning during training (ablation)SD3.5 flow matching · UDC unknownPSNR: 28.49 -> 28.60Improved
use RAR stopping criterion (2.4 rounds on average) vs fixed 4 rounds (ablation)RAR · UDC unknownPSNR: 28.60 -> 29.69Improved
use RAR iterative restoration vs AgenticIRAgenticIR · Composite Group Cwall clock time (s): 78.00 -> 6.92Improved
use RAR vs AutoDIR on single-degradation tasksAutoDIR · Single degradation averagePSNR: 27.81 -> 25.88Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy