Research map / Image generation / Image restoration

IRPO: Boosting Image Restoration via Post-training GRPO

NeurIPS 2026 · Image generation › Image restoration · approach: Efficient restoration architectures · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes IRPO, a GRPO-based post-training framework for deterministic image restoration that fine-tunes AdaIR on its worst 30% training samples with a composite fidelity/VLM/task-aware reward, improving in-domain and OOD PSNR.

Problem

Deterministic image restoration models trained with fixed pixel-wise losses tend to over-smooth and generalize poorly to out-of-domain and real-world degradations; post-training for low-level vision is underexplored.

Approach

Select the worst-performing 30% of training samples for post-training, turn a deterministic restorer into a stochastic policy via lightweight Beta-distribution heads on AdaIR's AFLB modules, and optimize it with GRPO under a composite reward (general fidelity/quality, a VLM expert score, and degradation-specific restoration cues) combined with supervised L1 and consistency losses.

Key results

• On the five-task all-in-one benchmark, IRPO reaches 31.13 dB average PSNR, improving AdaIR (30.20 dB) by 0.93 dB.
• On OOD real-world benchmarks, IRPO raises average PSNR from 20.30 to 23.73 dB (+3.43 dB) over AdaIR.
• Training on the 30% underperforming subset matches or beats full-data training (31.13 vs 30.95 dB) at roughly one third the cost.

Surprising findings

• Fine-tuning AdaIR on the hard subset with supervised loss only improves the five-degradation average but not OOD robustness (OOD 20.22 vs 20.30).
• Reward-ranked distillation over the same candidates and rewards falls short of full GRPO, suggesting group-relative policy learning adds value beyond candidate selection.

Limitations

• IRPO is instantiated only on AdaIR; transfer to other restoration backbones is asserted but not tested.
• Reward design relies on hand-crafted heuristics and a coarse VLM judge; reward weights tuned on a single task (SOTS).

Experimental setup

Datasets: SOTS Rain100L GoPro LOL BSD400 WED BSD68 Urban100

Benchmarks: SOTS Rain100L CBSD68 GoPro LOL RRSHID GT-Rain SIDD ReLoBlur LoLi-Street Urban100

Metrics: PSNR SSIM LPIPS DISTS NIQE MUSIQ

Baselines: AdaIR DFPIR VLU-Net InstructIR GridFormer PromptIR AirNet Restormer NAFNet MPRNet

Builds on: GRPO AdaIR PPO DeepSeek-R1

Closest papers by meaning

• UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.89)
• Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration (AAAI 2024 · similarity 0.88)
• IR-CM: The Fast and General-purpose Image Restoration Method Based on Consistency Model (NeurIPS 2024 · similarity 0.88)
• Boosting Image Restoration via Priors from Pre-trained Models (CVPR 2024 · similarity 0.87)
• Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.87)
• Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.87)