mlsanityData to 7 Oct 2026

GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution

arXiv 2603.16769 ↗First posted 16 Mar 2026cs.CVcited by 3

On the map: LLM alignment / RLHF / preference optimisation → Diffusion and multimodal model alignment

Abstract (arXiv)

Recently, reinforcement learning (RL) has been employed for improving generative image super-resolution (ISR) performance. However, the current efforts are focused on multi-step generative ISR, while one-step generative ISR remains underexplored due to its limited stochasticity. In addition, RL methods such as Direct Preference Optimization (DPO) require the generation of positive and negative sample pairs offline, leading to a limited number of samples, while Group Relative Policy Optimization (GRPO) only calculates the likelihood of the entire image, ignoring local details that are crucial for ISR. In this paper, we propose Group Direct Preference Optimization (GDPO), a novel approach to integrate RL into one-step generative ISR model training. First, we introduce a noise-aware one-step diffusion model that can generate diverse ISR outputs. To prevent performance degradation caused by noise injection, we introduce an unequal-timestep strategy to decouple the timestep of noise addition from that of diffusion. We then present the GDPO strategy, which integrates the principle of GRPO into DPO, to calculate the group-relative advantage of each online generated sample for model optimization. Meanwhile, an attribute-aware reward function is designed to dynamically evaluate the score of each sample based on its statistics of smooth and texture areas. Experiments demonstrate the effectiveness of GDPO in enhancing the performance of one-step generative ISR models. Code: https://github.com/Joyies/GDPO.

Key facts. Cited by 3 papers in the map (data through 7 Oct 2026). Placed in Diffusion and multimodal model alignment, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 57 of its 65 references that are in the map. First posted on arXiv 16 Mar 2026.

Builds on

57 of 65 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025DanceGRPO: Unleashing GRPO on Visual Generation12 May
  2. 2023Diffusion Model Alignment Using Direct Preference Optimization21 Nov

Other references · 55

  1. 2020Denoising Diffusion Probabilistic Modelscited by 20,851
  2. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
  3. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,084
  4. 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,029
  5. 2020Score-Based Generative Modeling through Stochastic Differential Equationscited by 8,964
  6. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  7. 2021Diffusion Models Beat GANs on Image Synthesiscited by 8,185
  8. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  9. 2018A Style-Based Generator Architecture for Generative Adversarial Networkscited by 5,986
  10. 2023Adding Conditional Control to Text-to-Image Diffusion Modelscited by 5,466
  11. 2016Perceptual Losses for Real-Time Style Transfer and Super-Resolutioncited by 2,678
  12. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
38 more
  1. 2016Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Networkcited by 1,898
  2. 2021SwinIR: Image Restoration Using Swin Transformercited by 1,869
  3. 2017Enhanced Deep Residual Networks for Single Image Super-Resolutioncited by 1,543
  4. 2018Image Super-Resolution Using Very Deep Residual Channel Attention Networkscited by 1,256
  5. 2018ESRGAN: Enhanced Super-Resolution Generative Adversarial Networkscited by 1,250
  6. 2023ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillationcited by 1,207
  7. 2021MUSIQ: Multi-scale Image Quality Transformercited by 1,058
  8. 2022Exploring CLIP for Assessing the Look and Feel of Imagescited by 974
  9. 2021Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Datacited by 961
  10. 2022Denoising Diffusion Restoration Modelscited by 951
  11. 2018Residual Dense Network for Image Super-Resolutioncited by 796
  12. 2020Image Quality Assessment: Unifying Structure and Texture Similaritycited by 792
  13. 2025Flow-GRPO: Training Flow Matching Models via Online RLcited by 605
  14. 2022Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Modelcited by 574
  15. 2021Designing a Practical Degradation Model for Deep Blind Image Super-Resolutioncited by 524
  16. 2022MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessmentcited by 514
  17. 2023Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levelscited by 489
  18. 2023Exploiting Diffusion Prior for Real-World Image Super-Resolutioncited by 486
  19. 2019Toward Real-World Single Image Super-Resolution: A New Benchmark and A New Modelcited by 400
  20. 2023ResShift: Efficient Diffusion Model for Image Super-resolution by Residual Shiftingcited by 360
  21. 2023SeeSR: Towards Semantics-Aware Real-World Image Super-Resolutioncited by 311
  22. 2023DiffBIR: Towards Blind Image Restoration with Generative Diffusion Priorcited by 255
  23. 2023Pixel-Aware Stable Diffusion for Realistic Image Super-resolution and Personalized Stylizationcited by 251
  24. 2020Component Divide-and-Conquer for Real-World Image Super-Resolutioncited by 243
  25. 2024One-Step Effective Diffusion Network for Real-World Image Super-Resolutioncited by 243
  26. 2023SinSR: Diffusion-Based Image Super-Resolution in a Single Stepcited by 235
  27. 2021SNIPS: Solving Noisy Inverse Problems Stochasticallycited by 191
  28. 2023Text-to-3D with Classifier Score Distillationcited by 127
  29. 2022Details or Artifacts: A Locally Discriminative Learning Approach to Realistic Image Super-Resolutioncited by 114
  30. 2024Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approachcited by 93
  31. 2024Arbitrary-steps Image Super-resolution via Diffusion Inversioncited by 85
  32. 2024TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolutioncited by 83
  33. 2025VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rankcited by 77
  34. 2024AddSR: Accelerating Diffusion-based Blind Super-Resolution with Adversarial Diffusion Distillationcited by 66
  35. 2024Degradation-Guided One-Step Image Super-Resolution with Diffusion Priorscited by 64
  36. 2024Towards Realistic Data Generation for Real-World Super-Resolutioncited by 56
  37. 2025DiT4SR: Taming Diffusion Transformer for Real-World Image Super-Resolutioncited by 50
  38. 2025Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modelingcited by 47

50 most cited of 55 All 55 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Bridging Restoration and Generation in One-step Diffusion for Real-World Image Super-Resolutioncited by 1
  2. 2026Hierarchical Image Tokenization for Multi-Scale Image Super Resolutioncited by 0
  3. 2026RelayVSR: Large-Small Model Collaboration for Efficient Real-World Video Super-Resolutioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025DP2O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution
  2. 2024RFSR: Improving ISR Diffusion Models via Reward Feedback Learning
  3. 2026V-GRPO: Online Reinforcement Learning for Denoising Generative Models Is Easier than You Think
  4. 2025ORL-LDM: Offline Reinforcement Learning Guided Latent Diffusion Model Super-Resolution Reconstruction
  5. 2026LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution
  6. 2025Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization
  7. 2025AR-GRPO: Training Autoregressive Image Generation Models via Reinforcement Learning
  8. 2023DFU: scale-robust diffusion model for zero-shot super-resolution image generation