CVPR 2025 超分辨率(super-resolution)方向上接收论文总结
CVPR 2025会议聚焦计算机视觉领域的最新进展,共接收论文2878篇(录用率22.1%)。在图像超分辨率方向,研究呈现三大趋势:1)扩散模型应用广泛,如FaithDiff实现8K图像恢复、PiSA-SR提出双LoRA架构;2)任意尺度超分取得突破,Oral论文DiffFNO通过傅里叶神经算子支持连续倍率超分;3)轻量化和多模态成为新方向,CATANet实现高效Token聚合,多篇工作探索文本/
CVPR 2025
CVPR 2025 会议于 2025 年 6 月 10 日至 17 日在美国纳什维尔举行。
今年共提交了 13008 份有效论文,2878篇论文被接收,录用率为22.1%。这一录取数量再次刷新大会纪录,反映出计算机视觉领域的研究热度持续高涨。
图像超分辨率(Super-Resolution, SR)是计算机视觉领域的经典任务,旨在从低分辨率(Low-Resolution, LR)图像中恢复出高分辨率(High-Resolution, HR)图像。随着深度学习技术的快速发展,超分辨率技术在医学影像、遥感图像、数字摄影、监控安防等领域展现出巨大的应用价值。
现将超分辨率方向上接收的论文汇总如下,遗漏之处还请大家斧正。
图像超分
扩散模型
- FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolution
- Paper: https://arxiv.org/abs/2411.18824
- Code: https://github.com/JyChen9811/FaithDiff
- Keywords: Diffusion Priors, Faithful Super-resolution, Latent Diffusion Models
- Features: 支持 8K 及以上超高清图像恢复,FP8 推理,CPU offloading,提供了 Real-Deg 数据集(238 张真实退化图像)
- Team: Junyang Chen, Jinshan Pan, Jiangxin Dong (南京理工大学)
- PiSA-SR: Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach
- Paper: https://arxiv.org/abs/2412.03017
- Code: https://github.com/csslc/PiSA-SR
- Keywords: Dual-LoRA, Pixel-level, Semantic-level, Adjustable Super-resolution
- Blog: CVPR 2025 | PiSA-SR: 像素级和语义级可调的超分辨率
- Features: 单步扩散完成超分辨率,可调节像素级保真度和语义级细节
- Team: Zhang Lei 团队 (香港理工大学,OPPO研究院)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion
- Paper: https://arxiv.org/abs/2412.09013
- Code: http://github.com/zsyOAOA/InvSR
- Blog: 【论文速读10】CVPR2025扩散超分辨率InvSR
- Keywords: 扩散反演
- Team: 西交大,南洋理工
- PassionSR: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution
- Paper: https://arxiv.org/abs/2411.17106
- Code: https://github.com/libozhu03/PassionSR
- Keywords: 后训练量化
- Team: 上交,港中文等
- Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model
- Paper: https://arxiv.org/abs/2503.18512
- Code: https://github.com/LabShuHangGU/UPSR
- Keywords: Uncertainty-guided, Perturbation, Diffusion Model
- Team: Gu Shuhang团队(电子科技大学)
任意尺度超分
- DiffFNO: Diffusion Fourier Neural Operator (Oral)
- Paper: https://arxiv.org/abs/2411.09911
- Project: https://jasonliu2024.github.io/difffno-diffusion-fourier-neural-operator/
- Keywords: Fourier Neural Operator, Arbitrary-Scale, ODE Solver, Real-time
- Blog: CVPR 2025 Oral | DiffFNO:傅里叶神经算子助力扩散,开启任意尺度超分辨率新篇章
- Features: 支持任意连续倍率(如 2.1、11.5 等),PSNR 提升 2-4dB,推理时间减少 30% 以上
- Team: Xiaoyi Liu, Hao Tang (圣路易斯华盛顿大学,北京大学)
- HIIF: Hierarchical Encoding based Implicit Image Function for Continuous Super-resolution
- Paper: https://arxiv.org/abs/2412.03748
- Keywords: Implicit Neural Representation, Hierarchical Encoding, Continuous Super-resolution
- Features: 层次化位置编码,多头线性注意力,平均 PSNR 提升 0.17dB
- Team: University of Bristol; Netflix Inc.
轻量级模型
- CATANet: Efficient Content-Aware Token Aggregation for Lightweight Image Super-Resolution
- Paper: https://arxiv.org/abs/2503.06896
- Code: https://github.com/EquationWalker/CATANet
- Blog: 【CVPR 2025】内容感知Token高效聚合,轻量超分网络CATANet,即插即用!
- Keywords: Lightweight, Content-Aware, Token Aggregation, Transformer
- Team: 南京大学软件新技术国家重点实验室
- Progressive Focused Transformer for Single Image Super-Resolution
- Paper: https://arxiv.org/abs/2503.20337
- Code: https://github.com/LabShuHangGU/PFT-SR
- Keywords: Progressive Focused Transformer, Attention Map Integration, Efficient Computation
- Features: 通过哈达玛乘积整合注意力图,显著降低计算成本
- Team: Gu Shuhang团队(电子科技大学)
- TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond
- Paper: openaccess
- Keywords: 旅行商问题,轻量级超分
- Team: SmartMore Corporation,SSE, CUHK-Shenzhen
多模态
- The Power of Context: How Multimodality Improves Image Super-Resolution
- Paper: https://arxiv.org/abs/2503.14503
- Project: https://mmsr.kfmei.com/
- Blog: CVPR2025 | 当扩散模型遇见多模态上下文:Google&约翰霍普金斯大学用深度+语义+文本实现精准超分
- Features: 融合深度、分割、边缘、文本等多模态信息
- Team: Google; 约翰霍普金斯大学
盲超分 / 真实世界 / 移动端
- Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution
- Paper: https://arxiv.org/abs/2506.12738
- Code: https://github.com/xuhang07/Adpative-Dropout
- Keywords: 盲超分,正则化方法
- Team: 中科大
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution
- Paper: https://arxiv.org/abs/2411.18263
- Code: https://github.com/Microtreei/TSD-SR
- Keywords: One-Step Diffusion, Target Score Distillation, Real-World
- Team: 浙江大学,Vivo,中国科学院大学
- Adversarial Diffusion Compression for Real-World Image Super-Resolution
- Paper: https://arxiv.org/abs/2411.13383
- Code: https://github.com/Guaishou74851/AdcSR
- Keywords: Adversarial Diffusion, Compression, Real-World Super-Resolution
- Features: 在A100 GPU上仅需0.03秒🚀即可将128×128分辨率的图像超分辨率提升至512×512
- Team: 陈杰、Zhang Lei 团队 (北大、香港理工大学,OPPO研究院)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning
- Paper: https://arxiv.org/abs/2412.06978
- Code: https://github.com/Microtreei/TSD-SR
- Keywords: On-device, Low Latency, Parameter Efficient, Bidirectional Conditioning
- Team: 三星,伦敦玛丽皇后大学
数据增强/感知
- ADD: A General Attribution-Driven Data Augmentation Framework for Boosting Image Super-Resolution
- Paper: openaccess
- Code: https://github.com/mizeyu/ADD
- Keywords: Data Augmentation
- Team: 南京大学
- Augmenting Perceptual Super-Resolution via Image Quality Predictors
- Paper: https://arxiv.org/abs/2504.18524
- Keywords: Perceptual Super-Resolution, Image Quality Predictors
- Features: 运用强大的非参考图像质量评估(NR-IQA)模型
- Team: 三星电子
- Auto-Encoded Supervision for Perceptual Image Super-Resolution
- Paper: https://arxiv.org/abs/2412.00124
- Code: https://github.com/2minkyulee/AESOP-Auto-Encoded-Supervision-for-Perceptual-Image-Super-Resolution
- Keywords: Auto-Encoded, Perceptual Super-Resolution
- Features: 像素级 Lp 损失
- Team: 三星电子
- Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment
- Paper: https://arxiv.org/abs/2503.19295
- Code: https://github.com/GuangluDong0728/SFD
- Keywords: Semantic Feature Discrimination, Perceptual Super-Resolution, Quality Assessment
- Team: 四川大学
其他方法
- AutoLUT: LUT-Based Image Super-Resolution with Automatic Sampling and Adaptive Residual Learning
- Paper: https://arxiv.org/abs/2503.01565
- Code: https://github.com/SuperKenVery/AutoLUT
- Keywords: LUT-Based, Automatic Sampling, Adaptive Residual Learning
- Features: 查找表(LUT)
- Team: 南大
- QMambaBSR: Burst Image Super-Resolution with Query State Space Model
- Paper: https://arxiv.org/abs/2408.08665
- Keywords: Burst, Mamba
- Team: 中国科学技术大学;华为诺亚方舟实验室
视频超分
- BF-STVSR: B-Splines and Fourier-Best Friends for High Fidelity Spatial-Temporal Video Super-Resolution
-
Keywords: B-Splines, Fourier, Spatial-Temporal, Video Super-Resolution
-
Team: 蔚山国立科学技术院(UNIST),韩国大学
- Efficient Video Super-Resolution for Real-time Rendering with Decoupled G-buffer Guidance
- Paper: openaccess
- Code: https://github.com/sunny2109/RDG
- Keywords: 实时渲染
- Team: 南京理工大学潘金山团队
- EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events (Highlight)
- Paper: https://arxiv.org/abs/2505.04657
- Code: https://github.com/W-Shuoyan/EvEnhancer
- Blog: CVPR 2025 | DORNet:一种面向降质和正则化的盲深度超分辨率网络
- Keywords: 连续时空视频超分辨率
- Team: 北京交通大学,合肥工业大学
- Event-based Video Super-Resolution via State Space Models
- Paper: openaccess
- Keywords: Event-based, Mamba
- Team: 新加坡国立大学
- PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution
- Paper: openaccess
- Keywords: Diffusion
- Team: 清华,快手,北京理工大学
- Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution
- Paper: https://arxiv.org/abs/2506.01037
- Code: https://github.com/ssj9596/SCST
- Keywords: Self-supervised, ControlNet, Spatio-Temporal Mamba, Video Super-Resolution
- Team: 江南大学,中国科学技术大学,北京大学,中国科学院
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution
- Paper: https://arxiv.org/abs/2404.12388
- Project: https://videogigagan.github.io
- Keywords: GigaGAN
- Team: 马里兰大学,Adobe
其他场景
- Decoupling Fine Detail and Global Geometry for Compressed Depth Map Super-Resolution
- Paper: https://arxiv.org/abs/2411.03239
- Code: https://github.com/Ian0926/GDNet
- Keywords: Depth Map, Super-Resolution, Fine Detail, Global Geometry
- Team: 澳门大学
- DifIISR: Diffusion Model with Gradient Guidance for Infrared Image Super-Resolution
- Paper: https://arxiv.org/abs/2503.01187
- Code: https://github.com/zirui0625/DifIISR
- Blog: 论文阅读笔记:DifIISR:具有梯度引导的红外图像超分辨率扩散模型 [CVPR 2025]
- Keywords: Diffusion Model,红外图像
- Team: 大连理工大学,西北工业大学,早稻田大学,大连海事大学
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution
- Paper: https://arxiv.org/abs/2410.11666
- Code: https://github.com/yanzq95/dornet
- Blog: CVPR 2025 | DORNet:一种面向降质和正则化的盲深度超分辨率网络
- Keywords: 盲深度超分辨率网络
- Team: 南京理工大学,南京邮电大学,南京大学
- Latent Space Super-Resolution for Higher-Resolution Image Generation with Diffusion Models
- Paper: https://arxiv.org/abs/2503.18446
- Code: https://github.com/3587jjh/LSRNA
- Features: 实现了 16 倍超分辨率(4096×4096)图像生成
- Keywords: Latent Space, Higher-Resolution, Image Generation
- Team: Yonsei University
- PIDSR: Complementary Polarized Image Demosaicing and Super-Resolution
- Paper: https://arxiv.org/abs/2504.07758
- Code: https://github.com/PRIS-CV/PIDSR
- Blog: 【CVPR 2025】偏振光学相关论文记录
- Keywords: 偏振相机,去马赛克
- Team: 北京邮电大学,日本国立信息学研究所,北京大学
- S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting
- Paper: https://arxiv.org/abs/2503.04314
- Project: https://jeasco.github.io/S2Gaussian/
- Keywords: 3D
- Team: 中国石油大学(华东),哈尔滨工业大学
- Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution
- Paper: https://arxiv.org/abs/2503.02261
- Keywords: 3D荧光显微镜成像
- Team: 香港城市大学,香港浸会大学等
总结
从本届接收的论文来看,CVPR 2025 超分辨率领域呈现以下几个明显趋势:
-
扩散模型持续主导:基于 Stable Diffusion 的方法仍然是主流,包括 FaithDiff、PiSA-SR、TSD-SR 等,这些方法在感知质量和真实感方面表现出色。
-
任意尺度超分成为热点:DiffFNO、HIIF 等方法在任意尺度超分辨率方面取得突破,支持连续倍率放大,特别在大倍率(8x、12x)下表现优异。
-
效率与质量的平衡:轻量级模型(CATANet)和快速推理方法(DiffFNO 的自适应 ODE 求解器)受到关注,旨在解决实际应用中的部署问题。
-
可调节性和交互性:PiSA-SR 等方法支持用户根据偏好调节像素级保真度和语义级细节,体现了个性化需求。
-
跨模态和语义理解:越来越多的方法结合语义信息和多模态特征,提升超分辨率结果的语义一致性和感知质量。
-
真实世界应用导向:针对真实世界退化、移动端部署、视频超分等实际应用场景的研究显著增加。
总体而言,CVPR 2025 超分辨率研究在保持高质量重建的同时,更加注重实用性、效率和用户体验,为该领域的进一步发展奠定了坚实基础。
参考资料
(注:文档部分内容由 AI 生成)
更多推荐

所有评论(0)