CVPR 2026 超分辨率（super-resolution）方向上接收论文总结

CVPR 2026

CVPR 2026 于 2026 年 6 月 3 日至 7 日在美国丹佛 Colorado Convention Center 举行，其中主会时间为 6 月 5 日至 7 日。

据公开统计，CVPR 2026 共有 16092 篇有效投稿进入评审流程，约 4090 篇论文被接收，录用率约 25.42%。本文件基于 CVPR 2026 Open Access 中的主会论文列表，汇总标题或内容与超分辨率（Super-Resolution, SR）相关的论文。

现将超分辨率方向上接收的论文汇总如下。

图像超分

扩散模型 / 生成式真实世界超分

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

  • Paper: https://arxiv.org/abs/2603.16769

  • Code: https://github.com/Joyies/GDPO

  • Keywords: One-Step Generative ISR, Group Direct Preference Optimization, Reinforcement Learning

  • Features: 将强化学习引入单步生成式图像超分，提出 noise-aware one-step diffusion 和局部细节感知的 GDPO 训练策略

  • Team: 香港理工大学，OPPO Research Institute

  1. HDW-SR: High-Frequency Guided Diffusion Model based on Wavelet Decomposition for Image Super-Resolution
  1. DVAR: Dynamic Visual Autoregressive Modeling for Image Super-Resolution

  • Paper: openaccess

  • Code: https://github.com/YuZheng9/DVAR

  • Keywords: Visual Autoregressive, Dynamic Scaling, Arbitrary Target Size

  • Features: 学习相对尺度动态而非固定尺度表，单模型适配不同目标分辨率

  • Team: 南京理工大学，南京大学

  1. VOSR: A Vision-Only Generative Model for Image Super-Resolution

  • Paper: https://arxiv.org/abs/2604.03225

  • Code: https://github.com/cswry/VOSR

  • Keywords: Vision-Only, Generative SR, Uni-modal Diffusion

  • Features: 不依赖文本先验和大规模图文预训练，使用视觉编码器注入语义引导以完成纯视觉 SR

  • Team: 香港理工大学，OPPO Research Institute

  1. FRAMER: Frequency-Aligned Self-Distillation with Adaptive Modulation Leveraging Diffusion Priors for Real-World Image Super-Resolution
  1. One-Step Diffusion Transformer for Controllable Real-World Image Super-Resolution

  • Paper: https://arxiv.org/abs/2511.17138

  • Code: https://github.com/RedMediaTech/ODTSR

  • Keywords: One-Step Diffusion Transformer, Controllable Real-ISR, Qwen-Image, Prompt Control

  • Features: 提出 ODTSR，通过 Noise-hybrid Visual Stream 和 Fidelity-aware Adversarial Training 实现单步、可控真实世界超分

  • Team: 上海交通大学，小红书

  1. DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer

  • Paper: https://arxiv.org/abs/2605.15682

  • Keywords: Ultra-High-Resolution, Diffusion Transformer, Receptive Field Enhancement

  • Features: 面向超高分辨率图像恢复，增强扩散 Transformer 的感受野以缓解高分辨率切块和长程依赖问题

  • Team: ByteDance Inc.

  1. IFCSR: Inference-Free Fidelity-Realism Control for One-Step Diffusion-based Real-World Image Super-Resolution

  • Paper: openaccess

  • Keywords: Inference-Free Control, Fidelity-Realism Trade-off, One-Step Diffusion

  • Features: 在一次生成后通过参数控制保真度和真实感，无需重复网络推理

  • Team: BLUEDOT Inc.

  1. Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution

  • Paper: https://arxiv.org/abs/2508.16557

  • Code: https://github.com/zty557/TADSR

  • Keywords: Time-Aware, One-Step Diffusion, Real-ISR, VSD

  • Features: 引入 Time-Aware VAE Encoder，使学生模型能利用 Stable Diffusion 不同时间步的生成先验

  • Team: 南开大学 VCIP，vivo Mobile Communication，NKIARI Shenzhen Futian

  1. DNF-SR: Dual-Input and Negative-Aware Feature Fine-Tuning for Real-World Image Super-Resolution

  • Paper: openaccess

  • Keywords: Dual-Input, Negative-Aware Fine-Tuning, Real-ISR, Diffusion Editing

  • Features: 同时输入原始 LR 和加噪 LR，降低 LR latent 与扩散模型输入分布的差异

  • Team: 南开大学 VCIP，NKIARI Shenzhen Futian，Independent Researcher

  1. Bridging Fidelity-Reality with Controllable One-Step Diffusion for Image Super-Resolution

  • Paper: https://arxiv.org/abs/2512.14061

  • Code: https://github.com/Chanson94/CODSR

  • Keywords: Controllable One-Step Diffusion, Fidelity-Reality Trade-off, Region-Adaptive Prior

  • Features: 提出 CODSR，通过 LQ-guided feature modulation 和区域自适应生成先验激活改善保真度与真实感平衡

  • Team: 南京理工大学

  1. FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution

  • Paper: https://arxiv.org/abs/2603.02692

  • Code: https://github.com/Ar0Kim/FiDeSR

  • Keywords: One-Step Diffusion, Detail-Preserving, Latent Residual Refinement

  • Features: 训练阶段使用 detail-aware weighting，推理阶段使用低/高频自适应增强器和 latent residual refinement

  • Team: Kyungpook National University，Korea Electronics Technology Institute

  1. Disentangled Textual Priors for Diffusion-based Image Super-Resolution

  • Paper: https://arxiv.org/abs/2603.07430

  • Code: https://github.com/JL6666JL/DTPSR

  • Keywords: Textual Priors, Diffusion SR, Spatial Hierarchy, Frequency Semantics

  • Features: 将文本先验拆分为全局/局部空间层级和低频/高频语义，提高可控性与可解释性

  • Team: 南京大学软件新技术全国重点实验室

  1. TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

  • Paper: openaccess

  • Code: https://github.com/wuer5/TUDSR

  • Keywords: Twice Upsampling-Diffusion, Higher SR, Looped Chunk-based Training

  • Features: 两阶段训练：先在 R 分辨率训练，再引入 NR 分辨率的循环分块训练以支持更高倍率/更高分辨率输出

  • Team: 华东师范大学，浙江大学

任意尺度 / 大倍率超分

  1. CASR: A Robust Cyclic Framework for Arbitrary Large-Scale Super-Resolution with Distribution Alignment and Self-Similarity Awareness

  • Paper: https://arxiv.org/abs/2602.22159

  • Keywords: Arbitrary-Scale SR, Cyclic SR, Distribution Alignment, Self-Similarity

  • Features: 将超大倍率放大拆解为一系列分布内尺度转换，使用 SSAM 和 SARM 抑制迭代漂移与 patch 不一致

  • Team: 北京邮电大学，北京大学

  1. Adaptive Anisotropic Gaussian Splatting for Multi-contrast MRI Arbitrary-Scale Super-Resolution with Anatomy Guidance

  • Paper: openaccess

  • Code: https://github.com/Qiuhai-CV/GaussM2ASR

  • Keywords: Arbitrary-Scale SR, MRI, Gaussian Splatting, Anatomy Guidance

  • Features: 将 MRI 高频解剖结构重建转化为各向异性 2D Gaussian kernel 参数优化问题

  • Team: 华东师范大学

轻量化 / 量化 / 高效模型

  1. Gradient Knows Best: Mixed-Precision Quantization via Gradient-Guided Bit Allocation for Super-Resolution

  • Paper: openaccess

  • Keywords: Mixed-Precision Quantization, Gradient-Guided Bit Allocation, Lightweight SR

  • Features: 用梯度引导层级 bit-width 分配，面向去 BN 的 SR 网络解决激活尺度不一致问题

  • Team: Dongguk University，Sogang University

  1. UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution

  • Paper: https://arxiv.org/abs/2603.11680

  • Code: https://github.com/hokiyoshi/UCAN

  • Keywords: Lightweight SR, Convolutional Attention, Expansive Receptive Field

  • Features: 统一卷积注意力结构，扩大轻量级 SR 模型有效感受野

  • Team: Ho Chi Minh City Open University，Ho Chi Minh City University of Technology，Code Mely AI Research Team，Shanghai Jiao Tong University，Ha Noi University of Science and Technology，University of Manitoba

  1. IAFMNet: Information-Aware Feature Modulation for Efficient Super-Resolution

  • Paper: openaccess

  • Keywords: Efficient SR, Information Density Map, Feature Modulation

  • Features: 估计 Information Density Map，将更多计算和注意力分配给纹理、边缘等高信息区域

  • Team: 西安电子科技大学

感知质量 / 评价 / 奖励模型

  1. FinPercep-RM: A Fine-grained Reward Model and Co-evolutionary Curriculum for RL-based Real-world Super-Resolution
  1. Bridging the Perception Gap in Image Super-Resolution Evaluation

  • Paper: https://arxiv.org/abs/2503.13074

  • Project: https://color.cvc.uab.cat/rqi/

  • Keywords: SR Evaluation, Human Preference, IQA Metrics, Perception Gap

  • Features: 系统分析 SR 评价指标与人类偏好的相关性，并提出面向 SR 评价可靠性的研究框架

  • Team: Computer Vision Center，Universitat Autonoma de Barcelona，INSAIT Sofia University

特定图像内容 / 采集设备

  1. Restore Text First, Enhance Image Later: Two-Stage Scene Text Image Super-Resolution with Glyph Structure Guidance

  • Paper: https://arxiv.org/abs/2510.21590

  • Keywords: Scene Text Image Super-Resolution, Glyph Structure Guidance, Two-Stage SR

  • Features: 先恢复文本字形再增强图像整体质量，缓解自然图像 SR 方法对文字结构的扭曲

  • Team: 南开大学 VCIP，vivo Mobile Communication，香港中文大学（深圳）

  1. RAW-Domain Degradation Models for Realistic Smartphone Super-Resolution

  • Paper: https://arxiv.org/abs/2603.12493

  • Keywords: Smartphone SR, RAW Domain, Realistic Degradation Modeling

  • Features: 从 RAW 域建模智能手机真实退化，更贴近移动摄影成像链路

  • Team: AI Center-Toronto, Samsung Electronics

  1. Edge-Focused Super-Resolution for Omnidirectional Images with Spherical Geometric Augmentation

  • Paper: openaccess

  • Keywords: Omnidirectional Image SR, Edge-Focused Block, Spherical Geometric Augmentation

  • Features: 面向全景图投影畸变和大倍率放大，结合边缘增强模块与球面几何增强

  • Team: 西南大学，University of Edinburgh，Hiroshima City University

视频超分

  1. Time Without Time: Pseudo-Temporal Representation for Space-Time Super-Resolution

  • Paper: openaccess

  • Keywords: Space-Time Video Super-Resolution, Pseudo-Temporal Pretraining, Image Dataset Pretraining

  • Features: 将单张图像复制为伪时序视频并做随机区域重建，用图像数据预训练 STVSR 网络

  • Team: Samsung，Seoul National University

  1. TextOVSR: Text-Guided Real-World Opera Video Super-Resolution

  • Paper: https://arxiv.org/abs/2603.15153

  • Code: https://github.com/ChangHua0/TextOVSR

  • Keywords: Opera Video SR, Text-Guided, Real-World VSR, CLIP

  • Features: 面向老旧戏曲视频，利用文本语义引导真实世界视频超分，降低复杂真实退化与纹理恢复难度

  • Team: 武汉科技大学，哈尔滨工业大学郑州研究院，广东省人工智能与数字经济实验室（深圳），华为

  1. Thermal Diffusion Matters: Infrared Spatial-Temporal Video Super-Resolution through Heat Conduction Priors

  • Paper: openaccess

  • Keywords: Infrared Video SR, Spatial-Temporal SR, Heat Conduction Priors

  • Features: 提出 THERIS，将红外像素强度时空演化建模为热扩散/热传导过程

  • Team: 北京理工大学，北京航空航天大学，Imperial College London，艾睿光电

  1. DUO-VSR: Dual-Stream Distillation for One-Step Video Super-Resolution
  1. DTG-Restore: Training-Free Diffusion Refinement for Generative Video Super-Resolution

  • Paper: openaccess

  • Keywords: Training-Free, Diffusion Refinement, Decoupled Time Guidance, Generative VSR

  • Features: 提出 Decoupled Time Guidance，在无需训练的情况下解耦条件/无条件分支以改善低清视频恢复

  • Team: Virginia Tech，Adobe

  1. VSRELL: A Simple Baseline for Video Super-Resolution and Enhancement in Low-Light Environment

  • Paper: openaccess

  • Keywords: Low-Light VSR, Video Enhancement, Simple Baseline

  • Features: 将低照度增强和视频超分统一建模，面向文物表面监测等低光场景

  • Team: 天津大学，SMARC

  1. STCDiT: Spatio-Temporally Consistent Diffusion Transformer for High-Quality Video Super-Resolution
  1. Rethinking Diffusion Model-Based Video Super-Resolution: Leveraging Dense Guidance from Aligned Features

  • Paper: https://arxiv.org/abs/2511.16928

  • Code: https://github.com/tszssong/DGAF-VSR

  • Keywords: Diffusion VSR, Dense Guidance, Aligned Features, Feature Compensation

  • Features: 重新审视 VSR 中对齐与补偿作用，使用特征域密集引导提升扩散模型保真度

  • Team: 北京航空航天大学，阿里巴巴淘宝天猫集团，虚拟现实技术与系统全国重点实验室

  1. Compressed-Domain-Aware Online Video Super-Resolution
  1. FlashVSR: Towards Real-time Diffusion-Based Streaming Video Super Resolution
  1. PS-SR: Pseudo-Single-Step Video Super-Resolution via Speculative Diffusion

  • Paper: openaccess

  • Keywords: Pseudo-Single-Step VSR, Speculative Diffusion, Frequency-Domain Update

  • Features: 强基模型先做单步全局生成，再由轻量 draft model 做高频推测式细化，兼顾速度与细节

  • Team: 中国科学技术大学，HiDream.ai Inc.

其他场景

  1. Dual Graph Regularized Deep Unfolding Network for Guided Depth Map Super-resolution

  • Paper: openaccess

  • Keywords: Guided Depth Map SR, Deep Unfolding, Dual Graph Laplacian Prior

  • Features: 将双图拉普拉斯先验的图优化展开为可解释深度网络 LapNet

  • Team: City University of Hong Kong，vivo BlueImage Lab

  1. Spectral Super-Resolution via Adversarial Unfolding and Data-Driven Spectrum Regularization: From Multispectral Satellite Data to NASA Hyperspectral Image

  • Paper: https://arxiv.org/abs/2603.00920

  • Code: https://github.com/IHCLab/UALNet

  • Keywords: Spectral Super-Resolution, Multispectral-to-Hyperspectral, Deep Unfolding, Remote Sensing

  • Features: 将 Sentinel-2 多光谱数据重建为 NASA AVIRIS-NG 风格高光谱图像，并统一空间分辨率

  • Team: National Cheng Kung University

  1. Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset

  • Paper: https://arxiv.org/abs/2603.04745

  • Code: https://github.com/JZD151/Real-IISR

  • Keywords: Infrared Image SR, Real-World Benchmark, Autoregressive Framework, FLIR-IISR

  • Features: 构建真实红外图像超分数据集 FLIR-IISR，并用热-结构引导的视觉自回归逐尺度恢复

  • Team: 西北工业大学，大连理工大学，浙江大学，大连海事大学

  1. EMR-Diff: Edge-aware Multimodal Residual Diffusion Model for Hyperspectral Image Super-resolution

  • Paper: openaccess

  • Keywords: Hyperspectral Image SR, Multimodal Residual Diffusion, Edge-Aware

  • Features: 面向高光谱超分的边缘感知多模态残差扩散模型

  • Team: 杭州电子科技大学，丽水学院，北京理工大学，浙江省低空泛在网络技术重点实验室

  1. Enhancing Unregistered Hyperspectral Image Super-Resolution via Unmixing-based Abundance Fusion Learning

  • Paper: https://arxiv.org/abs/2603.07918

  • Keywords: Hyperspectral Image SR, Unregistered Fusion, Spectral Unmixing, Abundance Map

  • Features: 通过解混 abundance fusion 缓解参考图像未配准对 HSI SR 的影响

  • Team: 北京理工大学，杭州电子科技大学

  1. Remote Sensing Image Super-Resolution for Imbalanced Textures: A Texture-Aware Diffusion Framework
  1. Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction

  • Paper: https://arxiv.org/abs/2603.00149

  • Code: https://github.com/lizhihao2022/ReMD

  • Keywords: Fluid SR, Physics-Consistent Diffusion, Multiscale Residual Correction, Multi-Wavelet

  • Features: 提出 ReMD，将数据一致性与轻量物理约束结合，通过多尺度残差校正提升流体场超分的谱保真度

  • Team: 香港科技大学（广州），香港科技大学，南方科技大学，SandGold AI Research

  1. VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution

  • Paper: openaccess

  • Keywords: Volumetric SR, Dataset, Domain Shift, Real Low-Resolution Scans

  • Features: 构建真实低分辨率/高分辨率体数据配对数据集，揭示下采样训练与真实低清扫描之间的域偏移

  • Team: Technical University of Denmark

  1. SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian Splatting

  • Paper: https://arxiv.org/abs/2512.02172

  • Project: https://splatsure.github.io

  • Keywords: 3D Gaussian Splatting, Multi-view Consistency, Selective Super-Resolution

  • Features: 根据相机位姿与场景几何选择性添加 SR 内容，避免逐视角独立 SR 带来的多视角不一致

  • Team: University of Maryland, College Park

  1. SR3R: Rethinking Super-Resolution 3D Reconstruction With Feed-Forward Gaussian Splatting

  • Paper: https://arxiv.org/abs/2602.24020

  • Project: https://xiangfeng66.github.io/SR3R/

  • Keywords: 3D Super-Resolution, Feed-Forward 3DGS, Sparse LR Views

  • Features: 将 3D 超分重建建模为从稀疏 LR 视角到 HR 3DGS 的前馈预测，减少单场景优化依赖

  • Team: 杭州电子科技大学，上海科技大学，Griffith University，北京大学

其他

  1. Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers

  • Paper: https://arxiv.org/abs/2507.08422

  • Keywords: Latent Upsampling, Diffusion Transformer, Spatial Acceleration, Training-Free

  • Features: 提出 Region-Adaptive Latent Upsampling，通过混合分辨率 latent 上采样加速 DiT 生成

  • Team: Seoul National University

  1. NAF: Zero-Shot Feature Upsampling via Neighborhood Attention Filtering

  • Paper: https://arxiv.org/abs/2511.18452

  • Code: https://github.com/valeoai/NAF

  • Keywords: Feature Upsampling, Zero-Shot, Neighborhood Attention Filtering, Vision Foundation Models

  • Features: 对视觉基础模型的低分辨率特征做零样本上采样，服务分割、深度估计、开放词汇等密集任务

  • Team: Valeo.ai，Sorbonne Universite，CNRS，ISIR，Institut Universitaire de France

  1. UPLiFT: Efficient Pixel-Dense Feature Upsampling with Local Attenders

  • Paper: https://arxiv.org/abs/2601.17950

  • Keywords: Feature Upsampling, Pixel-Dense Prediction, Local Attenders

  • Features: 面向密集预测的高效特征上采样，用局部 attenders 恢复像素级表示

  • Team: University of Maryland, College Park，Meta

  1. Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling

  • Paper: https://arxiv.org/abs/2511.16301

  • Keywords: Feature Upsampling, Test-Time Optimization, Anisotropic Gaussian Kernel, Joint Bilateral Upsampling

  • Features: 无训练的每图优化框架，用各向异性 Gaussian kernel 将低分辨率特征恢复为高分辨率像素级输出

  • Team: KAIST ，MIT ，Microsoft

总结

从 CVPR 2026 接收论文来看，超分辨率方向呈现以下趋势：

  1. 单步扩散和高效生成式 SR 成为主线：GDPO-SR、ODTSR、TADSR、FiDeSR、CODSR、IFCSR 等都围绕单步扩散、可控生成和真实世界 SR 展开。

  2. 视频超分强调实时性与扩散蒸馏：DUO-VSR、FlashVSR、PS-SR、STCDiT、DGAF-VSR 等集中解决扩散 VSR 的速度、时序一致性和高分辨率推理问题。

  3. 可控性与感知质量继续强化：文本先验、奖励模型、保真度-真实感控制、评价指标可靠性成为图像 SR 的重要研究点。

  4. 任意尺度和超高分辨率仍是重点：CASR、DreamSR、TUDSR、GaussM2ASR 等尝试在任意尺度、大倍率和超高分辨率输出上保持稳定质量。

  5. 应用场景明显扩展：医学 MRI、红外、遥感、高光谱、深度图、流体场、3D Gaussian Splatting、体数据等场景均有专门方法或数据集。

  6. 轻量化与部署问题受到持续关注：UCAN、IAFMNet、混合精度量化、压缩域在线 VSR 等工作都指向实际部署中的速度、显存和算力约束。

参考资料

  1. CVPR 2026 Official Website

  2. CVPR 2026 Dates and Deadlines

  3. CVPR 2026 Open Access

  4. Kobaayyy - CVPR 2026 Low-Level Vision Papers

  5. CVPR 2026 Statistics - Paper Copilot

  6. CVPR 2026 Accepted Papers: Trends, Big Tech Bets & Top Highlights

（注：文档内容由公开论文页面、arXiv、CVF Open Access、项目页/代码仓库信息整理，部分代码链接来自论文摘要或公开汇总的交叉检索，由Codex自动生成。）

©著作权归作者所有,转载或内容合作请联系作者
【社区内容提示】社区部分内容疑似由AI辅助生成，浏览时请结合常识与多方信息审慎甄别。
平台声明：文章内容（如有图片或视频亦包括在内）由作者上传并发布，文章内容仅代表作者本人观点，简书系信息发布平台，仅提供信息存储服务。

相关阅读更多精彩内容

友情链接更多精彩内容