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

AAAI 2026目录AAAI 2026图像超分真实世界 / 扩散模型 / 高倍率量化 / 高效部署特定图像场景视频超分医学影像超分遥感 / 高光谱 / 红外超分3D / Gaussian Splatting / 点云深度 / 事件 / 音频 / 科学计算总结参考资料AAAI 2026The Fortieth AAAI Conference on Artificial Intelligence于 2026 年 1 月 20 日至 27 日在新加坡举行。超分辨率Super-Resolution, SR旨在从低分辨率、低采样率或低质量观测中恢复高分辨率结果。AAAI 2026 中SR 相关论文覆盖图像、视频、医学影像、遥感高光谱/红外、3D Gaussian Splatting、点云、音频与科学计算等多个场景。现将本届 AAAI 2026 超分辨率方向论文汇总如下遗漏之处还请大家斧正。图像超分真实世界 / 扩散模型 / 高倍率Mixture of Ranks with Degradation-Aware Routing for One-Step Real-World Image Super-ResolutionPaper: https://arxiv.org/abs/2511.16024Code: 暂未检索到Keywords: One-Step Real-World SR, Mixture of Ranks, Degradation-Aware Routing, LoRAFeatures: 针对一步真实世界图像超分引入退化感知路由与不同秩的专家混合让模型按退化复杂度动态选择恢复能力兼顾效率与真实感。Blog: 暂未检索到Team: Xidian University; Huawei Noah’s Ark LabContinuous Degradation Modeling via Latent Flow Matching for Real-World Super-ResolutionPaper: https://arxiv.org/abs/2602.04193Code: https://github.com/present091/DegFlowKeywords: Real-World SR, Latent Flow Matching, Continuous Degradation, Synthetic LR GenerationFeatures: 提出 DegFlow从单张 HR 图像出发在潜在退化空间中连续建模真实退化可合成不同尺度和退化强度的 LR 图像用于训练传统和任意尺度 SR 模型。Blog: ResearchTrend.AI 解读Team: Hanyang UniversityRealism Control One-step Diffusion for Real-world Image Super ResolutionPaper: https://arxiv.org/abs/2509.10122Project: https://zongliang-wu.github.io/RCOD-SR/Code: https://github.com/Zongliang-Wu/RCODKeywords: One-Step Diffusion, Real-ISR, Fidelity-Realism Trade-off, Visual PromptBlog: ChatPaper 中文解读Features: 通过 latent domain grouping 和 degradation-aware sampling 控制一步扩散 SR 的保真度/真实感权衡并用视觉提示替代文本提示增强语义一致性。Team: Zhejiang University; Westlake University; vivo Mobile Communication Co., LtdSelective Diffusion Distillation for Real-World High-Scale Image Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/38351/42313Code: 暂未检索到Keywords: High-Scale SR, Diffusion Distillation, Real-UltraSR, x8/x10/x12/x14Blog: ChatPaper 解读Features: 面向真实世界高倍率超分提出 SDD 框架从低倍率扩散教师向高倍率学生蒸馏可靠知识并构建 Real-UltraSR 高倍率真实世界基准。Team: Beijing University of Posts and Telecommunications; Stony Brook University量化 / 高效部署HarmoQ: Harmonized Post-Training Quantization for High-Fidelity Image Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/37944/41906Code: https://github.com/Dreamzz5/HarmoQKeywords: Post-Training Quantization, Weight-Activation Coupling, Image SR, Edge DeploymentFeatures: 系统分析权重量化与激活量化对 SR 的不同影响通过结构残差校准、尺度优化与边界细化实现低比特 SR 部署。Blog: 暂未检索到Team: The University of Tokyo; Southern University of Science and Technology; The Hong Kong Polytechnic University; Jilin University特定图像场景Event-Guided Scene Text Image Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/37801/41763Code: https://github.com/codes81/EVTSRKeywords: Scene Text SR, Event Camera, Dual-Stream Frequency Boost, Text Recognition PriorFeatures: 首个将事件数据引入场景文字图像超分的框架 EvTSR利用事件高频边缘信息恢复笔画并结合文本识别先验保证字符一致性。Blog: 暂未检索到Team: Hefei University of Technology; National University of SingaporePortraitSR: Artist-Inspired Prior Learning for Progressive Face Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/37964/41926Code: https://github.com/amazingwmq/PortraitSRKeywords: Face SR, Progressive Prior, Sketching Structure Prior, Associative Texture PriorFeatures: 借鉴“先结构、后细节”的艺术绘制过程通过结构先验、纹理字典先验与整体先验融合提升大倍率人脸结构一致性和纹理真实感。Blog: 暂未检索到Team: Chongqing University of Posts and Telecommunications; Guangyang Bay Laboratory; Nanjing University of Science and TechnologySeeing Through the Rain: Resolving High-Frequency Conflicts in Deraining and Super-Resolution via Diffusion GuidancePaper: https://arxiv.org/abs/2511.12419Code: https://github.com/PRIS-CV/DHGMKeywords: Deraining, Super-Resolution, Diffusion Guidance, High-Frequency ConflictFeatures: 研究去雨与超分之间的高频冲突提出 DHGM 利用扩散先验与高通滤波同时去除雨纹并增强结构细节面向小目标检测等下游任务。Blog: 暂未检索到Team: Beijing University of Posts and Telecommunications; Nankai University; The University of Tokyo视频超分Spatio-Temporal Distortion Aware Omnidirectional Video Super-ResolutionPaper: https://arxiv.org/abs/2410.11506Code: https://github.com/nichenxingmeng/STDANKeywords: Omnidirectional Video, 360 Video, ODV-SR, Spatio-Temporal Distortion, Latitude-SaliencyBlog: The Moonlight 中文解读Features: 面向 360 度全景视频提出 STDAN联合处理 ERP 投影空间畸变和时间闪烁并构建 ODV-SR 数据集。Team: University of the Chinese Academy of Sciences; ByteDance Inc.; Peking UniversityQuantVSR: Low-Bit Post-Training Quantization for Real-World Video Super-ResolutionPaper: https://arxiv.org/abs/2508.04485Code: https://github.com/bowenchai/QuantVSRKeywords: Real-World VSR, Post-Training Quantization, Low-Bit Diffusion, STCA, LBABlog: Emergent Mind 解读Features: 针对扩散式真实世界 VSR 的低比特部署提出时空复杂度感知分配与可学习偏置对齐在低比特下保持接近全精度模型的恢复质量。Team: Shanghai Jiao Tong University; Joy Future Academy; Max Planck Institute for InformaticsMambaOVSR: Multiscale Fusion with Global Motion Modeling for Chinese Opera Video Super-ResolutionPaper: https://arxiv.org/abs/2511.06172Code: https://github.com/ChangHua0/MambaOVSRKeywords: Opera Video SR, Mamba, Global Motion Modeling, Multiscale FusionFeatures: 面向中国戏曲视频的低清/复杂运动退化利用 Mamba 建模长程运动并进行多尺度融合强调传统戏曲视频修复场景。Blog: 暂未检索到Team: Wuhan University of Science and Technology; Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System; Harbin Institute of TechnologyExploiting Blurry Representations for Event-guided Video Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/38081/42043Code: 暂未检索到Keywords: Event-guided VSR, Blurry Representation, Motion Saliency, RGB-Event FusionBlog: ChatPaper 中文解读Features: 提出 BluR-EVSR用事件流和相邻帧自监督学习模糊退化表示在视频超分中同时处理运动模糊和分辨率退化。Team: National University of SingaporeTemporal Inconsistency Guidance for Super-resolution Video Quality AssessmentPaper: https://arxiv.org/abs/2412.18933Code: https://github.com/Lighting-YXLI/TIG-SVQA-mainKeywords: SR Video Quality Assessment, Temporal Inconsistency, VQA, Human PerceptionFeatures: 不是生成式 VSR而是面向 SR 视频质量评价显式建模超分带来的时间不一致与闪烁伪影用于更准确评价 SR 视频感知质量。Blog: 暂未检索到Team: Beihang University; Jiangxi University of Finance and Economics; Eastern Institute of Technology; Dalian University of Technology; Cardiff University医学影像超分CD-DPE: Dual-Prompt Expert Network Based on Convolutional Dictionary Feature Decoupling for Multi-Contrast MRI Super-ResolutionPaper: https://arxiv.org/abs/2511.14014Code: https://github.com/xianming-gu/CD-DPEKeywords: Multi-Contrast MRI SR, Dual-Prompt Expert, Convolutional Dictionary, Feature DecouplingFeatures: 面向多对比度 MRI 超分利用卷积字典特征解耦区分跨对比度共享结构与对比度特有信息降低参考图像纹理误导。Blog: 暂未检索到Team: Guizhou University; Maastricht UniversityPINGS-X: Physics-Informed Normalized Gaussian Splatting with Axes Alignment for Efficient Super-Resolution of 4D Flow MRIPaper: https://arxiv.org/abs/2511.11048Code: https://github.com/SpatialAILab/PINGS-XKeywords: 4D Flow MRI, Physics-Informed, Gaussian Splatting, Spatiotemporal Flow FieldFeatures: 将 4D 时空血流速度场表示为轴对齐 Gaussian并结合物理约束和 Gaussian merging大幅减少逐患者训练时间。Blog: 暂未检索到Team: Hanyang University; Nanyang Technological University遥感 / 高光谱 / 红外超分GEWDiff: Geometric Enhanced Wavelet-based Diffusion Model for Hyperspectral Image Super-resolutionPaper: https://arxiv.org/abs/2511.07103Code: https://github.com/zhu-xlab/GEWDiffKeywords: Hyperspectral Image SR, Wavelet, Diffusion Model, Geometry PreservationFeatures: 采用小波编码器压缩高光谱数据并保留光谱-空间信息结合几何增强扩散过程和多级损失实现 4 倍高光谱图像超分。Blog: 暂未检索到Team: Technical University of Munich; Munich Center for Machine Learning; Universitat Autonoma de BarcelonaTRT: Harnessing Tensor Ring Transformer for Hyperspectral Image Super-ResolutionPaper: https://ojs.aaai.org/index.php/AAAI/article/download/38103/42065Code: 暂未检索到Keywords: Hyperspectral Image SR, Deep Unfolding, Tensor Ring Transformer, Multilinear ProductFeatures: 将深度展开网络与 Tensor Ring Transformer 结合用张量环多线性积替代传统注意力点积建模高光谱数据的高维结构先验。Blog: 暂未检索到Team: Taizhou University; Zhejiang University of Technology; Xiamen University; Zhejiang Wanli UniversityThermal-Physics Guided Infrared Image Super-Resolution with Dynamic High-Frequency AmplificationPaper: https://ojs.aaai.org/index.php/AAAI/article/download/38381/42343Code: 暂未检索到Keywords: Infrared Image SR, Thermal Physics, Dynamic High-Frequency Amplification, InfraredSRFeatures: 提出 ThesIS利用热辐射物理约束和动态高频增强同时保持红外热分布准确性和视觉纹理细节并构建 InfraredSR 数据集。Blog: 暂未检索到Team: Beijing Institute of Technology; Beihang University; Iray Technology Co., Ltd.HATIR: Heat-Aware Diffusion for Turbulent Infrared Video Super-ResolutionPaper: https://arxiv.org/abs/2601.04682Code: https://github.com/JZ0606/HATIRKeywords: Infrared Video SR, Atmospheric Turbulence, Heat-Aware Diffusion, FLIR-IVSRBlog: CatalyzeX 解读Features: 将热感知形变先验注入扩散采样过程联合建模湍流退化和结构细节恢复并构建首个湍流红外 VSR 数据集 FLIR-IVSR。Team: Northwestern Polytechnical University; Dalian University of Technology; Zhejiang University; Dalian Maritime UniversityMFmamba: A Multi-function Network for Panchromatic Image Resolution Restoration Based on State-Space ModelPaper: https://arxiv.org/abs/2511.18888Code: https://github.com/QianqianWang1325/MFmamba.gitKeywords: Panchromatic Image, Resolution Restoration, Mamba, Joint SR and ColorizationFeatures: 标题不含 Super-Resolution归入其他相关方向模型可在 PAN 图像 SR、光谱恢复、联合 SR光谱恢复三种输入设置下工作。Blog: 暂未检索到Team: Yunnan University; Wroclaw University of Science and Technology3D / Gaussian Splatting / 点云IE-SRGS: An Internal-External Knowledge Fusion Framework for High-Fidelity 3D Gaussian Splatting Super-ResolutionPaper: https://arxiv.org/abs/2511.22233Code: https://github.com/ChrisShuo/IE-SRGS检索到仓库当前内容较少Keywords: 3D Gaussian Splatting, Super-Resolution, Internal-External Knowledge, 2D SR PriorBlog: Hugging Face Papers 摘要Features: 融合外部 2D SR/深度估计先验与内部多尺度 3DGS 特征缓解 2D SR 跨视角不一致和 3D Gaussian 歧义。Team: Hangzhou Dianzi University; ShanghaiTech University; Griffith UniversitySRSplat: Feed-Forward Super-Resolution Gaussian Splatting from Sparse Multi-View ImagesPaper: https://arxiv.org/abs/2511.12040Project: https://xinyuanhu66.github.io/SRSplat/Code: https://github.com/XinyuanHu66/SRSplat_CodeKeywords: Feed-Forward 3DGS, Sparse Multi-View, Reference-Guided Feature Enhancement, Texture-Aware Density ControlBlog: alphaXiv 摘要Features: 从稀疏低分辨率多视图前馈重建高分辨率 3D 场景使用 MLLM/扩散生成场景参考图并结合内部纹理密度控制。Team: Hangzhou Dianzi University; Peking University; Li Auto Inc.Arbitrary-Scale 3D Gaussian Super-ResolutionPaper: https://arxiv.org/abs/2508.16467Project: https://huimin-zeng.github.io/3DASR/Code: https://github.com/huimin-zeng/Arbi-3DGSRKeywords: Arbitrary-Scale, 3DGS, Scale-Aware Rendering, Generative Prior, Progressive Super-ResolvingBlog: Paper Notes 解读Features: 单个 3DGS 模型支持整数和非整数任意倍率 HR 渲染通过尺度感知渲染、生成先验约束和渐进式训练保持实时速度。Team: Northeastern UniversityPUFM: Efficient Point Cloud Upsampling via Flow MatchingPaper: https://ojs.aaai.org/index.php/AAAI/article/download/37685/41647Code: https://github.com/Holmes-Alan/PUFMKeywords: Point Cloud Upsampling, Flow Matching, 3D GeometryFeatures: 标题不含 Super-Resolution归入其他相关方向用 flow matching 建模点云上采样从稀疏点云生成更稠密的高质量点云。Blog: 暂未检索到Team: Lappeenranta-Lahti University of Technology; The Hong Kong Polytechnic University; University of Leicester; Technical University of Munich; University of VirginiaPUNO: A Neural Operator Framework for Point Cloud UpsamplingPaper: https://ojs.aaai.org/index.php/AAAI/article/download/38082/42044Code: https://github.com/Rangiant5b72/PUNOKeywords: Point Cloud Upsampling, Neural Operator, Continuous GeometryFeatures: 标题不含 Super-Resolution归入其他相关方向用神经算子框架处理点云上采样强调连续几何映射和泛化能力。Blog: 暂未检索到Team: Southeast University深度 / 事件 / 音频 / 科学计算SpatioTemporal Difference Network for Video Depth Super-ResolutionPaper: https://arxiv.org/abs/2508.01259Code: https://github.com/yanzq95/STDNetKeywords: Video Depth SR, Spatial Difference, Temporal Difference, Long-Tailed DistributionFeatures: 针对视频深度超分中的空间非平滑区域和时间变化区域长尾问题设计空间差异分支和时间差异分支进行 RGB-D 聚合与运动补偿。Blog: 暂未检索到Team: Nanjing University of Science and Technology; Nankai University; National University of SingaporeUltralight Polarity-Split Neuromorphic SNN for Event-Stream Super-ResolutionPaper: https://arxiv.org/abs/2508.03244Code: 暂未检索到Keywords: Event-Stream SR, Spiking Neural Network, Polarity Split, Ultralight ModelFeatures: 面向事件流超分设计极性分离的轻量级 SNN利用事件相机正负极性变化实现更高分辨率事件流恢复。Blog: 暂未检索到Team: The University of SydneyInference-time Scaling for Diffusion-based Audio Super-resolutionPaper: https://arxiv.org/abs/2508.02391Code: 暂未检索到Keywords: Audio Super-Resolution, Diffusion Model, Inference-Time Scaling, Test-Time ComputeFeatures: 将 inference-time scaling 引入扩散式音频超分在推理阶段通过更多采样/搜索计算提升音频带宽扩展与细节恢复质量。Blog: 暂未检索到Team: Hong Kong University of Science and Technology; Meta AI; The Chinese University of Hong KongMultimodal Super-Resolution: Discovering Hidden Physics and Its Application to Fusion Plasmas (Abstract Reprint)Paper: https://arxiv.org/abs/2405.05908Code: 暂未检索到Keywords: Multimodal SR, Hidden Physics, Fusion Plasma, Scientific Machine LearningFeatures: AAAI Journal Track 摘要重印面向等离子体物理中的多模态超分通过不同诊断模态发现隐藏物理并提升时空分辨率。Blog: 暂未检索到Team: Princeton University; Princeton Plasma Physics Laboratory; Chung-Ang University; Columbia University; Seoul National University总结从 AAAI 2026 接收论文来看超分辨率方向呈现以下趋势扩散模型继续深入真实世界 SRRCOD、SDD、DegFlow、HATIR 等工作围绕一步扩散、蒸馏、退化建模和物理先验展开重点从“生成得好”转向“可控、可部署、面向真实退化”。场景化任务明显增多文字、人脸、全景视频、戏曲视频、去雨、红外、医学 MRI、4D flow MRI、等离子体等方向都出现专门设计说明 SR 正在从通用图像恢复走向任务和传感器定制。效率和部署成为核心议题QuantVSR 与 HarmoQ 分别从视频和图像角度研究低比特量化RCOD 等一步扩散方法也强调推理速度和可调节性。3D 与多视角超分升温IE-SRGS、SRSplat、Arbi-3DGSR 将 SR 从 2D 图像拓展到 3DGS 表示、稀疏多视图和任意尺度渲染适合 AR/VR、机器人与低带宽传输。物理先验与模态特性更受重视红外热物理、4D flow MRI 血流物理、事件相机高频边缘、点云连续几何等先验被显式纳入模型设计。总体而言AAAI 2026 的 SR 研究不再局限于 PSNR/SSIM 导向的单图像重建而是更关注真实退化、跨模态信息、可控生成、低成本部署以及 3D/科学计算等实际应用场景。参考资料AAAI 2026 ProceedingsAAAI OJS Proceedings: Vol. 40arXivGitHub注文档部分内容由 AI 辅助整理论文链接、PDF、作者单位以 AAAI OJS 与 arXiv 检索结果为准。