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

AAAI 2026（The 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

  1. Mixture of Ranks with Degradation-Aware Routing for One-Step Real-World Image Super-Resolution

  • Paper: https://arxiv.org/abs/2511.16024

  • Code: 暂未检索到

  • Keywords: One-Step Real-World SR, Mixture of Ranks, Degradation-Aware Routing, LoRA

  • Features: 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 Lab

  1. Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution
  1. Realism Control One-step Diffusion for Real-world Image Super Resolution
  1. Selective Diffusion Distillation for Real-World High-Scale Image Super-Resolution

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/38351/42313

  • Code: 暂未检索到

  • Keywords: High-Scale SR, Diffusion Distillation, Real-UltraSR, x8/x10/x12/x14

  • Blog: 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

  1. HarmoQ: Harmonized Post-Training Quantization for High-Fidelity Image Super-Resolution

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/37944/41906

  • Code: https://github.com/Dreamzz5/HarmoQ

  • Keywords: Post-Training Quantization, Weight-Activation Coupling, Image SR, Edge Deployment

  • Features: 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

  1. Event-Guided Scene Text Image Super-Resolution
  1. PortraitSR: Artist-Inspired Prior Learning for Progressive Face Super-Resolution

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/37964/41926

  • Code: https://github.com/amazingwmq/PortraitSR

  • Keywords: Face SR, Progressive Prior, Sketching Structure Prior, Associative Texture Prior

  • Features: 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 Technology

  1. Seeing Through the Rain: Resolving High-Frequency Conflicts in Deraining and Super-Resolution via Diffusion Guidance

  • Paper: https://arxiv.org/abs/2511.12419

  • Code: https://github.com/PRIS-CV/DHGM

  • Keywords: Deraining, Super-Resolution, Diffusion Guidance, High-Frequency Conflict

  • Features: 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

  1. Spatio-Temporal Distortion Aware Omnidirectional Video Super-Resolution
  1. QuantVSR: Low-Bit Post-Training Quantization for Real-World Video Super-Resolution

  • Paper: https://arxiv.org/abs/2508.04485

  • Code: https://github.com/bowenchai/QuantVSR

  • Keywords: Real-World VSR, Post-Training Quantization, Low-Bit Diffusion, STCA, LBA

  • Blog: 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 Informatics

  1. MambaOVSR: Multiscale Fusion with Global Motion Modeling for Chinese Opera Video Super-Resolution

  • Paper: https://arxiv.org/abs/2511.06172

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

  • Keywords: Opera Video SR, Mamba, Global Motion Modeling, Multiscale Fusion

  • Features: 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 Technology

  1. Exploiting Blurry Representations for Event-guided Video Super-Resolution
  1. Temporal Inconsistency Guidance for Super-resolution Video Quality Assessment

  • Paper: https://arxiv.org/abs/2412.18933

  • Code: https://github.com/Lighting-YXLI/TIG-SVQA-main

  • Keywords: SR Video Quality Assessment, Temporal Inconsistency, VQA, Human Perception

  • Features: 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

  1. CD-DPE: Dual-Prompt Expert Network Based on Convolutional Dictionary Feature Decoupling for Multi-Contrast MRI Super-Resolution

  • Paper: https://arxiv.org/abs/2511.14014

  • Code: https://github.com/xianming-gu/CD-DPE

  • Keywords: Multi-Contrast MRI SR, Dual-Prompt Expert, Convolutional Dictionary, Feature Decoupling

  • Features: 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 University

  1. PINGS-X: Physics-Informed Normalized Gaussian Splatting with Axes Alignment for Efficient Super-Resolution of 4D Flow MRI

  • Paper: https://arxiv.org/abs/2511.11048

  • Code: https://github.com/SpatialAILab/PINGS-X

  • Keywords: 4D Flow MRI, Physics-Informed, Gaussian Splatting, Spatiotemporal Flow Field

  • Features: 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

  1. GEWDiff: Geometric Enhanced Wavelet-based Diffusion Model for Hyperspectral Image Super-resolution

  • Paper: https://arxiv.org/abs/2511.07103

  • Code: https://github.com/zhu-xlab/GEWDiff

  • Keywords: Hyperspectral Image SR, Wavelet, Diffusion Model, Geometry Preservation

  • Features: 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 Barcelona

  1. TRT: Harnessing Tensor Ring Transformer for Hyperspectral Image Super-Resolution

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/38103/42065

  • Code: 暂未检索到

  • Keywords: Hyperspectral Image SR, Deep Unfolding, Tensor Ring Transformer, Multilinear Product

  • Features: 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 University

  1. Thermal-Physics Guided Infrared Image Super-Resolution with Dynamic High-Frequency Amplification

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/38381/42343

  • Code: 暂未检索到

  • Keywords: Infrared Image SR, Thermal Physics, Dynamic High-Frequency Amplification, InfraredSR

  • Features: 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.

  1. HATIR: Heat-Aware Diffusion for Turbulent Infrared Video Super-Resolution

  • Paper: https://arxiv.org/abs/2601.04682

  • Code: https://github.com/JZ0606/HATIR

  • Keywords: Infrared Video SR, Atmospheric Turbulence, Heat-Aware Diffusion, FLIR-IVSR

  • Blog: 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 University

  1. MFmamba: A Multi-function Network for Panchromatic Image Resolution Restoration Based on State-Space Model

  • Paper: https://arxiv.org/abs/2511.18888

  • Code: https://github.com/QianqianWang1325/MFmamba.git

  • Keywords: Panchromatic Image, Resolution Restoration, Mamba, Joint SR and Colorization

  • Features: 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 Technology

3D / Gaussian Splatting / 点云

  1. IE-SRGS: An Internal-External Knowledge Fusion Framework for High-Fidelity 3D Gaussian Splatting Super-Resolution
  1. SRSplat: Feed-Forward Super-Resolution Gaussian Splatting from Sparse Multi-View Images
  1. Arbitrary-Scale 3D Gaussian Super-Resolution
  1. PUFM: Efficient Point Cloud Upsampling via Flow Matching

  • Paper: https://ojs.aaai.org/index.php/AAAI/article/download/37685/41647

  • Code: https://github.com/Holmes-Alan/PUFM

  • Keywords: Point Cloud Upsampling, Flow Matching, 3D Geometry

  • Features: 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 Virginia

  1. PUNO: A Neural Operator Framework for Point Cloud Upsampling

深度 / 事件 / 音频 / 科学计算

  1. SpatioTemporal Difference Network for Video Depth Super-Resolution

  • Paper: https://arxiv.org/abs/2508.01259

  • Code: https://github.com/yanzq95/STDNet

  • Keywords: Video Depth SR, Spatial Difference, Temporal Difference, Long-Tailed Distribution

  • Features: 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 Singapore

  1. Ultralight Polarity-Split Neuromorphic SNN for Event-Stream Super-Resolution

  • Paper: https://arxiv.org/abs/2508.03244

  • Code: 暂未检索到

  • Keywords: Event-Stream SR, Spiking Neural Network, Polarity Split, Ultralight Model

  • Features: 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 Sydney

  1. Inference-time Scaling for Diffusion-based Audio Super-resolution

  • Paper: https://arxiv.org/abs/2508.02391

  • Code: 暂未检索到

  • Keywords: Audio Super-Resolution, Diffusion Model, Inference-Time Scaling, Test-Time Compute

  • Features: 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 Kong

  1. Multimodal Super-Resolution: Discovering Hidden Physics and Its Application to Fusion Plasmas (Abstract Reprint)

  • Paper: https://arxiv.org/abs/2405.05908

  • Code: 暂未检索到

  • Keywords: Multimodal SR, Hidden Physics, Fusion Plasma, Scientific Machine Learning

  • Features: 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：

  1. 扩散模型继续深入真实世界 SR：RCOD、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”。

  2. 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。

  3. 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。

  4. 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。

  5. 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

  1. AAAI 2026 Proceedings

  2. AAAI OJS Proceedings: Vol. 40

  3. arXiv

  4. GitHub

（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。）

©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