CVPR 2026 Papers — Page 33
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models
Yifei Zhao (University of Central Florida), Mengxin Zheng (University of Central Florida)
Safty and PrivacyTransformerVision Language ModelImageText
🎯 What it does: Propose a non-intrusive LVLM fingerprint identification framework SIF, which utilizes text watermarks to migrate into image inputs and detects fingerprints during black-box inference.
SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models
Sofian Chaybouti (Technology Innovation Institute), Hakim Hacid (Technology Innovation Institute)
ClassificationSegmentationRetrievalKnowledge DistillationMixture of ExpertsVision Language ModelImageMultimodality
🎯 What it does: Propose SigLino, an aggregated visual foundation model learned through multi-teacher distillation.
SIGMA: A Physics-Based Benchmark for Gas Chimney Understanding in Seismic Images
Bao Truong, Anh Nguyen (University of Arkansas)
Object DetectionSegmentationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImageBenchmarkPhysics Related
🎯 What it does: Proposed and made publicly available a physics-based synthetic seismic dataset named SIGMA for gas chimney detection and enhancement, conducting benchmark experiments on this dataset.
SIGMA: Selective-Interleaved Generation with Multi-Attribute Tokens
Xiaoyan Zhang (Creatly AI), Yiren Song (National University of Singapore)
GenerationTransformerDiffusion modelMultimodality
🎯 What it does: Propose the SIGMA unified framework, which supports multi-attribute cross-insertion in text and image conditions, enabling controllable multi-source image generation.
SignPR: A Progressive Vector-Quantized Diffusion Framework for Sign Language Production
Xiao Liu (Nanjing University), Sanglu Lu (Nanjing University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelAuto EncoderVideoTextSequential
🎯 What it does: Propose the SignPR framework, which generates text-to-sign language pose sequences using structured VQ-Diffusion and block-level causal reasoning.
Similarity-as-Evidence: Calibrating Overconfident VLMs for Interpretable and Label-Efficient Medical Active Learning
Zhuofan Xie (Xiamen University), Shuo Li (Case Western Reserve University)
Explainability and InterpretabilityData-Centric LearningTransformerPrompt EngineeringVision Language ModelBiomedical Data
🎯 What it does: Propose the Similarity-as-Evidence (SaE) framework, converting text-image similarity from Vision-Language Models (VLM) into Dirichlet evidence for medical image active learning, addressing VLM overconfidence and cold-start problems.
Similarity-Consistent Likelihood Diffusion enables Hidden Person Detection from Wall Reflections
Zhiwen Zheng (Hangzhou Dianzi University), Xingru Huang (Hangzhou Dianzi University)
RestorationObject DetectionDiffusion modelImage
🎯 What it does: Proposes the Similarity-Likelihood Diffusion Network (SLD-Net) for reconstructing and detecting hidden human figures under corner cameras by utilizing multi-exposure wall reflection differential images.
SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images
Aayush Dhakal (Washington University in St. Louis), Nathan Jacobs (Washington University in St. Louis)
Anomaly DetectionRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Propose the SimLBR scheme, which learns a compact decision boundary for real images by mixing a small amount of pseudo-fake information with real images in the DINOv3 latent space;
SIMPACT: Simulation-Enabled Action Planning using Vision-Language Models
Haowen Liu (University of Maryland), Yilun Du (Harvard)
SegmentationPose EstimationDepth EstimationOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision Language ModelVision-Language-Action ModelWorld ModelImageTextPoint Cloud
🎯 What it does: This paper proposes SIMPACT, a VLM action planning framework based on physics simulation during testing, which can accomplish fine-grained, physics-aware robotic manipulation without task-specific training.
Simple Agents Outperform Experts in Biomedical Imaging Workflow Optimization
Xuefei Wang, Jennifer J. Sun
OptimizationLarge Language ModelAgentic AIBiomedical Data
🎯 What it does: Studied the use of minimized AI agents to automatically adapt existing biomedical imaging tool code, constructed an evaluation framework, and verified its effectiveness surpassing human experts in three major cases (Polaris, Cellpose, MedSAM).
Simple but Effective Triplet-Based Compression Strategies for Compact Visual Localization
Torsten Sattler (Czech Technical University in Prague), Zuzana Kukelova (Czech Technical University in Prague)
Pose EstimationCompressionSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a structure compression method based on three-point matching, which evaluates the camera pose of database images using a P3P solver to select subsets of 3D points that ensure high-precision pose estimation, achieving memory compression for visual localization.
Simple-ViLMedSAM: Simple Text Prompts Meet Vision-Language Models for Medical Image Segmentation
Chengcan Qian (Nanjing University of Aeronautics and Astronautics), Xuyun Wen (Nanjing University of Aeronautics and Astronautics)
SegmentationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageBiomedical DataComputed Tomography
🎯 What it does: Propose a medical image segmentation framework called Simple-ViLMedSAM based on CLIP and SAM, which can perform zero-shot and few-shot segmentation using only simple text labels.
SIMPLEPOSTER: A SIMPLE BASELINE FOR PRODUCT POSTER GENERATION
Benlei Cui (Alibaba Group), Pipei Huang (Taobao & Tmall Group of Alibaba)
GenerationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes a product poster generation method called SimplePoster based on a simple inpainting framework.
SimRecon: SimReady Compositional Scene Reconstruction from Real Videos
Chong Xia (Tsinghua University), Yueqi Duan (Tsinghua University)
GenerationData SynthesisOptimizationVision Language ModelDiffusion modelNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Propose the SimRecon framework to achieve object-centric 3D scene reconstruction from chaotic real-world videos into physics-interactable environments within simulators.
SimScale: Learning to Drive via Real-World Simulation at Scale
Haochen Tian (University of Chinese Academy of Sciences), Hongyang Li (University of Hong Kong)
Domain AdaptationAutonomous DrivingGaussian Splatting
🎯 What it does: This study proposes the SimScale system, which leverages large-scale, scalable real-world simulation data to enhance the robustness and generalization of end-to-end autonomous driving models.
SIMSPINE: A Biomechanics-Aware Simulation Framework for 3D Spine Motion Annotation and Benchmarking
Muhammad Saif Ullah Khan (German Research Center for Artificial Intelligence), Didier Stricker (German Research Center for Artificial Intelligence)
Data SynthesisPose EstimationConvolutional Neural NetworkTransformerVideoBiomedical DataBenchmark
🎯 What it does: Designed a biomechanics-based keypoint simulation framework, utilizing models such as OpenSim to generate 3D spinal keypoints from existing full-body motion data and creating the SIMSPINE dataset.
SineProject: Machine Unlearning for Stable Vision-Language Alignment
Arpit Garg (Adelaide University), Simon Lucey (Adelaide University)
Representation LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed and validated a sine-modulated projector called SINEPROJECT, which maintains visual-language alignment in multi-modal large models during zero-shot (forgetting) scenarios, preventing the model from incorrectly rejecting harmless queries alongside harmful ones.
SinGeo: Unlock Single Model's Potential for Robust Cross-View Geo-Localization
Yang Chen (National University of Defense Technology), Tao Wu (National University of Defense Technology)
RetrievalConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposes the SinGeo framework, achieving robust cross-view geolocation with a single model across different perspectives and FoV through Dual Discriminative Learning and Curriculum Learning.
Single-Round Scalable Analytic Federated Learning
Alan T. L. Bacellar (University of Texas at Austin), Lizy K. John (University of Texas at Austin)
Federated LearningImage
🎯 What it does: Propose the SAFLe framework, utilizing bucketing and sparse embeddings with a nonlinear head to achieve one-round analytical training in federated learning.
Single-step Diffusion-based Video Coding with Semantic-Temporal Guidance
Naifu Xue (Communication University Of China), Yan Lu (Microsoft Research Asia)
CompressionKnowledge DistillationConvolutional Neural NetworkDiffusion modelOptical FlowVideo
🎯 What it does: Propose a single-step diffusion video encoder named SVC2, combining conditional compression with single-step diffusion generation to achieve high perceptual quality video compression at low bitrates.
SIR: Structured Image Representations for Explainable Robot Learning
Paul Mattes (Karlsruhe Institute of Technology), Rudolf Lioutikov (Karlsruhe Institute of Technology)
Explainability and InterpretabilityRepresentation LearningRobotic IntelligenceGraph Neural NetworkTransformerVision-Language-Action ModelImagePoint Cloud
🎯 What it does: Proposes the SIR method, which converts images into scene graphs (Scene Graph), generates task-related subgraphs through learnable sparsification, and inputs the subgraph as an intermediate state into the GCIL policy, thereby achieving interpretable robot learning.
SJD-PAC: Accelerating Speculative Jacobi Decoding via Proactive Drafting and Adaptive Continuation
Jialiang Kang (Peking University), Xinghao Chen (Huawei Technologies)
GenerationComputational EfficiencyVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Designed and implemented a training-free, lossless Speculative Jacobi Decoding acceleration framework, SJD-PAC, for efficient inference in text-to-image (T2I) models.
SkeletonContext: Skeleton-side Context Prompt Learning for Zero-Shot Skeleton-based Action Recognition
Ning Wang (Chang'an University), Liang Zhang (Donghai Lab)
RecognitionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: This paper proposes a framework called SkeletonContext, which enhances the zero-shot recognition performance of skeletal actions using a language-driven context prompting method.
Sketch2Colab: Sketch-Conditioned Multi-Human Animation via Controllable Flow Distillation
Divyanshu Daiya (Purdue University), Aniket Bera (Purdue University)
GenerationData SynthesisDiffusion modelRectified FlowAuto EncoderImageVideoOrdinary Differential Equation
🎯 What it does: Propose the Sketch2Colab framework, which uses sketches to control the generation of multi-agent human-object interaction animations, combining efficient rectified-flow sampling with stage planning using continuous-time Markov chains to precisely follow temporal, trajectory, and contact constraints.
Sketch2CT: Multimodal Diffusion for Structure-Aware 3D Medical Volume Generation
Delin An (University of Notre Dame), Chaoli Wang (University of Notre Dame)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelMultimodalityComputed Tomography
🎯 What it does: Propose a multimodal diffusion framework Sketch2CT that can generate structurally consistent 3D medical volumes based on user-drawn 2D sketches and textual descriptions.
SketchAssist: A Practical Assistant for Semantic Edits and Precise Local Redrawing
Han Zou (Baidu Inc.), Zhenpeng Zhan (Baidu Inc.)
GenerationTransformerPrompt EngineeringMixture of ExpertsVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Propose SketchAssist, a unified interactive sketch editing system that supports both text-based global semantic modifications and line-based local redrawing;
SketchDeco: Training-Free Latent Composition for Precise Sketch Colourisation
Chaitat Utintu (University of Surrey), Yi-Zhe Song (University of Surrey)
Image TranslationGenerationVision Language ModelDiffusion modelImageOrdinary Differential Equation
🎯 What it does: Propose a training-free sketch coloring framework called SketchDeco. First, at the global level, it generates a full-color image consistent with the sketch structure and the user-provided color palette using a pre-trained diffusion model and ControlNet. Subsequently, at the local level, it achieves precise color blending and smooth transitions in the latent space through diffusion inversion, Gaussian noise injection, and self-attention injection, ultimately producing high-quality color images that preserve sketch details and conform to the color palette.
SketchFaceGS: Real-Time Sketch-Driven Face Editing and Generation with Gaussian Splatting
Bo Li (Shandong Technology and Business University), Lin Gao (Institute Of Computing Technology Chinese Academy Of Sciences)
GenerationTransformerGenerative Adversarial NetworkGaussian SplattingImageMesh
🎯 What it does: Proposed a real-time sketch-driven facial generation and editing framework called SketchFaceGS based on 3D Gaussian Splatting, which can quickly generate and edit visual 3D facial models from a single hand-drawn sketch and reference image.
SketchRevive: Fine-Grained Pixel-to-Vector Sketch Completion with Diffusion-Prior-Guided Multimodal LLMs
Ran Zuo (Communication University of China), Yong-Jin Liu (Tsinghua University)
Image TranslationGenerationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageMultimodality
🎯 What it does: Propose a two-stage fine-grained pixel-to-vector sketch completion method called SketchRevive, which converts sparse pixel sketches into complete and editable SVG vector graphics.
SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More
Muye Huang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
OptimizationExplainability and InterpretabilityLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: This paper proposes SketchVL, a multimodal large language model based on the image reasoning paradigm. During inference, the model directly writes intermediate steps as drawing annotations on the image, then feeds the annotated image back into the model to achieve observable multi-step reasoning. The model enhances chart understanding ability by using the FinePO reinforcement learning framework combined with the FinePRM process reward model, which assigns fine-grained rewards to each drawing action.
SkillSight: Efficient First-Person Skill Assessment with Gaze
Chi Hsuan Wu (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
Computational EfficiencyKnowledge DistillationTransformerMultimodality
🎯 What it does: This paper proposes a skill assessment framework called SkillSight based on first-person eye movement information. It first trains a teacher model using video plus eye movement data, then obtains a student model that uses only eye movement through knowledge distillation, achieving low-power skill assessment.
Skullptor: High Fidelity 3D Head Reconstruction in Seconds with Multi-View Normal Prediction
Noé Artru (Ubisoft La Forge), Abdallah Dib (Ubisoft La Forge)
RestorationGenerationDepth EstimationTransformerNeural Radiance FieldAuto EncoderGaussian SplattingImageMeshBenchmark
🎯 What it does: This paper proposes a method that can recover a high-fidelity 3D head model from sparse multi-view images (fewer than 10 cameras) within 30 seconds;
Sky2Ground: A Benchmark for Site Modeling under Varying Altitude
Zengyan Wang (University of Central Florida), Yogesh Rawat (University of Central Florida)
Pose EstimationTransformerGaussian SplattingImageBenchmark
🎯 What it does: Propose the Sky2Ground dataset and SkyNet model for cross-view (satellite, aerial, ground) camera localization and 3D reconstruction.
Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning
Yifei Li (Tsinghua University), Jiwen Lu (Tsinghua University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes Skyra, a multimodal large language model specialized for AI-generated video detection, achieving signal-based explainable detection through the human-annotated ViF-CoT-4K dataset.
SkyReels-Text: Fine-Grained Font-Controllable Text Editing for Poster Design
Yunjie Yu (Skywork AI), Guibin Chen (Skywork AI)
RecognitionGenerationVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose SkyReels-Text, achieving fine-grained, font-controllable editing of multi-region text on posters.
SkySense-VITA: Towards Universal In-context Segmentation of Multi-modal Remote Sensing Imagery
Kang Wu (Wuhan University), Yansheng Li (Ant Group)
SegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose SkySense‑VITA, a unified in‑context segmentation model capable of simultaneously processing optical and SAR images, and supporting visual, textual, or fused prompts.
SLARM: Streaming and Language-Aligned Reconstruction Model for Dynamic Scenes
Zhicheng Qiu (Huawei Technologies Ltd), Zhan Xu (Huawei Technologies Ltd)
SegmentationGenerationDepth EstimationAutonomous DrivingComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerVision Language ModelGaussian SplattingVideoTextMultimodality
🎯 What it does: Propose SLARM, a unified 4D Gaussian reasoning model that can real-time infer the 3D geometry, scene flow, metric depth, and language-aligned semantic information of dynamic scenes without requiring optical flow annotations.
SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control
Arman Zarei (University of Maryland), Soheil Feizi (University of Maryland)
GenerationTransformerPrompt EngineeringImage
🎯 What it does: Propose the SliderEdit framework, which utilizes low-rank adapters to achieve continuous, fine-grained control over instruction-level edits, enabling direct sliding adjustment of the intensity of each editing instruction on instruction-based image editing models such as FLUX-Kontext and Qwen-Image-Edit.
SLVMEval: Synthetic Meta Evaluation Benchmark for Text-to-Long Video Generation
Ryosuke Matsuda (Tohoku University), Jun Suzuki (Tohoku University)
GenerationData SynthesisVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed the SLVMEval benchmark for meta-evaluation of text-to-long-video generation assessment systems.
Small Object, Great Challenge: A Benchmark for Small Object Visual Grounding
Wenqi Jia (Beijing University of Posts and Telecommunications), Xiaojie Wang (Beijing University of Posts and Telecommunications)
Object DetectionTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the Small Object Visual Grounding (SoVG) task and constructed the RefCOCOs dataset based on COCO
SMAP: Semantic Route Planning with Map-Grounded Multimodal Alignment
Wenjie Zhang (Xidian University), Pengbo Zhang (Amap, Alibaba Group)
Autonomous DrivingOptimizationTransformerLarge Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed SMAP, the first multimodal semantic route planning framework that combines user queries, POI metadata, and map tiles;
Smart Replay: Adaptive Scheduling of Memory Rehearsal for Computational Resource-Aware Incremental Learning
Jianting Chen (Tongji University), Irwin King (Chinese University of Hong Kong)
Computational EfficiencyImageBenchmark
🎯 What it does: Propose an algorithm called Smart Replay that dynamically adjusts the memory replay ratio in incremental learning scenarios with limited computational resources.
SmokeSVD: Smoke Reconstruction from A Single View via Progressive Novel View Synthesis and Refinement with Diffusion Models
Chen Li (Tianjin University of Technology), Kemeng Huang (University of Hong Kong)
GenerationConvolutional Neural NetworkDiffusion modelVideoPhysics Related
🎯 What it does: Proposes the SmokeSVD framework, which utilizes diffusion models to generate side views and progressively improves multi-view synthesis, enabling efficient reconstruction of 3D smoke density, velocity fields, and inflow states from single-view videos.
Smoothing the Score Function to Enhance Generalization in Diffusion Models
Xinyu Zhou (University of Wisconsin-Madison), Stephen J. Wright (University of Wisconsin-Madison)
GenerationDiffusion modelScore-based ModelImage
🎯 What it does: This paper investigates the memorization problem in diffusion models, proposing a theoretical framework that explains memorization as arising from the sharpness of the score function of the empirical distribution, and enhancing generalization through smoothing techniques.
SMRABooth: Subject and Motion Representation Alignment for Customized Video Generation
Xuancheng Xu (Nanjing University of Posts and Telecommunications), Bing-Kun Bao (Nanjing University of Posts and Telecommunications)
GenerationTransformerSupervised Fine-TuningDiffusion modelOptical FlowVideoText
🎯 What it does: Propose the SMRABooth framework to achieve personalized text-to-video generation, supporting separable control over subject appearance and motion trajectories, and capable of handling various motion types (linear, curved, rotational, conceptual motion).
SMV-EAR: Bring Spatiotemporal Multi-View Representation Learning into Efficient Event-Based Action Recognition
Rui Fan (Xidian University), Lin Gu (Tohoku University)
RecognitionComputational EfficiencyRepresentation LearningConvolutional Neural NetworkBenchmark
🎯 What it does: Propose a multi-view spatiotemporal representation learning framework SMV-EAR for event camera action recognition
SMVRT: Implicit Human 3D Modeling Using Sparse Multi-View Volumetric Reconstruction with Transformer Fusion
Chuanmao Fan (University of Missouri-Columbia), Ye Duan (Clemson University)
GenerationConvolutional Neural NetworkTransformerNeural Radiance FieldAuto EncoderImage
🎯 What it does: Propose SMVRT, an end-to-end implicit 3D human reconstruction framework specifically designed for sparse multi-view inputs.
SO-Bench: A Structural Output Evaluation of Multimodal LLM
Di Feng (Apple), Afshin Dehghan (Apple)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed SO-Bench to evaluate the ability of multimodal LLMs in visual structured output tasks and enhanced model performance through training.
SO(3)-Equivariant ViT-Adapter for Data-Efficient Zero-Shot Sim-to-Real Indoor Panoramic Depth Estimation
Ziyan He (Shenzhen University), Xu Wang (Shenzhen University)
Depth EstimationDomain AdaptationTransformerImage
🎯 What it does: Proposed an SO(3)-equivariant ViT-Adapter for data-efficient zero-shot indoor panoramic depth estimation, aiming to address the generalization problem of panoramic images in real-world scenarios.
SoC: Semantic Orthogonal Calibration for Test-Time Prompt Tuning
Leo Fillioux (Universite Paris-Saclay), Jose Dolz (LIVIA)
ClassificationExplainability and InterpretabilityTransformerPrompt EngineeringContrastive LearningImageText
🎯 What it does: To address the overconfidence issue in vision-language models (VLMs) during test-time prompt tuning (TPT), this paper proposes a semantic orthogonal calibration (SoC) regularization method based on Huber loss, which is validated on multiple datasets.
SoccerMaster: A Vision Foundation Model for Soccer Understanding
Haolin Yang (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
ClassificationObject DetectionObject TrackingRetrievalTransformerSupervised Fine-TuningVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: Propose SoccerMaster, a unified foundation model for football vision tasks, capable of simultaneously performing detection, tracking, field registration, event classification, and audio-visual alignment.
SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation
Ziyi Chen (Amap, Alibaba Group), Yu Zhang (Zhejiang University)
Explainability and InterpretabilityRobotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsFlow-based ModelImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposed a hierarchical foundation model called SocialNav, which integrates high-level semantic reasoning with low-level trajectory generation to achieve social, interpretable robot navigation.
Socratic-Geo: Synthetic Data Generation and Cross-Modal Geometric Reasoning via Multi-Agent Interaction
Zhengbo Jiao (Alibaba Group Holding Limited), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationData SynthesisLarge Language ModelReinforcement LearningAgentic AIDiffusion modelImageTextMultimodality
🎯 What it does: Built a self-driven multi-agent framework called Socratic-Geo, where the Teacher generates and verifies procedural geometry problems, the Solver performs reinforcement learning using GRPO on the generated training set, and the Generator learns high-quality geometric images from procedural instructions generated by the Teacher.
SODA: Sensitivity-Oriented Dynamic Acceleration for Diffusion Transformer
Tong Shao (Harbin Institute of Technology), Jingyong Su (Harbin Institute of Technology)
Computational EfficiencyTransformerDiffusion modelImageVideoBenchmark
🎯 What it does: Developed a dynamic acceleration framework called SODA based on fine-grained sensitivity-aware mechanisms, combining caching and pruning techniques to significantly improve inference efficiency while maintaining or enhancing the generation quality of Diffusion Transformers.
Soft Modality-Guided Expert Specialization in MoE-VLMs
Zi-Hao Bo (Li Auto Inc), Kaiwen Long
Computational EfficiencyMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes Soft Modality-Guided Expert Specialization (SMoES), achieving modality-guided expert differentiation in MoE-VLM through dynamic soft modality scores, expert binning, and mutual information regularization;
SoliReward: Mitigating Susceptibility to Reward Hacking and Annotation Noise in Video Generation Reward Models
Jiesong Lian (Huazhong University of Science and Technology), Junchi Yan (Shanghai Jiao Tong University)
GenerationReinforcement LearningVision Language ModelVideo
🎯 What it does: This paper proposes the SoliReward framework for training video reward models (RM), addressing issues such as label noise, reward hacking, and insufficient model expressiveness.
Solvability of the Viewing Graph Under the Affine Camera Model
Gabriele Pedroni (Politecnico di Milano), Federica Arrigoni (Politecnico di Milano)
Graph
🎯 What it does: This paper studies the solvability of the viewing graph under the affine camera model, and provides its linear system formulation and decision algorithm;
Solving a Nonlinear Blind Inverse Problem for Tagged MRI with Physics and Deep Generative Priors
Zhangxing Bian (Johns Hopkins University), Jerry L Prince (Johns Hopkins School of Medicine)
RestorationSuper ResolutionDiffusion modelBiomedical DataMagnetic Resonance ImagingPhysics Related
🎯 What it does: Propose a blind nonlinear inverse problem framework named InvTag, which combines the MR physics model with diffusion generative models to simultaneously recover high-resolution anatomical images, unlabeled cine images, three-dimensional Lagrangian motion fields, and the imaging point spread function from low-resolution 3D tagged MRI.
Solving Minimal Problems Without Matrix Inversion Using FFT-Based Interpolation
Haidong Wu (University of Oulu), Janne Heikkila (University of Oulu)
Optimization
🎯 What it does: Propose an FFT-based interpolation sampling method for constructing a minimal problem solver that does not require matrix inversion.
SonoWorld: From One Image to a 3D Audio-Visual Scene
Derong Jin (University of Maryland, College Park), Ruohan Gao (University of Maryland, College Park)
GenerationData SynthesisVision Language ModelGaussian SplattingImageVideoMultimodalityAudio
🎯 What it does: Propose a method to generate a freely navigable 3D audio-visual scene from a single image, integrating image, semantics, and spatial audio to construct a 3D audio-visual world that can be rendered in real-time.
SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs
Koonting Yip (University of Macau), Ka-Veng Yuen (University of Macau)
Representation LearningTransformerVision Language ModelPoint CloudBenchmark
🎯 What it does: Propose a position information encoding method called SoPE based on spherical coordinates to enhance the spatial perception capabilities of 3D large-scale audio-visual language models.
SOTA: Self-adaptive Optimal Transport for Zero-Shot Classification with Multiple Foundation Models
Zhanxuan Hu (Yunnan Normal University), Huafeng Li (Kunming University of Science and Technology)
ClassificationVision Language ModelImageMultimodalityBiomedical DataBenchmark
🎯 What it does: This paper proposes a training-agnostic adaptive optimal transport framework called SOTA, which can fuse outputs from various base models (VLMs and VFM) to enhance zero-shot classification performance.
Soul: Breathe Life into Digital Human for High-fidelity Long-term Multimodal Animation
Jiangning Zhang (Zhejiang University), Chengjie Wang (Youtu Lab, Tencent)
GenerationData SynthesisTransformerVision-Language-Action ModelDiffusion modelAuto EncoderImageTextMultimodalityAudio
🎯 What it does: Proposes the Soul framework, achieving high-fidelity long-term multimodal digital human animation generation based on single-frame images, text, and audio.
SounDiT: Geo-Contextual Soundscape-to-Landscape Generation
Junbo Wang (University of Texas at Austin), Yuhao Kang (University of Texas at Austin)
GenerationTransformerMixture of ExpertsDiffusion modelImageMultimodalityBenchmarkAudio
🎯 What it does: Proposed the GeoS2L task of generating geographic scene images from environmental soundscapes, and designed the SounDiT model and PSS evaluation framework based on it.
SOUPLE: Enhancing Audio-Visual Localization and Segmentation with Learnable Prompt Contexts
Khanh Binh Nguyen (Deakin University), Chae Jung Park (National Cancer Center)
SegmentationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningMultimodalityAudio
🎯 What it does: Propose SOUPLE, a prompt learning-based audio-visual source localization and segmentation framework, replacing fixed prompts with instance-conditional, learnable context tokens to enhance semantic correspondence between audio and visual modalities.
Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping
Junmyeong Lee (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
GenerationTransformerGaussian SplattingVideo
🎯 What it does: Propose a framework named MoGaF that utilizes 4D Gaussian Splatting for long-term prediction in dynamic scenes, achieving high-fidelity rendering of future frames through the construction of motion-aware Gaussian groups.
SpaceDrive: Infusing Spatial Awareness into VLM-based Autonomous Driving
Peizheng Li (Mercedes-Benz AG), Andreas Zell (University of Tbingen)
Depth EstimationAutonomous DrivingTransformerLarge Language ModelVision Language ModelImagePoint Cloud
🎯 What it does: Propose SpaceDrive, an end-to-end autonomous driving framework that integrates 3D spatial encoding into VLM, directly replacing digital tokens with a unified 3D position encoding for regression prediction;
SpaceMind: Camera-Guided Modality Fusion for Spatial Reasoning in Vision-Language Models
Ruosen Zhao, Zizhuang Wei (Huawei)
TransformerImageTextBenchmark
🎯 What it does: Propose a camera-guided modality fusion method called SpaceMind, significantly enhancing the performance of VLM in 3D spatial reasoning tasks.
SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time
Zhening Huang (University of Cambridge), Chun-Hao Huang (Adobe Research)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelVideo
🎯 What it does: Developed a video diffusion model called SpaceTimePilot, which can generate continuous and coherent multi-view videos from a single-camera video, independently controlling the camera perspective (space) and motion sequence (time).
SpaceTools: Tool-Augmented Spatial Reasoning via Double Interactive RL
Siyi Chen (University of Michigan), Jonathan Tremblay (NVIDIA)
Robotic IntelligenceSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelImageMultimodalityBenchmark
🎯 What it does: Propose a dual-interaction reinforcement learning (DIRL) framework and tool platform Toolshed, enabling Vision-Language Models (VLMs) to collaborate with multiple tools to accomplish precise spatial reasoning and robotic grasping tasks.
SPAN: Spatial-Projection Alignment for Monocular 3D Object Detection
Yifan Wang (Tsinghua University), Wenming Yang (Tsinghua University)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: Propose a pluggable geometric consistency constraint framework named SPAN for monocular 3D detectors, which simultaneously considers spatial point alignment and 3D-2D projection alignment.
SPAR: Single-Pass Any-Resolution ViT for Open-vocabulary Segmentation
Naomi Kombol (University of Zagreb), Giorgos Tolias (Czech Technical University in Prague)
SegmentationKnowledge DistillationTransformerImage
🎯 What it does: Leverage teacher-student distillation to transfer high-resolution ViT inference capabilities from a sliding-window-based approach to a single-channel ViT without architectural modifications, enabling efficient dense feature extraction at arbitrary resolutions for open-vocabulary segmentation.
SPARK: Sim-ready Part-level Articulated Reconstruction with VLM Knowledge
Yumeng He (Ucla), Chenfanfu Jiang (Ucla)
GenerationTransformerDiffusion modelRectified FlowImageMesh
🎯 What it does: Extract rough URDF parameters and component reference images using Vision-Language Models (VLM), and combine them with Diffusion Transformers (DiT) to achieve multi-level local-global-hierarchical component generation, ultimately producing kinematic component-level 3D models directly applicable for physical simulation; subsequently refine joint parameters through joint optimization of differentiable forward kinematics and rendering.
SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs
Mohamad Alansari (Khalifa University), Muzammal Naseer (Khalifa University)
Object DetectionObject TrackingSegmentationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposed SPARROW, a multimodal large language model capable of simultaneously achieving pixel-level video localization and cross-frame semantic consistency;
Sparse Spectral LoRA: Routed Experts for Medical VLMs
Omid Nejatimanzari, Hassan Rivaz (Concordia University)
Mixture of ExpertsVision Language ModelBiomedical Data
🎯 What it does: This paper proposes MedQwen, a medical vision-language model based on sparse spectral LoRA, which achieves parameter-efficient fine-tuning through SVD-structured Mixture-of-Experts (MoE).
Sparse Task Vector Mixup with Hypernetworks for Efficient Knowledge Transfer in Whole-Slide Image Prognosis
Pei Liu, Yiping Liu (Hunan University)
ClassificationDomain AdaptationImageBiomedical Data
🎯 What it does: Proposed the STEPH method, achieving efficient cross-cancer WSI prognosis knowledge transfer through sparse task vector mixing and hypernetworks;
Sparse-LaViDa: Sparse Multimodal Discrete Diffusion Language Models
Shufan Li (Adobe), Jason Kuen (Adobe)
GenerationVision Language ModelDiffusion modelMultimodality
🎯 What it does: Propose Sparse-LaViDa, improving the inference process of Masked Discrete Diffusion Models by dynamically truncating redundant mask tokens and using register tokens to compress information, significantly accelerating generation while maintaining quality.
Sparse-View Localization via Online Neural 3D Regression
Ludvig Dillén (Lund University), Viktor Larsson (Lund University)
Pose EstimationComputational EfficiencySupervised Fine-TuningSimultaneous Localization and MappingImage
🎯 What it does: Propose ON3R, an online-trained neural 3D regressor that utilizes sparse matching to predict 3D coordinates and solves camera poses in sparse views via P3P-RANSAC
SparseCam4D: Spatio-Temporally Consistent 4D Reconstruction from Sparse Cameras
Weihong Pan (Zhejiang University), Guofeng Zhang (Zhejiang University)
GenerationData SynthesisDiffusion modelGaussian SplattingVideo
🎯 What it does: Combined sparse cameras with generative video diffusion models to build a 4D scene reconstruction framework under sparse camera settings;
Sparsely Timing the Change: A Spiking Temporal Framework for Remote Sensing Interpretation
Shilong Li (University of Electronic Science and Technology of China), Guisong Liu (Southwestern University of Finance and Economics)
Explainability and InterpretabilitySpiking Neural NetworkTransformerImage
🎯 What it does: Proposed the SpikeAdapter framework, which combines sparse temporal encoding SNN with ANN to model and interpret spatiotemporal changes in remote sensing images with only two frames.
SparseOIT: Improving Order-Independent Transparency 3DGS via Active Set Method
Wentao Yang (Zhejiang University), Xiangru Huang (Westlake University)
OptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Propose SparseOIT, a 3D Gaussian splatting reconstruction method based on OIT, which significantly improves training speed using active set techniques and GPU acceleration while maintaining rendering quality comparable to traditional volume rendering methods.
SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction
Zicheng Zhang (Fudan University), Wenchao Ding (Fudan University)
GenerationGaussian SplattingImagePoint Cloud
🎯 What it does: Propose SparseSplat, a novel forward 3D Gaussian projection model that can adaptively allocate Gaussian primitives and perform local attribute prediction in 3D space, thereby generating sparse yet high-quality scene representations;
SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model
Jiayuan Du (Tongji University), Qijun Chen (Tongji University)
Autonomous DrivingTransformerWorld ModelImageBenchmark
🎯 What it does: Propose a pure-attention transformer architecture that directly predicts future multi-frame 3D occupancy scenes from raw image features using sparse anchors, achieving four-dimensional occupancy world modeling conditioned on trajectories.
Sparsity as a Key: Unlocking New Insights from Latent Structures for Out-of-Distribution Detection
Ahyoung Oh (Yonsei University), Songkuk Kim (Yonsei University)
Anomaly DetectionRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: Decompose the CLS embedding of ViT using a Top-k sparse autoencoder to obtain a class activation profile (CAP), and detect out-of-distribution (OOD) samples based on the energy profile divergence (EPD) of CAP's energy distribution.
Sparsity-Aware Voxel Attention and Foreground Modulation for 3D Semantic Scene Completion
Yu Xue (Xi'an Jiaotong University), Xiaoning Zhang (Xi'an Jiaotong University)
SegmentationConvolutional Neural NetworkAuto EncoderPoint CloudBenchmark
🎯 What it does: This paper proposes a 3D semantic scene completion framework called VoxSAMNet based on a single RGB image
SparVAR: Exploring Sparsity in Visual AutoRegressive Modeling for Training-Free Acceleration
Zekun Li (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: Proposed the SparVAR framework to achieve training-agnostic acceleration of visual autoregressive models (VAR) without skipping any scale, significantly improving inference speed while maintaining high-frequency details;
Spatia: Video Generation with Updatable Spatial Memory
Jinjing Zhao (University of Sydney), Yan Lu (Microsoft Research)
GenerationFlow-based ModelSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: Maintain 3D scene point clouds as updatable spatial memory during video generation to achieve dynamic-static separation, ensure long-term spatial consistency, and support explicit camera control and 3D perception interactive editing.
Spatial Matters: Position-Guided 3D Referring Expression Segmentation
Yabing Wang (Xi'an Jiaotong University), Sanping Zhou (Xi'an Jiaotong University)
SegmentationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningTextPoint Cloud
🎯 What it does: Propose Position3D, an end-to-end method capable of precisely segmenting target objects in point cloud scenes based on natural language descriptions.
Spatial Retrieval Augmented Autonomous Driving
Xiaosong Jia (Fudan University), Yu-Gang Jiang (Fudan University)
Autonomous DrivingConvolutional Neural NetworkTransformerWorld ModelImageRetrieval-Augmented Generation
🎯 What it does: Introduce a spatial retrieval paradigm, incorporating offline geographic images (e.g., Google Maps street view/satellite images) as additional input to enhance multi-task autonomous driving performance.
Spatial-Aware VLA Pretraining through Visual-Physical Alignment from Human Videos
Yicheng Feng (Peking University), Zongqing Lu (Peking University)
TransformerVision-Language-Action ModelVideoTextMultimodalityPoint Cloud
🎯 What it does: This paper proposes a spatial-aware vision-language-action (VIPA-VLA) pre-training framework based on human videos to bridge the gap between 2D visual inputs and 3D physical actions in traditional VLA models.
Spatial-Frequency Collaborative Learning for Occluded Visible-Infrared Person Re-Identification
Jian Yu (Jiangsu University of Technology), Yimu Ji (Nanjing University of Posts and Telecommunications)
RecognitionRetrievalTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose the Spatial-Frequency Collaborative Learning (SFCL) framework, leveraging collaborative learning between frequency domain information and spatial features to address occlusion and modality difference issues in visible-infrared pedestrian re-identification.
Spatial-SAM: Spatially Consistent 3D Electron Microscopy Segmentation with SDF Memory and Semi-Supervised Learning
Yikai Huang, Ligang Liu
SegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: Propose the Spatial-SAM framework, achieving an end-to-end pipeline from few interactive annotations to fully automatic 3D electron microscopy segmentation.
Spatial-Spectral Residuals Informed Diffusion Neural Operator for Pan-sharpening
Jiahan Huang (Southeast University), Liang-Jian Deng (University of Electronic Science and Technology of China)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: Proposed a function space diffusion model (SRINO) based on Neural Operators for spectral and spatial information fusion in remote sensing images (full-resolution multispectral reconstruction)
Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning
Yuhong Liu (Shanghai AI Laboratory), Jiaqi Wang (Shanghai AI Laboratory)
Reinforcement LearningVision Language ModelContrastive LearningImageVideo
🎯 What it does: Pretrain LVLM on ordinary RGB/RGB-D images through the self-supervised reinforcement learning framework Spatial-SSRL to enhance its spatial reasoning capabilities.
SpatialDiff: 3D-Aware Object Movement via Implicit Spatial Modeling
Zheng Liu (Sun Yat Sen University), Guanbin Li (Sun Yat Sen University)
GenerationTransformerSupervised Fine-TuningDiffusion modelImageMultimodality
🎯 What it does: Proposes a method for achieving precise spatial object movement in 2D image editing by leveraging implicit 3D space modeling.
SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models
Yuechen Xie (Zhejiang University), Jie Song (Zhejiang University)
SegmentationDepth EstimationVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the SpatiaLQA benchmark to evaluate the spatial reasoning capabilities of vision-language models, and introduced a recursive scene graph-assisted reasoning method to enhance model performance.
SpatialReward: Verifiable Spatial Reward Modeling for Fine-Grained Spatial Consistency in Text-to-Image Generation
Sashuai Zhou (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose SpatialReward, a verifiable reward model for evaluating fine-grained spatial consistency in text-to-image generation, and construct a new benchmark called SpatRelBench.
SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence
Haoning Wu (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
Large Language ModelSupervised Fine-TuningAgentic AIVision-Language-Action ModelMultimodalityPoint CloudBenchmark
🎯 What it does: This paper proposes the SpatialScore benchmark and SpatialCorpus training resources, constructs a multi-modal spatial reasoning evaluation framework, and enhances the model's spatial understanding capabilities through the multi-tool collaboration of SpatialAgent.
SpatialStack: Layered Geometry-Language Fusion for 3D VLM Spatial Reasoning
Jian Zhang, Zhiwen Fan (Texas A&M University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper proposes a hierarchical geometric-language fusion framework called SpatialStack, aiming to enhance the performance of large-scale vision-language models (VLM) in 3D spatial reasoning tasks.
SpatialTree: How Spatial Intelligence Branches Out in MLLMs
Yuxi Xiao (Bytedance Seed), Bingyi Kang (Bytedance Seed)
Autonomous DrivingRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision-Language-Action ModelImageVideoMultimodalityPoint CloudMeshBenchmark
🎯 What it does: Proposed a cognitive science-based hierarchical structure of spatial ability called SpatialTree, and constructed a corresponding capability center evaluation benchmark, SpatialTree-Bench, to systematically assess the performance of mainstream multi-modal large language models (MLLMs) on four levels of spatial ability (perception, mapping, simulation, action).