CVPR 2026 Papers — Page 34
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
SpatialVID: A Large-Scale Video Dataset with Spatial Annotations
Jiahao Wang (Nanjing University), Yao Yao (Nanjing University)
Data SynthesisLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: This paper constructs SpatialVID, a large-scale video dataset containing camera poses, depth maps, structured captions, and motion instructions, aiming to support spatial intelligence research;
Spatio-Temporal Conditional Denoising Transformer for Modality-Missing RGBT Tracking
Andong Lu (Anhui University), Bin Luo (Anhui University)
RestorationObject TrackingTransformerImageMultimodality
🎯 What it does: Propose a unified spatiotemporal conditional denoising transformer (SCDT) that simultaneously achieves reconstruction of missing modalities and feature enhancement of complete modalities in RGB-T tracking.
Spatio-Temporal Difference Guided Motion Deblurring with the Complementary Vision Sensor
Yapeng Meng (Tsinghua University), Rong Zhao (Tsinghua University)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a motion deblurring framework named STGDNet based on Complementary Vision Sensors (CVS), which can utilize spatial differences (SD) and temporal differences (TD) signals simultaneously collected by CVS during a single-frame RGB exposure period, significantly improving deblurring performance under extreme motion blur.
Spatiotemporal Pyramid Flow Matching for Climate Emulation
Jeremy A. Irvin (Stanford University), Duncan Watson-Parris (University of California, San Diego)
Data SynthesisTransformerFlow-based ModelTime SeriesPhysics Related
🎯 What it does: This paper proposes a novel spatiotemporal pyramid flow matching model (SPF) that can efficiently generate climate simulation results in parallel across multiple timescales, supporting direct sampling at different spatial and temporal resolutions;
SPDMark: Selective Parameter Displacement for Robust Video Watermarking
Samar Fares (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
GenerationConvolutional Neural NetworkDiffusion modelVideo
🎯 What it does: In video diffusion models, implicit watermarking during generation is achieved by selectively displacing model parameters to embed watermarks;
Spe-BEVHead: Rethinking the Detection Head Design for Bird's-Eye-View Object Detection
Junshu Zhang (Tsinghua University), Guiguang Ding (Tsinghua University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper proposes a new detection head called Spe-BEVHead for bird's-eye view (BEV) 3D object detection, addressing three major shortcomings of traditional center-based detection heads.
SPE-MVS: Spatial Position Encoding Enhanced Multi-View Stereo with Monocular Depth Priors
Shaoqian Wang (North China Electric Power University), Yuchao Dai (Northwestern Polytechnical University)
Depth EstimationConvolutional Neural NetworkImageBenchmark
🎯 What it does: This paper proposes the SPE-MVS framework, which enhances the quality of multi-view stereo reconstruction by integrating spatial position encoding (SPE) based on monocular depth prior.
SPEAR-1: Scaling Beyond Robot Demonstrations via 3D Understanding
Nikolay Nikolov (INSAIT), Danda Pani Paudel (INSAIT)
Robotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelFlow-based ModelImageVideoText
🎯 What it does: Built a robot foundation model named SPEAR-1, integrating 3D perception with language instruction control strategies.
Specificity-aware reinforcement learning for fine-grained open-world classification
Samuele Angheben (University of Trento), Yiming Wang (Fondazione Bruno Kessler)
ClassificationLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Propose SpeciaRL, a specificity-aware reinforcement learning framework designed to enhance model specificity in open-world fine-grained image classification tasks while maintaining accuracy.
Spectral Conformal Risk Control: Distribution-Free Tail Guarantees via Bayesian Quadrature
Mohammad Mahdi Kazemi Esfeh (University of British Columbia), Purang Abolmaesumi (University of British Columbia)
OptimizationImageTextBenchmark
🎯 What it does: This paper proposes the Bayesian-Quadrature Spectral Risk Control (BQ-SRC) framework, achieving distribution-agnostic tail risk control;
Spectral Mixture-of-Experts for Continual Learning
Chen Yin (Anhui University), Zhe Jin (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
Mixture of ExpertsImage
🎯 What it does: Propose the Spectral Mixture-of-Experts (Spectral MoE) framework, which addresses structural interference and routing drift between experts in continuous learning by separating expert parameters in the frequency domain.
Spectral Scalpel: Amplifying Adjacent Action Discrepancy via Frequency-Selective Filtering for Skeleton-Based Action Segmentation
Haoyu Ji (Harbin Institute of Technology), Honghai Liu (Harbin Institute of Technology)
SegmentationGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: Propose the Spectral Scalpel framework, which enhances temporal action segmentation in skeletal sequences through frequency domain filtering;
Spectral Super-Resolution via Adversarial Unfolding and Data-Driven Spectrum Regularization: From Multispectral Satellite Data to NASA Hyperspectral Image
Si-Sheng Young (National Cheng Kung University), Chia-Hsiang Lin (National Cheng Kung University)
Super ResolutionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposed an algorithm to convert Sentinel-2 multispectral images into NASA AVIRIS-level hyperspectral images, achieving joint reconstruction of spectral super-resolution from 12 bands to 186 bands and 5μm spatial resolution.
Spectral-Geometric Neural Fields for Pose-Free LiDAR View Synthesis
Yinuo Jiang (Huazhong University of Science and Technology), Cheng Cheng (Huazhong University of Science and Technology)
GenerationPose EstimationAutonomous DrivingNeural Radiance FieldGenerative Adversarial NetworkPoint Cloud
🎯 What it does: Designed the SG-NLF framework, a pose-free LiDAR NeRF based on a spectral-geometric hybrid representation, achieving high-quality view synthesis and pose estimation.
Spectrally Distilled Representations Aligned with Instruction-Augmented LLMs for Satellite Imagery
Minh Kha Do (La Trobe University), Ramana Rao Kompella (Cisco Research)
ClassificationSegmentationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed an RGB-only visual-language foundation model named SATtxt, which can retain multispectral priors and achieve tasks such as zero-shot classification, retrieval, and open-vocabulary segmentation without using multispectral inputs.
Spectrum from Defocus: Fast Spectral Imaging with Chromatic Focal Stack
M. Kerem Aydin (Northwestern University), Emma Alexander (Northwestern University)
TransformerImagePhysics Related
🎯 What it does: Propose a method that uses two ordinary lenses and a grayscale camera to capture multiple blurred images by moving the lens to achieve chromatic focal sweep, and rapidly restores hyperspectral images using a physics-based iterative reconstruction algorithm.
SpeeDe3DGS: Speedy Deformable 3D Gaussian Splatting with Temporal Pruning and Motion Grouping
Allen Tu (University of Maryland, College Park), Matthias Zwicker (University of Maryland, College Park)
Computational EfficiencyGaussian SplattingBenchmark
🎯 What it does: Propose the SpeeDe3DGS framework, combining Temporal Sensitivity Pruning (TSP), Temporal Sensitivity Sampling (TSS), and GroupFlow, significantly accelerating dynamic 3D Gaussian projection while maintaining the image quality of the original neural motion field.
SpeeDiff: Scalable Pixel-Anchored End-to-End Latent Diffusion Model
Bingliang Zhang (ByteDance Seed), Qiushan Guo (ByteDance Seed)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose an expandable, pixel-anchored end-to-end latent diffusion model called SpeeDiff, which can jointly train VAE and diffusion networks from scratch;
Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists
Jiaqi Liu (Wayne State University), Zhizhong Han (Wayne State University)
Computational EfficiencyGaussian SplattingImage
🎯 What it does: By periodically shrinking Gaussian scales (scale reset) during the 3D Gaussian Splatting (3DGS) training process and applying entropy constraints to α blending weights, the Gaussian list per pixel is significantly shortened, accelerating learning; combined with a resolution scheduler further enhances efficiency.
SPEGC: Continual Test-Time Adaptation via Semantic-Prompt-Enhanced Graph Clustering for Medical Image Segmentation
Xiaogang Du (Shaanxi University of Science and Technology), Yingbo Wang (Shaanxi University of Science and Technology)
SegmentationDomain AdaptationConvolutional Neural NetworkPrompt EngineeringBiomedical Data
🎯 What it does: Propose a semantic prompt enhanced graph clustering based continual test-time adaptation (SPEGC) framework for medical image segmentation; the framework first enhances local features through dual-pool semantic prompts, then generates high-order structural representations by solving a differentiable graph clustering solver (rewritten as an optimal transport problem), driving the model to gradually adapt on unlabeled target domains.
Spherical Leech Quantization for Visual Tokenization and Generation
Yue Zhao (University of Texas Austin), Philipp Kraehenbuehl (University of Texas Austin)
GenerationCompressionRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a spherical quantization method Λ24-SQ based on the Leech lattice, which can train visual tokenizers and autoregressive generative models with ~200K codebook without using regularization;
Spherical Voronoi: Directional Appearance as a Differentiable Partition of the Sphere
Francesco Di Sario (University of Torino), Andrea Tagliasacchi (Simon Fraser University)
GenerationNeural Radiance FieldGaussian SplattingImageBenchmark
🎯 What it does: In radiance field methods, view-dependent and reflection-dependent directional appearance modeling is achieved through a differentiable spherical Voronoi approximation.
SpiderCam: Low-Power Snapshot Depth from Differential Defocus
Marcos A. Ferreira, Emma Alexander (Northwestern University)
Depth EstimationComputational EfficiencyImage
🎯 What it does: Developed a low-power FPGA-based camera called SpiderCam, utilizing the differential defocus algorithm to achieve real-time sparse depth measurement.
Spike-driven Discrete Aggregation for Event-based Object Detection
Huaning Li (Zhejiang University), Huajin Tang (Zhejiang University of Technology)
Object DetectionSpiking Neural NetworkTime Series
🎯 What it does: Designed and implemented an event camera object detection method based on differentiable discrete aggregation (SDA-MTF), embedding it into a full-spike neural network.
SpikeTrack: A Spike-driven Framework for Efficient Visual Tracking
Qiuyang Zhang (Tongji University), Shangce Gao (University of Toyama)
Object TrackingSpiking Neural NetworkTransformerImageVideo
🎯 What it does: Designed and implemented a RGB visual tracking framework called SpikeTrack based on spiking neural networks, adopting an asymmetric Siamese structure and a memory retrieval module to significantly reduce energy consumption while maintaining high accuracy.
SpikeTrack: High-performance and Energy-efficient Event-Based Object Tracking with Spiking Neural Network
Yang Wang (Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology), Xin Yang (Key Laboratory of Social Computing and Cognitive Intelligence, Dalian University of Technology)
Object TrackingSpiking Neural NetworkTransformer
🎯 What it does: Proposes a pure SNN framework called SpikeTrack for single-target tracking based on event cameras.
SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras
Huanjing Yue (Tianjin University), Jingyu Yang (Tianjin University)
Image TranslationRestorationTransformerDiffusion modelImage
🎯 What it does: Proposed a RGB-to-RAW conversion method called SpiralDiff based on diffusion models, achieving a cross-camera unified model through signal-dependent noise weighting and CamLoRA adapter.
Spk2VidNet: A Hierarchical Recurrent Architecture for High-Fidelity Video Reconstruction from Long Spike-Camera Streams
Yuanlin Wang (Peking University), Tiejun Huang (Peking University)
Super ResolutionRecurrent Neural NetworkVideo
🎯 What it does: This paper proposes Spk2VidNet, which reconstructs high-resolution videos from long-term Spike camera data using a hierarchical recursive network.
Splat-Based Metal Artifact Reduction in Cone-Beam CT via Compact Attenuation Modeling
Kiseok Choi (KAIST), Min H. Kim (KAIST)
Gaussian SplattingBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a CBCT metal artifact reduction framework based on Gaussian scattering, achieving high-quality reconstruction without requiring a metal mask by incorporating a differentiable multicolor X-ray projection model into a continuous Gaussian representation;
Splatent: Splatting Diffusion Latents for Novel View Synthesis
Or Hirschorn (Amazon Prime Video), Lior Fritz (Amazon Prime Video)
GenerationData SynthesisDiffusion modelAuto EncoderGaussian SplattingImage
🎯 What it does: This paper proposes a new framework called Splatent for high-quality 3D view synthesis in the VAE latent space, constructing a latent radiance field using 3D Gaussian Splatting (3DGS) and enhancing the details of the rendered latent representation through a single-step diffusion model with a multi-view attention mechanism.
SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian Splatting
Pranav Asthana (University of Maryland), Amitabh Varshney (University of Maryland)
Super ResolutionGaussian SplattingImage
🎯 What it does: When training the 3D Gaussian Splatting model, super-resolution is applied only in regions lacking high-frequency information to avoid global consistency issues.
SplitFlux: Learning to Decouple Content and Style from a Single Image
Yitong Yang (Shanghai University of Finance and Economics), Shuting He (Shanghai University of Finance and Economics)
GenerationSupervised Fine-TuningVision Language ModelDiffusion modelImageText
🎯 What it does: For a single image, we propose SplitFlux, which achieves decoupling of image content and style by fine-tuning the single-stream blocks in the Flux model, and allows the decoupled content to be re-embedded into any new context.
Spot The Ball: A Benchmark for Visual Social Inference
Neha Balamurugan (Stanford University), Tobias Gerstenberg (Stanford University)
Convolutional Neural NetworkPrompt EngineeringVision Language ModelDiffusion modelImageBenchmarkChain-of-Thought
🎯 What it does: Developed a visual social reasoning benchmark named SPOT THE BALL, requiring Vision-Language Models (VLMs) to infer the location of the removed ball through social cues such as players' gaze, posture, and orientation.
SPOT: Spatiotemporal Prompt Optimization for Motion-Stabilized MLLM-Guided Video Segmentation
Jiayi Fan (Shandong University), Yilong Yin (Shandong University)
SegmentationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoMultimodality
🎯 What it does: Propose the SPOT framework, achieving spatiotemporally consistent video segmentation without video pre-training by applying spatiotemporal constraints on prompts generated by a multimodal large language model.
SpotEdit: Selective Region Editing in Diffusion Transformers
Zhibin Qin (National University of Singapore), Xinchao Wang (National University of Singapore)
Image HarmonizationGenerationTransformerDiffusion modelRectified FlowAuto EncoderImage
🎯 What it does: Proposes SpotEdit, a training-free, diffusion transformer-based local editing framework that performs reverse diffusion only on the regions requiring modification, while directly preserving original features in non-edited areas;
SPREAD: Spatial-Physical REasoning via geometry Aware Diffusion
Minzhang Li (ShanghaiTech University), Jingyi Yu (ShanghaiTech University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelPoint CloudGraph
🎯 What it does: Propose a diffusion-based 3D scene generation framework called SPREAD, which can simultaneously learn spatial and physical relationships, and achieve joint modeling of pose, shape, and interactions during generation through a geometry-aware Perceiver module.
SR3R: Rethinking Super-Resolution 3D Reconstruction With Feed-Forward Gaussian Splatting
Xiang Feng (Hangzhou Dianzi University), Yanming Zhu (Griffith University)
Super ResolutionTransformerGaussian SplattingImagePoint Cloud
🎯 What it does: Researchers propose reconstructing the 3D super-resolution task as a forward mapping from sparse low-resolution views to high-resolution 3D Gaussian Splatting.
SRA 2: Variational Autoencoder Self-Representation Alignment for Efficient Diffusion Training
Mengmeng Wang (Zhejiang University of Technology), Jingdong Wang (Baidu)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Propose the SRA 2 scheme, which achieves lightweight guidance for training diffusion Transformers (e.g., SiT) by performing self-representation alignment on the reconstructed features of a pre-trained VAE;
SRA-Det: Learning Omni-Grained Open-Vocabulary Detection Beyond Category Names
Li Yang (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
Object DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextRetrieval-Augmented Generation
🎯 What it does: Proposed Semantic-Retrieval-Augmented Detector (SRA-Det), achieving fine-grained open-vocabulary object detection through hierarchical retrieval of multi-level semantic features, and constructed an automated attribute-enhanced data pipeline based on LLM and CLIP.
SRGCD: Stability-Driven Region Growth Framework for 3D Change Detection
Yue Wu (Xidian University), Wenping Ma (Xidian University)
SegmentationConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Proposes SRGCD, a stability-driven region-growing framework for detecting changes from dual-time point clouds, with the core idea of starting from seeds of extremely high-confidence unchanged regions and then iteratively expanding stable areas to achieve fine-grained segmentation.
SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models
Senyu Fei (Tongji University), Xipeng Qiu (Fudan University)
OptimizationRobotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelWorld ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a self-referential policy optimization (SRPO) framework that leverages successful trajectories generated by the visual-language-action model itself as a reference, assigns progress rewards to failed trajectories through the latent representation of the world model, and achieves efficient reinforcement learning fine-tuning.
SSM-Aware Token-Efficient VMamba via Adaptive Patch Pruning and Merging for Person Re-Identification
Huiyuan Huang (Kookmin University), Sang Min Yoon (Kookmin University)
RecognitionComputational EfficiencyImage
🎯 What it does: Proposes an efficient person re-identification framework called TE-VMamba based on a variable state space model, which significantly reduces the number of tokens while maintaining recognition accuracy.
ST4R-Splat: Spatio-Temporal Referring Segmentation in 4D Gaussian Splatting
Yuming Meng (Peking University), Hongbin Zha (Peking University)
SegmentationLarge Language ModelVision Language ModelContrastive LearningGaussian SplattingVideoTextMultimodality
🎯 What it does: Proposes a spatiotemporal referential segmentation framework called ST4R-Splat based on 4D Gaussian Splatting, enabling precise segmentation and temporal localization of object instances in dynamic scenes through natural language.
Stability-Driven Motion Generation for Object-Guided Human-Human Co-Manipulation
Jiahao Xu (Tianjin University), Buzhen Huang (Tianjin University)
GenerationData SynthesisRobotic IntelligenceDiffusion modelFlow-based ModelGenerative Adversarial NetworkVideoPoint CloudMesh
🎯 What it does: Propose a flow-matching based framework to generate coordinated two-person manipulation actions that follow a given object trajectory, enhancing action naturalness and physical feasibility through object-affinity-driven contact strategies, dual-person interaction discriminators, and stability simulations.
Stabilizing Feature Geometry in Noisy Pretrained Models for Robust Downstream Tasks
Quanyu Zhang (Shandong University), Shuo Li (Case Western Reserve University)
Representation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: The study investigates the impact of pre-trained noise on the geometric structure of feature space, proposing Principal Directional Angle (PDA) to measure the rotation of the principal subspace caused by noise, and designing a Feature Geometry Stabilization (FGS) framework (comprising Feature Perturbation Consistency, Variance-Activation Regularization, and Feature Consistency Distillation) to suppress direction drift induced by noise during fine-tuning.
Stabilizing Streaming Video Geometry via Dynamic Feature Normalization
Xiaoyang Lyu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningVideoPoint Cloud
🎯 What it does: This paper proposes a dynamic feature normalization module (DyFN), which stabilizes geometric consistency in continuous video streams by recursively adjusting the mean and variance of latent features, while keeping the pre-trained single-frame geometric estimation model unchanged.
Stable and Efficient Single-Rollout RL for Multimodal Reasoning
Rui Liu (Tencent AI Lab), Dong Yu (Tencent AI Lab)
Reinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposes the MSSR framework for verifiable reward reinforcement learning in multimodal large language models, which requires only a single rollout for training.
Stable Mean Flow: Lyapunov-Inspired One-Step Flow Matching
Guangxun Zhang (New York University), Davi Geiger (New York University)
GenerationScore-based ModelFlow-based ModelImage
🎯 What it does: Propose a Stable Mean Flow algorithm based on Lyapunov non-expansion constraints, achieving stable training and sampling with single-step generation
Stable Spike: Dual Consistency Optimization via Bitwise AND Operations for Spiking Neural Networks
Yongqi Ding (University of Electronic Science and Technology of China), Lin Zuo (University of Electronic Science and Technology of China)
ClassificationSpiking Neural NetworkImage
🎯 What it does: To address the temporal pulse inconsistency issue in spiking neural networks (SNNs), the Stable Spike method is proposed, which extracts a stable pulse skeleton through bitwise operations, forces the original pulse diagram to converge to the skeleton, and enhances generalization by adding amplitude-adaptive pulse noise on the skeleton.
StableMaterials: Enhancing Diversity in Material Generation via Semi-Supervised Learning
Giuseppe Vecchio (Adobe Research)
GenerationData SynthesisVision Language ModelDiffusion modelImageTextStochastic Differential Equation
🎯 What it does: Developed StableMaterials, a semi-supervised learning method based on diffusion models, generating high-resolution, seamless PBR material maps using text or image prompts.
StableMTL: Repurposing Latent Diffusion Models for Multi-Task Learning from Partially Annotated Synthetic Datasets
Anh-Quan Cao (Valeo.ai), Raoul de Charette (Inria)
SegmentationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Train a multi-task model to learn seven different pixel-level tasks on partially labeled synthetic data and achieve cross-domain generalization on real datasets.
STAC: Plug-and-Play Spatio-Temporal Aware Cache Compression for Streaming 3D Reconstruction
Runze Wang (University of Science and Technology of China), Ligang Liu (University of Science and Technology of China)
Pose EstimationCompressionOptimizationComputational EfficiencyTransformerImagePoint CloudBenchmark
🎯 What it does: Proposed a no-training, plug-and-play STAC framework to compress the KV cache of Causal-VGGT, achieving long-term consistency and low memory consumption during online 3D reconstruction.
STAGE: Storyboard-Anchored Generation for Cinematic Multi-shot Narrative
Peixuan Zhang (Beijing University of Posts and Telecommunications), Boxin Shi (Peking University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: Propose a storyboard-based multi-shot video generation framework named STAGE, with the core being the STEP 2 model that predicts start and end frame pairs for each shot, achieving cross-shot consistency and cinematic language through multi-shot memory packages, dual encoding strategies, and two-phase training.
Stake the Points: Structure-Faithful Instance Unlearning
Kiseong Hong (Chung Ang University), Eunwoo Kim (Chung Ang University)
Safty and PrivacyLarge Language ModelVision Language ModelImage
🎯 What it does: Propose the structure-preserving instance-level machine forgetting framework STRUCTGUARD, which can achieve forgetting of specified instances while only accessing the pre-trained model and the data to be forgotten, without compromising the semantic structure of other instances.
StaMo: Unsupervised Learning of Generalizable Robot Motion from Compact State Representation
Mingyu Liu (Zhejiang University), Chunhua Shen (Zhejiang University)
Representation LearningRobotic IntelligenceVision-Language-Action ModelDiffusion modelFlow-based ModelAuto EncoderWorld ModelImage
🎯 What it does: Propose the StaMo method, which compresses a single-frame image into a state representation consisting of two 1024-dimensional tokens using a Diffusion Autoencoder; the state difference represents motion and can be applied to world modeling, action planning, and integration with Vision-Language Models (VLMs);
Stand-In: A Lightweight and Plug-and-Play Identity Control for Video Generation
Bowen Xue (Tencent Inc.), Jing Lyu (Nankai University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Propose a lightweight, plug-and-play identity-preserving video generation framework called Stand-In, which achieves identity control by adding a conditional image branch to a pre-trained video diffusion model, generating high-quality videos consistent with the text while keeping the reference image unchanged.
StaR-KVQA: Structured Reasoning Traces for Implicit-Knowledge Visual Question Answering
Zhihao Wen (Ant International Ant Group), Xin Zhang (Ant International Ant Group)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose the StaR-KVQA framework, which explicitly models the reasoning process in implicit knowledge visual question answering tasks by training through generating dual paths (text + visual relationship paths) and corresponding natural language explanations.
STAR-R1: Multi-View Spatial TrAnsformation Reasoning by Reinforcing Multimodal LLMs
Zongzhao Li (Renmin University of China), Wenbing Huang (Renmin University of China)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the STAR-R1 two-stage framework, first training a multi-modal LLM to learn spatial reasoning processes through Process-Supervised SFT using high-quality Chain-of-Thought (CoT), and then enhancing the selection of key referents and the correctness of final answers via Referential-Aware RL, thereby improving multi-perspective spatial reasoning capabilities.
STAR: Test-Time Adaptation Can Enhance Universal Prompt Learning for Vision-Language Models
Yiwei Fu (Peking University), Minghua Deng (Peking University)
Domain AdaptationPrompt EngineeringVision Language ModelImage
🎯 What it does: The study enhances the robustness of vision-language models to noisy unlabeled target data during testing through adaptive prompt learning.
STARFlow-V: End-to-End Video Generative Modeling with Autoregressive Normalizing Flows
Jiatao Gu (Apple), Shuangfei Zhai (Apple)
GenerationComputational EfficiencyTransformerScore-based ModelFlow-based ModelAuto EncoderImageVideoTextMultimodality
🎯 What it does: Proposes STARFlow-V, an end-to-end video generation framework based on autoregressive regularized flows, which adopts a global-local architecture, flow score matching denoising, and efficient sampling via Jacobi iteration, enabling multi-task generation (T2V, I2V, V2V) and streaming generation under text, image, and video conditions.
Statistical Characteristic-Guided Denoising for Rapid High-Resolution Transmission Electron Microscopy Imaging
Hesong Li (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
RestorationData SynthesisConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Aiming at the strong noise generated by high-resolution transmission electron microscopy (HRTEM) during rapid observation of nucleation processes, a statistical feature-guided denoising network (SCGN) and a corresponding dedicated synthetic dataset are proposed.
STAvatar: Soft Binding and Temporal Density Control for Monocular 3D Head Avatars Reconstruction
Jiankuo Zhao (Chinese Academy of Sciences), Zhen Lei (Chinese Academy of Sciences)
GenerationConvolutional Neural NetworkGaussian SplattingVideo
🎯 What it does: This paper proposes a monocular video 3D head avatar reconstruction method called STAvatar based on 3D Gaussian Splatting, which can generate high-fidelity, animatable head avatars from monocular video.
Stay in your Lane: Role Specific Queries with Overlap Suppression Loss for Dense Video Captioning
Seung Hyup Baek (Konkuk University), Jae Won Cho (Konkuk University)
GenerationRecurrent Neural NetworkTransformerVision Language ModelContrastive LearningVideoTextBenchmark
🎯 What it does: Propose a role-specific query-based Dense Video Captioning framework ROS-DVC, which uses independent localization queries and description queries, and improves localization and description accuracy through cross-task contrastive alignment, overlap suppression loss, and concept guide module.
STCast: Adaptive Boundary Alignment for Global and Regional Weather Forecasting
Hao Chen (Hong Kong University of Science and Technology), Lei Bai (Shanghai AI Laboratory)
TransformerMixture of ExpertsTime SeriesPhysics Related
🎯 What it does: Propose the STCast framework, combining adaptive regional boundaries and dynamic monthly expert allocation to unify global and regional weather prediction.
STCDiT: Spatio-Temporally Consistent Diffusion Transformer for High-Quality Video Super-Resolution
Junyang Chen (Nanjing University of Science and Technology), Jinshan Pan (Nanjing University of Science and Technology)
Super ResolutionTransformerSupervised Fine-TuningDiffusion modelAuto EncoderVideo
🎯 What it does: Propose an STCDiT framework based on a pre-trained video diffusion model for high-quality video super-resolution, addressing structural fidelity and temporal consistency issues.
Stealing Split Learning Bottom Models by Recovering Embedding Geometry
Qinbo Zhang (Stony Brook University), Jian Li (New York University)
Federated LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImageMultimodalityTabular
🎯 What it does: Proposes VENOM, a geometry-aware model stealing attack targeting split learning/vertical federated learning, capable of recovering the adversary's bottom model under various defense mechanisms.
Steering Where to Diffuse: Generative Modeling of Phenotypic Response Simulation with Steered Diffusion Bridge
Rongchao Zhang (Peking University), Yu Huang (Peking University)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelImageBiomedical DataStochastic Differential Equation
🎯 What it does: This paper proposes a cell phenotypic response simulation method based on guided diffusion bridges (SimuSDB), used to generate morphological changes in cells under chemical or genetic perturbations.
Stepwise Credit Assignment for GRPO on Flow-Matching Models
Yash Savani (Carnegie Mellon University), Krishna Kumar Singh (Adobe Research)
GenerationReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelImageTextBenchmarkStochastic Differential Equation
🎯 What it does: Improves the reinforcement learning training method for flow matching models by assigning credit at each step of the diffusion process, enabling more efficient image generation.
Stereo World Model: Camera-Guided Stereo Video Generation
Yang-Tian Sun (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
GenerationTransformerDiffusion modelAuto EncoderWorld ModelImageVideo
🎯 What it does: Built StereoWorld, a camera-controlled stereo video generation world model that can generate perspective-consistent and geometrically consistent stereo videos under the condition of given stereo images and camera trajectories.
StereoWorld: Geometry-Aware Monocular-to-Stereo Video Generation
Ke Xing (Beijing Jiaotong University), Yunchao Wei (Beijing Jiaotong University)
Image TranslationGenerationData SynthesisDepth EstimationTransformerDiffusion modelVideo
🎯 What it does: Directly converting ordinary monocular videos into high-quality, geometrically consistent stereo videos
Stitch-a-Demo: Creating Video Demonstrations from Multistep Descriptions
Chi Hsuan Wu, Kristen Grauman (University of Texas Austin)
GenerationData SynthesisRetrievalTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the Stitch-a-Demo method, which utilizes retrieval and evaluation networks to concatenate coherent demonstration videos from multiple textual steps;
STiTch: Semantic Transition and Transportation in Collaboration for Training-Free Zero-Shot Composed Image Retrieval
Miaoge Li (Hong Kong Polytechnic University), Jingcai Guo (Hong Kong Polytechnic University)
RetrievalLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes a training-agnostic, single-stage zero-shot compositional image retrieval framework called STiTch;
Stochastic Ray Tracing for the Reconstruction of 3D Gaussian Splatting
Peiyu Xu (University of Illinois Urbana Champaign), Iliyan Georgiev (Adobe Research)
GenerationOptimizationComputational EfficiencyGaussian SplattingImagePoint CloudBenchmark
🎯 What it does: Proposed a differentiable, unordered stochastic ray tracing framework for reconstruction and rendering of standard and relit 3D Gaussian Splatting scenes.
SToRe3D: Sparse Token Relevance in ViTs for Efficient Multi-View 3D Object Detection
Sandro Papais (University Of Toronto), Lingting Ge (Zoox Inc)
Object DetectionAutonomous DrivingComputational EfficiencyTransformerImageBenchmark
🎯 What it does: This paper proposes a sparsity framework called SToRe3D based on Vision Transformer for multi-view 3D object detection, which can significantly reduce inference latency while maintaining detection accuracy.
StoryTailor:A Zero-Shot Pipeline for Action-Rich Multi-Subject Visual Narratives
Jinghao Hu (Northwest University), Han Zhang (Northwest University)
GenerationTransformerPrompt EngineeringVision-Language-Action ModelDiffusion modelImageVideoText
🎯 What it does: Proposes a zero-tuning, zero-training multi-agent visual storytelling generation pipeline called StoryTailor, which can generate identity-preserving, action-rich, and cross-frame background coherent image sequences based on long text prompts, reference images, and bounding boxes for each agent.
StreamAvatar: Streaming Diffusion Models for Real-Time Interactive Human Avatars
Zhiyao Sun (Tsinghua University), Yong-Jin Liu (Tsinghua University)
GenerationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkVideoTextAudio
🎯 What it does: Proposes StreamAvatar, a framework capable of real-time, streaming generation of full-body human videos, supporting two interaction states: speaking and listening.
StreamDiT: Real-Time Streaming Text-to-Video Generation
Akio Kodaira (UC Berkeley), Yue Zhao (Meta)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderVideoText
🎯 What it does: Propose the StreamDiT model, enabling real-time streaming text-to-video generation and supporting infinite streaming generation of long videos.
Streaming Diffusion Model for Fast Infrared and Visible Video Fusion
Jinyuan Liu (Dalian University of Technology), Xin Fan (Dalian University of Technology)
GenerationConvolutional Neural NetworkDiffusion modelOptical FlowVideoMultimodality
🎯 What it does: This paper proposes an infrared-visible video fusion method based on a streaming diffusion model, which can achieve real-time inference while maintaining high quality.
Streaming Video Crime Anticipation with Spatio-Temporal Causal Reasoning
Yusong Wang (Institute of Artificial Intelligence, China Telecom), Qingsong Zhao (Fudan University)
Anomaly DetectionRecurrent Neural NetworkGraph Neural NetworkSupervised Fine-TuningVision Language ModelVideo
🎯 What it does: Propose a real-time video crime warning framework, construct the STCRC spatiotemporal causal reasoning dataset, and design the STCH hypergraph streaming module to convert implicit entity dynamics into explicit causal structures for VLM usage.
Streaming Video Instruction Tuning
Jiaer Xia (Hong Kong Baptist University), Kaiyang Zhou (Hong Kong Baptist University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Developed a real-time streaming video LLM named Streamo, unifying decision-making and generation, and constructed a large instruction dataset called Streamo-Instruct-465K.
StreamingTOM: Streaming Token Compression for Efficient Video Understanding
Xueyi Chen (Westlake University), Huan Wang (Westlake University)
CompressionTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: Designed StreamingTOM, a training-agnostic two-stage framework specifically for real-time streaming video understanding; significantly reduces forward computation and memory usage through causal pre-LLM token compression and online kv cache quantization;
Streamlined Knowledge Distillation
Hyeon-Jin Jeong (Yonsei University), Seok-Hwan Choi (Yonsei University)
Knowledge DistillationImage
🎯 What it does: Propose a simplified knowledge distillation framework SKD that uses only two forms of knowledge: instance-level and direction-level, eliminating redundant steps of multi-knowledge alignment and relationship modeling.
Streamlined Open-Vocabulary Human-Object Interaction Detection
Chang Sun (South China University of Technology), Changxing Ding (South China University of Technology)
ClassificationObject DetectionTransformerContrastive LearningImage
🎯 What it does: Propose a single-model SL-HOI to achieve one-stage open-vocabulary human-object interaction detection, utilizing the DINOv3 backbone for localization and the vision head for interaction classification.
StreamRAG: Enhancing Real-Time Video Understanding with Retrieval Augmentation
Junlin Xie (Chinese University of Hong Kong), Guanbin Li (Sun Yat-sen University)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the StreamRAG framework to address three major challenges in real-time video question answering: event segmentation, knowledge extraction acceleration, and query-aware dynamic knowledge injection.
StreamReady: Learning What to Answer and When in Long Streaming Videos
Shehreen Azad (University of Central Florida), Yogesh S Rawat (University of Central Florida)
RetrievalTransformerVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed StreamReady, a timely answering framework for streaming video that provides answers only when sufficient evidence is present in the video stream; and designed the Answer Readiness Score (ARS) metric to evaluate timeliness.
StreamVLO: Streaming Visual-LiDAR Odometry with Cumulative Drift Compensation
Mengmeng Liu (University of Twente), Hao Cheng (University of Twente)
Pose EstimationAutonomous DrivingSimultaneous Localization and MappingOptical FlowMultimodalityPoint Cloud
🎯 What it does: This paper proposes StreamVLO, an end-to-end visual-LiDAR odometry framework capable of fusing multi-modal information and estimating pose under real-time streaming conditions.
STRNet: Visual Navigation with Spatio-Temporal Representation through Dynamic Graph Aggregation
Hao Ren (Sun Yat-sen University), Hui Cheng (Sun Yat-sen University)
Autonomous DrivingRobotic IntelligenceGraph Neural NetworkContrastive LearningImageVideo
🎯 What it does: Propose the STRNet framework to achieve unified fusion of spatial and temporal information in visual navigation;
Stronger Normalization-Free Transformers
Mingzhi Chen (Princeton University), Zhuang Liu (Princeton University)
ClassificationGenerationTransformerImageTextBiomedical DataAudio
🎯 What it does: Propose a new point function called Dynamic erf (Derf) to replace the normalization layers in Transformers.
Structural Action Transformer for 3D Dexterous Manipulation
Xiaohan Lei (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Robotic IntelligenceTransformerLarge Language ModelDiffusion modelTextPoint CloudTime Series
🎯 What it does: Propose converting action segmentation from traditional time series (T×D_a) to structured joint trajectory sequences (D_a×T), achieving cross-body skill transfer for diverse high-DoF robotic hands via embedded joint codebooks;
Structural Graph Probing of Vision-Language Models
Haoyu He (Northeastern University), Qi R. Wang (Northeastern University)
Explainability and InterpretabilityGraph Neural NetworkVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper constructs a neural correlation graph (neural topology) for each layer to study group-level interactions and structures within VLMs, and employs graph models for behavior prediction, structural alignment, and causal intervention analysis.
Structural-Semantic Perception for Diffusion-Guided Temporal Forgery Localization
Ligong Cao (National University of Defense Technology), Haoang Chi (National University of Defense Technology)
Anomaly DetectionTransformerDiffusion modelVideoMultimodality
🎯 What it does: Propose a unified structure-semantic aware and diffusion-guided temporal forgery localization framework capable of precisely locating tampered time segments in videos or audio-video content.
Structure-Aware Representation Distillation for Tiny-Dense Object Segmentation
Xuesong Liu, Emmett Ientilucci
SegmentationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkScore-based ModelImageBenchmark
🎯 What it does: Proposed a Structure-Aware Representation Distillation (SARD) framework for efficiently transferring knowledge of fine-grained dense object segmentation, emphasizing structural importance rather than direct mask imitation.
Structure-to-Intensity Diffusion for Adverse-Weather LiDAR Generation
Peiyang Ni (University of Electronic Science and Technology of China), Ping Hu (University of Electronic Science and Technology of China)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelPoint Cloud
🎯 What it does: Proposes a LiDAR inverse weather generation framework based on structure-to-strength diffusion (SiD) and real prior weather simulation (RPWS), capable of synthesizing point clouds with geometric consistency and realistic reflectance intensity under adverse weather conditions.
StructXLIP: Enhancing Vision-language Models with Multimodal Structural Cues
Zanxi Ruan (University of Verona), Marco Cristani (University of Verona)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a fine-tuning framework called StructXLIP, which enhances CLIP's alignment effectiveness by extracting structural information (edge maps and structured text) from both visual and textual modalities, thereby improving performance in long-text retrieval tasks.
STUR3D: Spatio-Temporal Unified Representation Learning for 3D Object Detection
Huijie Fan (Shenyang Institute of Automation, Chinese Academy of Sciences), Liangqiong Qu (Nanjing University of Posts and Telecommunications)
Object DetectionDepth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint CloudTime SeriesBenchmark
🎯 What it does: Designed and implemented a camera-based 3D object detection framework called STUR3D, which projects and propagates 2D and 3D detection results from time series, injects geometric depth information into the 2D detection head, and finally maps the improved 2D results to precise 3D queries for detection.
Style-GRPO: Semantic-Aware Preference Optimization for Image Style Transfer Guided by Reward Modeling
Jianbin Zhao (Peking University), Yonghong Tian (Peking University)
Image TranslationOptimizationSupervised Fine-TuningReinforcement LearningVision Language ModelFlow-based ModelImageText
🎯 What it does: Proposed a semantic-aware preference optimization framework called Style-GRPO based on a reward model for text-guided image style transfer.
StyleDoctor: Towards Specialist Reward Model for Style-centric Generation Tasks
Xilin He (Mohamed bin Zayed University of Artificial Intelligence), Muhammad Haris Khan (Mohamed bin Zayed University of Artificial Intelligence)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose StyleDoctor, which jointly evaluates style consistency between images and text, and uses it as a reward model to enhance the quality of style generation under reinforcement learning.
StyleGallery: Training-free and Semantic-aware Personalized Style Transfer from Arbitrary Image References
Boyu He (National University of Defense Technology), Zhiping Cai (National University of Defense Technology)
Image TranslationGenerationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: This paper proposes an adaptive style transfer framework called StyleGallery, which requires no training or external semantic masks. It automatically performs semantic segmentation and semantic matching between content images and an arbitrary number of style images, generating personalized stylized images by guiding diffusion sampling through regional style loss and global content loss.
StyleTextGen: Style-Conditioned Multilingual Scene Text Generation
Zeyu Chen (Nankai University), Yu Zhou (Nankai University)
GenerationTransformerVision Language ModelDiffusion modelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose StyleTextGen, a style-controlled generation framework for multi-language scene text, capable of accurately replicating the font, color, and texture of reference text in complex backgrounds.
Submodel Extraction for Efficient and Personalized Federated Learning via Optimal Transport
Zheng Jiang (Tsinghua University), Lifeng Sun (Tsinghua University)
OptimizationFederated LearningComputational EfficiencyImageTextTime Series
🎯 What it does: Proposes a server-side personalized pruning framework called SubFLOT based on optimal transport, which can achieve personalized training for different clients while ensuring model sparsity.