ICCV 2025 Papers — Page 14
IEEE/CVF International Conference on Computer Vision · 2701 papers
Large Multi-modal Models Can Interpret Features in Large Multi-modal Models
Kaichen Zhang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
Explainability and InterpretabilityTransformerLarge Language ModelAuto EncoderImageMultimodality
🎯 What it does: By embedding a Sparse Autoencoder (SAE) in LLaVA-NeXT-8B and combining it with another larger LMM (LLaVA-OV-72B), an automatic explanation pipeline is constructed to identify and explain the open semantic features within the LMM, while guiding model behavior by modulating feature activations.
Large Scene Generation with Cube-Absorb Discrete Diffusion
Qianjiang Hu (Peking University), Wei Hu (Peking University)
SegmentationGenerationData SynthesisAutonomous DrivingTransformerDiffusion modelPoint Cloud
🎯 What it does: By establishing a cubic absorption discrete diffusion model (CADD) and a sparse cubic attention Transformer (SCDT), hierarchical generation from low-resolution rough scenes to high-resolution fine-grained 3D outdoor scenes is achieved;
Large-scale Pre-training for Grounded Video Caption Generation
Evangelos Kazakos (Czech Technical University in Prague), Josef Sivic (Inria)
Object DetectionGenerationTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: This paper proposes a framework called GROVE that combines video subtitle generation and object localization, and designs a large-scale automatic annotation process, constructing the HowToGround1M pre-training dataset and the high-quality manually annotated dataset iGround.
Lark: Low-Rank Updates After Knowledge Localization for Few-shot Class-Incremental Learning
Jinxin Shi (East China Normal University), Liang He (East China Normal University)
ClassificationPose EstimationTransformerImage
🎯 What it does: A method is proposed that first locates model parameters that are easy to fine-tune, and then uses low-rank (Rank-1) updates to complete few-shot class incremental learning.
Latent Diffusion Models with Masked AutoEncoders
Junho Lee (Seoul National University), Joonseok Lee (Seoul National University)
GenerationData SynthesisCompressionTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposed and implemented Variational Masked AutoEncoders (VMAE) and integrated them into Latent Diffusion Models (LDMAE), achieving higher quality and more efficient image generation; by adding masked reconstruction and hierarchical compression based on VAE, it addresses the three major bottlenecks of traditional autoencoders in LDM; extensive experiments on LDMAE across various datasets demonstrate its superiority in generation quality, speed, and model size compared to existing methods.
Latent Expression Generation for Referring Image Segmentation and Grounding
Seonghoon Yu (Gwangju Institute of Science and Technology), Jeany Son (POSTECH)
RecognitionObject DetectionSegmentationTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This study proposes a visual-oriented task framework based on latent expression generation, capable of generating diverse latent expressions from a single text description and utilizing them for target image segmentation and detection.
Latent Swap Joint Diffusion for 2D Long-Form Latent Generation
Yusheng Dai (University of Science and Technology of China), Jiefeng Ma (iFlytek Research)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelAuto EncoderImageMultimodalityAudio
🎯 What it does: This paper proposes the Swap Forward (SaFa) framework, which utilizes two types of latent swap operators to achieve seamless integration of long spectrum/panoramic generation and cross-view consistency.
Latent-Reframe: Enabling Camera Control for Video Diffusion Models without Training
Zhenghong Zhou (University of Rochester), Jiebo Luo (University of Rochester)
GenerationData SynthesisPose EstimationDiffusion modelVideoTextPoint Cloud
🎯 What it does: By reconstructing and repairing the latent codes during the sampling phase of the pre-trained text-to-video diffusion model, we achieved untrained video generation with camera trajectory control.
LATINO-PRO: LAtent consisTency INverse sOlver with PRompt Optimization
Alessio Spagnoletti (University of Paris Cite), Marcelo Pereyra (Heriot-Watt University)
RestorationSuper ResolutionOptimizationPrompt EngineeringAuto EncoderImage
🎯 What it does: This paper proposes two zero-shot Plug-and-Play inference frameworks, LATINO and LATINO-PRO, using the Latent Consistency Model (LCM) as a prior, and addresses the inverse problem through Langevin sampling and text prompt self-calibration.
Latte: Collaborative Test-Time Adaptation of Vision-Language Models in Federated Learning
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
Domain AdaptationFederated LearningTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes Latte, an adaptive framework for collaborative testing of pre-trained vision-language models in a federated learning environment, which combines local memory and external shared memory for online adaptation.
LawDIS: Language-Window-based Controllable Dichotomous Image Segmentation
Xinyu Yan (Tianjin University), Deng-Ping Fan (Nankai University)
SegmentationDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes LawDIS, a controllable binary image segmentation method implemented using a stable diffusion model, supporting two modes: language prompts and window refinement.
Lay-Your-Scene: Natural Scene Layout Generation with Diffusion Transformers
Divyansh Srivastava (University of California San Diego), Zhuowen Tu (University of California San Diego)
Object DetectionGenerationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes an open vocabulary text-to-scene layout generation framework called Lay-Your-Scene, which is based on a lightweight open-source language model and a diffusion Transformer. It can automatically generate two-dimensional bounding box layouts that comply with spatial and counting constraints based on natural language prompts.
Lay2Story: Extending Diffusion Transformers for Layout-Togglable Story Generation
Ao Ma (JD.com, Inc.), Zhanjie Zhang (JD.com, Inc.)
GenerationTransformerDiffusion modelImageVideoTextBenchmark
🎯 What it does: A layout-switchable story generation task is proposed, and the Lay2Story model is implemented based on the Diffusion Transformers architecture; simultaneously, a large-scale comic video story dataset Lay2Story-1M and its evaluation benchmark Lay2Story-Bench are constructed.
Layer-wise Vision Injection with Disentangled Attention for Efficient LVLMs
Xuange Zhang (Beijing University of Technology), Tongtong Yuan (Beijing University of Technology)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A hierarchical visual injection and decoupled attention mechanism (LVIDA) is proposed, which significantly reduces the redundant propagation of visual information by retaining only the full forward propagation of text in the language decoder and injecting visual information at each layer.
LayerAnimate: Layer-level Control for Animation
Yuxue Yang (University of Chinese Academy of Sciences), Zhaoxiang Zhang (University of Chinese Academy of Sciences)
GenerationData SynthesisDiffusion modelOptical FlowVideoMultimodality
🎯 What it does: The LayerAnimate framework is proposed to achieve hierarchical animation video generation and multimodal hierarchical control, supporting the combined editing of hand-drawn sketches, motion trajectories, and motion scores.
LayerD: Decomposing Raster Graphic Designs into Layers
Tomoyuki Suzuki (CyberAgent), Kota Yamaguchi (CyberAgent)
Image TranslationSegmentationVision Language ModelImage
🎯 What it does: An end-to-end automated framework called LayerD has been implemented, which can decompose raster graphic designs into editable layers.
LayerLock: Non-collapsing Representation Learning with Progressive Freezing
Goker Erdogan (Google DeepMind), Joao Carreira (Google DeepMind)
Depth EstimationRepresentation LearningTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: Proposes the LayerLock method, which freezes the Transformer layer by layer and dynamically switches prediction targets from pixel-level to deeper latent predictions;
LayerTracer: Cognitive-Aligned Layered SVG Synthesis via Diffusion Transformer
Yiren Song (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodality
🎯 What it does: The LayerTracer framework is proposed to achieve hierarchical SVG generation and vectorization from text or images for cognitive alignment.
LazyMAR: Accelerating Masked Autoregressive Models via Feature Caching
Feihong Yan (Beijing Institute of Technology), Linfeng Zhang (Shanghai Jiaotong University)
GenerationTransformerImage
🎯 What it does: This paper proposes LazyMAR, which accelerates the inference of the Masked Autoregressive (MAR) image generation model through feature caching technology.
LBM: Latent Bridge Matching for Fast Image-to-Image Translation
Clément Chadebec (Jasper Research), Benjamin Aubin (Jasper Research)
Image TranslationRestorationDepth EstimationAuto EncoderImageStochastic Differential Equation
🎯 What it does: This paper proposes a Latent Bridge Matching (LBM) method for bridge matching in latent space, aimed at achieving fast image-to-image translation in a single inference.
LD-RPS: Zero-Shot Unified Image Restoration via Latent Diffusion Recurrent Posterior Sampling
Huaqiu Li (Tsinghua University), Xiangxiang Chu (Alibaba Group)
RestorationLarge Language ModelDiffusion modelImage
🎯 What it does: A recursive posterior sampling framework based on latent diffusion models, LD-RPS, is proposed, which can achieve various image restoration tasks without the need for a training set and under zero-shot conditions.
LDIP: Long Distance Information Propagation for Video Super-Resolution
Michael Bernasconi (ETH Zurich), Christopher Schroers (Disney Research Studios)
RestorationSuper ResolutionOptical FlowVideo
🎯 What it does: A long-distance information propagation framework (LDIP) is proposed, which effectively transmits cross-frame long-distance information in video super-resolution (VSR) through a Long-Distance Reference Fusion module (LRRF), and optionally integrates additional high-resolution reference images; at the same time, arbitrary scale super-resolution is achieved by replacing pixel rearrangement with a custom scale module.
LDPose: Towards Inclusive Human Pose Estimation for Limb-Deficient Individuals in the Wild
Jiaying Ying (University of Queensland), Xin Yu (Australian National University)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proposes the task of 'human pose estimation for individuals with limb loss' and constructs the first pose dataset, LDPose, specifically for this population. It designs specialized key points for limb loss, a limb loss loss function (LDLoss), and evaluation metrics (LD Metrics) tailored for limb loss. The effectiveness of the task and dataset is validated through fine-tuning and training with LD Loss on mainstream methods such as YOLOPose, RTMPose, and ViTPose.
LeanVAE: An Ultra-Efficient Reconstruction VAE for Video Diffusion Models
Yu Cheng (Zhejiang University), Fajie Yuan (Westlake University)
GenerationCompressionComputational EfficiencyDiffusion modelAuto EncoderVideo
🎯 What it does: We propose LeanVAE, an efficient video VAE for video compression and decoding in Latent Video Diffusion Models (LVDM), significantly reducing FLOPs while maintaining high quality;
Leaps and Bounds: An Improved Point Cloud Winding Number Formulation for Fast Normal Estimation and Surface Reconstruction
Chamin Hewa Koneputugodage (Australian National University), Stephen Gould (Australian National University)
Point Cloud
🎯 What it does: This paper proposes a bounded point cloud winding number approximation (w_B(q)) and applies it to globally consistent normal estimation and surface reconstruction.
Learn2Synth: Learning Optimal Data Synthesis Using Hypergradients for Brain Image Segmentation
Xiaoling Hu (Massachusetts General Hospital and Harvard Medical School), Yaël Balbastre (University College London)
SegmentationData SynthesisDomain AdaptationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: We propose Learn2Synth, a framework that trains a segmentation network using only synthetic data while optimizing the synthetic data augmentation network through hypergradient optimization, allowing the model to achieve better performance on real data.
Learnable Feature Patches and Vectors for Boosting Low-light Image Enhancement without External Knowledge
Xiaogang Xu (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: A lightweight learnable feature patch and vector (LFPVs) is proposed as a reference to supplement the knowledge not learned by low-light image enhancement models, thereby improving the enhancement effect.
Learnable Fractional Reaction-Diffusion Dynamics for Under-Display ToF Imaging and Beyond
Xin Qiao (Xi'an Jiaotong University), Stefano Mattoccia (Anyang Institute of Technology)
RestorationDepth EstimationSuper ResolutionConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: A learnable fractional-order reaction-diffusion dynamics framework (LFRD) that combines physical models is proposed, which enhances the quality of Under-Display ToF depth maps through iterative optimization and validates its effectiveness in a broader range of depth recovery tasks.
Learnable Logit Adjustment for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch
Hyuck Lee (Krafton), Heeyoung Kim (KAIST)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An algorithm for learning adjustable Logit, called LLA, is proposed, specifically targeting the mismatch in class distribution between labeled and unlabeled data in class-imbalanced semi-supervised learning. It improves the quality of pseudo-labels by adaptively adjusting Logit and achieves fair predictions during testing. Additionally, an extended feature clustering compression (EFCC) is introduced to enhance the density and separability of minority class features.
Learnable Retrieval Enhanced Visual-Text Alignment and Fusion for Radiology Report Generation
Qin Zhou (East China University of Science and Technology), Sai Wu (Zhejiang University)
GenerationRetrievalTransformerVision Language ModelImageTextMultimodalityElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: A learnable retrieval-enhanced visual-text alignment and fusion framework (REVTAF) is proposed for generating radiology reports from lung X-ray images.
Learned Image Compression with Hierarchical Progressive Context Modeling
Yuqi Li (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
CompressionConvolutional Neural NetworkImage
🎯 What it does: A Hierarchical Progressive Context Model (HPCM) is proposed for end-to-end learned image compression.
Learning 3D Object Spatial Relationships from Pre-trained 2D Diffusion Models
Sangwon Baik, Hanbyul Joo
GenerationData SynthesisLarge Language ModelDiffusion modelScore-based ModelImagePoint Cloud
🎯 What it does: Utilizing a pre-trained 2D diffusion model to synthesize images and elevate them to 3D samples, learning and generating spatial relationships between object pairs (OOR), a score-based text-conditioned OOR diffusion model is proposed.
Learning 3D Scene Analogies with Neural Contextual Scene Maps
Junho Kim (Seoul National University), Young Min Kim (Seoul National University)
RecognitionDomain AdaptationTransformerContrastive LearningPoint CloudMesh
🎯 What it does: This study investigates methods for recognizing similar contextual regions in 3D scenes and generating dense mappings (3D scene analogs).
Learning 4D Embodied World Models
Haoyu Zhen (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
GenerationData SynthesisRobotic IntelligenceTransformerReinforcement LearningDiffusion modelWorld ModelOptical FlowVideoPoint Cloud
🎯 What it does: Constructed and trained a 4D embodied world model TesserAct, capable of generating RGB, depth, and normal videos given an initial image and text instructions, and subsequently reconstructing them into a spatiotemporally consistent 4D scene, which is then used for inverse dynamics motion planning.
Learning A Unified Template for Gait Recognition
Panjian Huang (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
RecognitionRepresentation LearningTransformerDiffusion modelImage
🎯 What it does: A unified template called Origins is designed to achieve generative and representation learning in gait recognition, addressing issues of semantic inconsistency and unification.
Learning an Implicit Physics Model for Image-based Fluid Simulation
Emily Yue-Ting Jia (University of Southern California), Yue Wang (University of Southern California)
GenerationData SynthesisGaussian SplattingOptical FlowVideoPhysics Related
🎯 What it does: This paper proposes a method for generating physically consistent four-dimensional (3D + time) animations from a single natural fluid image.
Learning Beyond Still Frames: Scaling Vision-Language Models with Video
Yiyuan Zhang (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
ClassificationRetrievalTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: This paper presents LLaVA-Prime, which combines video pre-training to extend the visual language model of LLaVA, using Causal Hierarchical Aggregation (CHA) to efficiently process videos and enhance temporal understanding of vision and language.
Learning Counterfactually Decoupled Attention for Open-World Model Attribution
Yu Zheng (Tsinghua University), Jie Zhou (Tsinghua University)
ClassificationGenerationAnomaly DetectionConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: Proposed Counterfactually Decoupled Attention Learning (CDAL), which separates model-specific traces from source content bias through causal intervention, thereby enhancing the attribution performance of open-world models.
Learning Deblurring Texture Prior from Unpaired Data with Diffusion Model
Chengxu Liu (Xi'an Jiaotong University), Ming-Hsuan Yang (University of California)
RestorationTransformerDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: The paper proposes an unsupervised image deblurring framework TP-Diff based on diffusion models, which can learn spatially varying texture priors from unpaired data and assist in deblurring through this prior.
Learning Dense Feature Matching via Lifting Single 2D Image to 3D Space
Yingping Liang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
Object DetectionData SynthesisDepth EstimationGaussian SplattingImage
🎯 What it does: A two-stage framework L2M is proposed, which elevates single-view 2D images to 3D space, utilizes multi-view synthesis and 3D Gaussian features to train a 3D perception encoder, and employs new view rendering with large-scale synthetic data to train a robust feature decoder for dense feature matching.
Learning Efficient and Generalizable Human Representation with Human Gaussian Model
Yifan Liu (Tsinghua University), Yueqi Duan (Tsinghua University)
Pose EstimationRepresentation LearningGraph Neural NetworkTransformerGaussian SplattingVideoMesh
🎯 What it does: By constructing a dual-layer Human Gaussian Graph, the 3D Gaussians predicted from multiple frames in the video are associated with the SMPL mesh, enabling the generation of animatable human Gaussian representations during inference.
Learning Few-Step Diffusion Models by Trajectory Distribution Matching
Yihong Luo (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
GenerationKnowledge DistillationDiffusion modelScore-based ModelImageVideoTextOrdinary Differential Equation
🎯 What it does: This paper proposes a unified diffusion model distillation framework TDM, which can generate high-quality images in just 4 steps and directly distill performance exceeding that of the teacher model without the need for additional image data.
Learning Hierarchical Line Buffer for Image Processing
Jiacheng Li (Sony Research), Daisuke Iso (Sony Research)
RestorationSuper ResolutionConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: A learnable hierarchical line buffer network is proposed for the on-chip line buffer scenario, utilizing layer-wise expanding strip convolution for row-wise feature enhancement, and efficiently capturing inter-row associations through a hierarchical compressed Long Short-Term Memory (LSTM) structure.
Learning Implicit Features with Flow-Infused Transformations for Realistic Virtual Try-On
Delong Zhang (Sun Yat-sen University), Wei Zhang (Shopee)
Image TranslationGenerationDiffusion modelOptical FlowImageBenchmark
🎯 What it does: A virtual try-on framework called FIA-VTON based on flow-infused attention is proposed, utilizing dense optical flow as implicit guidance to achieve precise garment deformation and detail reconstruction in a latent diffusion model.
Learning Interpretable Queries for Explainable Image Classification with Information Pursuit
Stefan Kolek (LMU Munich), René Vidal (University of Pennsylvania)
ClassificationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImage
🎯 What it does: The paper proposes a method for interpretable image classification by learning an interpretable query dictionary in the context of Information Tracking (IP) using the semantic embedding space of CLIP.
Learning Large Motion Estimation from Intermediate Representations with a High-Resolution Optical Flow Dataset Featuring Long-Range Dynamic Motion
Hoonhee Cho (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Knowledge DistillationOptical FlowImageVideo
🎯 What it does: A high-resolution 4K long-distance optical flow dataset, RelayFlow-4K, is proposed, along with a training strategy based on intermediate frames (matching cost distillation, incremental time step learning, and matching distance loss) to enhance high-resolution long-distance optical flow estimation.
Learning Neural Scene Representation from iToF Imaging
Wenjie Chang (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Depth EstimationRepresentation LearningNeural Radiance FieldPoint Cloud
🎯 What it does: Utilizing the amplitude and phase information from a multi-view iToF sensor, high-precision 3D reconstruction is achieved through neural implicit surface representation.
Learning Normal Flow Directly From Events
Dehao Yuan (University of Maryland), Cornelia Fermüller (University of Maryland)
Pose EstimationDomain AdaptationOptical FlowPoint Cloud
🎯 What it does: This paper proposes a supervised learning method based on point cloud encoding, which directly estimates the normal flow of each event from the raw events and uses it for camera motion estimation.
Learning Normals of Noisy Points by Local Gradient-Aware Surface Filtering
Qing Li (Southwest Jiaotong University), Yu-Shen Liu (Tsinghua University)
RestorationPoint Cloud
🎯 What it does: This paper proposes an unsupervised point cloud normal estimation method based on local gradient-aware surface filtering, which can directly recover high-quality normal vectors from noisy point clouds and achieve surface reconstruction and denoising.
Learning Null Geodesics for Gravitational Lensing Rendering in General Relativity
Mingyuan Sun (Northeastern University), Renjing Xu (Hong Kong University of Science and Technology)
ImagePhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper presents GravLensX, a method that utilizes physics-informed neural networks to learn the light geodesics in the spacetime around black holes, enabling efficient gravitational lens rendering.
Learning on the Go: A Meta-learning Object Navigation Model
Xiaorong Qin (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)
Meta LearningReinforcement LearningMultimodality
🎯 What it does: This paper proposes a meta-learning framework LOG that can learn in real-time during navigation, aligning environment-specific trajectory distributions with learned center distributions, enabling agents to quickly adapt and achieve target object navigation in unknown environments.
Learning Pixel-adaptive Multi-layer Perceptrons for Real-time Image Enhancement
Junyu Lou (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
RestorationImage
🎯 What it does: This paper proposes a pixel-adaptive multilayer perceptron (BPAM) framework based on a bilateral grid for real-time image enhancement.
Learning Precise Affordances from Egocentric Videos for Robotic Manipulation
Gen Li (University of Edinburgh), Laura Sevilla-Lara (University of Edinburgh)
SegmentationRobotic IntelligenceTransformerSupervised Fine-TuningVideo
🎯 What it does: A complete system has been constructed for automatically collecting precise grasping and functional component annotations from self-camera videos, training a geometry-guided affordance Transformer, and implementing a robotic grasping and interaction system based on it.
Learning Robust Image Watermarking with Lossless Cover Recovery
Jiale Chen (Beijing Institute of Technology), Xiping Hu (Beijing Institute of Technology)
RestorationCompressionAuto EncoderImage
🎯 What it does: A recoverable watermarking method named CRMark is proposed, achieving robust watermark embedding for images and complete recovery under lossless channels.
Learning Robust Stereo Matching in the Wild with Selective Mixture-of-Experts
Yun Wang (City University of Hong Kong), Junjie Hu (Chinese University of Hong Kong)
Depth EstimationDomain AdaptationAutonomous DrivingTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes SMoEStereo, a robust stereo matching framework that adapts to different scenes by introducing selective mixture of experts (MoE) and low-rank adaptation (LoRA) into the visual foundation model (VFM) along with a lightweight decision network.
Learning Separable Fine-Grained Representation via Dendrogram Construction from Coarse Labels for Fine-grained Visual Recognition
Guanghui Shi (Xidian University), Xiaoyu Lin (Xidian University)
RecognitionRepresentation LearningContrastive LearningImage
🎯 What it does: A framework called BuCSFR is proposed, which learns separable fine-grained representations using only coarse labels. It utilizes a bottom-up constructed hierarchical structure (dendrogram) to generate pseudo fine-grained labels for contrastive learning.
Learning Streaming Video Representation via Multitask Training
Yibin Yan (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
RecognitionSegmentationRetrievalTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes StreamFormer, a visual Transformer backbone network capable of processing real-time video streams frame by frame while maintaining historical context. It is jointly trained through a multi-task visual-language alignment framework to enhance the understanding of global semantics, temporal dynamics, and fine-grained spatial relationships.
Learning to Generalize without Bias for Open-Vocabulary Action Recognition
Yating Yu (Northwestern Polytechnical University), Yanning Zhang (Institute of Automation)
RecognitionMeta LearningTransformerContrastive LearningVideo
🎯 What it does: This paper proposes an open vocabulary action recognition framework called Open-MeDe based on meta-learning, which significantly suppresses the static bias of CLIP through cross-batch meta-optimization and Gaussian self-ensemble, achieving stronger known-to-open generalization and debiasing from image to video.
Learning to Inference Adaptively for Multimodal Large Language Models
Zhuoyan Xu (University of Wisconsin Madison), Yin Li (University of Wisconsin Madison)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper proposes AdaLLaVA, a learnable scheduler that dynamically reconfigures the computation path of multi-modal large language models (MLLM) during inference based on input content and latency budget, thereby maintaining high accuracy while ensuring latency constraints.
Learning to See in the Extremely Dark
Hai Jiang (Sichuan University), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a paired-to-paired data synthesis process to generate extremely low-light RAW images and high-quality sRGB reference images, constructing the SIED dataset, and enhancing RAW images in extremely dark scenes based on a multi-stage framework of diffusion models.
Learning to See Inside Opaque Liquid Containers using Speckle Vibrometry
Matan Kichler (Weizmann Institute of Science), Mark Sheinin (Weizmann Institute of Science)
ClassificationRecognitionTransformerImage
🎯 What it does: A system based on laser speckle vibration sensing is studied, which can estimate the internal liquid level of a container by remotely observing the micro-vibrations on the surface of the container without touching it.
Learning to Unlearn while Retaining: Combating Gradient Conflicts in Machine Unlearning
Gaurav Patel (Purdue University), Qiang Qiu (Purdue University)
GenerationOptimizationData-Centric LearningDiffusion modelImage
🎯 What it does: A new machine unlearning framework LUR is designed, which implicitly aligns the gradients of retention and forgetting by performing an intermediate gradient update on the retention objective before updating the forgetting objective, thereby reducing gradient conflicts and effectively deleting specified data while preserving the model's performance on the remaining data.
Learning Visual Hierarchies in Hyperbolic Space for Image Retrieval
Ziwei Wang (Amazon), Thalaiyasingam Ajanthan (Amazon)
Object DetectionRetrievalContrastive LearningImage
🎯 What it does: This paper proposes a new learning paradigm that utilizes hyperbolic space contrastive learning methods to embed user-defined multi-level visual hierarchies (image → object → part) into latent space without explicit hierarchical labels. By constructing inclusion relationship pairs (image contains box, hierarchical relationships between boxes, same category boxes across images as positive samples) and employing angular contrastive loss, the model can reflect the hierarchy in retrieval tasks. A hierarchical retrieval evaluation metric based on optimal transport is also proposed.
Learning Visual Proxy for Compositional Zero-Shot Learning
Shiyu Zhang (Tianjin University), Wenjun Wang (Tianjin University)
ClassificationRecognitionObject DetectionTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Based on the construction of a visual proxy center, cross-modal joint learning is implemented in conjunction with text prototypes to enhance the discrimination and generalization capabilities of compositional zero-shot learning.
Learning Yourself: Class-Incremental Semantic Segmentation with Language-Inspired Bootstrapped Disentanglement
Ruitao Wu (Beihang University), Jia Li (Beihang University)
ClassificationSegmentationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes a CLIP-based language-guided separation framework for the class-incremental semantic segmentation (CISS) task, utilizing text priors to achieve adaptive decoupling of prototype-feature and background-incremental semantic entanglement.
LEGION: Learning to Ground and Explain for Synthetic Image Detection
Hengrui Kang (Shanghai Jiao Tong University), Conghui He (Shanghai Artificial Intelligence Laboratory)
Object DetectionSegmentationGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper presents a new dataset called SynthScars and a multimodal large language model framework named LEGION, designed for high-precision detection, localization, and explanation of forgery traces in synthetic images, using LEGION as a controller to improve image generation quality.
LEGO-Maker: A Semantic-Driven Algorithm for Text-to-3D Generation
Yifei Zhang (Beijing Institute of Technology), Lei Chen (Beijing Institute of Technology)
GenerationDiffusion modelMesh
🎯 What it does: A three-stage text-to-3D generation framework named LEGO-Maker is proposed, which first generates the target image, then performs functional semantic decomposition, and finally constructs interactive 3D models through multi-task 3D generation and module fusion.
LeGrad: An Explainability Method for Vision Transformers via Feature Formation Sensitivity
Walid Bousselham (University of Tuebingen), Hilde Kuehne (University of Tuebingen)
Explainability and InterpretabilityTransformerImageAudio
🎯 What it does: An interpretable method named LeGrad is proposed, which performs sensitivity analysis on the attention maps of Vision Transformers (ViT) using gradients and aggregates multi-layer information to generate high-quality heatmaps.
Less is More: Empowering GUI Agent with Context-Aware Simplification
Gongwei Chen (Harbin Institute of Technology), Liqiang Nie (Huawei Noah's Ark Lab)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: A context-aware simplified GUI agent, SimpAgent, is proposed, which enhances efficiency and effectiveness through element occlusion removal and historical visual information compression.
Less is More: Improving Motion Diffusion Models with Sparse Keyframes
Jinseok Bae (Seoul National University), Mubbasir Kapadia (Roblox)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: A sparse keyframe diffusion model is designed, utilizing keyframe masks and linear interpolation to replace full-frame processing, significantly reducing training and inference costs.
Less Static, More Private: Towards Transferable Privacy-Preserving Action Recognition by Generative Decoupled Learning
Zhi-Wei Xia (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
RecognitionDomain AdaptationSafty and PrivacyAuto EncoderGenerative Adversarial NetworkVideo
🎯 What it does: A privacy-preserving action recognition framework called GenPriv based on generative decoupled learning is proposed, achieving cross-domain transfer on labeled videos in the source domain and unlabeled videos in the target domain.
Less-to-More Generalization: Unlocking More Controllability by In-Context Generation
Shaojin Wu (ByteDance), Qian He (ByteDance)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: The UNO framework is proposed, achieving high consistency and controllability in image generation under single/multiple subject conditions through model-data co-evolution, evolutionary cross-modal alignment, and full rotation position embedding.
Leveraging 2D Priors and SDF Guidance for Urban Scene Rendering
Siddharth Tourani (MBZUAI), Muhammad Haris Khan (MBZUAI)
Object TrackingGenerationDepth EstimationAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This work proposes a dynamic urban scene rendering method that combines SDF and 3D Gaussian Splatting, achieving high-precision dynamic object reconstruction and novel view synthesis without LiDAR and 3D motion annotations by utilizing 2D depth and point tracking information.
Leveraging BEV Paradigm for Ground-to-Aerial Image Synthesis
Junyan Ye (Sun Yat-Sen University), Conghui He (Sensetime Research)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes the SkyDiffusion method, which utilizes BEV perspective transformation and diffusion models to synthesize realistic and content-consistent aerial images from ground street view images.
Leveraging Debiased Cross-modal Attention Maps and Code-based Reasoning for Zero-shot Referring Expression Comprehension
Juntao Chen (Tongji University), Hongyun Zhang (Tongji University)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a zero-shot referential expression understanding framework that scores candidate boxes through interpretable programmatic reasoning and debiased cross-modal attention, achieving unsupervised object localization.
Leveraging Local Patch Alignment to Seam-cutting for Large Parallax Image Stitching
Tianli Liao (Henan University of Technology), Heling Cao (Henan University of Technology)
Image TranslationOptimizationOptical FlowImage
🎯 What it does: A Local Patch Alignment Module (LPAM) is proposed, which improves the quality of seam-cutting in large disparity image stitching by performing local SIFT-flow alignment and re-cropping on patches containing low-quality pixels at the seam.
Leveraging Panoptic Scene Graph for Evaluating Fine-Grained Text-to-Image Generation
Xueqing Deng (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: This paper proposes PSG-Bench and PSG-Score, which respectively construct a benchmark dataset containing 5,000 complex text prompts and design evaluation metrics based on panoramic scene graphs for systematically assessing the generation quality and text consistency of text-to-image models.
Leveraging Prior Knowledge of Diffusion Model for Person Search
Giyeol Kim (Chung-Ang University), Chanho Eom (Seoul National University)
RecognitionObject DetectionRetrievalDiffusion modelImage
🎯 What it does: Proposes the DiffPS framework, which utilizes knowledge from pre-trained diffusion models for person search.
Leveraging Spatial Invariance to Boost Adversarial Transferability
Zihan Zhou (Shanghai University), Guorui Feng (University of Shanghai for Science and Technology)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposes an attack method utilizing spatially invariant multi-scale and multi-position input transformations (SID) to enhance the transferability of adversarial samples against DNNs.
Leveraging the Power of MLLMs for Gloss-Free Sign Language Translation
Jungeun Kim (Yonsei University), Ha Young Kim (Yonsei University)
RecognitionGenerationTransformerLarge Language ModelPrompt EngineeringContrastive LearningVideoTextMultimodality
🎯 What it does: A gloss-free sign language translation framework called MMSLT is constructed, utilizing a multimodal large language model (MLLM) to generate sign language descriptions and fuse them with video features, ultimately outputting spoken sentences.
LGA-Net: Learning Local and Global Affinities for Sparse Scribble based Image Colorization
Hongjin Lyu (Cardiff University), Yu-Kun Lai (Cardiff University)
Image TranslationGenerationConvolutional Neural NetworkImage
🎯 What it does: LGA-Net is proposed to complete image coloring of sparse doodles by learning local and global pixel affinity relationships, rather than directly predicting color values; subsequently, color propagation is achieved by solving the maximum a posteriori problem using the Laplacian prior.
LHM: Large Animatable Human Reconstruction Model for Single Image to 3D in Seconds
Lingteng Qiu (Alibaba Group), Liefeng Bo (Alibaba Group)
GenerationData SynthesisPose EstimationTransformerGaussian SplattingVideo
🎯 What it does: A method is proposed to generate animatable 3D human models from a single image within seconds, using 3D Gaussian splatting representation, and directly completing texture and pose recovery in forward inference.
Liberated-GS: 3D Gaussian Splatting Independent from SfM Point Clouds
Weihong Pan (Zhejiang University), Guofeng Zhang (Zhejiang University)
Image TranslationData SynthesisDepth EstimationGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a 3D Gaussian flattening initialization framework called Librated-GS, which does not rely on SfM point clouds. It utilizes monocular depth priors, depth alignment, and progressive segmentation initialization to achieve high-quality view synthesis for dense image sequences.
Lidar Waveforms are Worth 40x128x33 Words
Dominik Scheuble (Mercedes-Benz AG), Felix Heide (Princeton University)
Object DetectionGenerationData SynthesisSuper ResolutionTransformerPoint Cloud
🎯 What it does: A Transformer-based neural DSP is proposed, which utilizes neighboring waveform features to process complete LiDAR waveforms simultaneously, directly outputting high-fidelity multi-echo point clouds and supporting super-resolution generation.
LIFT: Latent Implicit Functions for Task- and Data-Agnostic Encoding
Amirhossein Kazerouni (University of Toronto), Babak Taati (University of Toronto)
ClassificationRestorationGenerationSuper ResolutionMeta LearningDiffusion modelImagePoint Cloud
🎯 What it does: This paper proposes a multi-scale implicit function framework named LIFT, which efficiently encodes arbitrary data domains using a hierarchical latent generator and local parallel MLP through a meta-learning approach, achieving unified processing for generation, classification, and single-task inverse problems.
Lifting the Structural Morphing for Wide-Angle Images Rectification: Unified Content and Boundary Modeling
Wenting Luan (National University of Defense Technology), Kang Liao (Nanyang Technological University)
RestorationObject DetectionSegmentationImage
🎯 What it does: A one-stage ConBo-Net is proposed, which jointly corrects the distortion of wide-angle lenses and the image boundaries, achieving both geometric accuracy and rectangular boundaries for distorted images.
Light-A-Video: Training-free Video Relighting via Progressive Light Fusion
Yujie Zhou (Shanghai Jiao Tong University), Li Niu (Shanghai AI Laboratory)
RestorationGenerationDiffusion modelAuto EncoderVideo
🎯 What it does: Using a pre-trained image relighting model and a video diffusion model, we achieved untrained video relighting that unifies video lighting while maintaining motion.
LightBSR: Towards Lightweight Blind Super-Resolution via Discriminative Implicit Degradation Representation Learning
Jiang Yuan (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
RestorationSuper ResolutionKnowledge DistillationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A lightweight blind super-resolution model LightBSR is proposed, emphasizing the discernibility of implicit degradation representation (IDR);
LightCity: An Urban Dataset for Outdoor Inverse Rendering and Reconstruction under Multi-illumination Conditions
Jingjing Wang (State Key Laboratory of Computer Aided Design and Computer Graphics), Guofeng Zhang (State Key Laboratory of Computer Aided Design and Computer Graphics)
RestorationData SynthesisDomain AdaptationNeural Radiance FieldImageBenchmark
🎯 What it does: This paper constructs a high-quality synthetic urban dataset called LightCity and conducts benchmark evaluations on three major tasks: inverse rendering, intrinsic decomposition, and scene reconstruction based on this dataset.
LightsOut: Diffusion-based Outpainting for Enhanced Lens Flare Removal
Shr-Ruei Tsai (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)
RestorationGenerationDiffusion modelImage
🎯 What it does: This paper presents LightsOut, a diffusion model-based image extrapolation preprocessor designed to recover missing or off-screen light sources in images, significantly enhancing single-image glare removal (SIFR) performance.
LightSwitch: Multi-view Relighting with Material-guided Diffusion
Yehonathan Litman (Carnegie Mellon University), Shubham Tulsiani (Carnegie Mellon University)
RestorationGenerationDiffusion modelImage
🎯 What it does: This paper proposes the LightSwitch framework, which can achieve multi-view consistent relighting of any number of input images by utilizing inferred material information and multi-view attention.
Lightweight and Fast Real-time Image Enhancement via Decomposition of the Spatial-aware Lookup Tables
Wontae Kim (Seoul National University), Nam Ik Cho (Seoul National University)
Computational EfficiencySupervised Fine-TuningImage
🎯 What it does: A lightweight real-time image enhancement framework is proposed by decomposing 3D LUT into 2D LUT and utilizing SVD compression, balancing spatial awareness and achieving faster inference.
Lightweight Gradient-Aware Upscaling of 3D Gaussian Splatting Images
Simon Niedermayr (Technical University of Munich), Rüdiger Westermann
Image TranslationRestorationSuper ResolutionComputational EfficiencyGaussian SplattingImage
🎯 What it does: A 3D Gaussian Splatting (3DGS) image upsampling method for lightweight GPUs is proposed, utilizing Gaussian analytical gradients to achieve cubic spline interpolation.
LINR-PCGC: Lossless Implicit Neural Representations for Point Cloud Geometry Compression
Wenjie Huang (Shanghai Jiao Tong University), Zhu Li (University of Missouri-Kansas City)
CompressionPoint Cloud
🎯 What it does: A point cloud geometry lossless compression framework LINR-PCGC based on Implicit Neural Representation (INR) has been designed and implemented, supporting efficient encoding and decoding of point cloud sequences.
LiON-LoRA: Rethinking LoRA Fusion to Unify Controllable Spatial and Temporal Generation for Video Diffusion
Yisu Zhang (Zhejiang University), Jianke Zhu (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes the LiON-LoRA framework, which rethinks the integration method of LoRA to achieve precise and controllable generation of camera trajectories and object movements in video diffusion models.
LIRA: Inferring Segmentation in Large Multi-modal Models with Local Interleaved Region Assistance
Zhang Li (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A new framework called LIRA is proposed, which enhances segmentation accuracy of large models while maintaining text understanding and reducing hallucination generation.
LIRA: Reasoning Reconstruction via Multimodal Large Language Models
Zhen Zhou (Institute of Automation, Chinese Academy of Sciences), Fengshui Jing (Institute of Automation, Chinese Academy of Sciences)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelMultimodalityPoint CloudBenchmark
🎯 What it does: An online 3D reconstruction task aimed at implicit complex instructions is proposed, along with the LIRA framework to achieve this task.
LiT: Delving into a Simple Linear Diffusion Transformer for Image Generation
Jiahao Wang (Tsinghua University), Ping Luo
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelImage
🎯 What it does: Transform the pre-trained Diffusion Transformer (DiT) into a fully linear attention model LiT, achieving faster inference and lower memory usage without significant loss of performance.
LLaFEA: Frame-Event Complementary Fusion for Fine-Grained Spatiotemporal Understanding in LMMs
Hanyu Zhou (National University of Singapore), Gim Hee Lee (National University of Singapore)
TransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: The LLaFEA framework is proposed, which fuses video information from event cameras and frame cameras to achieve fine-grained spatiotemporal reasoning in LMM.
LLaVA-3D: A Simple yet Effective Pathway to Empowering LMMs with 3D Capabilities
Chenming Zhu (University of Hong Kong), Xihui Liu (University of Hong Kong)
RecognitionObject DetectionSegmentationGenerationTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
🎯 What it does: This paper proposes a simple and efficient framework called LLaVA-3D, which can extend existing 2D multimodal large models (LLaVA) to a model capable of understanding and reasoning about 3D scenes, supporting tasks such as 3D question answering, dense descriptions, and visual localization, while maintaining the original 2D video understanding capabilities.