IEEE/CVF Conference on Computer Vision and Pattern Recognition · 1047 papers
Bridging the 2D-3D Gap: A Hierarchical Semantic-Geometric Map for Vision Language Navigation
Kailing Li (East China Normal University), Liang He (East China Normal University)
CodeAutonomous DrivingRobotic IntelligenceTransformerVision Language ModelSimultaneous Localization and MappingImageTextMultimodalityPoint CloudBenchmarkChain-of-Thought
🎯 What it does: Proposes a Hierarchical Semantic-Geometric Map (HSGM), decoupling high-level semantic planning from low-level classical A* path planning by converting 3D environments into 2D BEV maps understandable by VLM;
🎯 What it does: Studied the problem of non-stationary modality imbalance in multi-modal continuous action quality assessment, proposing the BriMA method, which includes a memory-guided bridging completion module and a modality-aware replay module.
🎯 What it does: Designed a zero-shot, no-training 3D point cloud registration framework called C-GenReg, which generates multi-view consistent RGB images using a world generative model, extracts correspondences with a vision foundation model tailored for matching, and fuses them with geometric branch features.
CodeObject DetectionAnomaly DetectionTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Propose DualXrayBench and the GSR model, exploring the treatment of dual-view X-ray images as a modality similar to language to enhance prohibited item detection and cross-perspective reasoning.
🎯 What it does: Proposed a two-stage reference-based color grading framework called CanonCGT, which first canonicalizes the input image and then matches it to the reference image's color tone.
CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis
Di Zhang (Xi'an Jiaotong University), Zeyu Gao (University of Cambridge)
CodeClassificationSegmentationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes CARE, a foundational model for whole-slide images, which employs adaptive region partitioning and leverages molecular data to guide cross-modal pretraining, achieving efficient and clinically interpretable pathological analysis.
🎯 What it does: Achieve category-agnostic 4D (spatiotemporal) human-object interaction reconstruction from monocular RGB videos, outputting size, shape, pose, and contact information while maintaining spatial and temporal consistency throughout the entire video.
CC-VQA: Conflict- and Correlation-Aware Method for Mitigating Knowledge Conflict in Knowledge-Based Visual Question Answering
Yuyang Hong (School of Artificial Intelligence, UCAS), Jieping Ye (Alibaba Cloud Computing)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes a training-free framework that combines visual-centric conflict reasoning with relevance-guided encoding and decoding to alleviate conflicts between parameter knowledge and retrieved context in knowledge-based visual question answering.
CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning
Zhijiang Tang, Jianqiang Huang (Chinese Academy Of Sciences)
CodeGenerationLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose the CCCaption framework, which utilizes dual-reward reinforcement learning to achieve completeness and correctness of image captions, and constructs a multimodal query dataset named CCaption-44k with 44k samples;
CF-IPT: Cross-Modal Fusion Interactive Prompt Tuning of Vision-Language Pre-Trained Model for Multisource Remote Sensing Data Classification
Jinheng Ji (Xidian University), Yunsong Li (Xidian University)
CodeClassificationTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes an interactive prompt tuning framework based on CLIP, named CF-IPT, for multi-source remote sensing image classification.
CG-Reasoner: Centroid-Guided Positional Reasoning Segmentation for Medical Imaging with a Robust Visual-Text Consistency Metric
Lakshmikar Reddy Polamreddy, Ming Ma (Yeshiva University)
CodeSegmentationExplainability and InterpretabilityConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: Proposed a unified cross-modal framework CG-Reasoner that can simultaneously perform medical image segmentation and localization reasoning, achieving interpretable diagnostic reports through the fusion of visual and language information.
Chain of Event-Centric Causal Thought for Physically Plausible Video Generation
Zixuan Wang (Sichuan University), Yinjie Lei (Sichuan University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelVideoTextBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed an event-centric physically feasible video generation framework that decomposes complex physical phenomena into causal chain events and generates visual videos through cross-modal prompts.
🎯 What it does: Propose Chain-of-Models Pre-Training (CoM-PT), achieving lossless accelerated pre-training of visual foundation models through constructing model chains and reverse knowledge transfer.
Chart-FR1: Visual Focus-Driven Fine-Grained Reasoning on Dense Charts
Hongkun Pan (Zhejiang University), Wei Chen (State Key Lab Of Cad Cg Zhejiang University)
CodeReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed Chart-FR1, a high information density chart fine-grained reasoning model achieved through visual focused chain-of-thought reasoning and reinforcement learning.
CHEEM: Continual Learning by Reuse, New, Adapt and Skip - A Hierarchical Exploration-Exploitation Approach
Chinmay Savadikar (North Carolina State University), Tianfu Wu (North Carolina State University)
CodeComputational EfficiencyRepresentation LearningMeta LearningNeural Architecture SearchTransformerMixture of ExpertsImage
🎯 What it does: Propose a sample-free class-incremental continual learning framework named CHEEM, leveraging the internal parameter memory of Vision Transformers and external task centroid memory to achieve task-adaptive network structures.
CHIPS: Efficient CLIP Adaptation via Curvature-aware Hybrid Influence-based Data Selection
Xinlin Zhuang (MBZUAI), Imran Razzak (MBZUAI)
CodeClassificationRetrievalDomain AdaptationTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataBenchmark
🎯 What it does: This paper proposes an efficient data selection method called CHIPS for the CLIP model, aiming to achieve the effect of continuous pre-training with a small amount of data in vertical domains;
🎯 What it does: Propose a temporally modulated Gaussian representation method called ChronoGS, which can jointly reconstruct multi-period scenes and simultaneously capture invariant structures as well as time-varying geometry and appearance within the same model.
CIGMA: Causal Information-Gain Mechanistic Attribution of Attention Heads in Vision Transformers
Maisha Maliha (University of Oklahoma), Dean F. Hougen (University of Oklahoma)
CodeClassificationRecognitionExplainability and InterpretabilityComputational EfficiencyTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the CIGMA framework, which identifies and prunes misleading attention heads in Vision Transformers through training-agnostic head-level causal attribution.
CIGPose: Causal Intervention Graph Neural Network for Whole-Body Pose Estimation
Bohao Li (Northwestern Polytechnical University), Yangming Guo (Northwestern Polytechnical University)
CodePose EstimationGraph Neural NetworkImage
🎯 What it does: Propose a causal intervention-based graph neural network (CIGPose) for full-body pose estimation, addressing non-causal associations caused by visual context confusion.
🎯 What it does: Propose a two-stage Clay-to-Stone framework, leveraging 3D Gaussian Splatting to first learn object shape and motion from monocular RGB video through deformable and semantic-associated modulation, then recover joint parameters via rigid constraints, achieving high-quality geometry reconstruction and rendering for hand-object interactions.
CLIP Is Shortsighted: Paying Attention Beyond the First Sentence
Marc-Antoine Lavoie (University Of Toronto Robotics Institute), Steven L. Waslander (University Of Toronto Robotics Institute)
CodeRetrievalTransformerContrastive LearningText
🎯 What it does: Propose DeBias-CLIP, a parameter-free enhancement method that mitigates CLIP's early token bias for long texts by removing summary sentences, sub-sampling sentences, and text filling.
🎯 What it does: This paper reinterprets CLIP-like models as density ratio estimators and proposes two new applications based on this interpretation: importance-weighted learning and KL divergence estimation.
CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimization
Xinhai Hou (University of Michigan), Bryan Wang (Amazon.com)
CodeAI Code AssistantSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodality
🎯 What it does: Proposed CodeV, a code-based visual reasoning agent, and designed Tool-Aware Policy Optimization (TAPO) to explicitly reward the authenticity of tool usage;
Yuchen Che (Institute of Science Tokyo), Asako Kanezaki (Institute of Science Tokyo)
CodePose EstimationImage
🎯 What it does: Proposes a confidence-aware optimal transport (OT)-based unsupervised single-view novel object pose estimation framework named COG, achieving precise pose regression by leveraging soft correspondences and semantic priors.
CogniEdit: Dense Gradient Flow Optimization for Fine-Grained Image Editing
Yan Li (Hongkong University of Science and Technology), Qi Tian (Huawei Inc)
CodeGenerationOptimizationTransformerLarge Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the CogniEdit framework, integrating multimodal LLM, dynamic token focusing, and dense GRPO optimization to achieve instruction-driven fine-grained image editing.
ColaVLA: Leveraging Cognitive Latent Reasoning for Hierarchical Parallel Trajectory Planning in Autonomous Driving
Qihang Peng (Tsinghua University), Hongsheng Li (Chinese University Of Hong Kong)
CodeAutonomous DrivingExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelImageMultimodalityPoint CloudChain-of-Thought
🎯 What it does: Proposes ColaVLA, a unified vision-language-action framework that compresses multimodal inputs through cognitive latent reasoning and generates safe, interpretable continuous trajectories in real-time via a hierarchical parallel planner.
🎯 What it does: Propose a framework for colorizing old grayscale photographs based on the FLUX diffusion model, named ColorFLUX. It employs structure-color decoupling learning, visual semantic prompts, and progressive direct preference optimization (DPO), enabling the recovery of color, brightness, and saturation after denoising and scratch removal.
🎯 What it does: Train an unsupervised continuous latent motion model to learn continuous motion representations from internet videos for robot learning;
🎯 What it does: Proposed the Complementary Prototype Mapping (CPMAD) framework, which uses dynamically extracted consensus and complementary prototypes to guide multi-modal unsupervised defect detection, significantly reducing cross-modal mapping ambiguity and improving localization accuracy.
Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference Optimization
Zhuohan Liu (Fudan University), Zuxuan Wu (Fudan University)
CodeGenerationSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: In the text-to-image generation field, the BIDPO framework is proposed, which uses dual-modal preference optimization and region-aware guidance to perform post-training fine-tuning on Stable Diffusion XL, and constructs the BICOMP preference dataset through an automated pipeline.
CodeDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImage
🎯 What it does: This study proposes Concept-Guided Fine-Tuning (CFT), which significantly enhances the model's robustness under out-of-distribution conditions by aligning the internal feature maps of ViT with automatically generated concept-level semantic masks in an unsupervised manner.
ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval
Zixu Li (Shandong University), Liqiang Nie (Harbin Institute Of Technology)
CodeRetrievalVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper addresses the problem of noisy triplet correspondence in compositional image retrieval, proposing a three-stage network called ConeSep based on cone space. It achieves precise identification and processing of noisy samples through geometric fidelity quantization, negative boundary learning, and boundary-based directional unlearning.
Consensus Entropy: Harnessing Multi-VLM Agreement for Self-Verifying and Self-Improving OCR
Yulong Zhang (Fudan University), Gongshen Liu (East China Normal University)
CodeRecognitionVision Language ModelImageTextBenchmark
🎯 What it does: Propose an unsupervised Consensus Entropy (CE) metric and the CE-OCR framework, which is an OCR solution that uses outputs from multiple Vision-Language Models (VLMs) for self-inspection and self-correction.
Bozhao Li (Harbin Institute of Technology, Shenzhen), Jingyong Su (Harbin Institute of Technology, Shenzhen)
CodeObject DetectionData SynthesisTransformerLarge Language ModelDiffusion modelContrastive LearningImageTextBenchmark
🎯 What it does: This paper proposes the Contextual Consistency Learning (CCL) framework, which enhances the robustness of open-vocabulary object detection by strengthening internal consistency across different backgrounds.
🎯 What it does: Proposed the ConsisVLA-4D framework, achieving efficient consistency in robot manipulation through 3D perception and 4D spatiotemporal reasoning
🎯 What it does: Propose content-aware frequency encoding (CAFE) and its extended version CAFE+, which dynamically synthesize frequencies through parallel linear layers and Hadamard product, thereby mitigating the spectral bias and fixed Fourier basis limitations of INR.
Convexity-Aware Noise Calibration: A Self-Supervised Framework for Noise-Level-Unknown Image Denoising
Zhan Wang (China University of Petroleum (East China)), Yu Meng (China University of Petroleum (East China))
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a two-stage self-supervised image denoising framework that first accurately estimates the noise level from noisy images alone, then generates noise-clean pairs using this estimate for supervised training, achieving high-quality denoising of images with unknown noise levels.
COPE: Consistent Occlusion and Prompt Enhancement Network for Occluded Person Re-identification
Siyi Sun (Xiamen University), Zhiming Luo (Xiamen University)
CodeRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes a network called COPE to address feature interference and information loss in occluded person re-identification, covering three modules: Cross-Identity Consistent Occlusion (CICO), Prompt Background Filling (PBF), and Prompt Similarity Scoring (PSS).
CORE: Compact Object-centric REpresentations as a New Paradigm for Token Merging in LVLMs
Jingyu Lei (Zhejiang University), Der-Horng Lee (Zhejiang University)
CodeSegmentationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a CORE method for visual token compression based on object-centric principles, which uses a segmentation head to generate object masks, aggregates each object in the image into a single token, and inputs them into a language model after sorting by centroid.
Sen Liang (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
CodeGenerationLarge Language ModelVision Language ModelDiffusion modelVideoTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes the Plan-Guide-Edit framework, which integrates a Chain-of-Thought (CoT) enhanced multilingual large language model (MLLM) planner, box-guided mask branch, and diffusion editor, achieving text instruction-driven video editing.
CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models
Nan Zhou (Beihang University), Di Huang (Beihang University)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Proposed CoVFT, a context-aware visual fine-tuning framework for multi-modal large language models, enabling adaptive fine-tuning of the visual encoder.
🎯 What it does: Proposed the CRAFT-LoRA framework, achieving decoupling and efficient fusion of content and style during LoRA training, and enabling high-quality personalized image generation on SDXL.
CREval: An Automated Interpretable Evaluation for Creative Image Manipulation under Complex Instructions
Chonghuinan Wang (Harbin Institute of Technology), Hongxun Yao (Harbin Institute of Technology)
CodeImage TranslationExplainability and InterpretabilityLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a fully automatic, explainable QA evaluation framework called CREval for assessing the effectiveness of creative image editing under complex instructions, and constructed the CREval-Bench benchmark, which includes over 800 images, 13K evaluation questions, and covers 9 sub-dimensions across 3 categories of creative dimensions.
🎯 What it does: A framework named CPS-Prompt is proposed for continual learning on edge devices, which significantly reduces memory and computational overhead during training while maintaining accuracy through critical patch sampling and decoupled prompt/classifier training.
🎯 What it does: Proposes a multi-view progressive adaptation framework (MPA) for cross-domain few-shot segmentation (CD-FSS) tasks, capable of adapting and improving segmentation performance even when only a minimal number of samples are available in the target domain.
Cross-modal Fuzzy Alignment Network for Text-Aerial Person Retrieval and A Large-scale Benchmark
Yifei Deng (Anhui University), Jin Tang (Anhui University)
CodeRetrievalTransformerVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a cross-modal fuzzy alignment network (CFAN) for the text-aerial person retrieval task and constructs a large-scale AERI-PEDES dataset.
Cross-Scale Pansharpening via ScaleFormer and the PanScale Benchmark
Ke Cao (HFIPS, Chinese Academy of Sciences), Jie Zhang (HFIPS, Chinese Academy of Sciences)
CodeSuper ResolutionTransformerImageBenchmark
🎯 What it does: Propose a cross-scale panchromatic image fusion method called ScaleFormer and the PanScale benchmark to address challenges in high-resolution cross-scale fusion.
Cross-Slice Knowledge Transfer via Masked Multi-Modal Heterogeneous Graph Contrastive Learning for Spatial Gene Expression Inference
Zhiceng Shi (Yunnan University), Wenwen Min (Yunnan University)
CodeGraph Neural NetworkContrastive LearningMultimodalityGraphBiomedical Data
🎯 What it does: Predict spatial gene expression of tissue slices by constructing cross-slice multimodal heterogeneous graphs combined with contrastive learning.
CrossHOI-Bench: A Unified Benchmark for HOI Evaluation across Vision-Language Models and HOI-Specific Methods
Qinqian Lei (National University of Singapore), Robby T. Tan (National University of Singapore)
CodeObject DetectionVision Language ModelImageBenchmark
🎯 What it does: This paper proposes CrossHOI-Bench, a unified multiple-choice benchmark for human-object interaction (HOI) detection, designed to fairly evaluate large vision-language models (VLMs) and specialized HOI methods.
CrossVL: Complexity-Aware Feature Routing and Paired Curriculum for Cross-View Vision-Language Detection
Zhipeng Liu (University of Exeter), Chunbo Luo (University of Exeter)
CodeObject DetectionTransformerVision Language ModelMultimodality
🎯 What it does: Designed a cross-view visual language detection framework named CrossVL, aiming to address the issue of performance degradation caused by geometric differences between ground and aerial perspectives.
🎯 What it does: Proposed the CROWn framework for 3D medical image segmentation, specifically addressing the challenges of aliasing undersampling and cross-scale phase misalignment caused by directional spacing and varying reconstructions.
🎯 What it does: This paper proposes a Curriculum Guided by Reward Variance (CGPO), which enhances the efficiency of reinforcement learning in text-to-image generation by dynamically adjusting the sample sampling probability and class weighting.
🎯 What it does: This paper proposes a curvature-aware zeroth-order optimization (CAZO) method, achieving memory-efficient adaptation of pre-trained models during testing, avoiding high memory consumption caused by backpropagation.
🎯 What it does: This paper proposes a closed-loop dynamic network (CLDyN), which couples a visualization fusion network (VFN) with a task-specific semantic compensation module (RSC) to achieve adaptive semantic compensation after infrared-visible image fusion. This allows the generation of multi-task customizable fusion results according to different downstream task requirements.
Cut to the Chase: Training-free Multimodal Summarization via Chain-of-Events
Xiaoxing You (Hangzhou Dianzi University), Jun Yu (Harbin Institute of Technology)
CodeGenerationGraph Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose an unsupervised multi-modal summarization framework called CoE, which generates text summaries by constructing a Hierarchical Event Graph (HEG) to achieve hierarchical event modeling, cross-modal alignment, and temporal reasoning.
🎯 What it does: Designed a regularized framework called CycleBEV based on view cycle consistency to enhance the performance of view transformation (VT) models in bird's-eye-view (BEV) semantic segmentation.
🎯 What it does: This paper proposes a threshold-free convex prior based on a function called quasi-convexity, which is transformed into differentiable low-order inequalities for end-to-end trained image segmentation networks.
🎯 What it does: Propose a dual-channel dual-branch network D FER, which jointly models expression recognition under dual challenges of visual perturbations and label noise through a unified framework combining weak/strong augmentation, dynamic queue, momentum-updated Query-Key architecture, and adversarial contrastive learning.
D3D-VLP: Dynamic 3D Vision-Language-Planning Model for Embodied Grounding and Navigation
Zihan Wang (National University of Singapore), Gim Hee Lee (National University of Singapore)
CodeRobotic IntelligenceLarge Language ModelReinforcement LearningVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposed a dynamic 3D vision-language-planning model (D3D-VLP), unifying planning, localization, and navigation within a single autoregressive 3D VLM;
🎯 What it does: Propose a forward-promoting continuous test-time adaptation (CTTA) paradigm, which dynamically aligns generated class representative images with real-time target domain samples at input, statistical, and representation levels through dynamic style bridging, providing reliable supervision signals in unsupervised online environments and significantly enhancing model adaptability to evolving distributions.
🎯 What it does: Designed a large-scale RGB-thermal low-light action recognition dataset named DarkAct, and proposed the DarkAct-Net two-stage fusion framework, which utilizes spatiotemporal differencing, Motion-Aware Attention, and Light-Adaptive Fusion to achieve high-precision action recognition in low-light environments.
DASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples
Abdullah Al Nomaan Nafi (University of Maine), Prabuddha Chakraborty (University of Maine)
CodeAdversarial AttackImage
🎯 What it does: Constructed a differentiable meta-attack framework called DASH, which learns weighted combinations across multiple baseline l_p attacks using soft attention and generates adversarial examples through multi-stage chained iterative processes.
Haoru Tan (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
CodeKnowledge DistillationImageMultimodality
🎯 What it does: This paper proposes a dataset distillation method based on Influence Matching, which directly aligns the impact of synthetic data on the final model parameters rather than the intermediate training process;
🎯 What it does: This paper proposes a model fusion method called DC-Merge based on directional consistency, aiming to preserve the directional information of each task vector when merging multi-task models, thereby retaining the performance of the original tasks.
Shuai Wang (Nanjing University), Limin Wang (Nanjing University)
CodeGenerationTransformerDiffusion modelImage
🎯 What it does: Designed and trained a diffusion Transformer (DDT) with separated low-frequency encoder and high-frequency decoder, achieving faster convergence and higher quality image generation.
🎯 What it does: This paper proposes a decision boundary-aware generation framework (DBG) for long-tail learning, which analyzes the boundary ambiguity caused by head-tail migration. It designs a generator to produce information-rich samples near the boundary and combines a classifier-driven dual-branch cleaning process to remove harmful samples, thereby improving the decision space and enhancing tail class accuracy.
DeCo: Frequency-Decoupled Pixel Diffusion for End-to-End Image Generation
Zehong Ma (Peking University), Qi Tian (Huawei Inc)
CodeGenerationTransformerDiffusion modelImageText
🎯 What it does: Propose a frequency-decoupled pixel diffusion framework DeCo, using a downsampled DiT for low-frequency semantic modeling, a lightweight pixel decoder to generate high-frequency details at full resolution, and introducing a frequency-aware flow matching loss based on JPEG quantization tables.
🎯 What it does: Proposed Decoupled Residual Denoising Diffusion Models (DRDD), achieving unified and data-efficient image-to-image translation. The method maintains domain harmony and semantic mapping by splitting the diffusion process into two stages: noise diffusion and residual diffusion.
Decoupling Bias, Aligning Distributions: Synergistic Fairness Optimization for Deepfake Detection
Feng Ding (Nanchang University), Shu Hu (Purdue University)
CodeAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Propose a dual-mechanism collaborative optimization framework that integrates structural fairness decoupling and global distribution alignment to enhance the fairness and accuracy of deepfake detection models.
🎯 What it does: Proposed a diagnostic-relief framework DASP for multi-modal test-time adaptation (MM-TTA), which can adapt to distribution drift in an unsupervised manner.
Deeper Thought, Weaker Aim: Understanding and Mitigating Perceptual Impairment during Reasoning in Multimodal Large Language Models
Ruiying Peng (Tsinghua Shenzhen International Graduate School), Xiao-Hui Li
CodeTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a training-free Visual Region Guided Attention (VRGA) framework for multi-modal large language models (MLLMs) to automatically identify visually attentive heads and reweight their attention during chain-of-thought (CoT) reasoning, prompting the model to focus on problem-related visual regions and thus alleviate perceptual decay and visual drift during reasoning.
DeepProtect: Proactive Face-Swapping Defense using Identity Blending and Attribute Distortion
Eungi Lee (Chonnam National University), Seok Bong Yoo (Chonnam National University)
CodeSafty and PrivacyComputational EfficiencyPrompt EngineeringVision Language ModelGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: Propose an active facial deepfake defense method called DeepProtect, which first dilutes facial identity features by performing identity blending in the W+ space of StyleGAN, and then suppresses the identity mapping of face-swap models without significantly degrading the original image quality by embedding adversarial watermarks generated through CLIP-based text prompts for localized attribute perturbations.
DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models
Yangfu Li, Yue Lu
CodeExplainability and InterpretabilityComputational EfficiencyPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a training-free framework called DeepScan, which assists large vision-language models (LVLMs) in explicitly locating, calibrating, and integrating visual evidence during inference to enhance fine-grained understanding of high-resolution images.
🎯 What it does: Propose the DeGO framework, which achieves weakly supervised occupancy prediction in dynamic 3D scenes using a deformable Gaussian occupancy model, capable of separating rigid and non-rigid motions;
🎯 What it does: Propose a test-time adaptation framework called DCTTA based on degradation consistency, enabling all-in-one image restoration models to adapt online under unseen degradation distributions.
Demo2Tutorial: From Human Experience to Multimodal Software Tutorials
Zechen Bai (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeSegmentationGenerationLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: Propose the Demo2Tutorial framework, which automatically converts raw human-computer usage demonstrations (screen recordings + operation logs) into structured, editable multimodal software tutorials.
🎯 What it does: Propose an adaptive framework called DAPASS for panoramic semantic segmentation under source-free conditions, which significantly improves cross-domain generalization performance.
🎯 What it does: Proposes DPCache, a training-free diffusion model sampling acceleration framework that significantly reduces the number of sampling steps while preserving generation quality through global path planning.
🎯 What it does: This paper proposes Patch Forcing, which utilizes adaptive denoising schedules for each image patch to complete easy-to-generate regions first and provide context for difficult-to-generate regions, thereby improving the generation quality of diffusion/flow models.
Depth Any Endoscopy: Towards Self-Supervised Generalizable Depth Estimation in Monocular Endoscopy
Shuwei Shao (Shandong University), Zhe Min (Shandong University)
CodeDepth EstimationConvolutional Neural NetworkMixture of ExpertsBiomedical Data
🎯 What it does: Proposed Depth Any Endoscopy (DAE), a unified self-supervised monocular endoscopy depth estimation framework capable of achieving cross-domain depth prediction in various surgical environments (e.g., laparoscopy, colonoscopy).
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelAuto EncoderMultimodality
🎯 What it does: Designed and implemented a unified autoregressive model, Uni-AdGen, for one-stop generation of advertising images and text, achieving personalized ad generation through foreground awareness, instruction tuning, and coarse-to-fine prioritized preference extraction.
DETACH : Decomposed Spatio-Temporal Alignment for Exocentric Video and Ambient Sensors with Staged Learning
Junho Yoon (Korea Advanced Institute of Science and Technology), Dongman Lee (Korea Advanced Institute of Science and Technology)
CodeRecognitionContrastive LearningVideoMultimodalityTime Series
🎯 What it does: This paper proposes an unsupervised pre-training framework (DETACH) to align outdoor videos with environmental sensors, thereby enhancing the performance of non-intrusive multimodal action recognition.
Chenfan Qu (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Constructed a large-scale AI forged text-image dataset named DanceText, and proposed DS-Net, a unified detection network, for classifying and localizing three types of forgeries in text images: full-image generation, region editing, and removal.
Qing Jiang (South China University of Technology), Lei Zhang (Peking University)
CodeObject DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper proposes a 3B parameter multimodal large language model named Rex-Omni, which can perform object detection and various visual perception tasks by predicting the next token following coordinate words.
🎯 What it does: Proposes a single forward framework DGGT based on 3D Gaussian partitioning, which can reconstruct 4D dynamic driving scenes directly from sparse uncalibrated images without requiring camera pose input, providing editable outputs such as camera poses, dynamic Gaussians, and depth.
DialogueVPR: Towards Conversational Visual Place Recognition
Yukun Song (Beijing University of Posts and Telecommunications), Pengyang Wang (University of Macau)
CodeRetrievalTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the Dialogic Visual Place Recognition (DlgPR) framework, transforming geolocation into an interactive reasoning process, and created a new dataset called DlgQuest-Cities.
🎯 What it does: Propose Diff-SemiER, a diffusion model framework for removing semi-transparent glasses, integrating generative priors and adaptive fusion.
🎯 What it does: Propose the LDAC task and implement the DiffDecompose framework to achieve hierarchical decomposition of a single semi-transparent/transparent image.
🎯 What it does: Propose RDVQ, a differentiable vector quantization (VQ) compression framework that uses a soft distribution instead of hard quantization indices to achieve end-to-end rate-distortion (RD) optimization, and employs a self-attentive Transformer for entropy modeling during decoding.
DiffSoup: Direct Differentiable Rasterization of Triangle Soup for Extreme Radiance Field Simplification
Kenji Tojo (University of Tokyo), Nobuyuki Umetani (University of Tokyo)
CodeGenerationNeural Radiance FieldImageMesh
🎯 What it does: Reconstructing a simplified 3D scene from multi-view RGB images using a minimal number of triangles with neural textures and binary opacity, while achieving a directly differentiable rasterization pipeline.
🎯 What it does: Propose the Diffusion Probe framework, which utilizes early cross-attention features from diffusion models to predict the final image quality, achieving efficient early quality assessment and generation optimization.
🎯 What it does: Propose a metadata-free sRGB noise generation framework PNG, which extracts input noise features through Prompt Autoencoder (PAE) and generates images conforming to real noise distribution in latent space via Prompt DiT (P-DiT).