CVPR 2026 Papers — Page 9

DiffSoup: Direct Differentiable Rasterization of Triangle Soup for Extreme Radiance Field Simplification

Kenji Tojo (University of Tokyo), Nobuyuki Umetani (University of Tokyo)

🎯 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.

Diffusion Forcing Planner: History-Annealed Planning with Time-Dependent Guidance for Autonomous Driving

Zehan Zhang (University of Science and Technology of China), Jia Cai (Yinwang Intelligent Technology Co Ltd)

🎯 What it does: Propose Diffusion Forcing Planner (DFP), a history-guided diffusion planning framework that enables controllable trade-offs between maintaining trajectory continuity and adapting to environmental changes.

Diffusion Guided Chain-of-Vision for Large Autoregressive Vision Models

Xinyang Wang (Zhejiang University), Wei Chen

🎯 What it does: Propose the Diffusion Guided Chain-of-Vision framework, which decomposes pure vision tasks into a multi-step chain generation process.

Diffusion Mental Averages

Phonphrm Thawatdamrongkit, Supasorn Suwajanakorn

🎯 What it does: Propose a 'Mental Average' (DMA) method based on diffusion models, generating a single clear and realistic concept prototype image by aligning multiple noise trajectories during the diffusion process; and extend to multi-modal concepts, achieving mode-level averaging through CLIP clustering and text/LoRA fine-tuning.

Diffusion MRI Transformer with a Diffusion Space Rotary Positional Embedding (D-RoPE)

Gustavo Chau Loo Kung (Stanford University), Ehsan Adeli (Stanford University)

🎯 What it does: This paper proposes a Transformer model for diffusion magnetic resonance imaging (dMRI), further utilized for downstream tasks such as age prediction, gender classification, MCI classification, and ADAS-Cog regression through self-supervised Masked AutoEncoder pre-training.

Diffusion Probe: Generated Image Result Prediction Using CNN Probes

Bukun Huang (Zhejiang Gongshang University), Jingqun Tang (ByteDance Inc.)

🎯 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.

Diffusion Sampling Path Tells More: An Efficient Plug-and-Play Strategy for Sample Filtering

Sixian Wang (Chinese University of Hong Kong Shenzhen), Tsung-Hui Chang (Chinese University of Hong Kong Shenzhen)

🎯 What it does: Utilizing the accumulated score difference (ASD) during the classifier-free guidance process of diffusion models as an internal quality signal, constructing CFG-Rejection to reject low-quality trajectories early in the generation process, thereby improving the quality and efficiency of generated samples.

Diffusion with a Linguistic Compass: Steering the Generation of Clinically Plausible Future sMRI Representations for Early MCI Conversion Prediction

Zhihao Tang (Beijing University of Posts and Telecommunications), Xi Zhang (Beijing University of Posts and Telecommunications)

🎯 What it does: Propose an MCI-Diff framework based on diffusion models, which generates clinically interpretable MRI features for future time points by utilizing only baseline sMRI data, enabling immediate and high-precision prediction of MCI conversion.

Diffusion-Based Makeup Transfer with Facial Region-Aware Makeup Features

Zheng Gao (Queen Mary University Of London), Jifei Song (Queen Mary University Of London)

🎯 What it does: Propose a facial makeup transfer framework FRAM based on diffusion models, divided into two stages: ① generate diverse makeup descriptions using GPT-3 and synthesize makeup images with text-driven editing models like FLUX, while fine-tuning the CLIP visual encoder into a makeup feature encoder; ② learn to inject identity features (ControlNet Union) and regional makeup features (queryable regional makeup embeddings) into the diffusion denoising network through generated pre- and post-makeup pairs, achieving global and regional makeup transfer.

Diffusion-Based Native Adversarial Synthesis for Enhanced Medical Segmentation Generalization

Hongyu Zhang (Jilin University), Yingda Lyu (Jilin University)

🎯 What it does: In medical image segmentation, a method is proposed to leverage the local adversariality of diffusion models for data augmentation, aiming to enhance the model's generalization performance on unseen devices or modalities.

Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning

Jaekyun Ko (Hanyang University), Tae Hyun Kim (Hanyang University)

🎯 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).

DiffusionFF: A Diffusion-based Framework for Joint Face Forgery Detection and Fine-Grained Artifact Localization

Siran Peng (Institute of Automation, Chinese Academy of Sciences), Zhen Lei (Institute of Automation, Chinese Academy of Sciences)

🎯 What it does: Propose the DiffusionFF framework, leveraging a pre-trained facial forgery detector and a denoising diffusion model to achieve forgery detection and fine-grained forgery trace localization in facial images.

DiffusionHarmonizer: Bridging Neural Reconstruction and Photorealistic Simulation with Online Diffusion Enhancer

Yuxuan Zhang (NVIDIA), Zan Gojcic (NVIDIA)

🎯 What it does: Propose DiffusionHarmonizer, an online generator enhancer framework that transforms neural reconstructed video frames into temporally consistent, lighting realistic rendering results.

DiffuView: Multi-View Diffusion Pretraining for 3D Aware Robotic Manipulation

Kaizhao Zhang (Fudan University), Wenchao Ding (Fudan University)

🎯 What it does: Propose a 3D consistent visual representation framework called DiffuView based on multi-view diffusion pre-training, embedding the pre-trained diffusion network as a visual encoder into the action diffusion policy of imitation learning, achieving perspective-robust robotic manipulation.

DiG: Differential Grounding for Enhancing Fine-Grained Perception in Multimodal Large Language Models

Zhou Tao (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)

🎯 What it does: By proposing the difference localization task DiG, using Blender 3D rendering to generate controllable image pairs, and performing reinforcement learning fine-tuning on multimodal large language models, thus enhancing fine-grained visual perception and spatial reasoning capabilities.

DiGraphHal-Bench: Evaluating Multimodal Large Language Models on Complex Directed Graphs

Yixin Fan (Fudan University), Wei Wang (Fudan University)

🎯 What it does: Established a large-scale visual question-answering benchmark, DiGraphHal-Bench, to assess hallucinations and fine-grained reasoning capabilities of multimodal large language models on complex directed graphs.

DIMOS: Disentangling Instance-level Moving Object Segmentation

Hongxiang Huang (Hong Kong University of Science and Technology), Bojun Cheng (Hong Kong University of Science and Technology)

🎯 What it does: This paper proposes the DIMOS framework, which uses a dual-decomposition approach to simultaneously extract appearance and motion features from both image and event modalities, and enhances the performance of small target motion instance segmentation through multi-granularity cross-modal alignment.

DINO Eats CLIP: Adapting Beyond Knowns for Open-set 3D Object Retrieval

Xinwei He (Huazhong Agricultural University), Xiang Bai (Huazhong University of Science and Technology)

🎯 What it does: To address the open-set 3D object retrieval task, this paper proposes combining the self-supervised model DINO with CLIP to construct an adaptive framework called DEC.

DiP: Taming Diffusion Models in Pixel Space

Zhennan Chen (Nanjing University), Ying Tai (Nanjing University)

🎯 What it does: Training diffusion models in pixel space using large image patches to construct global structures, supplemented by a lightweight Patch Detailer Head for local detail reconstruction;

Direct Segmentation without Logits Optimization for Training-Free Open-Vocabulary Semantic Segmentation

Jiahao Li (Xiamen University), Yanyun Qu (Hanjiang National Laboratory)

🎯 What it does: Propose a training-agnostic open-vocabulary semantic segmentation method that does not require logit iterative optimization, directly solving the analytical solution of distribution differences to achieve pixel-level segmentation.

DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization

Zhengxian Yang (Tsinghua University), Tao Yu (Tsinghua University)

🎯 What it does: Propose DirectFisheye-GS, which directly supports fisheye image input within the 3D Gaussian Splatting framework, and design a cross-view joint optimization strategy based on this;

Direction-aware 3D Large Multimodal Models

Quan Liu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)

DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching

Chang Zou (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)

🎯 What it does: This paper significantly accelerates the sampling process of video diffusion transformers by training a learnable feature cache predictor combined with Restricted MeanFlow distillation, achieving a maximum speedup of 11.8× with almost no loss in generation quality.

Disco-GS: Gaussian Splatting in Dynamic Color Lighting

Ashish Kumar (Indian Institute of Technology Madras), A. N. Rajagopalan (Indian Institute of Technology Madras)

🎯 What it does: For videos captured under time-varying colored lighting (disco lights), Gaussian Splatting is used to simultaneously perform 3D scene reconstruction and recovery of the scene's canonical (non-colored lighting) appearance, while allowing control of overall brightness during inference.

Discover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation

Yilong Yang (Xiamen University), Liujuan Cao (Xiamen University)

🎯 What it does: Propose a three-stage framework DSS (Discover–Segment–Select) for zero-shot camouflaged object segmentation, generating candidate boxes through visual feature clustering, refining masks using SAM, and finally selecting masks via inference with a multi-modal large language model.

Discovering Adaptive Task Dependencies for Efficient Multi-Task Representation Compression

Zhimeng Huang (Peking University), Chuanmin Jia (Peking University)

🎯 What it does: For multi-task representation compression, the Adaptive Task Dependency Compression (ATDC) framework is proposed, which can dynamically construct task dependency graphs for each image and perform predictive coding in that order;

Discriminative Perception via Anchored Description for Reasoning Segmentation

Tao Yang (Northwestern Polytechnical University), Qi Wang (Northwestern Polytechnical University)

🎯 What it does: Proposes the DPAD framework, introducing anchored descriptions and discriminative rewards in semantic segmentation tasks under reinforcement learning, enabling the model to actively distinguish objects from the background and generate more focused and concise reasoning chains;

Disentangle-then-Align: Non-Iterative Hybrid Multimodal Image Registration via Cross-Scale Feature Disentanglement

Chunlei Zhang (University of Technology Sydney), Jian Zhang (University of Technology Sydney)

🎯 What it does: Proposes a non-iterative hybrid multimodal image registration network HRNet, which can predict rigid and non-rigid transformations simultaneously in a shared feature space.

Disentangled Textual Priors for Diffusion-based Image Super-Resolution

Lei Jiang (Nanjing University), Gangshan Wu (Nanjing University)

🎯 What it does: This paper proposes DTPSR, a diffusion model-based image super-resolution framework that achieves interpretable and controllable step-by-step recovery through decoupled text priors;

Disentanglement-wise Image Dehazing through Cross-Domain Manifold Consensus

Tianyi Lyu (Nanjing University of Posts and Telecommunications), Kai-Kuang Ma (Nanjing University of Aeronautics and Astronautics)

🎯 What it does: Propose a joint framework CIM-D, which unifies multi-modal perceptual domain features through cross-domain invariant manifolds (CIM), and designs a physics-guided color decomposition network in the HSV space to achieve color decoupling, addressing the domain inconsistency and color distortion issues in traditional dehazing methods.

Disentangling to Re-couple: Resolving the Similarity-Controllability Paradox in Subject-Driven Text-to-Image Generation

Shuang Li (Tencent), Jie Jiang (Tencent)

🎯 What it does: Propose a framework (DisCo) that first decouples topic identity from text instructions and then recouples them through reinforcement learning, addressing the similarity-controllability paradox in topic-driven text-to-image generation.

Distilling Balanced Knowledge from a Biased Teacher

Seonghak Kim (Agency for Defense Development)

🎯 What it does: Propose and implement a knowledge distillation framework called LTKD tailored for long-tailed distributions, addressing the issue of student model performance degradation caused by the teacher model's bias toward head classes.

Distilling Quasi-Conformal Mapping: A Generalizable and Efficient Solution for Wide-Angle Correction

Chengyang Liu (University of Hong Kong), Huibin Li (Xi'an Jiaotong University)

🎯 What it does: Proposed a wide-angle image correction framework that distills quasi-conformal mapping knowledge into deep networks, divided into two stages: the teacher generates high-quality unannotated pairs and the student network learns the flow field.

Distilling Unsigned Distance Function for Surface Reconstruction from 3D Gaussian Splatting

Qian Li (Hohai University), Fan Liu (Hohai University)

🎯 What it does: Propose a method that distills knowledge from a locally shape-based UDF teacher to a student UDF network of 3D Gaussian points, combining visibility and geometry-adaptive weights to achieve high-quality open surface reconstruction.

Distributed Image Compression with Multimodal Side Information at Extremely Low Bitrates

Guojun Xu (Wuhan University of Technology), Junwei Zhou (Wuhan University of Technology)

🎯 What it does: Propose a multi-modal distributed image compression (MDIC) framework that leverages both visual and textual side information to achieve high perceptual quality reconstruction at extremely low bit rates.

Distribution-Aligned Multimodal Fusion for Robust Object Detection

Xiaohui Hao (Beihang University), Rui She (Beihang University)

🎯 What it does: The paper proposes a distribution alignment framework for RGB-IR multi-modal object detection, which freezes the pre-trained detector and trains a lightweight fusion module to reduce feature distribution drift caused by degradation.

DiT-Distill: Open-Set Fine-Grained Retrieval via Generative Curriculum Knowledge

Xin Jiang (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)

🎯 What it does: This paper addresses the open-set fine-grained retrieval task by proposing the DiT-Distill framework, which distills the generative curriculum knowledge of the pre-trained text-image diffusion model DiT into a lightweight retrieval network, achieving efficient retrieval without requiring DiT inference.

DiT-IC: Aligned Diffusion Transformer for Efficient Image Compression

Junqi Shi (Nanjing University), Zhan Ma (Nanjing University)

🎯 What it does: Proposed a DiT-IC image compression framework based on Diffusion Transformer (DiT), achieving efficient compression through one-time decoding in a deep latent space downsampled at 32×.

DiT360: High-Fidelity Panoramic Image Generation via Hybrid Training

Haoran Feng (Insta360 Research), Lu Qi (Insta360 Research)

🎯 What it does: Designed and implemented the DiT360 framework, leveraging hybrid training with perspective images and panoramic images to generate high-fidelity, photorealistic 360° panoramic images.

Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization

Tahira Kazimi (Virginia Tech), Pinar Yanardag (Virginia Tech)

🎯 What it does: This paper proposes a framework called DPP-GRPO, based on Determinantal Point Processes (DPP) and Group Relative Policy Optimization (GRPO), to generate diverse video collections;

DiverseDiT: Towards Diverse Representation Learning in Diffusion Transformers

Mengping Yang (Fudan University), Hao Li (Fudan University)

🎯 What it does: Explores the internal representation learning mechanisms of Diffusion Transformers (DiTs) and proposes DiverseDiT, which enhances representation diversity and generation quality through long residual connections and representation diversity loss.

DiverseGRPO: Mitigating Mode Collapse in Image Generation via Diversity-Aware GRPO

Henglin Liu (Tsinghua University), Xiangyang Ji (Tsinghua University)

🎯 What it does: To address the mode collapse problem caused by GRPO in image generation, this paper proposes the DiverseGRPO method, which integrates distribution layer creativity rewards with structural-aware regularization;

Diversity over Uniformity: Rethinking Representation in Generated Image Detection

Qinghui He (Chongqing University of Posts and Telecommunications), Bin Xiao (Chongqing University of Posts and Telecommunications)

🎯 What it does: By constructing an 'anti-feature-collapse learning' framework, the method discriminates between generated images and real images, focusing on preserving diverse discriminative information to enhance robustness across different generative models.

Divide, Conquer, and Aggregate: Asymmetric Experts for Class-Imbalanced Semi-Supervised Medical Image Segmentation

Yajun Liu (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes a 'Divide, Conquer, and Aggregate' (DCA) framework to address the class imbalance problem in medical image segmentation;

Divide, then Ground: Adapting Frame Selection to Query Types for Long-Form Video Understanding

Jialuo Li (Tsinghua University), Yan Lu (Tsinghua University)

🎯 What it does: Propose a training-free, query-type-based frame selection framework called DIG, which automatically distinguishes between global and local queries, and respectively employs uniform sampling or content-adaptive plus reward-guided refinement processes to efficiently select frames from long videos, significantly enhancing the video understanding performance of LMMs.

DK-DDIL: Adaptive Knowledge Retention for Dynamic Domain-Incremental Learning in Medical Imaging

Yuxi Ma (Xiamen University), Liansheng Wang (Xiamen University)

🎯 What it does: Explored the application of dynamic domain incremental learning (DIL) in medical imaging, proposing the DK-DDIL framework without replay.

DLVP-CLIP: Enhancing Fine-Grained Zero-Shot Anomaly Detection via Dynamic Local Visual Prompting

Gaowei Zhang (Dalian University of Technology), Lihe Zhang (Dalian University of Technology)

🎯 What it does: Proposed a dynamic local visual prompt framework based on CLIP, named DLVP-CLIP, for zero-shot anomaly detection, combining Semantic-Aware Local Feature Selector (SLFS), Multi-Modal Local Prompt (MLoP), and High-Low Frequency Decomposition (HFD) to enhance local detail capture and cross-modal alignment;

DLWM: Dual Latent World Models enable Holistic Gaussian-centric Pre-training in Autonomous Driving

Yiyao Zhu (Hong Kong University of Science and Technology), Shaojie Shen (Hong Kong University of Science and Technology)

🎯 What it does: This paper proposes DLWM (Dual Latent World Models), a two-stage self-supervised pre-training framework. It first learns efficient 3D Gaussian semantic representations by reconstructing multi-view depth and semantic maps, then separately trains potential world models based on Gaussian flows and driving trajectories to enhance 3D occupancy prediction, 4D forecasting, and motion planning performance.

DMAligner: Enhancing Image Alignment via Diffusion Model Based View Synthesis

Xinglong Luo (University Of Electronic Science And Technology Of China), Shuaicheng Liu

🎯 What it does: Propose a diffusion model-based image alignment framework DMAligner, which directly generates aligned images using view synthesis;

DMGD: Train-Free Dataset Distillation with Semantic-Distribution Matching in Diffusion Models

Qichao Wang (Zhejiang University), Min Zhang (Zhejiang University)

🎯 What it does: Achieve dataset distillation through untrained diffusion models, generating synthetic small datasets with information similar to the original dataset.

dMLLM-TTS: Self-Verified and Efficient Test-Time Scaling for Diffusion Multi-Modal Large Language Models

Yi Xin (Shanghai AI Lab), Xiaohong Liu (Shanghai AI Lab)

🎯 What it does: Proposes a test-time scaling (TTS) framework for diffusion-based multimodal large language models (dMLLM), named dMLLM-TTS, which combines two-dimensional scaling through trajectory exploration and iterative refinement.

DNF-SR: Dual-Input and Negative-Aware Feature Fine-Tuning for Real-World Image Super-Resolution

Shuhao Han (Nankai University), Chongyi Li (Nankai University)

🎯 What it does: Propose a single-step real-world image super-resolution framework DNF-SR, which enhances image quality by utilizing dual inputs (noisy LR + original LR) and post-training negative feature fine-tuning.

Do Less, Achieve More: Do We Need Every-Step Optimization for RL Fine-tuning of Diffusion Models?

Renye Yan (Peking University), Yimao Cai (Tsinghua University)

🎯 What it does: Propose AdaScope, an RL-based plugin that dynamically identifies the optimal RL training interval during diffusion model fine-tuning, reducing computational costs and improving generation quality.

Do Vision-Language Models Leak What They Learn? Adaptive Token-Weighted Model Inversion Attacks

Ngoc-Bao Nguyen (Singapore University of Technology and Design), Ngai-Man Cheung (Singapore University of Technology and Design)

🎯 What it does: Studied model inversion attacks on vision-language models (VLM), systematically evaluated the privacy leakage risks of VLM, and proposed four token- and sequence-based inversion strategies, particularly introducing adaptive token weighting (SMI-AW) to enhance reconstruction effectiveness.

Do Vision-Language Models Measure Up? Benchmarking Visual Measurement Reading with MeasureBench

Fenfen Lin (Beijing Academy of Artificial Intelligence), Xi Yang (Beijing Academy of Artificial Intelligence)

🎯 What it does: This paper constructs the MeasureBench benchmark, providing 2,442 image-question pairs for 26 types of measuring instruments (including 1,272 real-world images and 1,170 synthetic images), and designs a controllable 2D/3D synthetic pipeline to expand the data.

Do VLMs Perceive or Recall? Probing Visual Perception vs. Memory with Classic Visual Illusions

Xiaoxiao Sun (Stanford University), Serena Yeung-Levy (Stanford University)

🎯 What it does: Designed and released the VI-Probe framework, which utilizes adjustable classic visual illusion images along with accompanying language prompts to systematically evaluate the visual perception and memory capabilities of Vision-Language Models (VLMs).

Do You Have Freestyle? Expressive Humanoid Locomotion via Audio Control

Zhe Li (Nanyang Technological University), Shanghang Zhang (BAAI)

🎯 What it does: Proposes the RoboPerform framework, achieving audio-driven full-body robotic dancing and voice gesture control.

Do You See What I Am Pointing At? Gesture-Based Egocentric Video Question Answering

Yura Choi (Imperial College London), Stefanos Zafeiriou (Imperial College London)

🎯 What it does: Proposed the EGOPOINTVQA dataset and the HINT model, specifically addressing first-person video question answering tasks based on gesture pointing.

DocPrune: Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning

Joonmyung Choi (Korea University), Hyunwoo J. Kim (KAIST)

🎯 What it does: Proposes a training-free, progressive document question answering Token pruning framework DocPrune, significantly reducing computational costs of visual Transformers on long documents.

DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding

Hao Yan (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)

🎯 What it does: Propose DocSeeker, an ALR (Analyze-Locate-Reason) workflow for long document visual question answering (VQA) and its implementation.

Does YOLO Really Need to See Every Training Image in Every Epoch?

Xingxing Xie (Northwestern Polytechnical University), Gong Cheng (Northwestern Polytechnical University)

🎯 What it does: Propose the Anti-Forgetting Sampling Strategy (AFSS), which dynamically selects and reviews training images during YOLO training, significantly reducing redundant computations;

Domain Sensitive Federated Learning with Fisher-Informed Pruning

Chenchen Lin (Sun Yat-sen University), Xuehe Wang (Sun Yat-sen University)

🎯 What it does: Designed and implemented a domain-sensitive federated learning pruning framework (FEDFIP), which generates a shared pruning mask by evaluating channel importance per domain and allows clients to reactivate domain-specific channels on this basis, achieving structure alignment and personalized sparse models.

Domain-Skewed Federated Learning with Feature Decoupling and Calibration

Huan Wang (University of Wollongong), Guansong Pang (Singapore Management University)

🎯 What it does: This paper proposes a feature decoupling and calibration framework (F2DC) for domain-skewed federated learning, which enhances cross-domain generalization performance by separating local features into domain-robust and domain-related components and calibrating the latter.

Don't Show Pixels, Show Cues: Unlocking Visual Tool Reasoning in Language Models via Perception Programs

Muhammad Kamran Janjua (Huawei Technologies), Bahador Rashidi (Huawei Technologies)

🎯 What it does: Proposed and implemented Perception Programs (P2), a zero-training, model-agnostic interface that converts dense pixel-level outputs from visual tools into compact, symbolized, and language-friendly structured summaries, enabling multimodal large language models to directly read visual cues;

Downscaling Intelligence: Exploring Perception and Reasoning Bottlenecks in Small Multimodal Models

Mark Endo (Stanford University), Serena Yeung-Levy (Stanford University)

🎯 What it does: This paper systematically studies the impact of scaling down large language models (LLMs) on the performance of multimodal models and finds that visual tasks are more sensitive to scaling down; further analysis reveals that both perception and reasoning become bottlenecks; to address this, we propose the EXTRACT+THINK two-stage framework, combining visual extraction tuning and step-by-step reasoning, which significantly improves the performance of small-scale multimodal models.

DP-FedAdamW: An Efficient Optimizer for Differentially Private Federated Large Models

Jin Liu (Xidian University), Junkang Liu (Tianjin University)

🎯 What it does: Propose DP-FedAdamW, an AdamW optimizer for differential privacy federated learning, aiming to address variance, bias in second-moment estimation, and client drift issues under privacy noise.

DPAR: Dynamic Patchification for Efficient Autoregressive Visual Generation

Divyansh Srivastava (University of California San Diego), Joshua Kimball (Dolby Laboratories)

🎯 What it does: Propose a self-attention visual generation model named DPAR that dynamically aggregates image tokens into variable-sized patches to reduce the number of tokens and computational cost.

DPGF-Net: Dual-Prior Guided Fusion Network for Joint Assessment of Perceptual Quality and Semantic Consistency in AI-Generated Images

Tao Li (Chongqing University), Mingliang Zhou (Chongqing University)

🎯 What it does: Proposes DPGF-Net, a dual-prior guided fusion network capable of simultaneously evaluating the perceptual quality of AI-generated images and their textual semantic consistency.

DPL: Decoupled Prototype Learning for Enhancing Robustness of Vision-Language Transformers to Missing Modalities

Jueqing Lu (Monash University), Lan Du (Monash University)

🎯 What it does: Designed and implemented a prediction head called Decoupled Prototype Learning (DPL) that adaptively handles missing modalities by separating and modality-specific decomposition of category prototypes, enhancing the robustness of Vision-Language Transformers under missing modality conditions.

Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design

Haoxiang Sun (Sichuan University), Jiancheng Lv (Sichuan University)

🎯 What it does: Propose the Dr. Seg framework, using GRPO to improve the training of vision-language models in visual perception tasks

Dr.Occ: Depth- and Region-Guided 3D Occupancy from Surround-View Cameras for Autonomous Driving

Xubo Zhu (Wuhan University), Huai Yu (Horizon Robotics)

🎯 What it does: This study proposes the Dr.Occ framework, combining depth-guided dual-projection view transformer and region-guided expert transformer to achieve 3D semantic occupancy prediction from surround-view camera inputs;

Draft and Refine with Visual Experts

Sungheon Jeong (University of California Irvine), Mohsen Imani (University of California Irvine)

🎯 What it does: Proposed the Draft and Refine framework, which evaluates and enhances the visual dependency and answer quality of large vision-language models (LVLM) by leveraging a visual utilization metric under problem conditions.

Drainage: A Unifying Framework for Addressing Class Uncertainty

Yasser Taha (Robert Koch Institute), Nils Körber (Robert Koch Institute)

🎯 What it does: Designed and implemented the Drainage framework, adding an extra drainage node at the output of neural networks and defining a corresponding drainage loss to address issues such as label noise and class uncertainty.

DRAMA: Next-Gen Dynamic Orchestration for Resilient Multi-Agent Ecosystems in Flux

Xinkui Zhao (Zhejiang University), Jianwei Yin (Zhejiang University)

🎯 What it does: Propose a multi-level DRAMA framework that unifies agents and tasks as resource entities, employing affinity-driven dual-capacity Hungarian assignment, hierarchical trust chains, and collective spatial reasoning to achieve adaptive scheduling and robust execution for multi-agent systems in dynamic environments.

DREAM: Document Recognition with Explicit Adaptive Memory

Tianqi Zhao (Tsinghua University), Yuyang Li (Huawei Noah Ark Lab)

🎯 What it does: This paper proposes the DREAM module, which supplements the visual encoder with explicit multi-scale prototype memory to achieve end-to-end recognition of documents and handwritten text.

DreamingComics: A Story Visualization Pipeline via Subject and Layout Customized Generation using Video Models

Patrick Kwon (University of Central Florida), Chen Chen (University of Central Florida)

🎯 What it does: Generate comic-style story visualizations based on text scripts, supporting multi-character, cross-frame identity and style consistency, and providing controllable layouts.

DreamOmni2: Multimodal Instruction-based Generation and Editing

Bin Xia (The Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)

🎯 What it does: This paper proposes the DreamOmni2 framework, enabling multi-modal image editing and generation based on text+image instructions.

DreamSAC: Learning Hamiltonian World Models via Symmetry Exploration

Jinzhou Tang (University Of California San Diego), Keze Wang (University Of California San Diego)

🎯 What it does: This paper proposes the DreamSAC framework, achieving self-supervised learning from pixels to physical states through self-driven symmetry exploration and a Hamiltonian world model.

DreamShot: Personalized Storyboard Synthesis with Video Diffusion Prior

Junjia Huang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)

🎯 What it does: Propose DreamShot, a controllable personalized storyboard generation framework based on video diffusion prior, supporting three generation modes: reference image, text, and previous frame, capable of generating multi-shot storyboards that maintain consistency in character identity, scene layout, and perspective switching.

DreamSR: Towards Ultra-High-Resolution Image Super-Resolution via a Receptive-Field Enhanced Diffusion Transformer

Qingji Dong (ByteDance Inc), Yitong Wang (ByteDance Inc)

🎯 What it does: Proposed DreamSR, a two-stage ultra-high-resolution image super-resolution model based on Diffusion Transformer, which suppresses over-generation and enhances detail reconstruction in patch-wise inference through dual-branch MM-ControlNet and Restoration Acceleration LoRA;

DreamStereo: Towards Real-Time Stereo Inpainting for HD Videos

Yuan Huang (ByteDance), Shaohui Jiao (ByteDance)

🎯 What it does: Propose an HD video real-time stereo inpainting framework based on gradient-aware disparity warping (GAPW), dual projection data generation (PBDP), and sparse attention (SASI).

DreamStyle: A Unified Framework for Video Stylization

Mengtian Li (ByteDance), Qian He (ByteDance)

🎯 What it does: Proposed DreamStyle, a unified video stylization framework that supports three style conditions: text, style images, and the first frame, achieving high-quality video stylization through a two-stage training process.

DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease

Runsheng Bai (MIT), Yangdong Deng (Tsinghua University)

🎯 What it does: Proposed DRiffusion, a draft-refinement parallel sampling framework that achieves parallel inference for diffusion models via jump transitions.

Drift-Resilient Temporal Priors for Visual Tracking

Yuqing Huang (Harbin Institute of Technology), Xin Li (Pengcheng Laboratory)

🎯 What it does: Proposes a lightweight, plug-and-play module called DTPTrack to suppress model drift in multi-frame visual tracking.

Drive My Way: Preference Alignment of Vision-Language-Action Model for Personalized Driving

Zehao Wang (University of California, Riverside), Jiachen Li (University of California, Riverside)

🎯 What it does: Propose Drive My Way (DMW), an end-to-end Vision-Language-Action (VLA) driving framework that achieves long-term driving habit alignment through learning user embeddings and enables short-term driving style adaptation via reinforcement learning fine-tuning and style instructions.

DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving

Enhui Ma (Zhejiang University), Kaicheng Yu (Li Auto Inc)

🎯 What it does: Propose the DriveCombo benchmark to evaluate the performance of multimodal large language models in multi-rule traffic regulation reasoning.

DriveLaW: Unifying Planning and Video Generation in a Latent Driving World

Tianze Xia (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)

🎯 What it does: Propose DriveLaW, a unified framework for video generation and trajectory planning, where the latent representations from the video generator drive trajectory generation;

DriveMoE: Mixture-of-Experts for Vision-Language-Action Model in End-to-End Autonomous Driving

Zhenjie Yang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes an end-to-end autonomous driving framework called DriveMoE, which improves driving performance across multiple scenarios by introducing Mixture-of-Experts (MoE) in the vision and decision modules to achieve dynamic perspective selection and skill specialization.

DrivePI: Spatial-aware 4D MLLM for Unified Autonomous Driving Understanding, Perception, Prediction and Planning

Zhe Liu (University of Hong Kong), Hengshuang Zhao (Yinwang Intelligent Technology Co. Ltd.)

🎯 What it does: Propose DrivePI, a unified 4D multimodal large language model capable of simultaneously performing spatial understanding, 3D perception, prediction, and planning.

DrivePTS: A Progressive Learning Framework with Textual and Structural Enhancement for Driving Scene Generation

Zhechao Wang (XPeng Motors), Cheng Lu (XPeng Motors)

🎯 What it does: Propose the DrivePTS framework, leveraging phased progressive learning, Vision-Language models to generate multi-perspective hierarchical text, frequency-guided structural loss, and generating controllable driving scenarios based on Stable Diffusion.

DriverGaze360: OmniDirectional Driver Attention with Object-Level Guidance

Shreedhar Govil (German Research Center for Artificial Intelligence), Jason Rambach (German Research Center for Artificial Intelligence)

🎯 What it does: Created the first 360° driver attention dataset, DriverGaze360, and proposed the DriverGaze360-Net prediction model.

DriveVLN: Towards Mapless Vision-and-Language Navigation in Autonomous Driving

Dongqian Guo (University of Macau), Jianbing Shen (University of Macau)

🎯 What it does: Propose the DriveVLN task, enabling vehicles to achieve point-to-point navigation in map-free environments based solely on short natural language descriptions;

Driving on Registers

Ellington Kirby (valeo.ai), Matthieu Cord (valeo.ai)

🎯 What it does: Propose DrivoR, an end-to-end autonomous driving model based on Vision Transformer, which uses camera-specific registration tokens to compress multi-camera features and separate trajectory generation from scoring modules;

DRM: Diffusion-based Reward Model With Step-wise Guidance

Jaxon Zhang (Peking University), Jing LYU (WeChat Vision, Tencent Inc)

🎯 What it does: Proposed a reward model based on pre-trained diffusion models (DRM), and applied it to alignment of diffusion models via reinforcement learning (Step-GRPO) and sampling improvement (Step-Sampling).

DROID-SLAM in the Wild

Moyang Li (Eth Zurich), Daniel Barath (Eth Zurich)

🎯 What it does: Propose a real-time monocular dynamic SLAM system DROID-W, which estimates pixel-level dynamic uncertainty through differentiable uncertainty-aware bundle adjustment, achieving robust tracking and high-quality geometric reconstruction in dynamic environments.

Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting

Shuangkang Fang (Beihang University), Takeo Igarashi (University of Tokyo)

🎯 What it does: In sparse view 3D Gaussian Splatting, a new dropout strategy called DropAnSH-GS is proposed, which removes a group of adjacent Gaussians through anchor-based dropout and reduces overfitting by applying dropout to high-order spherical harmonic coefficients.

DRS-GUI: Dynamic Region Search for Training-Free GUI Grounding

Yichao Liu (Nankai University), Yu Zhou (Nankai University)

🎯 What it does: Proposed a training-free dynamic region search framework called DRS-GUI, which utilizes a lightweight UI Perceptor and human-perceptual actions (Focus, Shift, Scatter) to first locate appropriate GUI regions in multimodal large language models, and then perform coordinate prediction;

DSCA: Dynamic Subspace Concept Alignment for Lifelong VLM Editing

Gyanendra Das (Zynix AI), Sai Jena (Zynix AI)

🎯 What it does: Designed a VLM knowledge editing framework (DSCA) based on dynamic subspace concept alignment, achieving precise editing of pre-trained VLMs without global fine-tuning, with concept-blocked modifications.

DSERT-RoLL: Robust Multi-Modal Perception for Diverse Driving Conditions with Stereo Event-RGB-Thermal Cameras, 4D Radar, and Dual-LiDAR

Hoonhee Cho (KAIST), Kuk-Jin Yoon (KAIST)

🎯 What it does: Propose the DSERT-RoLL multimodal driving dataset, integrating binocular event cameras, RGB cameras, thermal imaging cameras, 4D radar, and dual LiDAR, and provide a unified 3D/2D detection benchmark and fusion method.

DSFlash: Comprehensive Panoptic Scene Graph Generation in Realtime

Julian Lorenz (University of Augsburg), Rainer Lienhart (University of Augsburg)

🎯 What it does: Proposed DSFlash, a panoramic scene graph generation model for real-time applications, capable of completing full instance localization and relationship prediction at 56 FPS on standard GPUs.