CVPR 2026 Papers

Large vision-language models struggle with medical video understanding, where spatial precision, temporal reasoning, and clinical semantics are critical. To address this, we first introduce MedVidBench, a large-scale benchmark of 531,850 video-instruction pairs across 8 medical sources spanning video, segment, and frame-level tasks, curated through a rigorous quality assurance pipeline with expert-guided prompting and dual-model validation. While supervised fine-tuning on MedVidBench yields noticeable gains, standard Reinforcement Learning (RL) fails due to imbalanced reward scales across datasets, which destabilizes optimization and leads to training collapse. To overcome this, we introduce MedGRPO, a novel RL framework for balanced multi-dataset training with two key innovations: (1) cross-dataset reward normalization that maps each dataset's median performance to a common reward value, ensuring fair optimization regardless of difficulty, and (2) a medical LLM judge that evaluates caption quality on five clinical dimensions through comparative similarity scoring. Supervised fine-tuning Qwen2.5-VL-7B on MedVidBench substantially outperforms GPT-4.1 and Gemini-2.5-Flash across all tasks, demonstrating MedVidBench's efficacy, while our MedGRPO framework further improves upon the SFT baseline across grounding and captioning tasks. Our work establishes a foundational benchmark and robust training methodology for advancing vision-language models in medical domains.

Instruction-based multimodal image manipulation has recently made rapid progress. However, existing evaluation methods lack a systematic and human-aligned framework for assessing model performance on complex and creative editing tasks. To address this gap, we propose CREval, a fully automated question-answer (QA)-based evaluation pipeline that overcomes the incompleteness and poor interpretability of opaque Multimodal Large Language Models (MLLMs) scoring. Simultaneously, we introduce CREval-Bench, a comprehensive benchmark specifically designed for creative image manipulation under complex instructions. CREval-Bench covers three categories and nine creative dimensions, comprising over 800 editing samples and 13K evaluation queries. Leveraging this pipeline and benchmark, we systematically evaluate a diverse set of state-of-the-art open and closed-source models. The results reveal that while closed-source models generally outperform open-source ones on complex and creative tasks, all models still struggle to complete such edits effectively. In addition, user studies demonstrate strong consistency between CREval's automated metrics and human judgments. Therefore, CREval provides a reliable foundation for evaluating image editing models on complex and creative image manipulation tasks, and highlights key challenges and opportunities for future research.

Hyperbolic spaces provide a natural geometry for representing hierarchical and tree-structured data due to their exponential volume growth. To leverage these benefits, neural networks require intrinsic and efficient components that operate directly in hyperbolic space. In this work, we lift two core components of neural networks, Multinomial Logistic Regression (MLR) and Fully Connected (FC) layers, into hyperbolic space via Busemann functions, resulting in Busemann MLR (BMLR) and Busemann FC (BFC) layers with a unified mathematical interpretation. BMLR provides compact parameters, a point-to-horosphere distance interpretation, batch-efficient computation, and a Euclidean limit, while BFC generalizes FC and activation layers with comparable complexity. Experiments on image classification, genome sequence learning, node classification, and link prediction demonstrate improvements in effectiveness and efficiency over prior hyperbolic layers. The code is available at https://github.com/GitZH-Chen/HBNN.

The deployment of multimodal models in high-stakes domains, such as self-driving vehicles and medical diagnostics, demands not only strong predictive performance but also reliable mechanisms for detecting failures. In this work, we address the largely unexplored problem of failure detection in multimodal contexts. We propose Adaptive Confidence Regularization (ACR), a novel framework specifically designed to detect multimodal failures. Our approach is driven by a key observation: in most failure cases, the confidence of the multimodal prediction is significantly lower than that of at least one unimodal branch, a phenomenon we term confidence degradation. To mitigate this, we introduce an Adaptive Confidence Loss that penalizes such degradations during training. In addition, we propose Multimodal Feature Swapping, a novel outlier synthesis technique that generates challenging, failure-aware training examples. By training with these synthetic failures, ACR learns to more effectively recognize and reject uncertain predictions, thereby improving overall reliability. Extensive experiments across four datasets, three modalities, and multiple evaluation settings demonstrate that ACR achieves consistent and robust gains. The source code will be available at https://github.com/mona4399/ACR.

Recent diffusion-based one-step methods have shown remarkable progress in the field of image super-resolution, yet they remain constrained by three critical limitations: (1) inferior fidelity performance caused by the information loss from compression encoding of low-quality (LQ) inputs; (2) insufficient region-discriminative activation of generative priors; (3) misalignment between text prompts and their corresponding semantic regions. To address these limitations, we propose CODSR, a controllable one-step diffusion network for image super-resolution. First, we propose an LQ-guided feature modulation module that leverages original uncompressed information from LQ inputs to provide high-fidelity conditioning for the diffusion process. We then develop a region-adaptive generative prior activation method to effectively enhance perceptual richness without sacrificing local structural fidelity. Finally, we employ a text-matching guidance strategy to fully harness the conditioning potential of text prompts. Extensive experiments demonstrate that CODSR achieves superior perceptual quality and competitive fidelity compared with state-of-the-art methods while maintaining efficient one-step inference.

Recent video editing models have achieved impressive results, but most still require large-scale paired datasets. Collecting such naturally aligned pairs at scale remains highly challenging and constitutes a critical bottleneck, especially for local video editing data. Existing workarounds transfer image editing to video through global motion control for pair-free video editing, but such designs struggle with background and temporal consistency. In this paper, we propose NOVA: Sparse Control & Dense Synthesis, a new framework for video editing. Specifically, the sparse branch provides semantic guidance through user-edited keyframes distributed across the video, and the dense branch continuously incorporates motion and texture information from the original video to maintain high fidelity and coherence. Moreover, we introduce a degradation-simulation training strategy that enables the model to learn motion reconstruction and temporal consistency by training on artificially degraded videos, thus eliminating the need for paired data. Our extensive experiments demonstrate that NOVA outperforms existing approaches in edit fidelity, motion preservation, and temporal coherence.

Stereo vision is fundamental for enabling machines to perceive and interact with the world. While monocular stereo methods offer hardware compactness, they struggle with generalization due to reliance on data-driven priors. Binocular and multi-view systems improve accuracy but incur higher hardware complexity and data inefficiency. In this paper, we introduce a monocular solution for high-framerate stereo vision via temporal optical modulation. The modulation directs light from two views onto a single sensor in a mixed manner, while periodically attenuating one view at 60 Hz. To capture the temporal variations introduced by this modulation, we employ a high-speed spike camera that records the mixed scene as temporally dense spikes. The high temporal resolution of these spikes enables the construction of a linear system for efficient binocular video decoupling. Consequently, we introduce a two-stage decoding methodology for achieving high-quality stereo vision: An efficient least-squares-based baseline reconstruction followed by a deep learning refinement module. Experimental results demonstrate that our approach achieves 240FPS binocular video reconstruction with superior accuracy compared to monocular systems, while maintaining the hardware compactness and data efficiency. Code is available at https://github.com/yongqiye00/MonoSpikeStereo.

Performance estimation under distribution shift aims to predict how a model behaves on an unlabeled test set whose distribution differs from the training data, a scenario that requires reliable indicators that can faithfully reflect model behavior without ground-truth labels. Existing approaches rely solely on the outputs of the given model whose biases are amplified once the distribution shifts, weakening the correlation with the true performance. Motivated by this limitation, we propose Fused Reference Alignment Prediction (FRAP), which leverages the complementary strengths of an external foundation model and the base model to construct a more reliable surrogate of the ground-truth labels. FRAP aligns the prediction distribution of the foundation model with that of the base model by applying temperature-scaled calibration that minimizes their divergence. The aligned predictions are fused through confidence-based weighting into a refined reference distribution that integrates robustness from the foundation model and domain-specific expertise from the base model, and performance estimation is obtained by measuring how closely the base model predictions agree with this reference. Extensive experiments across diverse datasets and architectures show that FRAP provides consistent and substantial improvements over representative performance-estimation methods under distribution shift.

While multimodal large language models (MLLMs) have achieved remarkable success in recent advancements, their susceptibility to jailbreak attacks has come to light. In such attacks, adversaries exploit carefully crafted prompts to coerce models into generating harmful or undesirable content. Existing defense mechanisms often rely on external inference steps or safety alignment training, both of which are less effective and impractical when facing sophisticated adversarial perturbations in white-box scenarios. To address these challenges and bolster MLLM robustness, we introduce SAFEMLLM by adopting an adversarial training framework that alternates between an attack step for generating adversarial noise and a model updating step. At the attack step, SAFEMLLM generates adversarial perturbations through a newly proposed contrastive embedding attack (CoE-Attack), which optimizes token embeddings under a contrastive objective. SAFEMLLM then updates model parameters to neutralize the perturbation effects while preserving model utility on benign inputs. We evaluate SAFEMLLM across multiple MLLMs and six jailbreak methods spanning multiple modalities. Experimental results show that SAFEMLLM effectively defends against diverse attacks, maintaining robust performance and utilities.

Visual models are deployed on many Internet-of-Things (IoT) devices to power a variety of visual applications at the network edge. These models often need to be fine-tuned on-device continually to adapt to changing operating environments timely. However, the computing and energy overheads incurred are often overwhelming for resource-constrained IoT devices. Existing methods score sample importance based on the entire model via multi-epoch training, incurring overhead that may even exceed the training overhead reduction. To reduce these fine-tuning overheads, this paper presents LoPrune, a novel data pruning method that identifies and removes samples with negligible contributions to model adaptation. The key idea is to evaluate sample importance via a Trainable Subspace Alignment (TSA) Score to align the importance estimation with accurate update directions of the learnable adapter, i.e., Low-Rank Adaptation (LoRA). Specifically, LoPrune projects the influence function onto the LoRA subspace, enforcing consistency between the importance score and the model's updatable directions while substantially reducing the problem's dimensionality. It then leverages Kronecker-Factored Approximate Curvature to approximate the change of learnable adapter induced by a sample as its TSA score, retaining higher-scoring samples. Experiments with four representative visual models fine-tuned on three datasets demonstrate that compared with the best state-of-the-art data pruning baselines, LoPrune can reduce fine-tuning overhead by up to 72.9%, achieving a 3.69 xtraining speedup while improving fine-tuning accuracy by 3.50%.

Visual-language grounding aims to establish semantic correspondences between natural language and visual entities, enabling models to accurately identify and localize target objects based on textual instructions. Existing VLG approaches focus on coarse-grained, object-level localization, while traditional robotic grasping methods rely predominantly on geometric cues and lack language guidance, which limits their applicability in language-driven manipulation scenarios. To address these limitations, we propose the RealVLG framework, which integrates the RealVLG-11B dataset and the RealVLG-R1 model to unify real-world visual-language grounding and grasping tasks. RealVLG-11B dataset provides multi-granularity annotations including bounding boxes, segmentation masks, grasp poses, contact points, and human-verified fine-grained language descriptions, covering approximately 165,000 images, over 800 object instances, 1.3 million segmentation, detection, and language annotations, and roughly 11 billion grasping examples. Building on this dataset, RealVLG-R1 employs Reinforcement Fine-tuning on pretrained large-scale vision-language models to predict bounding boxes, segmentation masks, grasp poses, and contact points in a unified manner given natural language instructions. Experimental results demonstrate that RealVLG supports zero-shot perception and manipulation in real-world unseen environments, establishing a unified semantic-visual multimodal benchmark that provides a comprehensive data and evaluation platform for language-driven robotic perception and grasping policy learning. All data and code are publicly available at https://github.com/lif314/RealVLG-R1.

We propose POLAR, a novel radar-guided depth estimation method that introduces polynomial fitting to efficiently transform scaleless depth predictions from pretrained monocular depth estimation (MDE) models into metric depth maps. Unlike existing approaches that rely on complex architectures or expensive sensors, our method is grounded in a fundamental insight: although MDE models often infer reasonable local depth structure within each object or local region, they may misalign these regions relative to one another, making a linear scale and shift (affine) transformation insufficient given three or more of these regions. To address this limitation, we use polynomial coefficients predicted from cheap, ubiquitous radar data to adaptively adjust predictions non-uniformly across depth ranges. In this way, POLAR generalizes beyond affine transformations and is able to correct such misalignments by introducing inflection points. Importantly, our polynomial fitting framework preserves structural consistency through a novel training objective that enforces local monotonicity via first-derivative regularization. POLAR achieves state-of-the-art performance across three datasets, outperforming existing methods by an average of 24.9% in MAE and 33.2% in RMSE, while also achieving state-of-the-art efficiency in terms of latency and computational cost.

Recent advances in large vision models (LVMs) have shifted from modality-specific designs toward unified architectures that jointly process images, videos, and 3D data. However, existing unified LVMs primarily pursue functional integration, while overlooking the deeper goal of cross-vision synergy: the ability to reason over complementary priors across visual modalities. To address this, we present PolyV, a unified LVM that achieves cross-vision synergy at both the architectural and training levels. Architecturally, PolyV adopts a sparse Mixture-of-Experts LVM coordinated by a dynamic modality router, allowing each expert to specialize in modality-specific priors while enabling bidirectional interaction and mutual refinement across modalities. Training-wise, a synergy-aware paradigm combines modality-specific pretraining with coarse-to-fine synergy tuning via knowledge distillation and object-/relation-level alignment. Extensive experiments on 10 benchmarks spanning image, video, and 3D understanding, including synergy-focused datasets requiring spatial or temporal priors, demonstrate that PolyV consistently outperforms existing models, achieving over 10% average improvement over its backbone. Overall, PolyV establishes a unified framework for synesthetic visual reasoning, advancing toward truly synergistic LVMs.

Autoregressive models (ARMs) have long dominated the landscape of biomedical vision-language models (VLMs). Recently, masked diffusion models such as LLaDA have emerged as promising alternatives, yet their application in the biomedical domain remains largely underexplored. To bridge this gap, we introduce LLaDA-MedV, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning. LLaDA-MedV achieves relative performance gains of 7.855% over LLaVA-Med and 1.867% over LLaDA-V in the open-ended biomedical visual conversation task, and sets new state-of-the-art accuracy on the closed-form subset of three VQA benchmarks: 84.93% on VQA-RAD, 92.31% on SLAKE, and 95.15% on PathVQA. Furthermore, a detailed comparison with LLaVA-Med suggests that LLaDA-MedV is capable of generating reasonably longer responses by explicitly controlling response length, which can lead to more informative outputs. We also conduct an in-depth analysis of both the training and inference stages, highlighting the critical roles of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition. The code and model weight is released at https://github.com/LLM-VLM-GSL/LLaDA-MedV.

Vision-language models such as CLIP have achieved remarkable zero-shot recognition capabilities, yet their robustness against adversarial perturbations remains limited. Test-time counterattack (TTC) was recently proposed to improve CLIP's robustness by perturbing an input image to steer it away from a corrupted state during inference. However, TTC remains fragile under strong attacks because its counterattack relies on a directly corrupted original view and employs a noise-driven hard-gating scheme that cannot adapt to varying corruption severity. To address these limitations, we introduce Multi-view guided Adaptive Counterattack (MAC), which performs counterattacks for multi-view with corruption-aware soft weighting. Specifically, MAC first constructs augmented views of an input image to obtain diverse embeddings. It then performs counterattacks to refine corrupted embeddings of views. Next, MAC adaptively scales the counterattack intensity for each view based on its estimated corruption degree. Finally, the adaptively counterattacked views are aggregated to yield a robust final prediction. Extensive experiments across 20 datasets and diverse attack scenarios demonstrate that MAC substantially improves robustness while preserving high inference speed and memory efficiency with its tuning-free design. Our code is available at https://github.com/sunoh-kim/MAC.

The rapid adoption of vision-language models (VLMs) has heightened the demand for robust intellectual property (IP) protection of these high-value pretrained models. Effective IP protection should proactively confine model deployment within authorized domains and prevent unauthorized transfers. However, existing methods rely on static training-time definitions, limiting flexibility in dynamic environments and often producing opaque responses to unauthorized inputs. To address these limitations, we propose a novel dynamic authorization with legality-aware intellectual property protection (AoD-IP) for VLMs, a framework that supports authorize-on-demand and legality-aware assessment. AoD-IP introduces a lightweight dynamic authorization module that enables flexible, user-controlled authorization, allowing users to actively specify or switch authorized domains on demand at deployment time. This enables the model to adapt seamlessly as application scenarios evolve and provides substantially greater extensibility than existing static-domain approaches. In addition, AoD-IP incorporates a dual-path inference mechanism that jointly predicts input legality-aware and task-specific outputs. Comprehensive experimental results on multiple cross-domain benchmarks demonstrate that AoD-IP maintains strong authorized-domain performance and reliable unauthorized detection, while supporting user-controlled authorization for adaptive deployment in dynamic environments.

The rapid progress of Multimodal Large Language Models (MLLMs) marks a significant step toward artificial general intelligence, offering great potential for augmenting human capabilities. However, their ability to provide effective assistance in dynamic, real-world environments remains largely underexplored. Existing video benchmarks predominantly assess passive understanding through retrospective analysis or isolated perception tasks, failing to capture the interactive and adaptive nature of real-time user assistance. To bridge this gap, we introduce LifeEval, a multimodal benchmark designed to evaluate real-time, task-oriented human-AI collaboration in daily life from an egocentric perspective. LifeEval emphasizes three key aspects: task-oriented holistic evaluation, egocentric real-time perception from continuous first-person streams, and human-assistant collaborative interaction through natural dialogues. Constructed via a rigorous annotation pipeline, the benchmark comprises 4,075 high-quality question-answer pairs across 6 core capability dimensions. Extensive evaluations of 26 state-of-the-art MLLMs on LifeEval reveal substantial challenges in achieving timely, effective and adaptive interaction, highlighting essential directions for advancing human-centered interactive intelligence.

Crafting prompts via Prompt Engineering that steer a model's internal representations toward specific and pre-defined outcomes can be time-consuming, often requiring multiple iterations. Hard Prompt Inversion offers a complementary workflow: start from a reference image and generate a prompt that conditions a text-to-image (T2I) model to reconstruct the reference image. Existing inversion methods either yield incoherent text, or produce prompts that are overly sensitive to downstream token edits. We propose a dLLM-based prompt inversion framework that yield prompts that are (i) more interpretable to humans, (ii) better aligned with the reference image, and (iii) designed for downstream token swap and token append operations (aka edit-friendly prompts). The method is plug-and-play, requiring no finetuning of either the T2I model or the dLLM. Experiments across three datasets show a ~10xreduction in inversion time relative to existing prompt-inversion baselines, higher interpretability scores, and significantly higher prompt editability, as measured by TIFA, GPT-V preference scoring, and controlled user studies, all while preserving high-fidelity image generation. By coupling diffusion-time sampling with token-similarity control inside a dLLM decoder, our approach extends prompt inversion beyond reconstruction to downstream token-editing tasks, enabling faster, more transferable prompts that generalize across multiple T2I models.

Text-to-image (T2I) models such as Stable Diffusion and DALLE remain susceptible to generating harmful or Not-Safe-For-Work (NSFW) content under jailbreak attacks despite deployed safety filters. Existing jailbreak attacks either rely on proxy-loss optimization instead of the true end-to-end objective, or depend on large-scale and costly RL-trained generators. Motivated by these limitations, we propose JANUS , a lightweight framework that formulates jailbreak as optimizing a structured prompt distribution under a black-box, end-to-end reward from the T2I system and its safety filters. JANUS replaces a high-capacity generator with a low-dimensional mixing policy over two semantically anchored prompt distributions, enabling efficient exploration while preserving the target semantics. On modern T2I models, we outperform state-of-the-art jailbreak methods, improving ASR-8 from 25.30% to 43.15% on Stable Diffusion 3.5 Large Turbo with consistently higher CLIP and NSFW scores. JANUS succeeds across both open-source and commercial models. These findings expose structural weaknesses in current T2I safety pipelines and motivate stronger, distribution-aware defenses. Warning: This paper contains model outputs that may be offensive.