CVPR 2026 Papers

Recent Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language understanding tasks, yet their inference efficiency is often hampered by the large number of visual tokens, particularly in high-resolution or multi-image scenarios. To address this issue, we propose EvoComp, a visual token compression framework that significantly reduces token count while preserving task accuracy. EvoComp introduces a lightweight encoder-only transformer-based compressor that selects the most informative and non-redundant visual tokens by jointly considering visual and textual contexts. A core challenge lies in providing effective supervision for training the compressor. To this end, we design an evolutionary labeling strategy that searches for token subsets minimizing the MLLM's output loss, while enforcing semantic diversity through vocabulary-based token grouping. We further train the compressor using a tailored loss function combining the GHM loss to mitigate class and difficulty imbalance, and a cosine similarity regularization to encourage semantic separation between retained and discarded tokens. Extensive experiments across multiple vision-language benchmarks show that EvoComp outperforms existing methods based on attention or similarity heuristics. Notably, it retains 99.3% of the original accuracy under 3x token compression and delivers up to 1.6x speedup on mobile devices.

Reasoning segmentation has recently expanded from ground-level scenes to remote-sensing imagery, yet UAV data poses distinct challenges, including oblique viewpoints, ultra-high resolutions, and extreme scale variations. To address these issues, we formally define the UAV Reasoning Segmentation task and organize its semantic requirements into three dimensions: Spatial, Attribute, and Scene-level reasoning. Based on this formulation, we construct DRSeg, a large-scale benchmark for UAV reasoning segmentation, containing 10k high-resolution aerial images paired with Chain-of-Thought QA supervision across all three reasoning types. As a benchmark companion, we introduce PixDLM, a simple yet effective pixel-level multimodal language model that serves as a unified baseline for this task. Experiments on DRSeg establish strong baseline results and highlight the unique challenges of UAV reasoning segmentation, providing a solid foundation for future research. All datasets, models, and code are available at https://github.com/XIEFOX/PixDLM.

Event cameras provide superior temporal resolution, dynamic range, energy efficiency, and pixel bandwidth. Spiking Neural Networks (SNNs) naturally complement event data through discrete spike signals, making them ideal for event-based tracking. However, current approaches combining Artificial Neural Networks (ANNs) and SNNs suffer from suboptimal architectures that compromise energy efficiency and limit tracking performance. To address these limitations, we propose the first Transformer-based Spike-Driven Tracking (SDTrack) pipeline. It incorporates a novel event frame aggregation method called Global Trajectory Prompt (GTP) and a Transformer-based tracker. The GTP method effectively captures global trajectory information and aggregates it with event streams into event frames to enhance spatiotemporal representation. The Transformer-based tracker comprises a fully spike-driven SNN backbone and a simple tracking head. The SDTrack pipeline operates end-to-end without data augmentation or post-processing. Extensive experiments demonstrate that our SDTrack-Tiny pipeline achieves competitive accuracy with only 19.61M parameters and 8.16mJ energy consumption, while our Base version achieves state-of-the-art accuracy across three datasets. Our work establishes a solid foundation for future neuromorphic vision research.

Under-display cameras (UDCs) allow for full-screen designs by positioning the imaging sensor underneath the display. Nonetheless, light diffraction and scattering through the various display layers result in spatially varying and complex degradations, which significantly reduce high-frequency details. Current PSF-based physical modeling techniques and frequency-separation networks are effective at reconstructing low-frequency structures and maintaining overall color consistency. However, they still face challenges in recovering fine details when dealing with complex, spatially varying degradation. To solve this problem, we propose a lightweight Uncertainty-aware Context-Memory Network (UCMNet), for UDC image restoration. Unlike previous methods that apply uniform restoration, UCMNet performs uncertainty-aware adaptive processing to restore high-frequency details in regions with varying degradations. The estimated uncertainty maps, learned through an uncertainty-driven loss, quantify spatial uncertainty induced by diffraction and scattering, and guide the Memory Bank to retrieve region-adaptive context from the Context Bank. This process enables effective modeling of the non-uniform degradation characteristics inherent to UDC imaging. Leveraging this uncertainty as a prior, UCMNet achieves state-of-the-art performance on multiple benchmarks with 30% fewer parameters than previous models. Project page: https://kdhrick2222.github.io/projects/UCMNet.

Existing set-based gait recognition methods achieve remarkable performance by capturing global semantic context.However, their order-invariant nature prevents them from modeling the fine-grained kinematic patterns that unfold over time.To unify the global and process-level representations, we propose GaitMax, a framework that captures both semantic context and kinematic motion.GaitMax leverages attention-based spatiotemporal modeling to dynamically represent detailed part-level trajectories.While this detailed representation is more powerful, it also captures more nuisance factors (e.g., clothing, viewpoint), leading to potential shortcuts.To mitigate this, we introduce CDLoss, a Conditional Decorrelation Loss that explicitly disentangles the gait embeddings from nuisance factors using vision-language supervision.This loss requires high-quality nuisance descriptions. We therefore construct GCaption, a new resource that provides natural language annotations for multiple gait datasets, moving beyond simple categorical labels. GCaption not only enables CDLoss but also serves as a foundation for future context-aware gait analysis.The superiority of GaitMax is validated through extensive experiments on multiple large-scale gait benchmarks. Models, code, and resources will be released upon publication.

Ultra-high-resolution (UHR) remote sensing (RS) images offer rich fine-grained information but also present challenges in effective processing. Existing dynamic resolution and token pruning methods are constrained by a passive perception paradigm, suffering from increased redundancy when obtaining finer visual inputs. In this work, we explore a new active perception paradigm that enables models to revisit information-rich regions. First, we present LRS-GRO, a large-scale benchmark dataset tailored for active perception in UHR RS processing, encompassing 17 question types across global, region, and object levels, annotated via a semi-automatic pipeline. Building on LRS-GRO, we propose ZoomEarth, an adaptive cropping-zooming framework with a novel Region-Guided reward that provides fine-grained guidance. Trained via supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), ZoomEarth achieves state-of-the-art performance on LRS-GRO and, in the zero-shot setting, on three public UHR remote sensing benchmarks. Furthermore, ZoomEarth can be seamlessly integrated with downstream models for tasks such as cloud removal, denoising, segmentation, and image editing through simple tool interfaces, demonstrating strong versatility and extensibility. All data and code will be released at https://earth-insights.github.io/ZoomEarth.

In this paper, we present Chain-of-Models Pre-Training (CoM-PT), a novel performance-lossless training acceleration method for vision foundation models (VFMs). This approach fundamentally differs from existing acceleration methods in its core motivation: rather than optimizing each model individually, CoM-PT is designed to accelerate the training pipeline at the model family level, scaling efficiently as the model family expands. Specifically, CoM-PT establishes a pre-training sequence for the model family, arranged in ascending order of model size, called model chain. In this chain, only the smallest model undergoes standard individual pre-training, while the other models are efficiently trained through sequential inverse knowledge transfer from their smaller predecessors by jointly reusing the knowledge in the parameter space and the feature space. As a result, CoM-PT enables all models to achieve performance that is mostly superior to standard individual training while significantly reducing training cost, and this is extensively validated across 45 datasets spanning zero-shot and fine-tuning tasks. Notably, its efficient scaling property yields a remarkable phenomenon: training more models even results in higher efficiency. For instance, when pre-training on CC3M: i) given ViT-L as the largest model, progressively prepending smaller models to the model chain reduces computational complexity by up to 72%; ii) within a fixed model size range, as the VFM family scales across 3, 4, and 7 models, the acceleration ratio of CoM-PT exhibits a striking leap: from 4.13x to 5.68x and 7.09x. Since CoM-PT is naturally agnostic to specific pre-training paradigms, we open-source the code to spur further extensions in more computationally intensive scenarios, such as large language model pre-training.

Modern video-language models are typically trained on videos downsampled to low frames-per-second (FPS), and the most commonly used evaluation benchmarks are designed for low-FPS input as well. To address this shortcoming, we present FPS-Bench, a large video question-answering benchmark designed to evaluate VLMs' capabilities to understand video at high-frame rates. We introduce a new metric, the minimum frames-per-second (minFPS), which measures the minimum frame-rate required to solve a given question. While existing benchmarks require <1 minFPS, we rigorously curate more than 1000 questions from a diverse source of videos and manually verify minFPS for each example, leading to a benchmark that requires watching videos at on average 7 FPS to solve. Our evaluation of several state-of-the-art VLMs shows that they are severely lacking, achieving QA accuracy of 30% in the FPS-Bench multiple-choice task, while humans achieve 72% accuracy.

Scaling up multimodal models has enabled remarkable advances in visual understanding and reasoning, but practical demands call for smaller, efficient systems. In this work, we conduct a principled analysis of downscaling intelligence in multimodal models, examining how reduced large language model (LLM) capacity affects multimodal capabilities. Our initial findings reveal an interesting trend: LLM downscaling disproportionately affects visual capabilities, rather than abilities inherited from the LLM. We then examine whether this drop mainly reflects the expected decline in visual reasoning or a more fundamental loss of perceptual abilities. Isolating the effect of LLM downscaling on perception, we find performance still drops sharply, often matching or exceeding the impact on reasoning. To address this bottleneck, we introduce visual extraction tuning, which explicitly trains the model to extract instruction-relevant visual details consistently across tasks. With these extracted visual details, we then apply step-by-step reasoning to generate answers. Together, these components form our Extract+Think approach, setting a new standard for efficiency and performance in this space.

Vision-language models (VLMs) have achieved remarkable multimodal understanding and reasoning capabilities, yet remain computationally expensive due to dense visual tokenization. Existing efficiency approaches either merge redundant visual tokens or drop them progressively in language backbone, often trading accuracy for speed. In this work, we propose DUET-VLM, a versatile plug-and-play dual compression framework that consists of (a) vision-only redundancy aware compression of vision encoder's output into information-preserving tokens, followed by (b) layer-wise, salient text-guided dropping of visual tokens within the language backbone to progressively prune less informative tokens. This coordinated token management enables aggressive compression while retaining critical semantics. On LLaVA-1.5-7B, our approach maintains over 99% of baseline accuracy with 67% fewer tokens, and still retains >97% even at 89% reduction. With this dual-stage compression during training, it achieves 99.7% accuracy at 67% and 97.6% at 89%, surpassing prior SoTA visual token reduction methods across multiple benchmarks. When integrated into Video-LLaVA-7B, it even surpasses the baseline--achieving >100% accuracy with a substantial 53.1% token reduction and retaining 97.6% accuracy under an extreme 93.4% setting. These results highlight end-to-end training with DUET-VLM, enabling robust adaptation to reduced visual (image/video) input without sacrificing accuracy, producing compact yet semantically rich representations within the same computational budget. Our code is available at https://github.com/AMD-AGI/DUET-VLM.

Recent vision-language-action (VLA) models for multi-task robot manipulation often rely on fixed camera setups and shared visual encoders, which limit their performance under occlusions and during cross-task transfer. To address these challenges, we propose Task-aware Virtual View Exploration (TVVE), a framework that learns to select task-relevant virtual camera viewpoints and dynamically re-render observations from a reconstructed scene representation using the selected viewpoints. To enable efficient view selection, we train an exploration policy in a pseudo-environment. In addition, we introduce a Task-aware Mixture-of-Experts (TaskMoE) visual encoder that routes visual features to task-specialized experts, mitigating interference in multi-task learning. To evaluate robustness under distribution shifts, we construct RLBench-OG, an out-of-distribution benchmark with visual perturbations and camera pose variations. Experiments on RLBench and RLBench-OG demonstrate that TVVE achieves higher success rates than strong baselines, while real-robot experiments further confirm its robustness to visual disturbances and unseen instructions. Code and visualizations are available at: https://hcplab-sysu.github.io/TAVP.

Learning-based real image dehazing methods have achieved notable progress, yet they still face adaptation challenges in diverse real haze scenes. These challenges mainly stem from the lack of effective unsupervised mechanisms for unlabeled data and the heavy cost of full model fine-tuning. To address these challenges, we propose the haze-to-clear text-directed loss that leverages CLIP's cross-modal capabilities to reformulate real image dehazing as a semantic alignment problem in latent space, thereby providing explicit unsupervised cross-modal guidance in the absence of reference images. Furthermore, we introduce the Bilevel Layer-positioning LoRA (BiLaLoRA) strategy, which learns both the LoRA parameters and automatically search the injection layers, enabling targeted adaptation of critical network layers. Extensive experiments demonstrate our superiority against state-of-the-art methods on multiple real-world dehazing benchmarks.

OmniLottie is a versatile framework that generates high-quality vector animations from multi-modal instructions, including interleaved texts, images, and videos. To fully parameterize vector animations for flexible motion and visual content control, we seek help from the Lottie representation, which encodes both shapes and animated behaviors in a single JSON file. Building upon a pretrained vision-language model (VLM), OmniLottie produces vivid, semantically aligned vector animations that adhere closely to multi-modal conditions. To avoid the complexity and irregularity of raw JSON structures, we introduce a dedicated Lottie tokenizer that transforms Lottie files into structured sequences of function calls representing shapes, animation commands, and their parameters. This design enables the model to directly learn the underlying shape and animation priors from data, substantially improving generation stability and controllability. To further advance research in vector animation generation, we curate MMLottie-2M, a large-scale dataset of professionally designed vector animations paired with textual and visual annotations. Leveraging the well-designed tokenizer and our newly established dataset, OmniLottie demonstrates strong multi-modal conditional generation capabilities using a simple next-token prediction objective. For qualitative results, please refer to the generated animations rendered through standard Lottie players on the supplementary website.

Recent advances have markedly improved the cross-scene generalization of relative depth estimation, yet its practical applicability remains limited by the absence of metric scale, local inconsistencies, and low computational efficiency. To address these issues, we present Midas Touch for Depth (MTD), a mathematically interpretable approach that converts relative depth into metric depth using only extremely sparse 3D data. To eliminate local scale inconsistencies, it applies a segment-wise recovery strategy via sparse graph optimization, followed by a pixel-wise refinement strategy using a discontinuity-aware geodesic cost. MTD exhibits strong generalization and achieves substantial accuracy improvements over previous depth completion and depth estimation methods. Moreover, its lightweight, plug-and-play design facilitates deployment and integration on diverse downstream 3D tasks.

The rise of vision-language models (VLMs) has driven the initial exploration of open-vocabulary remote sensing image semantic segmentation (OVRSIS), enabling recognition of unseen categories in complex Earth observation scenes. However, existing methods primarily focus on enhancing visual representations of domain-specific remote sensing images, while overlooking the effect of textual information. In this paper, we argue that there exists a crucial issue of textual ambiguity in OVRSIS task, limiting final segmentation performance. Therefore, we propose a plug-and-play yet effective Test-time Multi-Prompt Adaptation (TMPA) method to mitigate textual ambiguity in OVRSIS. Specifically, TMPA first generates diverse, context-aware descriptions for each category instead of the naive class name by executing a large language model with a task-driven prompt, which can effectively avoid some textual ambiguity, i.e., background class has different meanings in various tasks. Furthermore, TMPA develops a visual-guided test-time adaptation strategy for the generated multi-prompts, which adaptively refines the prompt representations of each category with high-confidence visual features for the uncertain predictions with high entropy, making TMPA better applicable to different scenarios. Particularly, a pixel-level loss with entropy minimization is proposed to optimize the text prompt with a bias during inference, where prompt bias is constructed based on a weighted combination of high-confidence visual features. Our TMPA can be flexibly integrated into existing methods for boosting their performance. Extensive experiments are conducted on 17 remote sensing datasets, and the results show our TMPA can significantly improve its counterparts, while achieving state-of-the-art performance.

We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent's own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.

Manual font design is an intricate process that transforms a stylistic visual concept into a coherent glyph set. This challenge persists in automated Few-shot Font Generation (FFG), where models struggle to preserve both structural integrity and stylistic fidelity from limited references. While autoregressive (AR) models have demonstrated impressive generative capabilities, their application to FFG is constrained by conventional patch-level tokenization, which neglects global dependencies crucial for coherent font synthesis. Moreover, existing FFG methods remain within the image-to-image paradigm, relying solely on visual references and overlooking the role of language in conveying stylistic intent during font design. To address these limitations, we propose GAR-Font, a novel AR framework for multimodal few-shot font generation. GAR-Font introduces a global-aware tokenizer that effectively captures both local structures and global stylistic patterns, a multimodal style encoder offering flexible style control through a lightweight language-style adapter without requiring intensive multimodal pretraining, and a post-refinement pipeline that further enhances structural fidelity and style coherence. Extensive experiments show that GAR-Font outperforms existing FFG methods, excelling in maintaining global style faithfulness and achieving higher-quality results with textual stylistic guidance. Project Page: https://xtryer-s.github.io/projects_pages/GAR_Font.

Recent advances in image editing models have shown remarkable progress. A common architectural design couples a multimodal large language model (MLLM) encoder with a diffusion decoder, as seen in systems such as Step1X-Edit and Qwen-Image-Edit, where the MLLM encodes both the reference image and the instruction but remains frozen during training.In this work, we demonstrate that unlocking the reasoning capabilities of MLLM can further push the boundaries of editing models. Specifically, we explore two reasoning mechanisms, thinking and reflection, which enhance instruction understanding and editing accuracy.Based on that, our proposed framework enables image editing in a thinking-editing-reflection loop: the thinking mechanism leverages the world knowledge of MLLM to interpret abstract instructions, while the reflection reviews editing results, automatically corrects unintended manipulations, and identifies the stopping round.Extensive experiments demonstrate that our reasoning approach achieves significant performance gains, with improvements of ImgEdit(+4.4%), GEdit(+3.1%), and Kris(+11.5%) when initializing our DiT from the Step1X-Edit(ReasonEdit-S), and also outperforms previous open-source methods on both GEdit and Kris when integrated with Qwen-Image-Edit(ReasonEdit-Q).Code and checkpoints will be open-sourced.

Full-stack multimodal interaction in real-time is a central goal in building intelligent embodied agents capable of natural, dynamic communication. However, existing systems are either limited to unimodal generation or suffer from degraded reasoning and poor cross-modal alignment, preventing coherent and perceptually grounded interactions. In this work, we introduce U-Mind, the first unified system for high-intelligence multimodal dialogue that supports real-time generation and jointly models language, speech, motion, and video synthesis within a single interactive loop. At its core, U-Mind implements a Unified Alignment and Reasoning Framework that addresses two key challenges: enhancing cross-modal synchronization via a segment-wise alignment strategy, and preserving reasoning abilities through Rehearsal-Driven Learning. During inference, U-Mind adopts a text-first decoding pipeline that performs internal chain-of-thought planning followed by temporally synchronized generation across modalities. To close the loop, we implement a real-time video rendering framework conditioned on pose and speech, enabling expressive and synchronized visual feedback. Extensive experiments demonstrate that U-Mind achieves state-of-the-art performance on a range of multimodal interaction tasks, including question answering, instruction following, and motion generation, paving the way toward intelligent, immersive conversational agents.