CVPR 2026 Papers

Most existing 3D keypoint estimation methods rely on manual annotations or calibrated multi-view images, both of which are expensive to collect.This paper introduces KeyDiff3D, a framework that can accurately predict 3D keypoints from a single image, thus eliminating the need for such expensive data acquisitions.To achieve this, we leverage powerful geometric priors embedded in a pretrained multi-view diffusion model.In our framework, the diffusion model generates multi-view images from a single image, serving as supervision signals to provide 3D geometric cues to our model.We also introduce a 3D feature extractor that transforms implicit 3D priors embedded in the diffusion features into explicit 3D feature volumes.Beyond accurate keypoint estimation, we further introduce a pipeline that enables manipulation of 3D objects generated by the diffusion model.Experimental results on diverse datasets, including Human3.6M, CUB-200-2011, Stanford Dogs, and several in-the-wild and out-of-domain inputs, highlight the effectiveness of our method in terms of accuracy, generalization, and its ability to enable manipulation of 3D objects generated by the diffusion model from a single image.

This paper focuses on embodied task planning, where an agent acquires visual observations from the environment and executes atomic actions to accomplish a given task. Although recent Vision-Language Models (VLMs) have achieved impressive results in multimodal understanding and reasoning, their performance remains limited when applied to embodied planning that involves multi-turn interaction, long-horizon reasoning, and extended context analysis. To bridge this gap, we propose RoboAgent, a capability-driven planning pipeline in which the model actively invokes different sub-capabilities. Each capability maintains its own context, and produces intermediate reasoning results or interacts with the environment according to the query given by a scheduler. This framework decomposes complex planning into a sequence of basic vision-language problems that VLMs can better address, enabling a more transparent and controllable reasoning process. The scheduler and all capabilities are implemented with a single VLM, without relying on external tools. To train this VLM, we adopt a multi-stage paradigm that consists of: (1) behavior cloning with expert plans, (2) DAgger training using trajectories collected by the model, and (3) reinforcement learning guided by an expert policy. Across these stages, we exploit the internal information of the environment simulator to construct high-quality supervision for each capability, and we further introduce augmented and synthetic data to enhance the model's performance in more diverse scenarios. Extensive experiments on widely used embodied task planning benchmarks validate the effectiveness of the proposed approach.

Nighttime environments pose significant challenges for camera-based perception, as existing methods passively rely on the scene lighting. We introduce Lighting-driven Dynamic Active Sensing (LiDAS), a closed-loop active illumination system that combines off-the-shelf visual perception models with high-definition headlights. Rather than uniformly brightening the scene, LiDAS dynamically predicts an optimal illumination field that maximizes downstream perception performance, i.e., decreasing light on empty areas to reallocate it on object regions. LiDAS enables zero-shot nighttime generalization of daytime-trained models through adaptive illumination control. Trained on synthetic data and deployed zero-shot in real-world closed-loop driving scenarios, LiDAS enables +18.7% mAP50 and +5.0% mIoU over standard low-beam at equal power. It maintains performances while reducing energy use by 40%. LiDAS complements domain-generalization methods, further strengthening robustness without retraining. By turning readily available headlights into active vision actuators, LiDAS offers a cost-effective solution to robust nighttime perception. Project page: https://simondemoreau.github.io/LiDAS/

Scientific images often require accurate numerical representations and correct object attributes. However, current faithfulness metrics are primarily tailored toward photorealistic, real-life imagery, rendering them ill-suited for scientific image evaluation. To address this gap, we introduce a novel evaluation model, SCIEval (Scientific Image Evaluation), which aims to capture faithfulness through three key dimensions: (i) Relevance, measuring overall text-image correspondence; (ii) Accuracy, examining the technical details of scientific objects; and (iii) Explainability, which isolates unfaithful elements within the generated content. To address these dimensions, we curate a specialized dataset of scientific text-image pairs to train three evaluation modules. For the Relevance and Accuracy modules, we propose a CLIP-based strategy that enhances scientific image perception through intra- and cross-modal contrastive learning. Concurrently, the Explainability module is developed by fine-tuning a high-performance Large Multimodal Model (LMM) using supervised rationale signals. Finally, we present SCIEval-Bench, a human-annotated evaluation benchmark consisting of 3,000 samples for scientific text-to-image and 3,000 samples for scientific image captioning. Extensive experiments on SCIEval-Bench demonstrate that our SCIEvalmodel is significantly more reliable than 24 competing models--including GPT-4o--exhibiting a superior correlation with human judgments.

Recent breakthroughs in Diffusion Transformers (DiTs) have revolutionized the field of visual synthesis due to their superior scalability. To facilitate DiTs' capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment. However, the underlying mechanisms governing representation learning within DiTs are not well understood. To this end, we first systematically investigate the representation dynamics of DiTs. Through analyzing the evolution and influence of internal representations under various settings, we reveal that representation diversity across blocks is a crucial factor for effective learning. Based on this key insight, we propose DiverseDiT, a novel framework that explicitly promotes representation diversity. DiverseDiT incorporates long residual connections to diversify input representations across blocks and a representation diversity loss to encourage blocks to learn distinct features. Extensive experiments on ImageNet 256x256 and 512x512 demonstrate that our DiverseDiT yields consistent performance gains and convergence acceleration when applied to different backbones with various sizes, even when tested on the challenging one-step generation setting. Furthermore, we show that DiverseDiT is complementary to existing representation learning techniques, leading to further performance gains. Our work provides valuable insights into the representation learning dynamics of DiTs and offers a practical approach for enhancing their performance. Our code is available at https://github.com/kobeshegu/DiverseDiT.

GUI agents powered by Multimodal Large Language Models (MLLMs) have demonstrated impressive capability in understanding and executing user instructions. However, accurately grounding instruction-relevant elements from high-resolution screenshots cluttered with irrelevant UI components remains challenging for existing approaches. Inspired by how humans dynamically adjust their perceptual scope to locate task-related regions on complex screens, we propose DRS-GUI, a training-free dynamic region search framework for GUI grounding that can be seamlessly integrated into existing MLLMs. DRS-GUI introduces a lightweight UI Perceptor that performs three human-like perceptual actions (Focus, Shift, and Scatter) to progressively explore the interface and generate region proposals. To dynamically schedule these actions, we further design an Action Planner based on Monte Carlo Tree Search (MCTS). A region quality reward is employed to evaluate and select the highly instruction-relevant region, efficiently pruning redundant UI elements. Experiments demonstrate that DRS-GUI yields a 14% improvement on ScreenSpot-Pro for general and GUI-specific MLLMs (Qwen2.5-VL-7B and UGround-V1-7B), significantly enhancing grounding performance and generalization.

Reliable long-horizon trajectory prediction requires both high positional accuracy and physically plausible temporal motion consistency. However, existing methods suffer from two fundamental limitations. First, they overlook the inherent difference in prediction logic: near-future trajectories are primarily governed by historical dynamics, whereas distant-future behaviors are driven by high-level semantic context. Yet, most methods employ a unified decoding pathway that blurs the temporal distinction.Second, although the near future is relatively easier to predict, existing methods lack mechanisms for coherent trajectory propagation across time horizons, often resulting in kinematically implausible predictions with inconsistent heading evolution and degraded long-horizon performance. To address these challenges, we propose NDPNet, a dual-stage architecture that decouples near- and distant-horizon modeling into specialized pathways, with a dedicated transition module ensuring smooth temporal bridging. Furthermore, we introduce a novel motion-aware coherence loss that explicitly embeds kinematic priors to enforce trajectory consistency. Extensive experiments show that NDPNet achieves SOTA performance on Argoverse 2 and WOMD. Notably, on WOMD, it ranks 1^ \text st in both minFDE _6 and minADE _6 across all standard horizons (3s, 5s, 8s) without ensemble learning or NMS post-processing, and is the first to achieve sub-1.75 minFDE _6 for 8s prediction, surpassing prior methods by a large margin. The code will be released subsequently.

Capturing both geometry and rigid motion for structured dynamic objects, like multi-part assemblies or jointed mechanisms, remains a key challenge. Existing dynamic methods, such as deformable meshes or 3DGS, rely on unstructured representations and fail to jointly model suitable geometry and articulated motion. Primitive-based methods excel at structured static scenes, but their dynamic potential is still unexplored. We propose D-Prism, the first framework to achieve high-fidelity structured dynamic modeling by extending differentiable primitives to the dynamic domain.Specifically, we bind 3DGS to primitive surfaces, leveraging their respective strengths in appearance and geometry. We introduce a deformation network to control primitive motion, ensuring it accurately matches the object's movement. Furthermore, we design a novel adaptive control strategy to dynamically adjust primitive counts, better matching objects' true spatial footprint.Experiments confirm that our method excels at structured dynamic modeling, providing both structured geometry and precise motion tracking. Project page: https://zju3dv.github.io/d-prism/.

3D Gaussian Splatting, known for enabling high-quality static scene reconstruction with fast rendering, is increasingly being applied to multi-view dynamic scene reconstruction. A common strategy involves learning a deformation field to model the temporal changes of a canonical set of 3D Gaussians. However, these deformation-based methods often produce blurred renderings and lose fine motion details in highly dynamic regions due to the inherent limitations of a single, unified model in representing diverse motion patterns. To address these challenges, we introduce Motion-Aware Partitioning of Deformable 3D Gaussian Splatting (MAPo), a novel framework for high-fidelity dynamic scene reconstruction. Its core is a dynamic score-based partitioning strategy that distinguishes between high- and low-dynamic 3D Gaussians. For high-dynamic 3D Gaussians, we recursively partition them temporally and duplicate their deformation networks for each new temporal segment, enabling specialized modeling to capture intricate motion details. Concurrently, low-dynamic 3DGs are treated as static to reduce computational costs. However, this temporal partitioning strategy for high-dynamic 3DGs can introduce visual discontinuities across frames at the partition boundaries. To address this, we introduce a cross-frame consistency loss, which not only ensures visual continuity but also further enhances rendering quality. Extensive experiments demonstrate that MAPo achieves superior rendering quality compared to baselines while maintaining comparable computational costs, particularly in regions with complex or rapid motions.

The effectiveness of Contrastive Language-Image Pre-training (CLIP) models critically depends on the semantic diversity and quality of their training data. However, while existing synthetic data generation methods primarily focus on increasing data volume, such emphasis often leads to limited semantic diversity and redundant or shallow captions. To address this limitation, we propose Role-SynthCLIP, a novel data synthesis framework that leverages multi-perspective role-playing prompts (e.g., a compositional analyst, an interpreter of image context) to guide Multimodal Large Language Models (MLLMs) in generating semantically diverse captions from distinct viewpoints. This mechanism enhances the semantic diversity and fine-grained image-text alignment of synthetic pairs by enriching each image with multiple complementary captions, while keeping the number of training images fixed. Experimental results demonstrate the effectiveness and efficiency of our method. A CLIP-B/16 model trained on only 1 million images achieves a Recall@1 of 64.1% on the MS COCO validation set, surpassing the best existing synthetic data baseline (trained on 5M pairs) by 2.8 percentage points. The code and trained models can be found at https://github.com/huangfu170/Role-SynthCLIP.

We present DeSOPE, a large-scale dataset for 6DoF deformed objects. Most 6D object pose methods assume rigid or articulated objects, an assumption that fails in practice as objects deviate from their canonical shapes due to wear, impact, or deformation. To model this, we introduce the DeSOPE dataset, which features high-fidelity 3D scans of 26 common object categories, each captured in one canonical state and three deformed configurations, with accurate 3D registration to the canonical mesh. Additionally, it features an RGB-D dataset with 133K frames across diverse scenarios and 665K pose annotations produced via a semi-automatic pipeline. We begin by annotating 2D masks for each instance, then compute initial poses using an object pose method, refine them through an object-level SLAM system, and finally perform manual verification to produce the final annotations. We evaluate several object pose methods and find that performance drops sharply with increasing deformation, suggesting that robust handling of such deformations is critical for practical applications. The project page and dataset are available at https://desope-6d.github.io/.

The performance of multimodal learning systems, particularly in high-stakes domains like automated depression recognition, is fundamentally constrained by the challenge of learning robust visual representations from limited and complex clinical data. To overcome this, we introduce Cross-Modal Guided Visual Synthesis (CMG-VS), a novel training framework that internally enhances the learning process by synthesizing new, task-relevant visual features. At its core, CMG-VS leverages the rich context from audio and text modalities to guide a conditional generative model. This model learns the intricate mapping from speech and language to visual expression, generating a diverse manifold of plausible visual behaviors to enrich the training distribution. Crucially, this synthesis is not a separate pre-processing step. Through a task-guided joint optimization scheme, the generative process is dynamically steered by the downstream multimodal recognizer's performance. This closed-loop feedback mechanism ensures the synthesized visual features are optimized to be maximally discriminative for the recognition task, rather than merely realistic. Comprehensive experiments on the widely-used DAIC-WOZ and E-DAIC benchmark datasets demonstrate that CMG-VS significantly outperforms existing state-of-the-art methods across all standard regression and classification metrics. Ablation studies further validate that our task-guided synthesis is the key driver of this performance gain, proving its effectiveness as a new paradigm for robust multimodal representation learning.

This work introduces a new task, text-conditioned selective video-to-audio (V2A) generation, which produces only the user-intended sound from a multi-object video. This capability is especially crucial in multimedia production, where audio tracks are handled individually for each sound source for precise editing, mixing, and creative control. We propose SELVA, a novel text-conditioned V2A model that treats the text prompt as an explicit selector to distinctly extract prompt-relevant sound-source visual features from the video encoder. To suppress text-irrelevant activations with efficient video encoder finetuning, the proposed supplementary tokens promote cross-attention to yield robust semantic and temporal grounding. SELVA further employs an autonomous video-mixing scheme in a self-supervised manner to overcome the lack of mono audio track supervision. We evaluate SELVA on VGG-MONOAUDIO, a curated benchmark of clean single-source videos for such a task. Extensive experiments and ablations consistently verify its effectiveness across audio quality, semantic alignment, and temporal synchronization.

The heavy computational burden of Large Vision-Language Models (LVLMs) stems primarily from the lengthy visual token sequences generated by their vision encoders. To mitigate this, recent work has shifted towards pruning tokens within the vision encoder. However, we observe that these methods predominantly rely on a suboptimal decoupled heuristic method. This strategy is conceptually flawed: it is prone to sampling collapse, fails to fundamentally eliminate token redundancy, and tends to systematically discard secondary yet important semantic clusters. Addressing this limitation, this paper proposes to formalize visual token pruning as a unified Representativeness Optimization Problem. We introduce SCoRe (Salience-Coverage Reduction), a unified optimization method theoretically grounded in the Weighted k-Center Problem. SCoRe constructs the final token set greedily: at each iteration, it selects the token that maximizes the unified representativeness score, thereby optimizing global representativeness. Extensive experiments demonstrate that SCoRe achieves State-of-the-Art (SOTA) performance across multiple benchmarks. Notably, with negligible computational overhead, our method reduces tokens by 94.4% while retaining 95% of the full performance.

Despite the strong multimodal performance, large vision-language models (LVLMs) are vulnerable during fine-tuning to backdoor attacks, where adversaries insert trigger-embedded samples into the training data to implant behaviors that can be maliciously activated at test time. Existing defenses typically rely on retraining backdoored parameters (e.g., adapters or LoRA modules) with clean data, which is computationally expensive and often degrades model performance. In this work, we provide a new mechanistic understanding of backdoor behaviors in LVLMs: the trigger does not influence prediction through low-level visual patterns, but through abnormal cross-modal attention redistribution, where trigger-bearing visual tokens steal attention away from the textual context -- a phenomenon we term attention stealing. Motivated by this, we propose CleanSight, a training-free, plug-and-play defense that operates purely at test time. CleanSight (i) detects poisoned inputs based on the relative visual-text attention ratio in selected cross-modal fusion layers, and (ii) purifies the input by selectively pruning the suspicious high-attention visual tokens to neutralize the backdoor activation. Extensive experiments show that CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving the model's utility on both clean and poisoned samples.

Unsupervised domain adaptation (UDA) for semantic segmentation seeks to transfer models from a labeled source domain to an unlabeled target domain. While auxiliary self-supervised tasks such as contrastive learning have enhanced feature discriminability, masked modeling remains underexplored due to architectural constraints and misaligned objectives. We propose Masked Representation Modeling (MRM), an auxiliary task that performs representation masking and reconstruction directly in the latent space. Unlike prior masked modeling methods that reconstruct low-level signals (e.g., pixels or visual tokens), MRM targets high-level semantic features, aligning its objective with segmentation and integrating seamlessly into standard architectures like DeepLab and DAFormer. To support efficient reconstruction, we design a lightweight auxiliary module, Rebuilder, which is jointly trained with the segmentation network but removed during inference, introducing zero test-time overhead. Extensive experiments demonstrate that MRM consistently improves segmentation performance across diverse architectures and UDA benchmarks. When integrated with four representative baselines, MRM achieves an average gain of +2.3 mIoU on GTA -> Cityscapes and +2.8 mIoU on Cityscapes -> Synthia, establishing it as a simple, effective, and generalizable strategy for unsupervised domain-adaptive semantic segmentation.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in visual reasoning across diverse tasks. These tasks place different demands on visual representations: some prioritize high-level global context, while others emphasize fine-grained local details. However, most existing methods operate on visual representations primarily in the spatial domain, lacking an explicit mechanism for distinguishing between high-frequency local details and low-frequency global context. This limitation hinders fine-grained control of visual representations and complicates their hierarchical alignment with language. To address this issue, we introduce Language-guided Frequency Modulation (LFM), a plug-and-play approach that adaptively refines visual signals in the frequency domain under linguistic guidance. By selectively enhancing critical regions and details, LFM enables more structured and precise visual processing. Crucially, it adds no extra training parameters, relying solely on a lightweight learnable projector to refine visual tokens before integration into the LLM, thereby ensuring minimal computational overhead. Extensive experiments across diverse vision-language benchmarks highlight LFM's scalability, effectiveness, and broad applicability to LVLMs.

Training-free zero-shot composed image retrieval models are recently gaining increasing research interest due to their generalizability and flexibility in unseen multimodal retrieval. Recent LLM-based advances focus on generating the expected target caption by exploring the compositional ability behind the LLMs. Although efficient, we find that 1) the generated captions tend to introduce unexpected features from the reference image due to the semantic gap between the input image and text modification, where the image contains much more details than the text; 2) the point-to-point alignment during the retrieval stage fails to capture diverse compositions.To address these challenges, we introduce a novel Semantic Transition and Transportation in Collaboration framework for training-free zero-shot CIR tasks. Specifically, given the composed caption inferred by an LLM, we aim to refine it through a transition vector in the embedding space and make it closer to the target image. Combining LLMs with user instruction, the refined caption concentrates more on the core modification intent and thus filters out unnecessary noise. Moreover, to explore diverse alignment during the retrieval stage, we model the caption and image as discrete distributions and reformulate the retrieval task as a set-to-set alignment task. Finally, a bidirectional transportation distance is developed to consider fine-grained alignments across modalities and calculate the retrieval score.Extensive experimentsdemonstrate that our method can be general, effective, and beneficial for many CIR tasks.The code is attached in the supplementary material.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in vision-language understanding. Yet, human perception is inherently multisensory, integrating sight, sound, and motion to reason about the world. Among these modalities, sound provides indispensable cues about spatial layout, off-screen events, and causal interactions, particularly in egocentric settings where auditory and visual signals are tightly coupled. To this end, we introduce EgoSound, the first benchmark designed to systematically evaluate egocentric sound understanding in MLLMs. EgoSound unifies data from Ego4D and EgoBlind, encompassing both sighted and sound-dependent experiences. It defines a seven-task taxonomy spanning intrinsic sound perception, spatial localization, causal inference, and cross-modal reasoning. Constructed through a multi-stage auto-generative pipeline, EgoSound contains 7315 validated QA pairs across 900 videos. Comprehensive experiments on nine state-of-the-art MLLMs reveal that current models exhibit emerging auditory reasoning abilities but remain limited in fine-grained spatial and causal understanding. EgoSound establishes a challenging foundation for advancing multisensory egocentric intelligence, bridging the gap between seeing and truly hearing the world. Project page: https://groolegend.github.io/EgoSound/.