CVPR 2026 论文列表

Text-to-video diffusion models have enabled open-ended video synthesis, but often struggle with generating the correct number of objects specified in a prompt. We introduce NUMINA, a training-free identify-then-guide framework for improved numerical alignment. NUMINA identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout. It then refines this layout conservatively and modulates cross-attention to guide regeneration. On the introduced CountBench, NUMINA improves counting accuracy by up to 7.4% on Wan2.1-1.3B, and by 4.9% and 5.5% on 5B and 14B models, respectively. Furthermore, CLIP alignment is improved while maintaining temporal consistency. These results demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion. The code is available at https://github.com/H-EmbodVis/NUMINA.

Embodied Visual Reasoning (EVR) seeks to follow complex, free-form instructions based on egocentric video, enabling semantic understanding and spatiotemporal reasoning in dynamic environments. Despite its promising potential, EVR encounters significant challenges stemming from the diversity of complex instructions and the intricate spatiotemporal dynamics in long-term egocentric videos. Prior solutions either employ Large Language Models (LLMs) over static video captions, which often omit critical visual details, or rely on end-to-end Vision-Language Models (VLMs) that struggle with stepwise compositional reasoning. Considering the complementary strengths of LLMs in reasoning and VLMs in perception, we propose CLiViS. It is a novel training-free framework that leverages LLMs for high-level task planning and orchestrates VLM-driven open-world visual perception to iteratively update the scene context. Building on this synergy, the core of CLiViS is a dynamic Cognitive Map that evolves throughout the reasoning process. This map constructs a structured representation of the embodied scene, bridging low-level perception and high-level reasoning. Extensive experiments across multiple benchmarks demonstrate the effectiveness and generality of CLiViS, especially in handling long-term visual dependencies. Code is available at https://github.com/Teacher-Tom/CLiViS.

Large Vision-Language Models (VLMs) often answer classic visual illusions "correctly" on original images, yet persist with the same responses when illusion factors are inverted, even though the visual change is obvious to humans. This raises a fundamental question: do VLMs perceive visual changes or merely recall memorized patterns? While several studies have noted this phenomenon, the underlying causes remain unclear. To move from observations to systematic understanding, this paper introduces VI-Probe, a controllable visual-illusion framework with graded perturbations and matched visual controls (without illusion inducer) that disentangles visually grounded perception from language-driven recall. Unlike prior work that focuses on averaged accuracy, we measure stability and sensitivity using Polarity-Flip Consistency, Template Fixation Index, and an illusion multiplier normalized against matched controls. Experiments across different families reveal that response persistence arises from heterogeneous causes rather than a single mechanism. For instance, GPT-5 exhibits memory override, Claude-Opus-4.1 shows perception-memory competition, while Qwen variants suggest visual-processing limits. Our findings challenge single-cause views and motivate probing-based evaluation that measures both knowledge and sensitivity to controlled visual change. Data and code are available at https://sites.google.com/view/vi-probe/.

Visual AutoRegressive (VAR) modeling has garnered significant attention for its innovative next-scale prediction paradigm. However, mainstream VAR paradigms attend to all tokens across historical scales at each autoregressive step. As the next scale resolution grows, the computational complexity of attention increases quartically with resolution, causing substantial latency. Prior accelerations often skip high-resolution scales, which speeds up inference but discards high-frequency details and harms image quality. To address these problems, we present SparVAR, a training-free acceleration framework that exploits three properties of VAR attention: (i) strong attention sinks, (ii) cross-scale activation similarity, and (iii) pronounced locality. Specifically, we dynamically predict the sparse attention pattern of later high-resolution scales from a sparse decision scale, and construct scale self-similar sparse attention via an efficient index-mapping mechanism, enabling high-efficiency sparse attention computation at large scales. Furthermore, we propose cross-scale local sparse attention and implement an efficient block-wise sparse kernel, which achieves > 5xfaster forward speed than FlashAttention. Extensive experiments demonstrate that the proposed SparVAR can reduce the generation time of an 8B model producing 1024x1024 high-resolution images to the 1s, without skipping the last scales. Compared with the VAR baseline accelerated by FlashAttention, our method achieves a 1.57xspeed-up while preserving almost all high-frequency details. When combined with existing scale-skipping strategies, SparVAR attains up to a 2.28xacceleration, while maintaining competitive visual generation quality.

When obtaining visual illustrations from text descriptions, today's methods take a description with a single text context--a caption, or an action description--and retrieve or generate the matching visual context. However, prior work does not permit visual illustration of multistep descriptions, e.g. a cooking recipe or a gardening instruction manual, and simply handling each step description in isolation would result in an incoherent demonstration. We propose Stitch-a-Demo, a novel retrieval-based method to assemble a video demonstration from a multistep description. The resulting video contains clips, possibly from different sources, that accurately reflect all the step descriptions, while being visually coherent. We formulate a training pipeline that creates large-scale weakly supervised data containing diverse procedures and injects hard negatives that promote both correctness and coherence. Validated on in-the-wild instructional videos, Stitch-a-Demo achieves state-of-the-art performance, with gains up to 29% as well as dramatic wins in a human preference study.

The surge of highly realistic synthetic videos produced by contemporary generative systems has significantly increased the risk of malicious use, challenging both humans and existing detectors. Against this backdrop, we take a generator-side view and observe that internal cross-attention mechanisms in these models encode fine-grained speech-motion alignment, offering useful correspondence cues for forgery detection. Building on this insight, we propose X-AVDT, a robust and generalizable deepfake detector that probes generator-internal audio-visual signals accessed via DDIM inversion to expose these cues. X-AVDT extracts two complementary signals: (i) a video composite capturing inversion-induced discrepancies, and (ii) audio-visual cross-attention feature reflecting modality alignment enforced during generation. To enable faithful, cross-generator evaluation, we further introduce MMDF, a new multi-modal deepfake dataset spanning diverse manipulation types and rapidly evolving synthesis paradigms, including GANs, diffusion, and flow-matching. Extensive experiments demonstrate that X-AVDT achieves leading performance on MMDF and generalizes strongly to external benchmarks and unseen generators, outperforming existing methods with accuracy improved by +13.1%. Our findings highlight the importance of leveraging internal audio-visual consistency cues for robustness to future generators in deepfake detection. The code and dataset will soon be available.

In the context of long-term video understanding with large multimodal models, many frameworks have been proposed. Although transformer-based visual compressors and memory-augmented approaches are often used to process long videos, they usually compress each frame independently and therefore fail to achieve strong performance on tasks that require understanding complete events, such as temporal ordering tasks in MLVU and VNBench. This motivates us to rethink the conventional one-way scheme from perception to memory, and instead establish a feedbackdriven process in which past visual contexts stored in the context memory can benefit ongoing perception. To this end, we propose Question-guided Visual Compression with Memory Feedback (QViC-MF), a framework for long-term video understanding. At its core is a Question-guided Multimodal Selective Attention (QMSA), which learns to preserve visual information related to the given question from both the current clip and the past related frames from the memory. The compressor and memory feedback work iteratively for each clip of the entire video. This simple yet effective design yields large performance gains on longterm video understanding tasks. Extensive experiments show that our method achieves significant improvement over current state-of-the-art methods by 6.1% on MLVU test, 8.3% on LVBench, 18.3% on VNBench Long, and 3.7% on VideoMME Long. The code will be available at https://github.com/FujitsuResearch/QViC-MF/.

Cross-modal alignment aims to learn semantically consistent latent representations across diverse modalities. Prevailing methods rely on a text-guided aggregation paradigm to achieve fine-grained alignment, while they suffer from redundant patch-word correlations and high computational costs. To address these issues, we propose CoV-Align, an effective and efficient fine-grained cross-modal alignment framework with cohesive visual semantics priority. Through a semantically convergent attention mechanism, it progressively aggregates meaningful visual patches in a text-free manner. We design a coarse visual semantic feature extractor that integrates deformable attention and consistent assign attention to group patches with semantic consistency. A cohesive and discriminative feature optimization is presented to enhance intra-semantic cohesion and inter-semantic discriminability of visual region features, resulting in explicit improvements in cross-modal alignment. Extensive experiments demonstrate that CoV-Align achieves state-of-the-art performance on the Flickr30K and MS-COCO benchmarks. Notably, it delivers a 3-5xcomputational speedup compared to pioneer approaches, offering compelling advantages for large-scale multi-modal tasks.

Large Vision-Language Models (LVLMs) have achieved remarkable success across diverse multimodal tasks, yet their practical deployment remains constrained by the computational burden arising from lengthy visual tokens. While visual token pruning has emerged as a promising solution, existing methods suffer from a fundamental limitation: once tokens are pruned at a specific layer, they become inaccessible to all subsequent layers, leading to premature information loss that can compromise model performance. Through empirical studies, we observe that different layers exhibit distinct visual region focus, indicating a varying optimal token subset across layers. Motivated by this insight, we propose Adaptive Layer-wise Visual Token Selection (ALVTS), a novel framework that breaks away from the conventional static token pruning paradigm. ALVTS incorporates a lightweight token selector to identify and route important tokens for further processing, while allowing less important tokens to skip the layer, thus minimizing computational redundancy. These two streams of tokens are seamlessly reintegrated before being fed into subsequent layers, facilitating adaptive compression across the entire model. Grounded in our importance consistency constrained low-rank approximation, the proposed token selection module closely emulates the full attention mechanism, effectively capturing its essential patterns without requiring model retraining. Extensive experiments on LLaVA-1.5, LLaVA-NeXT, and Qwen2.5-VL validate the effectiveness of our method. With an 89% token compression ratio, ALVTS retains 96.7% of the original model's accuracy, achieving a superior efficiency-accuracy trade-off for LVLM inference.

Current feed-forward 3D/4D reconstruction systems rely on dense geometry and pose supervision - expensive to obtain at scale and particularly scarce for dynamic real-world scenes. We present Flow3r, a framework that augments visual geometry learning with dense 2D correspondences ('flow') as supervision, enabling scalable training from unlabeled monocular videos. Our key insight is that the flow prediction module should be factored: predicting flow between two images using geometry latents from one and pose latents from the other. This factorization directly guides the learning of both scene geometry and camera motion, and naturally extends to dynamic scenes. In controlled experiments, we show that factored flow prediction outperforms alternative designs and that performance scales consistently with unlabeled data. Integrating factored flow into existing visual geometry architectures and training with 800K unlabeled videos, Flow3r achieves state-of-the-art results across eight benchmarks spanning static and dynamic scenes, with its largest gains on in-the-wild dynamic videos where labeled data is most scarce.

The threat of Audio-Video (AV) forgery is rapidly evolving beyond human-centric deepfakes to include more diverse manipulations across complex natural scenes. However, existing benchmarks are still confined to DeepFake-based forgeries and single-granularity annotations, thus failing to capture the diversity and complexity of real-world forgery scenarios. To address this, we introduce AVFakeBench, the first comprehensive audio-video forgery detection benchmark that spans rich forgery semantics across both human subject and general subject. AVFakeBench comprises 12K carefully curated audio-video questions, covering seven forgery types and four levels of annotations. To ensure high-quality and diverse forgeries, we propose a multi-stage hybrid forgery framework that integrates proprietary models for task planning with expert generative models for precise manipulation. The benchmark establishes a multi-task evaluation framework covering binary judgment, forgery types classification, forgery detail selection, and explanatory reasoning. We evaluate 11 Audio-Video Large Language Models (AV-LMMs) and 2 prevalent detection methods on AVFakeBench, demonstrating the potential of AV-LMMs as emerging forgery detectors while revealing their notable weaknesses in fine-grained perception and reasoning.

Multimodal large language models have demonstrated remarkable capabilities in 2D vision, motivating their extension to 3D scene understanding. Recent studies represent 3D scenes as 3D spatial videos composed of image sequences with depth and camera pose information, enabling pre-trained video-language models to perform 3D reasoning tasks. However, the large number of visual tokens in spatial videos remains a major bottleneck for efficient inference. Existing pruning methods overlook the view consistency of spatial videos and the spatial diversity of the remaining tokens, which prevents them from effectively removing inter-frame redundancy and preserving scene completeness. In this paper, we propose Geo3DPruner, a Geometry-Guided 3D Visual Token Pruning framework. Geo3DPruner first models cross-frame relevance through geometry-aware global attention, and then performs a two-stage pruning process. The intra-voxel stage selects representative multi-view features within each voxel, while the inter-voxel stage preserves spatial diversity by selecting a globally distributed subset of voxels. Extensive experiments on multiple 3D scene understanding benchmarks demonstrate that Geo3DPruner retains over 90% of the original performance while pruning 90% of visual tokens, significantly outperforming existing text-guided and vision-guided pruning methods.

Visual Deformation Measurement (VDM) aims to recover dense deformation fields by tracking surface motion from camera observations. Traditional image-based methods rely on minimal inter-frame motion to constrain the correspondence search space, which limits their applicability to highly dynamic scenes or necessitates high-speed cameras at the cost of prohibitive storage and computational overhead. We propose an event-frame fusion framework that exploits events for temporally dense motion cues and frames for spatially dense precise estimation. Revisiting the solid elastic modeling prior, we propose an Affine Invariant Simplicial (AIS) framework. It partitions the deformation field into linearized sub-regions with low-parametric representation, effectively mitigating motion ambiguities arising from sparse and noisy events. To speed up optimization and reduce error accumulation, a neighborhood-greedy optimization strategy is introduced, enabling well-converged sub-regions to guide their poorly-converged neighbors, effectively suppress local error accumulation in long-term dense tracking. To evaluate the proposed method, a benchmark dataset with temporally aligned event streams and frames is established, encompassing over 120 sequences spanning diverse deformation scenarios. Experimental results show that our method outperforms the state-of-the-art baseline by 1.6x in survival rate. Remarkably, it achieves this using only 18.9% of the data storage and processing resources of high-speed video methods.

Real-world visual recognition faces the fundamental challenge of long-tailed distributions. While state-of-the-art methods often employ multi-expert models to address different frequency categories, we find that the mutual knowledge distillation used in these models enhances collaboration at the cost of introducing two critical limitations: indiscriminate knowledge transfer leads to bias propagation, where a single expert's error can spread and contaminate others, and error consolidation, where mutual reinforcement of incorrect predictions solidifies erroneous consensus. To overcome these issues, we propose Trust-calibrated Collaborative Learning (TCL). Our framework introduces the trustworthy knowledge orchestration module, which enables reliable distillation and precise collaboration through a knowledge quality gate that blocks erroneous information and a tail-class compensation mechanism that alleviates knowledge scarcity for tail categories. Furthermore, we design a consensus error calibration module that suppresses consensus high-confidence negative classes to correct collective misjudgments and steer optimization in the right direction. Extensive experiments on five long-tailed benchmarks demonstrate that TCL achieves the best performance, raising Top-1 accuracy on CIFAR100-LT to 58.7%, a gain of 2.4% over previous SOTA methods.

We propose Confidence-Guided Token Merging (Co-Me), an acceleration mechanism for visual geometric transformers without retraining or finetuning the base model. Co-Me distilled a light-weight confidence predictor to rank tokens by uncertainty and selectively merge low-confidence ones, effectively reducing computation while maintaining spatial coverage. Compared to similarity-based merging or pruning, the confidence signal in Co-Me reliably indicates regions emphasized by the transformer, enabling substantial acceleration without degrading performance. Co-Me applies seamlessly to various multi-view and streaming visual geometric transformers, achieving speedups that scale with sequence length. When applied to VGGT and Pi3, Co-Me achieves up to 21.5x and 20.4x speedup, making visual geometric transformers practical for real-time 3D perception and reconstruction.

This paper provides a simplification on OpenVision's architecture and loss design for enhancing its training efficiency. Following the prior vision-language pretraining works CapPa and AIMv2, as well as modern multimodal designs like LLaVA, our changes are straightforward: we remove the text encoder (and therefore the contrastive loss), retaining only the captioning loss as a purely generative training signal. We name this new version OpenVision 2. The initial results are promising: despite this simplification, OpenVision 2 competitively matches the original model's performance on a broad set of multimodal benchmarks while substantially cutting both training time and memory consumption. For example, with ViT-L/14, it reduces training time by about 1.5x (from 83h to 57h), and memory usage by about 1.8x (from 24.5GB to 13.8GB, equivalently allowing the maximum batch size to grow from 2k to 8k). This superior training efficiency also allows us to scale far beyond the largest vision encoder used in OpenVision, reaching more than 1 billion parameters. We hold a strong belief that this lightweight, generative-only paradigm is compelling for future vision encoder development in multimodal foundation models.

We introduce OpenVO, a novel framework for Open-world Visual Odometry (VO) with temporal awareness under limited input conditions. OpenVO effectively estimates real-world-scale ego-motion from monocular dashcam footage with varying observation rates and uncalibrated cameras, enabling robust trajectory dataset construction from rare driving events recorded in dashcam.Existing VO methods are trained on fixed observation frequency (e.g., 10Hz or 12Hz), completely overlooking temporal dynamics information. Many prior methods also require calibrated cameras with known intrinsic parameters. Consequently, their performance degrades when (1) deployed under unseen observation frequencies or (2) applied to uncalibrated cameras. These significantly limit their generalizability to many downstream tasks, such as extracting trajectories from dashcam footage.To address these challenges, OpenVO (1) explicitly encodes temporal dynamics information within a two-frame pose regression framework and (2) leverages 3D geometric priors derived from foundation models. We validate our method on three major autonomous-driving benchmarks - KITTI, nuScenes, and Argoverse 2 - achieving more than 20% performance improvement over state-of-the-art approaches. Under varying observation rate settings, our method is significantly more robust, achieving 46%-92% lower errors across all metrics.These results demonstrate the versatility of OpenVO for real-world 3D reconstruction and diverse downstream applications.

Humans do not just see attribute similarity -- we also see relational similarity. An apple is like a peach because both are reddish fruit, but the Earth is also like a peach: its crust, mantle, and core correspond to the peach's skin, flesh, and pit. This ability to perceive and recognize relational similarity, is arguable by cognitive scientist to be what distinguishes humans from other species. Yet, all widely used visual similarity metrics today (e.g., LPIPS, CLIP, DINO) focus solely on perceptual attribute similarity and fail to capture the rich, often surprising relational similarities that humans perceive. How can we go beyond the visible content of an image to capture its relational properties? How can we bring images with the same relational logic closer together in representation space? To answer these questions, we first formulate relational image similarity as a measurable problem: two images are relationally similar when their internal relations or functions among visual elements correspond, even if their visual attributes differ. We then curate 114k image-caption dataset in which the captions are anonymized -- describing the underlying relational logic of the scene rather than its surface content. Using this dataset, we finetune a Vision-Language model to measure the relational similarity between images. This model serves as the first step toward connecting images by their underlying relational structure rather than their visible appearance. Our study shows that while relational similarity has a lot of real-world applications, existing image similarity models fail to capture it -- revealing a critical gap in visual computing.

Recent Video-to-Audio (V2A) methods have achieved remarkable progress, enabling the synthesis of realistic, high-quality audio. However, they struggle with fine-grained temporal control in multi-event scenarios or when visual cues are insufficient, such as small regions, off-screen sounds, or occluded/partially visible objects.In this paper, we propose FoleyDirector, a framework that, for the first time, enables precise temporal guidance in DiT-based V2A generation while preserving the base model's audio quality and allowing seamless switching between V2A generation and temporally controlled synthesis. FoleyDirector introduces Structured Temporal Scripts(STS), a set of captions corresponding to short temporal segments, to provide richer temporal information. These features are integrated via the Script-Guided Temporal Fusion Module, which employs Temporal Script Attention to fuse STS features coherently. To handle complex multi-event scenarios, we further propose Bi-Frame Sound Synthesis, enabling parallel in-frame and out-of-frame audio generation and improving controllability.To support training and evaluation, we construct the DirectorSound dataset and introduce VGGSound-Director and DirectorBench. Experiments demonstrate that FoleyDirector substantially enhances temporal controllability while maintaining high audio fidelity, empowering users to act as Foley directors and advancing V2A toward more expressive and controllable.