CVPR 2026 Papers

Text-to-image models are increasingly utilized in design workflows, but articulating nuanced design intentions solely through text remains a challenge. This work proposes a method that extracts visual attributes from a reference image and injects them directly into the generation pipeline. Specifically, the method optimizes a text token to selectively represent the target attribute using a custom training prompt and two novel embeddings, termed distilled and residual. A wide range of attributes can be extracted through this approach, including shape, material, pose, and camera shot and angle. The effectiveness of the method is validated on various target attributes and text prompts drawn from a newly constructed dataset. The method outperforms existing approaches in selectively extracting and applying target attributes across diverse contexts. Ultimately, the proposed method enables intuitive and controllable text-to-image generation, streamlining the design process.

Recent advancements in video generation have enabled the development of "world models" capable of simulating potential futures for robotics and planning. However, specifying precise goals for these models remains a challenge; text instructions are often too abstract to capture physical nuances, while target images are frequently infeasible to specify for dynamic tasks. To address this, we introduce Goal Force, a novel framework that allows users to define goals via explicit force vectors and intermediate dynamics, mirroring how humans conceptualize physical tasks. We train a video generation model on a curated dataset of synthetic causal primitives--such as elastic collisions and falling dominos--teaching it to propagate forces through time and space. Despite being trained on simple physics data, our model exhibits remarkable zero-shot generalization to complex, real-world scenarios, including tool manipulation and multi-object causal chains. Our results suggest that by grounding video generation in fundamental physical interactions, models can emerge as implicit neural physics simulators, enabling precise, physics-aware planning without reliance on external engines.

Humor, as both a creative human activity and a social binding mechanism, has long posed a major challenge for AI generation. Although producing humor requires complex cognitive reasoning and social understanding, theories of humor suggest that it follows learnable patterns and structures, making it theoretically possible for generative models to acquire them implicitly. In recent years, multimodal humor has become a prevalent form of online communication, especially among Gen Z, highlighting the need for AI systems capable of integrating visual understanding with humorous language generation. However, existing data-driven approaches lack explicit modeling or theoretical grounding of humor, often producing literal descriptions that fail to capture its underlying cognitive mechanisms, resulting in the generated image descriptions that are fluent but lack genuine humor or cognitive depth. To address this limitation, we propose HUMORCHAIN (HUmor-guided Multi-step Orchestrated Reasoning Chain for Image Captioning), a theory-guided multi-stage reasoning framework. It integrates visual semantic parsing, humor- and psychology-based reasoning, and a fine-tuned discriminator for humor evaluation, forming an interpretable and controllable cognitive reasoning chain. To the best of our knowledge, this is the first work to explicitly embed cognitive structures from humor theories into multimodal humor generation, enabling a structured reasoning process from visual understanding to humor creation. Experiments on Meme-Image-No-Text, Oogiri-GO, and OxfordTVG-HIC datasets show that HUMORCHAIN outperforms state-of-the-art baselines in human humor preference, Elo/BT scores, and semantic diversity, demonstrating that theory-driven structured reasoning enables large language models to generate humor aligned with human perception.

We address the problem of making 3D scenes interactive by asking: what would objects feel like if touched in a virtual environment? While neural scene representations achieve high visual realism, they fall short in modeling physical interactions, such as the vibration patterns produced when surfaces are explored with different actions. We propose a framework that learns the correspondence between scene, user action, and haptic response, enabling the generation of tactile signals from simulated interactions in 3D scenes. Our approach leverages a neural field representation conditioned on geometry information and interaction dynamics to synthesize material-specific vibrotactile signals. Experiments show that the generated signals reliably convey both material properties and the user action. This paves the way toward interactive, touch-aware virtual environments with realistic haptic feedback. Project page available here.

Recent Vision-Language Models (VLMs) have made remarkable progress in multimodal understanding tasks, yet their evaluation on long video understanding remains unreliable. Due to limited frame inputs, key frames necessary for answering the question may be missing from the model's input. However, models that truthfully refuse to answer under such uncertainty are marked as incorrect, while those that guess may coincidentally produce the correct answer and thus obtain deceptively higher accuracy, leading to misleading evaluation results and encouraging models to guess rather than respond honestly. To address this issue, we introduce VirtueBench, a benchmark explicitly designed to assess model trustworthiness under uncertainty. VirtueBench constructs multiple frame-sampling levels for each video and provides ground truths that distinguish between answerable and unanswerable cases. Evaluations on 25 open-source and commercial VLMs reveal distinct refusal behaviors across different model families, with refusal accuracy ranging from over 70% in the best models to nearly 0% in the worst. Moreover, most models exhibit a substantial drop in refusal when the prompt does not explicitly require them to do so. These findings highlight the need for developing trustworthy VLMs for multimodal understanding, guided by benchmarks and leaderboards that emphasize reliability and trustworthiness.

Pre-training on large-scale videos to improve reinforcement learning efficiency is promising yet remains challenging. Existing methods typically treat the agent as an indivisible entity, modeling motion patterns globally. Such global modeling is tightly coupled with the morphology, hindering transfer across domains. In contrast, despite the vast disparity in global motions, the local components exhibit similar motion patterns across different agents. Building on this insight, we propose a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations. Specifically, in the Deconstruct phase, we identify multiple local points and track their frame-wise motions, defining each as an Atomic Action. We introduce a Dual-Attention Encoder (DAE) to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships. In the Recompose phase, we compose local motion representations with a learnable Motion Aggregation Token '[MAT]' via latent dynamics model learning. Additionally, an adapter bridges local motion and downstream action-specific dynamics to accelerate policy learning. Extensive experiments demonstrate that our method effectively transfers to diverse robotic control and manipulation tasks, significantly improving sample efficiency and performance.

Unlike offline processing, streaming video vision-language models face two fundamental constraints: causality and accumulation. Causality prevents access to future frames that offline methods exploit, while accumulation causes tokens to grow unbounded, creating efficiency bottlenecks. However, existing approaches only regulate post-LLM kv-cache, leaving costly pre-LLM prefill unchanged. We introduce StreamingTOM, a training-free, plug-and-play two-stage framework that addresses both pre-LLM and post-LLM bottlenecks. Causal Temporal Reduction imposes a fixed per-frame budget and selects tokens based on adjacent-frame changes and token saliency, drastically reducing per-frame prefill cost by processing only a compact subset of visual tokens, ensuring predictable latency. Online Quantized Memory stores tokens in 4-bit format, retrieves relevant groups on demand, and dequantizes them, keeping the active kv-cache bounded regardless of stream length. Experiments demonstrate our method achieves 15.7x kv-cache compression ratio; compared to prior SOTA (LiveVLM), it delivers 1.2x lower peak memory and 2x faster TTFT. StreamingTOM achieves state-of-the-art accuracy among training-free methods with an average of 63.8% on offline benchmarks and 55.8% accuracy and 3.7 score on RVS. These results demonstrate that real-time streaming video understanding with bounded active memory is achievable without model retraining.

We present SWIM (See What I Mean), a novel training strategy that aligns vision and language representations to enable fine-grained object understanding solely from textual prompts. Unlike existing approaches that require explicit visual prompts, such as masks or points, SWIM leverages mask supervision only during training to guide cross-modal attention, allowing the model to automatically attend to the user-specified object at inference.Our cross-attention analysis of pretrained multimodal large language models (MLLMs) reveals a systematic discrepancy: Attribute words produce sharp, localized activations in the visual modality, whereas object nouns yield diffuse and scattered patterns due to semantic reference bias and distributed high-level representations. To address this misalignment, we construct NL-Refer, an enriched dataset, in which each object mask is paired with a precise natural language referring expression. SWIM extracts multi-layer cross-attention maps from object nouns and enforces spatial consistency with ground-truth masks. Experimental results demonstrate that SWIM substantially improves text-visual alignment and achieves superior performance over visual-prompt-based methods on fine-grained object understanding benchmarks. The code and data are available at https://github.com/HumanMLLM/SWIM.

Accurately modelling human attention is essential for numerous computer vision applications, particularly in the domain of automotive safety. Existing methods typically collapse gaze into saliency maps or scanpaths, treating gaze dynamics only implicitly. We instead formulate gaze modelling as an autoregressive dynamical system and explicitly unroll raw gaze trajectories over time, conditioned on both gaze history and the evolving environment. Driving scenes are represented as gaze-centric graphs processed by the Affinity Relation Transformer (ART), a heterogeneous graph transformer that models interactions between driver gaze, traffic objects, and road structure. We further introduce the Object Density Network (ODN) to predict next-step gaze distributions, capturing the stochastic and object-centric nature of attentional shifts in complex environments. We also release Focus100, a new dataset of raw gaze data from 30 participants viewing egocentric driving footage. Trained directly on raw gaze, without fixation filtering, our unified approach produces more natural gaze trajectories, scanpath dynamics, and saliency maps than existing attention models, offering valuable insights for the temporal modelling of human attention in dynamic environments.

Synthesis of diverse driving scenes serves as a crucial data augmentation technique for validating the robustness and generalizability of autonomous driving systems. Current methods aggregate high-definition (HD) maps and 3D bounding boxes as geometric conditions in diffusion models for conditional scene generation. However, implicit inter-condition dependency causes generation failures when control conditions change independently. Additionally, these methods suffer from insufficient details in both semantic and structural aspects. Specifically, brief and view-invariant captions restrict semantic contexts, resulting in weak background modeling. Meanwhile, the standard denoising loss with uniform spatial weighting neglects foreground structural details, causing visual distortions and blurriness. To address these challenges, we propose DrivePTS, which incorporates three key innovations. Firstly, our framework adopts a progressive learning strategy to mitigate inter-dependency between geometric conditions, reinforced by an explicit mutual information constraint. Secondly, a Vision-Language Model is utilized to generate multi-view hierarchical descriptions across six semantic aspects, providing fine-grained textual guidance. Thirdly, a frequency-guided structure loss is introduced to strengthen the model's sensitivity to high-frequency elements, improving foreground structural fidelity. Extensive experiments demonstrate that our DrivePTS achieves state-of-the-art fidelity and controllability in generating diverse driving scenes. Notably, DrivePTS successfully generates rare scenes where prior methods fail, highlighting its strong generalization ability.

Diffusion-native watermarking provides a more secure and reliable way to trace images from latent diffusion models (LDMs) by embedding information directly into the generative process. However, existing methods suffer from a fundamental limitation: their embedding capacity is extremely small. We introduce MaxMark, a high-capacity watermarking framework that supports embed rich watermark messages into generated images. MaxMark uses two components: a robust watermark embedding module that enhance the secret message and places them into reliable regions of the latent noise, and a distribution transformation module that maps the watermarked latent back to an approximate Gaussian, ensuring compatibility with the diffusion process and preserving image fidelity. The distribution transformation is implemented with an invertible neural network (INN), whose exactly reversible structure enables precise recovery and efficient training. Experiments show that MaxMark surpasses prior methods in capacity, robustness, and imperceptibility, achieving up to a 46% improvement in bit accuracy for large watermark payloads.

Monocular metric depth estimation (MMDE) is a core challenge in computer vision, playing a pivotal role in real-world applications that demand accurate spatial understanding. Although prior works have shown promising zero-shot performance in MMDE, they often struggle with generalization across diverse camera types, such as fisheye and 360^\circ cameras. Recent advances have addressed this through unified camera representations or canonical representation spaces, but they require either including large FoV camera data during training or separately trained models for different domains. We propose UniDAC, an MMDE framework that presents universal robustness in all domains and generalizes across diverse cameras using a single model. We achieve this by decoupling metric depth estimation into relative depth prediction and spatially varying scale estimation, enabling robust performance across different domains. We propose a lightweight Depth-Guided Scale Estimation module that upsamples a coarse scale map to high resolution using the relative depth map as guidance to account for local scale variations. Furthermore, we introduce RoPE-\phi, a distortion-aware positional embedding that respects the spatial warping in Equi-Rectangular Projections (ERP) via latitude-aware weighting. UniDAC achieves state of the art (SoTA) in cross-camera generalization by consistently outperforming prior methods across all datasets.

Current audio pre-training seeks to learn unified representations for broad audio understanding tasks, but it remains fragmented and is fundamentally bottlenecked by its reliance on weak, noisy, and scale-limited labels. Drawing lessons from vision's foundational pre-training blueprint, we argue that the audio field must first establish its own large-scale, strong supervision framework. We introduce a new data-centric pipeline that leverages a high-fidelity captioner to create SOTA-quality captions and the first Unified Tag System (UTS) that bridges speech, music, and environmental sounds. We then conduct a systematic comparative study of different pre-training objectives on these strong source data. Our experiments suggest that data quality and coverage are the primary drivers of performance, while the choice of objective dictates downstream task specialization.

Real-world robotic tasks are long-horizon and often span multiple floors, demanding rich spatial reasoning. However, existing embodied benchmarks are largely confined to single-floor in-house environments, failing to reflect the complexity of real-world tasks. We introduce MANSION, the first language-driven framework for generating building-scale, multi-floor 3D environments. Being aware of vertical structural constraints, MANSION generates realistic, navigable whole-building structures with diverse, human-friendly scenes, enabling the development and evaluation of cross-floor long-horizon tasks. Building on this framework, we release MansionWorld, a dataset of over 1,000 diverse buildings ranging from hospitals to offices, alongside a Task-Semantic Scene Editing Agent that customizes these environments using open-vocabulary commands to meet specific user needs. Benchmarking reveals that state-of-the-art agents degrade sharply in our settings, establishing MANSION as a critical testbed for the next generation of spatial reasoning and planning.

The robust execution of long-horizon manipulation tasks remains a central challenge in embodied intelligence, necessitating both coherent high-level planning and reliable low-level control. Existing approaches often encounter two critical limitations: the accumulation of prediction errors in subgoal planning, leading to compounding deviations over time; and the planning-execution gap, where high-level abstract plans fail to be effectively grounded in the continuous perception-action space. To address these challenges, we propose a novel unified framework, Affordance-Grounded Bidirectional Latent Planning (AGiLe). AGiLe introduces a bidirectional latent planning mechanism that jointly optimizes a backward planner and a forward critic. The backward planner generates goal-directed subgoals from the final objective, while the forward critic assesses their reachability, ensuring temporal robustness through sustained consistency in long-horizon planning. Furthermore, AGiLe bridges the planning-execution gap by leveraging affordance as structural guidance, grounding abstract subgoals into dense, pixel-level visual affordances that drive action. This enhances spatial robustness, enabling the system to effectively adapt to semantic and visual distractors. Extensive empirical evaluations across both simulation and real-world settings confirm that AGiLe significantly outperforms strong baselines, achieving an 8.5% improvement over prior state-of-the-art methods and demonstrating strong effectiveness and robustness in long-horizon manipulation tasks. Project website: https://agile-long.github.io.

Graphical user interface (GUI) agents built on multimodal large language models (MLLMs) have recently demonstrated strong decision-making abilities in screen-based interaction tasks. However, they remain highly vulnerable to pop-up-based environmental injection attacks, where malicious visual elements divert model attention and lead to unsafe or incorrect actions. Existing defense methods either require costly retraining or perform poorly under inductive interference. In this work, we systematically study how such attacks alter the attention behavior of GUI agents and uncover a layer-wise attention divergence pattern between correct and incorrect outputs. Based on this insight, we propose LaSM, a Layer-wise Scaling Mechanism that selectively amplifies attention and MLP modules in critical layers. LaSM improves the alignment between model saliency and task-relevant regions without additional training. Extensive experiments across multiple datasets demonstrate that our method significantly improves the defense success rate and exhibits strong robustness, while having negligible impact on the model's general capabilities. Our findings reveal that attention misalignment is a core vulnerability in MLLM agents and can be effectively addressed through selective layer-wise modulation.

With the rapid advancement of generative models, generated image detection has become an important task in visual forensics. Although existing methods have achieved remarkable progress, they often rely, after training, on only a small subset of highly salient forgery cues, which limits their ability to generalize to unseen generative mechanisms. We argue that reliably generated image detection should not depend on a single decision path but should preserve multiple judgment perspectives, enabling the model to understand the differences between real and generated images from diverse viewpoints. Based on this idea, we propose an anti-feature-collapse learning framework that filters task-irrelevant components and suppresses excessive overlap among different forgery cues in the representation space, preventing discriminative information from collapsing into a few dominant feature directions. This design maintains diverse and complementary evidence within the model, reduces reliance on a small set of salient cues, and enhances robustness under unseen generative settings. Extensive experiments on multiple public benchmarks demonstrate that the proposed method significantly outperforms the state-of-the-art approaches in cross-model scenarios, achieving an accuracy improvement of 5.02% and exhibiting superior generalization and detection reliability. The source code is available at https://github.com/Yanmou-Hui/DoU.

Articulated 3D objects are critical for embodied AI, robotics, and scene understanding, yet creating simulation-ready assets remains labor-intensive and requires expert modeling of part hierarchies and motion structures. We introduce SPARK, a framework for reconstructing physically consistent, kinematic part-level articulated objects from a single RGB image. Given the image, we first leverage VLMs to extract coarse URDF parameters and generate part-level reference images. We then integrate the part-image guidance and the inferred structure graph into a diffusion transformer to synthesize consistent part and complete shapes of articulated objects. To further refine the URDF parameters, we explore a VLM-based reprediction strategy for discrete attribute refinement and a differentiable forward kinematics module for continuous parameter optimization under VLM-generated open-state supervision. Extensive experiments show that SPARK produces high-quality, simulation-ready articulated assets across diverse categories, enabling downstream applications such as robotic manipulation and interaction modeling.

Diffusion Transformers (DiTs) have achieved state-of-the-art image and video generation performance, but sampling remains expensive due to repeated transformer forward passes over many timesteps. Feature caching offers a training-free way to accelerate inference by reusing or forecasting hidden representations, yet recent forecasting-based methods derive their coefficients from hand-crafted formulas (e.g., Taylor expansion), which ultimately reduce to fixed linear combinations of a few historical features. Such fixed coefficients are suboptimal and fragile under aggressive skipping. In this paper, we first show that existing forecasting-based caching methods can be unified in a common linear form, and then analyze DiT feature trajectories, finding that for most denoising steps the current feature can be reconstructed from past features with projection fidelity above 0.95, indicating that accurate linear prediction is feasible. Motivated by this, we propose L^2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces hand-designed coefficients with learnable per-timestep weights trained on a small set of cached trajectories using a mean-squared error loss, converging in about 20 seconds on a single GPU. Extensive experiments on state-of-the-art DiTs demonstrate that L^2P consistently outperforms existing caching baselines: on FLUX.1-dev, L^2P achieves a 4.55xFLOPs reduction and 4.15xlatency speedup with a PSNR of 31.459, and on Qwen-Image and Qwen-Image-Lightning, it maintains high visual fidelity even under up to 7.18xacceleration, where prior methods suffer from noticeable quality degradation. These results show that learning linear predictors is a practical and effective alternative to designing increasingly complex forecasting formulas for efficient diffusion model inference. The implementation code is available at https://github.com/Aredstone/L2P-Cache.