CVPR 2026 论文列表

Autonomous driving needs fast, scalable 4D reconstruction and re-simulation for training and evaluation, yet most methods for dynamic driving scenes still rely on per-scene optimization, known camera calibration, or short frame windows, making them slow and impractical. We revisit this problem from a feedforward perspective and introduce Driving Gaussian Grounded Transformer (DGGT), a unified framework for pose-free dynamic scene reconstruction. We note that the existing formulations, treating camera pose as a required input, limit flexibility and scalability. Instead, we reformulate pose as an output of the model, enabling reconstruction directly from sparse, unposed images and supporting an arbitrary number of views for long sequences. Our approach jointly predicts per-frame 3D Gaussian maps and camera parameters, disentangles dynamics with a lightweight dynamic head, and preserves temporal consistency with a lifespan head that modulates visibility over time. A diffusion-based rendering refinement further reduces motion/interpolation artifacts and improves novel-view quality under sparse inputs. The result is a single-pass, pose-free algorithm that achieves state-of-the-art performance and speed. Trained and evaluated on large-scale driving benchmarks (Waymo, nuScenes, Argoverse2), our method outperforms prior work both when trained on each dataset and in zero-shot transfer across datasets, and it scales well as the number of input frames increases.

Existing methods for deepfake detection aim to develop generalizable detectors. Although "generalizable" could be the ultimate target once and for all, with limited training forgeries and domains, it appears idealistic to expect generalization that covers entirely unseen variations, especially given the diversity, advancement, and vast volume of real-world deepfakes. Therefore, introducing large-scale multi-domain data for training can be feasible and important for real-world applications.However, within such a multi-domain scenario, the differences between multiple domains, rather than the subtle real/fake distinctions, dominate the feature space. As a result, despite detectors being able to relatively separate real and fake within each domain (i.e., high AUC), they struggle with single-image real/fake judgments in domain-unspecified conditions (i.e., low ACC).In this paper, we first define a new research paradigm named Multi-In-Domain Face Forgery Detection (MID-FFD), which includes sufficient volumes of real-fake domains for training. Then, the detector should provide definitive real-fake judgments to the domain-unspecified inputs, which simulate the frame-by-frame independent detection scenario in the real world. Meanwhile, to address the domain-dominant issue, we propose a two-stage, model-agnostic framework termed DevDet (\underline Dev eloper for \underline Det ector) to amplify real/fake differences and make them dominant in the feature space. DevDet consists of a Face Forgery Developer (FFDev) and a Dose-Adaptive detector Fine-Tuning strategy (DAFT). Experiments demonstrate our superiority in effectively predicting real-fake under the MID-FFD scenario while maintaining original generalization ability to unseen data.

Unsupervised point cloud segmentation is critical for embodied artificial intelligence and autonomous driving, as it mitigates the prohibitive cost of dense point-level annotations required by fully supervised methods. While integrating 2D pre-trained models such as the Segment Anything Model (SAM) to supplement semantic information is a natural choice, yet this approach faces a fundamental mismatch between discrete 3D points and continuous 2D images. This mismatch leads to inevitable projection overlap and complex modality alignment, resulting in compromised semantic consistency across 2D-3D transfer. To address these limitations, this paper proposes PointGS, a simple yet effective pipeline for unsupervised 3D point cloud segmentation. PointGS leverages 3D Gaussian Splatting as a unified intermediate representation to bridge the discrete-continuous domain gap. Input sparse point clouds are first reconstructed into dense 3D Gaussian spaces via multi-view observations, filling spatial gaps and encoding occlusion relationships to eliminate projection-induced semantic conflation. Multi-view dense images are rendered from the Gaussian space, with 2D semantic masks extracted via SAM, and semantics are distilled to 3D Gaussian primitives through contrastive learning to ensure consistent semantic assignments across different views. The Gaussian space is aligned with the original point cloud via two-step registration, and point semantics are assigned through nearest-neighbor search on labeled Gaussians. Experiments demonstrate that PointGS outperforms state-of-the-art unsupervised methods, achieving +0.9% mIoU on ScanNet-V2 and +2.8% mIoU on S3DIS.

Cross-domain panoramic semantic segmentation has attracted growing interest as it enables comprehensive 360^ \circ scene understanding for real-world applications. However, it remains particularly challenging due to severe geometric Field of View (FoV) distortions and inconsistent open-set semantics across domains. In this work, we formulate an open-set domain adaptation setting, and propose Extrapolative Domain Adaptive Panoramic Segmentation (EDA-PSeg) framework that trains on local perspective views and tests on full 360^ \circ panoramic images, explicitly tackling both geometric FoV shifts across domains and semantic uncertainty arising from previously unseen classes. To this end, we propose the Euler-Margin Attention (EMA), which introduces an angular margin to enhance viewpoint-invariant semantic representation, while performing amplitude and phase modulation to improve generalization toward unseen classes. Additionally, we design the Graph Matching Adapter (GMA), which builds high-order graph relations to align shared semantics across FoV shifts while effectively separating novel categories through structural adaptation. Extensive experiments on four benchmark datasets under camera-shift, weather-condition, and open-set scenarios demonstrate that EDA-PSeg achieves state-of-the-art performance, robust generalization to diverse viewing geometries, and resilience under varying environmental conditions. The code is available at https://github.com/zyfone/EDA-PSeg.

Latent diffusion models have emerged as the dominant framework for high-fidelity and efficient image generation, owing to their ability to learn diffusion processes in compact latent spaces. However, while previous research has focused primarily on reconstruction accuracy and semantic alignment of the latent space, we observe that another critical factor, robustness to sampling perturbations, also plays a crucial role in determining generation quality. Through empirical and theoretical analyses, we show that the commonly used b-VAE-based tokenizers in latent diffusion models, tend to produce overly compact latent manifolds that are highly sensitive to stochastic perturbations during diffusion sampling, leading to visual degradation. To address this issue, we propose a simple yet effective solution that constructs a latent space robust to sampling perturbations while maintaining strong reconstruction fidelity. This is achieved by introducing a Variance Expansion loss that counteracts variance collapse and leverages the adversarial interplay between reconstruction and variance expansion to achieve an adaptive balance that preserves reconstruction accuracy while improving robustness to stochastic sampling. Extensive experiments demonstrate that our approach consistently enhances generation quality across different latent diffusion architectures, confirming that robustness in latent space is a key missing ingredient for stable and faithful diffusion sampling. Our project page: https://github.com/CVL-UESTC/VE-Loss .

Recent advances in video generation have been remarkable, enabling models to produce visually compelling videos with synchronized audio. While existing video generation benchmarks provide comprehensive metrics for visual quality, they lack convincing evaluations for audio-video generation, especially for models aiming to generate synchronized audio-video outputs. To address this gap, we introduce VABench, a comprehensive and multi-dimensional benchmark framework designed to systematically evaluate the capabilities of synchronous audio-video generation. VABench encompasses three primary task types: text-to-audio-video (T2AV), image-to-audio-video (I2AV), and stereo audio-video generation. It further establishes two major evaluation modules covering 15 dimensions. These dimensions specifically assess pairwise similarities (text-video, text-audio, video-audio), audio-video synchronization, lip-speech consistency, and carefully curated audio and video question-answering (QA) pairs, among others. Furthermore, VABench covers seven major content categories: animals, human sounds, music, environmental sounds, synchronous physical sounds, complex scenes, and virtual worlds. We provide a systematic analysis and visualization of the evaluation results, aiming to establish a new standard for assessing video generation models with synchronous audio capabilities and to promote the comprehensive advancement of the field.

Monocular 3D pose estimation is fundamentally ill-posed due to depth ambiguity and occlusions, thereby motivating probabilistic methods that generate multiple plausible 3D pose hypotheses. In particular, diffusion-based models have recently demonstrated strong performance, but their iterative denoising process typically requires many timesteps for each prediction, making inference computationally expensive. In contrast, we leverage Flow Matching (FM) to learn a velocity field defined by an Ordinary Differential Equation (ODE), enabling efficient generation of 3D pose samples with only a few integration steps. We propose a novel generative pose estimation framework, FMPose3D, that formulates 3D pose estimation as a conditional distribution transport problem. It continuously transports samples from a standard Gaussian prior to the distribution of plausible 3D poses conditioned only on 2D inputs. Although ODE trajectories are deterministic, FMPose3D naturally generates various pose hypotheses by sampling different noise seeds. To obtain a single accurate prediction from those hypotheses, we further introduce a Reprojection-based Posterior Expectation Aggregation (RPEA) module, which approximates the Bayesian posterior expectation over 3D hypotheses. FMPose3D surpasses existing methods on the widely used human pose estimation benchmarks Human3.6M and MPI-INF-3DHP, and further achieves state-of-the-art performance on the 3D animal pose datasets Animal3D and CtrlAni3D, demonstrating strong performance across both 3D pose domains. The code is available at https://github.com/AdaptiveMotorControlLab/FMPose3D.

Embodied Question Answering (EQA) requires an agent to interpret language, perceive its environment, and navigate within 3D scenes to produce responses. Existing EQA benchmarks assume that every question must be answered, but embodied agents should know when they do not have sufficient information to answer. In this work, we focus on a minimal requirement for EQA agents, abstention: knowing when to withhold an answer. From an initial study of 500 human queries, we find that 32.4% contain missing or underspecified context. Drawing on this initial study and cognitive theories of human communication errors, we derive five representative categories requiring abstention: actionability limitation, referential underspecification, preference dependence, information unavailability, and false presupposition. We augment OpenEQA by having annotators transform well-posed questions into ambiguous variants outlined by these categories. The resulting dataset, AbstainEQA, comprises 1,636 annotated abstention cases paired with 1,636 original OpenEQA instances for balanced evaluation. Evaluating on AbstainEQA, we find that even the best frontier model only attains 42.79% abstention recall, while humans achieve 91.17%. We also find that scaling, prompting, and reasoning only yield marginal gains, and that fine-tuned models overfit to textual cues. Together, these results position abstention as a fundamental prerequisite for reliable interaction in embodied settings and as a necessary basis for effective clarification.

The growing demand for oriented object detection (OOD) across various domains has driven significant research in this area. However, the high cost of dataset annotation remains a major concern. Current mainstream OOD algorithms can be mainly categorized into three types: (1) fully supervised methods using complete oriented bounding box (OBB) annotations, (2) semi-supervised methods using partial OBB annotations, and (3) weakly supervised methods using weak annotations such as horizontal boxes or points. However, these algorithms inevitably increase the cost of models in terms of annotation speed or annotation cost. To address this issue, we propose: (1) the first Partial Weakly-Supervised Oriented Object Detection (PWOOD) framework based on partially weak annotations (horizontal boxes or single points), which can efficiently leverage large amounts of unlabeled data, significantly outperforming weakly supervised algorithms trained with partially weak annotations, also offers a lower cost solution; (2) Orientation-and-Scale-aware Student (OS-Student) model capable of learning orientation and scale information with only a small amount of orientation-agnostic or scale-agnostic weak annotations; and (3) Class-Agnostic Pseudo-Label Filtering strategy (CPF) to reduce the model's sensitivity to static filtering thresholds. Comprehensive experiments on DOTA-v1.0/v1.5/v2.0 and DIOR datasets demonstrate that our PWOOD framework performs comparably to, or even surpasses traditional semi-supervised algorithms.

Training machine learning models on massive datasets is expensive and time-consuming. Dataset distillation addresses this by creating a small synthetic dataset that achieves the same performance as the full dataset. Recent methods use diffusion models to generate distilled data, either by promoting diversity or matching training gradients. However, existing approaches produce redundant training signals, where samples convey overlapping information. Empirically, disjoint subsets of distilled datasets capture 80-90% overlapping signals. This redundancy stems from optimizing visual diversity or average training dynamics without accounting for similarity across samples, leading to datasets where multiple samples share similar information rather than complementary knowledge. We propose learnability-driven dataset distillation, which constructs synthetic datasets incrementally through successive stages. Starting from a small set, we train a model and generate new samples guided by learnability scores that identify what the current model can learn from, creating an adaptive curriculum. We introduce Learnability-Guided Diffusion (LGD), which balances training utility for the current model with validity under a reference model to generate curriculum-aligned samples. Our approach reduces redundancy by 39.1%, promotes specialization across training stages, and achieves state-of-the-art results on ImageNet-1K (60.1%), ImageNette (87.2%), and ImageWoof (72.9%).

Multimodal learning hinges on capturing redundant, unique, and synergistic information across modalities, which collectively constitute multimodal interactions. A critical yet underexplored challenge is that these implicit interactions vary dynamically across samples. In this work, we present the first systematic, information-theoretic analysis highlighting why learning these dynamic, sample-specific interactions is critical for effective multimodal learning. Our analysis further reveals deficits in conventional paradigms at learning these distinct interaction types: modality ensemble approaches struggle to capture synergy, while joint learning paradigms often under-utilize redundant information. This highlights the need for an approach that can adaptively learn from different interaction types on a per-sample basis. To this end, we propose Decomposition-based Multimodal Interaction Learning (DMIL), a novel paradigm that explicitly models and learns from sample-specific interactions. First, we design a variational decomposition architecture to isolate the constituent interaction components. Second, we employ a new learning strategy that leverages these explicit interaction components in a fine-tuning process to achieve comprehensive interaction learning. Extensive experiments across diverse tasks and architectures demonstrate that DMIL consistently achieves superior performance by adapting to holistic sample-specific interactions. Our framework is flexible and broadly applicable, establishing an interaction-centric paradigm for multimodal learning.

Micro-gestures are subtle and transient movements triggered by unconscious neural and emotional activities, holding great potential for human-computer interaction and clinical monitoring. However, their low amplitude, short duration, and strong inter-subject variability make existing deep models prone to degradation under low-sample, noisy, and cross-subject conditions. This paper presents an active inference-based framework for micro-gesture recognition, featuring Expected Free Energy (EFE)-guided temporal sampling and uncertainty-aware adaptive learning. The model actively selects the most discriminative temporal segments under EFE guidance, enabling dynamic observation and information gain maximization. Meanwhile, sample weighting driven by predictive uncertainty mitigates the effects of label noise and distribution shift. Experiments on the SMG dataset demonstrate the effectiveness of the proposed method, achieving consistent improvements across multiple mainstream backbones. Ablation studies confirm that both the EFE-guided observation and the adaptive learning mechanism are crucial to the performance gains. This work offers an interpretable and scalable paradigm for temporal behavior modeling under low-resource and noisy conditions, with broad applicability to wearable sensing, HCI, and clinical emotion monitoring.

We present IntrinsicWeather, a diffusion-based framework for controllable weather editing in intrinsic space. Our framework includes two components based on diffusion priors: an inverse renderer that estimates material properties, scene geometry, and lighting as intrinsic maps from an input image, and a forward renderer that utilizes these geometry and material maps along with a text prompt that describes specific weather conditions to generate a final image. The intrinsic maps enhance controllability compared to traditional pixel-space editing approaches. We propose an intrinsic map-aware attention mechanism that improves spatial correspondence and decomposition quality in large outdoor scenes. For forward rendering, we leverage CLIP-space interpolation of weather prompts to achieve fine-grained weather control. We also introduce a synthetic and a real-world dataset, containing 38k and 18k images under various weather conditions, each with intrinsic map annotations. IntrisicWeather outperforms state-of-the-art pixel-space editing approaches, weather restoration methods, and rendering-based methods, showing promise for downstream tasks such as autonomous driving, enhancing the robustness of detection and segmentation in challenging weather scenarios.

User interface to code (UI2Code) aims to generate executable code that can faithfully reconstruct a given input UI. Prior work focuses largely on web pages and mobile screens, leaving app widgets underexplored. Unlike web or mobile UIs with rich hierarchical context, widgets are compact, context-free micro-interfaces that summarize key information through dense layouts and iconography under strict spatial constraints. Moreover, while (image, code) pairs are widely available for web or mobile UIs, widget designs are proprietary and lack accessible markup. We formalize this setting as the Widget-to-Code (Widget2Code) and introduce an image-only widget benchmark with fine-grained, multi-dimensional evaluation metrics. Benchmarking shows that although generalized multimodal large language models (MLLMs) outperform specialized UI2Code methods, they still produce unreliable and visually inconsistent code. To address these limitations, we develop a baseline that jointly advances perceptual understanding and structured code generation. At the perceptual level, we follow widget design principles to assemble atomic components into complete layouts, equipped with icon retrieval and UI composition modules. At the system level, we design an end-to-end infrastructure, WidgetFactory, which includes a framework-agnostic widget-tailored domain-specific language (WidgetDSL) and a compiler that translates it into multiple front-end implementations (e.g., React, HTML). An adaptive rendering module further refines spatial dimensions to satisfy compactness constraints. Together, these contributions substantially enhance visual fidelity, establishing a strong baseline and unified infrastructure for future Widget2Code research.

Recent advances in Multimodal Large Language Models (MLLMs) have enabled promising progress in table reasoning from visual table inputs. Despite their ability to capture rich visual cues such as color and layout, MLLMs still underperform compared to text-only models.We argue that a major limitation lies in the pre-training process, which inadvertently weakens the model's intrinsic reasoning ability and consequently hinders the effectiveness of reinforcement fine-tuning on table reasoning tasks.In this paper, we introduce TableMix, a novel framework that tackles this challenge from a data-centric perspective. At the core of TableMix is a principled data mixing strategy. Specifically, TableMix constructs a hybrid dataset that combines: (1) multimodal table reasoning data to improve task-specific reasoning, (2) text-only mathematical reasoning data to revive the model's logical competence, and (3) simple multimodal perception data to preserve visual grounding.Recognizing the non-uniform difficulty of mixed data, we further propose a Difficulty-Aware Reward Shaping (DRS) mechanism, which enables the Group Relative Policy Optimization (GRPO) algorithm to adaptively reward concise reasoning for easy problems while encouraging more elaborate reasoning for complex ones, thereby reducing redundant computation and errors.Extensive experiments show that TableMix markedly enhances the reasoning ability of MLLMs, outperforming strong multimodal baselines and even rivaling state-of-the-art text-only models.

Text-to-video generation has significantly enriched content creation and holds the potential to evolve into powerful world simulators. However, modeling the vast spatiotemporal space remains computationally demanding, particularly when employing Transformers, which incur quadratic complexity in sequence processing and thus limit practical applications. Recent advancements in linear-time sequence modeling, particularly the Mamba architecture, offer a more efficient alternative. Nevertheless, its plain design limits its direct applicability to multimodal and spatiotemporal video generation tasks. To address these challenges, we introduce M4V, a multimodal Mamba framework for efficient text-to-video generation. Specifically, a MultiModal diffusion Mamba (MM-DiM) block is designed within the framework to enable seamless integration of multimodal information and spatiotemporal modeling. In detail, we introduce a novel multimodal token re-composition design, which employs a bidirectional scheme for multimodal information integration through simple token arrangement, along with visual registers to enhance spatial-temporal consistency. As a result, the MM-DiM blocks in M4V reduce FLOPs by 45% compared with the attention-based alternative when generating videos at 768x1280 resolution. Additionally, several training strategies are explored in this work to provide a better understanding of training text-to-video models using only publicly available datasets. Extensive experiments on text-to-video benchmarks demonstrate M4V's ability to produce high-quality videos while significantly lowering computational costs. Code will be made publicly available.

Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding tasks. However, the increasing demand for high-resolution image and long-video understanding results in substantial token counts, consequently leading to reduced inference efficiency. Token compression offers a direct solution by reducing the number of tokens to be processed, thereby improving computational efficiency without architectural changes. Through extensive analysis, we identify two critical limitations in existing inner-LLM token compression methods: positional bias and incompatibility with efficient operators, which critically hinder their practical deployment for LVLM acceleration. This paper presents the first approach from a dynamic token variation perspective, revealing that visual token variations within LLMs exhibit task-agnostic properties. We propose Variation-aware Vision Token Dropping (i.e., V2Drop), which progressively removes visual tokens with minimal variation during LVLM inference, thereby enhancing computational efficiency. Extensive experiments across multiple models and benchmarks consistently demonstrate that V2Drop maintains 94.0% and 98.6% of the original performance for image and video understanding tasks respectively, while reducing LLM generation latency by 31.5% and 74.2%.

Large Vision-Language Models (LVLMs) are foundational to modern multimodal applications, yet their susceptibility to adversarial attacks remains a critical concern. Prior white-box attacks rarely generalize across tasks, and black-box methods depend on expensive transfer, which limits efficiency. The vision encoder, standardized and often shared across LVLMs, provides a stable gray-box pivot with strong cross-model transfer. Building on this premise, we introduce PA-Attack (Prototype-Anchored Attentive Attack). PA-Attack begins with a prototype-anchored guidance that provides a stable attack direction towards a general and dissimilar prototype, tackling the attribute-restricted issue and limited task generalization of vanilla attacks. Building on this, we propose a two-stage attention enhancement mechanism: (i) leverage token-level attention scores to concentrate perturbations on critical visual tokens, and (ii) adaptively recalibrate attention weights to track the evolving attention during the adversarial process. Extensive experiments across diverse downstream tasks and LVLM architectures show that PA-Attack achieves an average 75.1% score reduction rate (SRR), demonstrating strong attack effectiveness, efficiency, and task generalization in LVLMs.

Radiology report generation (RRG) aims to automatically produce clinically faithful reports from chest X-ray images. Prevailing work typically follows a scale-driven paradigm, by multi-stage training over large paired corpora and oversized backbones, making pipelines highly data- and compute-intensive. In this paper, we propose Oracle-educated GRPO (OraPO) with a FactScore-based reward (FactS) to tackle the RRG task under constrained budgets. OraPO enables single-stage, RL-only training by converting failed GRPO explorations on rare or difficult studies into direct preference supervision via a lightweight oracle step. FactS grounds learning in diagnostic evidence by extracting atomic clinical facts and checking entailment against ground-truth labels, yielding dense, interpretable sentence-level rewards. Together, OraPO and FactS create a compact and powerful framework that significantly improves learning efficiency on clinically challenging cases, setting the new SOTA performance on the CheXpert Plus dataset (0.341 in F1) with 2-3 orders of magnitude less training data using a small base VLM on modest hardware.