AAAI 2026 论文列表

High-scale image super-resolution (SR) has become increasingly important with the rapid growth of mobile devices and high-resolution displays. However, current SR methods primarily focus on lower scales and generalize poorly to high-scale scenarios due to severe information loss and complex real-world degradations. In this paper, we propose a novel Selective Diffusion Distillation (SDD) framework for real-world high-scale SR, which distills reliable knowledge from a low-scale diffusion teacher to a high-scale student. Specifically, considering severe information loss in high-scale inputs, directly distilling from low-scale models may result in feature misalignment. To address this, we introduce a Degradation-aware Metric Learning (DML) approach to align feature distributions across different degradation levels. In addition, since the diffusion-based teacher may hallucinate artifacts in ambiguous regions, blindly imitating these unreliable outputs can degrade the student’s fidelity. To tackle this, we propose a Region-aware Selective Distillation (RSD) strategy to filter out uncertain predictions and adaptively supervise only on reliable areas. To evaluate the effectiveness of our method, we introduce Real-UltraSR, a new real-world benchmark that contains diverse high-scale LR-HR pairs, including x8, x10, x12, and x14. Extensive experiments demonstrate that our SDD framework achieves state-of-the-art performance across multiple benchmarks.

Recent advances in multi-instance learning (MIL) have demonstrated impressive performance in whole slide image (WSI) analysis. However, current methods search for cues and draw conclusions from all instances or regions, resulting in excessive redundant computation and suboptimal representation quality due to irrelevant and uninformative feature interference. To address these issues, we propose CICS, an efficient and general framework that performs compact information compression and selection for high-efficiency WSI analysis. In particular, CICS features two key components: (1) context-aware compression (CAC), which partitions the instance space into sub-regions and applies learnable compression to discard irrelevant components, reduce computational complexity while facilitating information selection, and (2) global-proximity selective attention (GPSA), which cherry-picks the most informative representation with a proximity-assisted global dynamic selection strategy. Building upon these innovations, CICS forms a plug-and-play module that reduces computational complexity through compact instance representations while improving feature quality by preserving the most informative cues. Extensive experiments on six WSI classification and survival prediction datasets show that CICS consistently improves the performance of multiple representative MIL methods. It achieves 2.5%, 7.7%, and 3.9% accuracy gain over the state-of-the-art Transformer-based TransMIL, Mamba-based MambaMIL, and graph-based WIKG methods on the ESCA dataset.

Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To reduce their substantial computational costs, feature caching techniques have been proposed to accelerate inference by reusing hidden representations from previous timesteps. However, current methods often struggle to maintain generation quality at high acceleration ratios, where prediction errors increase sharply due to the inherent instability of long-step forecasting. In this work, we adopt an ordinary differential equation (ODE) perspective on the hidden-feature sequence, modeling layer representations along the trajectory as a feature-ODE. We attribute the degradation of existing caching strategies to their inability to robustly integrate historical features under large skipping intervals. To address this, we propose FoCa (Forecast-then-Calibrate), which treats feature caching as a feature-ODE solving problem. Extensive experiments on image, video generation, and super-resolution tasks demonstrate the effectiveness of FoCa, especially under aggressive acceleration. Without additional training, FoCa achieves near-lossless speedups of 5.50× on FLUX, 6.45× on HunyuanVideo, 3.17× on Inf-DiT, and maintains high quality with a 4.53× speedup on DiT.

3D object detection in adverse weather is crucial for autonomous driving, especially in smoke where LiDAR data becomes sparse and noisy. Due to the lack of real smoke data, this paper introduces a physics-based simulation framework to generate realistic LiDAR point clouds of smoke and augment large-scale driving datasets. First, we present a 3D fluid dynamics-based smoke simulation framework in Unity, which models the realistic spatial diffusion and temporal evolution of smoke particles. Coupled with a physically accurate LiDAR perception module, our system captures complex light interactions—such as beam attenuation, scattering, and multi-path effects—to generate high-fidelity, physically consistent smoke point clouds. Second, we propose a range image-based data fusion strategy that seamlessly integrates the simulated smoke point clouds into large-scale real-world LiDAR datasets (e.g., Waymo). This approach accurately emulates LiDAR scanning characteristics and naturally incorporates occlusion effects, enabling realistic smoke integration without compromising spatial consistency. To validate our approach, we collect a real-world LiDAR smoke dataset (LiSmoke) and conduct extensive experiments using state-of-the-art 3D detectors. Results demonstrate that models trained with our augmented synthetic data achieve significant improvements in smoke-affected scenarios, while maintaining competitive performance in clear-weather conditions. Our work provides a cost-effective solution for enhancing perception robustness in safety-critical environments.

3D object detection is critical for autonomous driving, yet it remains fundamentally challenging to simultaneously maximize computational efficiency and capture long-range spatial dependencies.We observed that Mamba-based models, with their linear state-space design, capture long-range dependencies at lower cost, offering a promising balance between efficiency and accuracy.However, existing methods rely on axis-aligned scanning within a fixed window, inevitably discarding spatial information. To address this problem, we propose WinMamba, a novel Mamba-based 3D feature-encoding backbone composed of stacked WinMamba blocks. To enhance the backbone with robust multi-scale representation, the WinMamba block incorporates a window-scale-adaptive module that compensates voxel features across varying resolutions during sampling. Meanwhile, to obtain rich contextual cues within the linear state space, we equip the WinMamba layer with a learnable positional encoding and a window-shift strategy.Extensive experiments on the KITTI and Waymo datasets demonstrate that WinMamba significantly outperforms the baseline. Ablation studies further validate the individual contributions of the WSF and AWF modules in improving detection accuracy. The code will be made publicly available.

Foundation segmentation models, such as SAM and its video-oriented variant SAM2, have achieved remarkable success in natural image and video segmentation. However, their direct application to echocardiography video is challenged by structural uncertainty arising from severe speckle noise and blurry anatomical boundaries. To address this, we propose E³SAM2, a lightweight adaptation framework that introduces a novel entropy-based methodology to explicitly model and mitigate such uncertainty. Specifically, an entropy-guided attention mechanism is introduced to steer the model’s focus toward structurally reliable features, particularly in speckle-dominated regions. Additionally, an entropy regularization loss is introduced to further enhance target-background discrimination. To better resolve indistinct anatomical contours, an edge-aware supervision module is incorporated to inject explicit boundary priors for sharper delineation. These components are efficiently integrated through a global-local feature adapter. Experiments on CAMUS and EchoNet-Dynamic datasets demonstrate that E³SAM2 achieves state-of-the-art segmentation and clinical estimation performance, while maintaining high computational efficiency.

Stable Diffusion (SD) often produces degraded outputs when the training dataset contains adversarial noise. Adversarial purification offers a promising solution by removing adversarial noise from contaminated data. However, existing purification methods are primarily designed for classification tasks and fail to address SD-specific adversarial strategies, such as attacks targeting the VAE encoder, UNet denoiser, or both. To address the gap in SD security, we propose Universal Diffusion Adversarial Purification (UDAP), a novel framework tailored for defending adversarial attacks targeting SD models. UDAP leverages the distinct reconstruction behaviors of clean and adversarial images during Denoising Diffusion Implicit Models (DDIM) inversion to optimize the purification process. By minimizing the DDIM metric loss, UDAP can effectively remove adversarial noise. Additionally, we introduce a dynamic epoch adjustment strategy that adapts optimization iterations based on reconstruction errors, significantly improving efficiency without sacrificing purification quality. Experiments demonstrate UDAP’s robustness against diverse adversarial methods, including PID (VAE-targeted), Anti-DreamBooth (UNet-targeted), MIST (hybrid), and robustness-enhanced variants like Anti-Diffusion (Anti-DF) and MetaCloak. UDAP also generalizes well across SD versions and text prompts, showcasing its practical applicability in real-world scenarios.

Pan-sharpening aims to generate high-resolution multispectral images by integrating the spectral richness of low-resolution multispectral images with the spatial details of high-resolution panchromatic images. Although frequency-domain modeling shows great potential in this field, most existing methods are still limited to spatial-domain processing or fail to effectively capture the contextual interactions between frequency and spatial features. To address these issues, we propose a novel multi-scale frequency-spatial collaborative fusion approach. A Frequency-Spatial U-Net is introduced as the backbone network, in which frequency-spatial modeling blocks are embedded to progressively enhance the frequency-guided spatial contextual modeling capability across layers. To this end, we design a Dual Branch Frequency Attention module that adaptively enhances high- and low-frequency information. In addition, we introduce fine-mid-coarse resolution branches and devise a main-auxiliary multi-scale reconstruction loss to facilitate collaborative optimization. The effectiveness of the proposed model is validated through extensive experiments, demonstrating superior performance in both qualitative and quantitative evaluations. Moreover, our model achieves the fastest inference time among all compared methods, striking an excellent balance between accuracy and efficiency.

Camera-based temporal 3D object detection has shown impressive results in autonomous driving, with offline models improving accuracy by using future frames. Knowledge distillation (KD) can be an appealing framework for transferring rich information from offline models to online models. However, existing KD methods overlook future frames, as they mainly focus on spatial feature distillation under strict frame alignment or on temporal relational distillation, thereby making it challenging for online models to effectively learn future knowledge. To this end, we propose a sparse query-based approach, Future Temporal Knowledge Distillation (FTKD), which effectively transfers future frame knowledge from an offline teacher model to an online student model. Specifically, we present a future-aware feature reconstruction strategy to encourage the student model to capture future features without strict frame alignment. In addition, we further introduce future-guided logit distillation to leverage the teacher's stable foreground and background context. FTKD is applied to two high-performing 3D object detection baselines, achieving up to 1.3 mAP and 1.3 NDS gains on the nuScenes dataset, as well as the most accurate velocity estimation, without increasing inference cost.

The limited availability of high-quality training data poses a persistent challenge for synthetic aperture radar (SAR) target classification. Existing data augmentation methods mainly adopt a simplistic application of GAN-based style transfer techniques to directly synthesize pseudo-SAR images from optical images. However, our in-depth analysis of this cross-modal conversion reveals that such straightforward strategies primarily focus on transferring high-level semantic information (e.g., target shapes), thus failing to adequately capture the essential low-level features unique to SAR imagery (e.g., scattering textures). To address this inherent trade-off between high-level semantic preservation and low-level feature authenticity, we propose a Hierarchical Feature-Constrained GAN (HiFC-GAN) tailored for optical-to-SAR style transfer. Specifically, HiFC-GAN enhances the representation of low-level SAR features by introducing local texture contrast constraints at shallow layers, while introducing explicit feature mapping constraints at deeper layers to maintain high-level semantic consistency throughout the reconstruction process. Experimental results demonstrate that HiFC-GAN significantly outperforms existing GAN-based techniques in image generation quality, particularly improving the low-level feature authenticity of pseudo-SAR images. Moreover, the generated pseudo-SAR images further improve the performance of downstream target classification tasks, yielding accuracy gains ranging from 3.56% to 5.90% on average with mainstream CNN-based models.

The proliferation of synthetic facial imagery has intensified the need for robust Open-World DeepFake Attribution (OW-DFA), which aims to attribute both known and unknown forgeries using labeled data for known types and unlabeled data containing a mixture of known and novel types. However, existing OW-DFA methods face two critical limitations: 1) A confidence skew that leads to unreliable pseudo-labels for novel forgeries, resulting in biased training. 2) An unrealistic assumption that the number of unknown forgery types is known a priori. To address these challenges, we propose a Confidence-aware Asymmetric Learning (CAL) framework, which adaptively balances model confidence across known and novel forgery types. CAL mainly consists of two components: Confidence-aware Consistency Regularization (CCR) and Asymmetric Confidence Reinforcement (ACR). CCR mitigates pseudo-label bias by dynamically scaling sample losses based on normalized confidence, gradually shifting the training focus from high- to low-confidence samples. ACR complements this by separately calibrating confidence for known and novel classes through selective learning on high-confidence samples, guided by their confidence gap. Together, CCR and ACR form a mutually reinforcing loop that significantly improves the model's OW-DFA performance. Moreover, we introduce a Dynamic Prototype Pruning (DPP) strategy that automatically estimates the number of novel forgery types in a coarse-to-fine manner, removing the need for unrealistic prior assumptions and enhancing the scalability of our methods to real-world OW-DFA scenarios. Extensive experiments on the standard and OW-DFA benchmark and a newly extended benchmark incorporating advanced manipulations demonstrate that CAL consistently outperforms previous methods, achieving new state-of-the-art performance on both known and novel forgery attribution.

The rise of Vision Transformers (ViTs) as cornerstone models in safety-critical applications like autonomous driving and medical diagnosis has shifted the focus from pure accuracy to verifiable trustworthiness. However, the very mechanisms used to explain these models, their internal attention maps, are themselves vulnerable. This creates a critical "trust gap," as the model's apparent reasoning can be maliciously manipulated. To systematically investigate this vulnerability, we introduce A-SAGE (Attention-based Steering Adversarial Generation by Corrupting Explanations), a dual-objective attack framework that forces a model to misclassify an input while simultaneously corrupting its internal attention patterns to generate a misleading explanation. A-SAGE achieves this by optimizing a unified loss that combines a standard classification objective with two explanation-specific terms: an attention entropy loss to diffuse the model's focus and an attention map distortion loss to steer the corrupted explanation towards a desired target. Our primary finding is A-SAGE's exceptional black-box transferability. Using a CaiT-S as a white-box surrogate, adversarial examples generated with imperceptible perturbations achieve attack success rates of 79.4% on ViT-B, 49.7% on ResNet-50, and over 81.5% on other transformers (DeiT-B,TNT-S). Crucially, these successful attacks do not merely destroy the explanation; they generate a coherent but false attention map that deceptively "justifies" the wrong prediction. These results reveal a systemic vulnerability in the core reasoning of modern foundation models, establishing A-SAGE as a critical benchmark for auditing the robustness of AI explainability.

Aerial-Ground Person Re-IDentification (AGPReID) aims to extract identity-discriminative representations from heterogeneous perspectives across different platforms in complex real-world environments. However, existing methods primarily focus on visual appearance modeling and make insufficient use of semantic attribute priors, which limits their ability to bridge the aerial-ground view gap. To address this limitation, we propose a Semantic-driven Visual Progressive Refinement framework for AGPReID (SVPR-ReID), which effectively leverages textual attribute priors to guide the extraction of fine-grained visual cues. Specifically, we design a View-Decoupled Feature Extractor that incorporates view-aware textual prompts to decouple view-invariant identity features. Then, to alleviate inter-class ambiguity, we propose an Attribute-Scattered Mixture-of-Experts module that integrates attribute semantics into the visual space, thereby improving discrimination among visually similar pedestrians. Finally, we design a Context-Vision Progressive Refinement module for progressive refinement of attribute and view-invariant features, obtaining robust cross-view identity representations. In particular, we contribute a comprehensive benchmark for AGPReID, named CP2108, which contains 142,817 images of 2,108 identities annotated with 22 attributes. Notably, it includes 191 identities captured across different times, enabling both short- and long-term ReID evaluation, addressing the limitation of existing datasets that focus only on short-term scenarios. Extensive experimental results validate the effectiveness of our SVPR-ReID on four AGPReID datasets.

In the field of multi-spectral object re-identification (ReID), multi-modal knowledge and modal-specific knowledge exhibit complementary advantages when handling hard samples, but existing methods rarely integrate this collaborative information. Knowledge distillation is a direct approach for transferring information, however, heterogeneity in model architectures and variations in sample hardness can undermine the stability and controllability of knowledge transfer. To alleviate these limitations, we propose the novel Progressive Multi-modal Knowledge Distillation (PMKD) framework that enables multi-stage knowledge transfer guided by hard sample awareness. In the multi-modal knowledge transfer stage, the source model (pre-trained on multi-modal data) disseminates its learned multi-modal collaborative knowledge to multiple independently modal-specific target models, guiding their adaptation to hard samples within training batches. In the modal-specific knowledge retention stage, the independent models enriched with multi-modal knowledge guide the training phase. The architectural consistency between source-target models ensures more lossless knowledge transfer, effectively mitigating the risk of capability drift, and preserving inherent competence. Moreover, the entire progressive multi-modal knowledge distillation is regulated by the proposed hardness-aware distillation loss, which automatically adapts distillation intensity through hard sample mining, thereby ensuring stable transfer of hard sample handling capabilities. Extensive experiments on benchmark multi-spectral ReID datasets validate the effectiveness and superior performance of the proposed method.

Multi-modal object re-identification (ReID) aims to retrieve specific targets by leveraging complementary cues from different sensing modalities. Despite recent progress, two key challenges remain: (1) the limited ability to jointly address both modality and viewpoint discrepancies, and (2) the difficulty of effectively leveraging reliable target-domain data to improve generalization. To address these challenges, we propose Proxy-driven Test-Time Training (ProxyTTT), a unified framework that enhances both multi-modal identity representation learning and model generalization. During training, we propose a Multi-Proxy Learning (MPL) mechanism to address the representation bias across different views and modalities. MPL disentangles fine-grained modality-specific and modality-common identity proxies as semantic anchors to align identity features across diverse perspectives and sensing modalities. This alignment strategy enables the model to learn robust and discriminative global identity representations under heterogeneous modality conditions. At test time, to reliably exploit target domain data, we propose Proxy-guided Entropy-based Selective Adaptation (PESA) for test-time training. Specifically, PESA leverages the semantic structure encoded by identity proxies to estimate prediction uncertainty via entropy, and selectively adapts the model using only high-confidence samples. This selective adaptation effectively mitigates the domain shift between training and deployment environments, improving the model’s generalization in real-world scenarios. Extensive experiments on four public multi-modal ReID benchmarks (RGBNT201, RGBNT100, MSVR310, and WMVeID863) demonstrate the effectiveness of ProxyTTT.

Despite advancements in Video Large Language Models (Vid-LLMs) improving multimodal understanding, challenges persist in streaming video reasoning due to its reliance on contextual information. Existing paradigms feed all available historical contextual information into Vid-LLMs, resulting in a significant computational burden for visual data processing. Furthermore, the inclusion of irrelevant context distracts models from key details. This paper introduces a challenging task called Context-guided Streaming Video Reasoning (CogStream), which simulates real-world streaming video scenarios, requiring models to identify the most relevant historical contextual information to deduce answers for questions about the current stream. To support CogStream, we present a densely annotated dataset featuring extensive and hierarchical question-answer pairs, generated by a semi-automatic pipeline. Additionally, we present CogReasoner as a baseline model. It effectively tackles this task by leveraging visual stream compression and historical dialogue retrieval. Extensive experiments prove the effectiveness of this method.

Existing text-video retrieval methods mainly focus on singlemodal video content (i.e., visual entities), often overlooking heterogeneous scene text that is ubiquitous in human environments. Although scene text in videos provides finegrained semantics for cross-modal retrieval, effectively utilizing it presents two key challenges: (1) Temporally dense scene text disrupts sync with sparse video frames, obstructing video understanding;(2) Redundant scene text and irrelevant video frames hinder the learning of discriminative temporal clues for retrieval. To address them, we propose a temporal scene-text calibrating and distilling (TCD) network for textvideo retrieval. Specifically, we first design a window-OCR captioner that aggregates dense scene text into OCR captions to facilitate feature interaction. Next, we devise a heterogeneous semantics calibration module that leverages scene text as a self-supervised signal to temporally align window-level OCR captions and frame-level video features. Further, we introduce a context-guided temporal clue distillation module to learn the complementary and relevant details between scene text and video modalities, thereby obtaining discriminative temporal clues for retrieval. Extensive experiments show that our TCD achieves state-of-the-art performance on three scene-text related benchmarks.

Addressing missing modalities is a critical challenge in multimodal brain tumor segmentation. Most existing approaches merely handle modality-incomplete inputs during inference, assuming a full set of modalities for all training samples. However, this unrealistic assumption limits the usage of abundant modality-incomplete data commonly observed in clinical practice. In this paper, we explore a more practical task of tackling missing modalities during both training and inference. We propose a universal model featuring robust modality reconstruction and prompt-guided modality adaptation. Our mask-reconstruction pre-training enables robust modality-invariant representation learning, during which we design a novel distribution approximation method that supervises the reconstruction of absent modalities without requiring full-modal training data. Afterwards, when adapting our model to the segmentation task, we introduce the complete-then-distill (CTD) paradigm, which first estimates missing modalities in training samples from the available ones, and then distills the knowledge from the reconstructed full-modal representations to enhance learning from modality-incomplete data. Moreover, we propose prompt-guided modality adaptation to personalize a subset of model parameters during CTD, enabling the model to adapt to each distinct modality input scenario by using prompts with rich visual-textual information. Extensive experiments on two brain tumor segmentation benchmarks show our method consistently surpasses previous state-of-the-art approaches under dual-stage missing modality settings across various missing ratios.

Generating high-quality, controllable, and structurally consistent 3D scenes in complex multi-object environments remains a fundamental challenge. We present SceneGenesis, a unified framework that synthesizes 3D scenes by combining semantic structural priors with mesh-guided video–geometry fusion. SceneGenesis first employs large language models to convert textual descriptions into category-aware object specifications, which are transformed into structured meshes using procedural approximations and pretrained asset generators, enabling precise layout control and scalable scene construction. To obtain rich and style-controllable appearances, SceneGenesis generates multi-view video representations conditioned on the initialized structure. A mesh-guided video–geometry fusion module then consolidates video evidence with mesh priors through mesh-conditioned fragment initialization, progressive geometric refinement, and structure-aware optimization, substantially improving global geometric fidelity and visual realism. Experiments demonstrate that SceneGenesis supports flexible style variation and object-level editing while achieving strong controllability, scalability, and structural quality.