AAAI 2026 论文列表

While adapting pretrained vision models to downstream dense prediction tasks is widely used, current methods often overlook adaptation efficiency, especially in the context of multi-task learning (MTL). Although parameter-efficient fine-tuning (PEFT) methods can enhance parameter efficiency, broader aspects such as GPU memory and training time efficiency remain underexplored. In this paper, we propose a new paradigm that simultaneously achieves efficiency in Parameters, GPU Memory, and Training Time for Multi-Task Dense Vision Adaptation. Specifically, we propose a dual-branch framework, in which a frozen pretrained backbone serves as the generic main branch, and the proposed Bi-Directional Task Adaptation (BDTA) modules are integrated in parallel to form a task bypass branch that extracts adaptation features required by multiple specific tasks. This adaptation module is lightweight, efficient, and does not require backpropagation through the large pre-trained backbone, thus avoiding resource-intensive gradient computations. Moreover, a Mixture of Task Experts mechanism (MoTE) is further proposed to integrate adaptation features across tasks and scales, thereby obtaining more robust representations tailored for dense prediction tasks. On the PASCAL-Context benchmark, our method achieves over 2× relative performance improvement compared to the best prior multi-task PEFT method, while using only ~30% of the parameters, ~50% of the memory, and ~60% of the training time, demonstrating superior overall adaptation efficiency.

Nucleus detection and classification (NDC) in histopathology analysis is a fundamental task that underpins a wide range of high-level pathology applications. However, existing methods heavily rely on labor-intensive nucleus-level annotations and struggle to fully exploit large-scale unlabeled data for learning discriminative nucleus representations. In this work, we propose MUSE (MUlti-scale denSE self-distillation), a novel self-supervised learning method tailored for NDC. At its core is NuLo (Nucleus-based Local self-distillation), a coordinate-guided mechanism that enables flexible local self-distillation based on predicted nucleus positions. By removing the need for strict spatial alignment between augmented views, NuLo allows critical cross-scale alignment, thus unlocking the capacity of models for fine-grained nucleus-level representation. To support MUSE, we design a simple yet effective encoder-decoder architecture and a large field-of-view semi-supervised fine-tuning strategy that together maximize the value of unlabeled pathology images. Extensive experiments on three widely used benchmarks demonstrate that MUSE effectively addresses the core challenges of histopathological NDC. The resulting models not only surpass state-of-the-art supervised baselines but also outperform generic pathology foundation models.

Camera-based 3D semantic scene completion (SSC) plays a crucial role in autonomous driving, enabling voxelized 3D scene understanding for effective scene perception and decision-making. Existing SSC methods have shown efficacy in improving 3D scene representations, but suffer from the inherent input-output dimension gap and annotation-reality density gap, where the 2D planar view from input images with sparse annotated labels leads to inferior prediction of real-world dense occupancy with a 3D stereoscopic view. In light of this, we propose the corresponding High-Dimension High-Density Semantic Scene Completion (HD²-SSC) framework with expanded pixel semantics and refined voxel occupancies. To bridge the dimension gap, a High-dimension Semantic Decoupling module is designed to expand 2D image features along a pseudo third dimension, decoupling coarse pixel semantics from occlusions, and then identify focal regions with fine semantics to enrich image features. To mitigate the density gap, a High-density Occupancy Refinement module is devised with a ``detect-and-refine" architecture to leverage contextual geometric and semantic structures for enhanced semantic density with the completion of missing voxels and correction of erroneous ones. Extensive experiments and analyses on the SemanticKITTI and SSCBench-KITTI-360 datasets validate the effectiveness of our HD²-SSC framework.

The increasing popularity of long Text-to-Image (T2I) generation has created an urgent need for automatic and interpretable models that can evaluate the image-text alignment in long prompt scenarios. However, the existing T2I alignment benchmarks predominantly focus on short prompt scenarios and only provide MOS or Likert scale annotations. This inherent limitation hinders the development of long T2I evaluators, particularly in terms of the interpretability of alignment. In this study, we contribute LongT2IBench, which comprises 14K long text-image pairs accompanied by graph-structured human annotations. Given the detail-intensive nature of long prompts, we first design a Generate-Refine-Qualify annotation protocol to convert them into textual graph structures that encompass entities, attributes, and relations. Through this transformation, fine-grained alignment annotations are achieved based on these granular elements. Finally, the graph-structed annotations are converted into alignment scores and interpretations to facilitate the design of T2I evaluation models. Based on LongT2IBench, we further propose LongT2IExpert, a LongT2I evaluator that enables multi-modal large language models (MLLMs) to provide both quantitative scores and structured interpretations through an instruction-tuning process with Hierarchical Alignment Chain-of-Thought (CoT). Extensive experiments and comparisons demonstrate the superiority of the proposed LongT2IExpert in alignment evaluation and interpretation.

Existing paradigms for remote sensing change detection are caught in a trade-off: CNNs excel at efficiency but lack global context, while Transformers capture long-range dependencies at a prohibitive computational cost. This paper introduces ChangeRWKV, a new architecture that reconciles this conflict. By building upon the Receptance Weighted Key Value (RWKV) framework, our ChangeRWKV uniquely combines the parallelizable training of Transformers with the linear-time inference of RNNs. Our approach core features two key innovations: a hierarchical RWKV encoder that builds multi-resolution feature representation, and a novel Spatial-Temporal Fusion Module (STFM) engineered to resolve spatial misalignments across scales while distilling fine-grained temporal discrepancies. ChangeRWKV not only achieves state-of-the-art performance on the LEVIR-CD benchmark, with an 85.46% IoU and 92.16% F1 score, but does so while drastically reducing parameters and FLOPs compared to previous leading methods. This work demonstrates a new, efficient, and powerful paradigm for operational-scale change detection.

Vision-Language Models (VLMs) excel at extracting salient visual features from query images, thus exhibiting promising visual recognition performance. However, VLMs would encounter significant degradation in fine-grained scenarios due to their deficiency in distinguishing nuanced differences among candidate categories. As a remedy, we draw inspiration from the ``System 1 & System 2" cognitive theory of humans, paving the way to achieve fine-grained recognition for VLMs. To be specific, we observe that VLMs naturally align with System 1, quickly identifying candidate categories but leaving easily-confused ones unresolved. Based on the observation, we propose System-2 enhanCed visuAl recogNition (SCAN), a novel plug-and-play approach that makes VLMs aware of nuanced differences. In brief, SCAN first specifies and abstracts the discriminative attributes for the confused candidate categories and query images by resorting to off-the-shelf large foundation models, respectively. After that, SCAN adaptively integrates the salient visual features from System 1 with the nuanced differences derived from System 2, resolving confusion in candidates with estimated uncertainty. Extensive experiments on eight widely used fine-grained recognition benchmarks against 10 state-of-the-art baselines verify the effectiveness and superiority of SCAN.

The high cost of synthetic aperture radar (SAR) data acquisition motivates SAR image generation research. However, the data scarcity and SAR's inherent azimuth sensitivity make generative models suffer from severe azimuth overfitting. Most existing methods require supplementary data to work effectively, limiting their practicality. In this paper, we propose SAR-DisentDM, a novel semantic-disentangled diffusion model for limited-data SAR image generation, without requiring any auxiliary resources. We develop a physics-aware diffusion architecture that explicitly models semantic knowledge of SAR images, including intrinsic characteristics, contextual diversity, and measurement randomness. A key innovation is the attention-guided semantic disentanglement (AGSD) module, designed to decouple category-specific features from azimuth-variable scattering patterns. This is achieved by aid of a dual disentangled loss with time-step-adaptive optimization. Furthermore, we introduce an azimuth angle perturbation augmentation (AAPA) mechanism, to enhance the model's robustness to minor azimuth angle errors. Extensive evaluations validate that SAR-DisentDM enables controllable SAR image synthesis with designated attributes, significantly improving representation and generalization abilities under limited data. Synthetic imagery from our approach boosts automatic target recognition (ATR) accuracy beyond state-of-the-art methods.

The success of 3DGS in generative and editing applications has sparked growing interest in 3DGS-based style transfer. However, current methods still face two major challenges: (1) multi-view inconsistency often leads to style conflicts, resulting in appearance smoothing and distortion; and (2) heavy reliance on VGG features, which struggle to disentangle style and content from style images, often causing content leakage and excessive stylization. To tackle these issues, we introduce FantasyStyle, a 3DGS-based style transfer framework, and the first to rely entirely on diffusion model distillation. It comprises two key components: (1) Multi-View Frequency Consistency. We enhance cross-view consistency by applying a 3D filter to multi-view noisy latent, selectively reducing low-frequency components to mitigate stylized prior conflicts. (2) Controllable Stylized Distillation. To suppress content leakage from style images, we introduce negative guidance to exclude undesired content. In addition, we identify the limitations of Score Distillation Sampling and Delta Denoising Score in 3D style transfer and remove the reconstruction term accordingly. Building on these insights, we propose a controllable stylized distillation that leverages negative guidance to more effectively optimize the 3D Gaussians. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches, achieving higher stylization quality and visual realism across various scenes and styles.

Multimodal 3D object detection for autonomous driving, a task for real-world applications, poses substantial challenges in maintaining robust performance under various perturbations and complex environmental conditions. However, most existing approaches primarily focus on performance optimization under relatively ideal scenarios or focus on one or few disturbing conditions (or adverse conditions), lacking systematic exploration of robustness against real-world factors, including high class imbalance, adverse weather conditions, sensor jitter and failures, and significant scene variations. To address this issue, we propose a robust multimodal 3D detector, termed RobusTor3D, which integrates robustness at both the structural and supervisory levels by blending the knowledge from Vision-Language Models (VLMs). Structurally, textual descriptions are incorporated to enhance the semantic richness and diversity of rare classes. This novel semantic injection operation compensates for the inherent class imbalance and modality weakness in conventional visual features. Furthermore, semantic alignment capability and robust representation by Vision-Language Knowledge Extraction (V-LKE) serve as semantic priors to complement modality-specific representations, significantly improving model adaptability. At the supervisory level, we propose a Scene-level Multimodal Consistency Learning (SMCL) strategy, which jointly enforces global semantic constraints across modalities, encouraging the learning of stable and abundant semantic representations. This special design specifically reduces the impact of spatial alignment, while notably enabling semantic compensation under modality-loss conditions. Extensive robustness experiments conducted on KITTI, KITTI-C, and CADC benchmarks evaluate five robustness aspects, including long-tail problem, adverse weather (rain, snow, fog, strong sunlight), sensor spatial misalignment and motion blur, modality loss, and cross-domain scenarios. The results show that RobusTor3D demonstrates superior robustness across all five evaluated aspects. It consistently outperforms the state-of-the-art methods under various challenging conditions.

This paper focuses on the task of speech-driven 3D facial animation, which aims to generate realistic and synchronized facial motions driven by speech inputs. Recent methods have employed audio-conditioned diffusion models for 3D facial animation, achieving impressive results in generating expressive and natural animations. However, these methods process the whole audio sequences in a single pass, which poses two major challenges: they tend to perform poorly when handling audio sequences that exceed the training horizon and will suffer from significant latency when processing long audio inputs. To address these limitations, we propose a novel autoregressive diffusion model that outputs facial motions in a streaming manner. This design ensures flexibility with varying audio lengths and achieves low latency independent of audio duration. Specifically, we select a limited number of past frames as historical motion context and combine them with the audio input to create a dynamic condition. This condition guides a lightweight diffusion head to iteratively generate facial motion frames, enabling real-time synthesis with high-quality results. Experiments conducted on public datasets demonstrate that our approach outperforms recent baseline methods.

Universal medical image segmentation models have emerged as a promising paradigm due to their strong generalizability across diverse tasks, showing great potential for a wide range of clinical applications. This potential has been partly driven by the success of general-purpose vision models such as the Segment Anything Model (SAM), which has inspired the development of various fine-tuned variants for medical segmentation tasks. However, fine-tuned variants like MedSAM are trained on comparatively limited medical imaging data that often suffers from heterogeneity, scarce annotations, and distributional shifts. These challenges limit their ability to generalize across a wide range of medical segmentation tasks. In this regard, we propose MedSAMix, a training-free model merging method that integrates the strengths of both generalist models (e.g., SAM) and specialist models (e.g., MedSAM) for medical image segmentation. In contrast to traditional model merging approaches that rely on manual configuration and often result in suboptimal outcomes, we propose a zero-order optimization method to automatically discover optimal layer-wise merging solutions. Furthermore, for clinical applications, we develop two regimes to meet the demand of domain-specificity and generalizability in different scenarios by single-task optimization and multi-objective optimization respectively. Extensive evaluations on 25 medical segmentation tasks demonstrate that MedSAMix effectively mitigates model bias and consistently improves performance in both domain-specific accuracy and generalization, achieving improvements of 6.67% on specialized tasks and 4.37% on multi-task evaluations.

Cross-Domain Few-Shot Object Detection (CD-FSOD) faces significant challenges due to the dual issues of domain shift and limited labeled samples. One major challenge is style bias, caused by limited support samples that fail to represent the target domain’s style diversity. Another is feature confusion, which stems from distribution shifts and limited supervision, manifesting as both object-background ambiguity and object-object confusion. To address these challenges, we propose Style-Augmented Prototype Learning (StyleProto), which constructs style-aware prototypes from support samples with diverse visual styles, and refines them via spatial weighting and discriminative fusion. Specifically, our StyleProto consists of three components: (1) Style Generation Augmentation (SGA); (2) Semantic-Focused Prototype Construction (SPC); (3) Hierarchical Prototype Fusion Aggregator (HPFA). SGA synthesizes style-diverse yet semantically consistent training samples by recombining style statistics from the support set, thus improving robustness to unseen styles. SPC aggregates support features using spatial attention to highlight object semantics and suppress background noise, yielding cleaner and more distinctive class prototypes. HPFA leverages query-guided attention to integrate discriminative support features, enhancing prototype representations with richer class-specific details. Extensive experiments on multiple benchmarks demonstrate that StyleProto consistently outperforms existing state-of-the-art methods.

JPEG, as a widely used image compression standard, often introduces severe visual artifacts when achieving high compression ratios. Although existing deep learning-based restoration methods have made considerable progress, they often struggle to recover complex texture details, resulting in over-smoothed outputs. To overcome these limitations, we propose SODiff, a novel and efficient semantic-oriented one-step diffusion model for JPEG artifacts removal. Our core idea is that effective restoration hinges on providing semantic-oriented guidance to the pre-trained diffusion model, thereby fully leveraging its powerful generative prior. To this end, SODiff incorporates a semantic-aligned image prompt extractor (SAIPE). SAIPE extracts rich features from low-quality (LQ) images and projects them into an embedding space semantically aligned with that of the text encoder. Simultaneously, it preserves crucial information for faithful reconstruction. Furthermore, we propose a quality factor-aware time predictor that implicitly learns the compression quality factor (QF) of the LQ image and adaptively selects the optimal denoising start timestep for the diffusion process. Extensive experimental results show that our SODiff outperforms recent leading methods in both visual quality and quantitative metrics.

Arbitrary style transfer (AST), a popular AI-powered photo editing function, aims to strike an optimal balance between content and style injection from two images in order to generate a novel high-fidelity stylised image. Recently, diffusion models have been applied to AST due to their high generation quality as well as flexibility to embed conditions. However, these models are still not satisfactory and may exhibit inferior performance compared to non-diffusion based methods. This is due to the diffusion process not being purposely designed for AST, leading to suboptimal solutions to trade-off content preservation and style embedding. In this paper, we propose ACID-Style, a novel adaptive condition injection diffusion-based AST framework for improved content/style feature injection to address this research challenge. Using two lightweight adapters, a content and a style injection module, and an adaptive injection mechanism, our approach is able to fully exploit a pre-trained stable diffusion model for AST-specific adaptation and our diffusion model thus learns the most effective timing for content and style injection in the diffusion sampling process. Comprehensive evaluations demonstrate that our method achieves superior style transfer performance, both quantitatively and qualitatively, compared to other state-of-the-art style transfer methods.

Monocular 3D object detection offers a cost-effective solution for autonomous driving, but it suffers from the ill-posed depth and a limited field of view. These constraints lead to the lack of geometric cues and reduced accuracy in occluded or truncated scenes. While recent approaches incorporate additional depth information to address geometric ambiguity, they overlook the importance of visual cues essential for robust object recognition. In this paper, we propose MonoCLUE that enhances monocular 3D detection by leveraging both local clustering and generalized scene memory of visual features. First, we perform K-means clustering on visual features to capture distinct object-level appearance visual parts (e.g., bonnet, car roof), which improves the detection of partially visible objects. The clustered features are then propagated across the entire region to capture objects with similar appearances. Second, we construct a generalized scene memory by aggregating clustered features across images, providing consistent appearance representations that generalize scenes. This improves the consistency of object-level features, enabling stable detection across varying environments. Lastly, we integrate both local cluster features and generalized scene memory into object queries, guiding attention toward informative regions in the feature map. Exploiting an unified local clustering and generalized scene memory strategy, MonoCLUE enables robust monocular 3D detection under occlusion and limited visibility. Our proposed model achieves state-of-the-art performance on the KITTI benchmark.

Generating realistic and coordinated 3D human motion for multiple individuals within complex environments remains a significant challenge. Existing text-to-motion methods are often ``blind'' to the physical scene, leading to implausible motions, while scene-conditioned (HSI) approaches demand cumbersome full 3D data and largely neglect multi-person dynamics. To address these limitations, we introduce the VL2Motion paradigm and its embodiment, MMG-VL, a hierarchical framework that generates coordinated multi-person motions from the most accessible inputs: a single 2D image and natural language. MMG-VL first employs a Scene-Aware Intent Planner (SAIP) to interpret the visual context and decompose the user's command into a set of spatially-grounded, multi-person action blueprints. Subsequently, a Coordinated Motion Synthesizer (CMS) translates these blueprints into high-fidelity 3D motion sequences. The synergy between these stages is driven by two novel loss functions: a Spatial-Semantic Grounding Loss to ensure the planner's output is grounded in visual reality, and a Coordinated Environmental Realism Loss that enforces physical constraints and coherent group dynamics during synthesis. To facilitate this research, we introduce HumanVL, the first large-scale dataset featuring multi-person activities in multi-room scenes, providing aligned images, text, blueprints, 3D motions, and scene geometry. Extensive experiments demonstrate that MMG-VL significantly outperforms existing methods in generating spatially coherent, physically realistic, and coordinated multi-person motions, paving the way for more scalable and intuitive creation of dynamic virtual worlds.

Vector quantization (VQ) transforms continuous image features into discrete representations, providing compressed, tokenized inputs for generative models. However, VQ-based frameworks suffer from several issues, such as non-smooth latent spaces, weak alignment between representations before and after quantization, and poor coherence between the continuous and discrete domains. These issues lead to unstable codeword learning and underutilized codebooks, ultimately degrading the performance of both reconstruction and downstream generation tasks. To this end, we propose VAEVQ, which comprises three key components: (1) Variational Latent Quantization (VLQ), replacing the AE with a VAE for quantization to leverage its structured and smooth latent space, thereby facilitating more effective codeword activation; (2) Representation Coherence Strategy (RCS), adaptively modulating the alignment strength between pre- and post-quantization features to enhance consistency and prevent overfitting to noise; and (3) Distribution Consistency Regularization (DCR), aligning the entire codebook distribution with the continuous latent distribution to improve utilization. Extensive experiments on two benchmark datasets demonstrate that VAEVQ outperforms state-of-the-art methods.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in multimodal reasoning. However, they often excessively rely on textual information during the later stages of inference, neglecting the crucial integration of visual input. Current methods typically address this by explicitly injecting visual information to guide the reasoning process. In this work, through an analysis of MLLM attention patterns, we made an intriguing observation: with appropriate guidance, MLLMs can spontaneously re-focus their attention on visual inputs during the later stages of reasoning, even without explicit visual injection. This spontaneous shift in focus suggests that MLLMs are intrinsically capable of performing visual fusion reasoning. Building on this insight, we introduce Look-Back, an implicit approach designed to guide MLLMs to look back at visual information in a self-directed manner during reasoning. Look-Back empowers the model to autonomously determine when, where, and how to re-focus on visual inputs, eliminating the need for explicit model-structure constraints or additional input. We demonstrate that Look-Back significantly enhances the model's reasoning and perception capabilities, as evidenced by extensive empirical evaluations on multiple multimodal benchmarks.

The multi-modality remote sensing foundation model (MM-RSFM) has made notable progress recently. However, most existing approaches remain limited to medium-resolution, single-modality, restricting their performance in fine-grained downstream applications such as disaster response and urban planning. In this work, MaRS is proposed, a multi-modality very-high-resolution (VHR) remote sensing foundation model designed for cross-modality granularity interpretation of complex scenes. To achieve this, a multi-modality VHR SAR-optical dataset, MaRS-16M, is constructed through large-scale collection and semi-automated processing, comprising over 16 million paired samples. Unlike previous work, MaRS tackles two fundamental challenges in VHR SAR-optical self-supervised learning (SSL) techniques. Cross-granularity contrastive learning (CGCL) is introduced to alleviate alignment inconsistencies caused by imaging differences, and meta-modality attention (MMA) is designed to unify heterogeneous physical characteristics across modalities. Compared to existing remote sensing foundation models (RSFMs) and general vision foundation models (VFMs), MaRS performs better as a pre-trained backbone across nine multi-modality VHR downstream tasks.