AAAI 2026 论文列表

Recent advancements in diffusion-based generative priors have enabled visually plausible image compression at extremely low bit rates. However, existing approaches suffer from slow sampling processes and suboptimal bit allocation due to fragmented training paradigms. In this work, we propose Accelerate Diffusion-based Image Compression via Consistency Prior Refinement (DiffCR), a novel compression framework for efficient and high-fidelity image reconstruction. At the heart of DiffCR is a Frequency-aware Skip Estimation (FaSE) module that refines the epsilon-prediction prior from a pre-trained latent diffusion model and aligns it with compressed latents at different timesteps via Frequency Decoupling Attention (FDA). Furthermore, a lightweight consistency estimator enables fast two-step decoding by preserving the semantic trajectory of diffusion sampling. Without updating the backbone diffusion model, DiffCR achieves substantial bitrate savings (27.2% BD-rate(LPIPS) and 65.1% BD-rate(PSNR)) and over 10 times speed-up compared to SOTA diffusion-based compression baselines.

A 3D point cloud completion task is to generate completed 3D objects given partial observations. Auto-encoder-based models suffer from poor generalization ability to untrained 3D data. Current diffusion-based models add isotropic noise with the same variance in three x, y, z axes. More importantly, these models ignore real-world anisotropic evolution properties of 3D particles from a non-equilibrium state to thermodynamic equilibrium in the real physical world due to the velocity and energy thermodynamics of the particles, leading to unstable completions of 3D object topology. This paper presents a novel physically-based anisotropic 3D diffusion model (3DDM) to address these issues. We also present derivations of our proposed forward and reverse processes and a loss function in closed form, thus reproducibility. The 3DDM contains anisotropic energy-aware forward and reverse processes with a novel anisotropic quadratic loss function. The forward process adds anisotropic 3D Gaussian noises per-axis and mimics the thermal non-equilibrium evolution towards Maxwellian equilibrium based on velocity and kinetic energy evolutions of 3D particles in the real physical space. The reverse process learns to denoise along per-axis and per-timestep anisotropically. The anisotropic quadratic loss function penalizes errors along certain axes, yielding a highly flexible and anisotropic reverse diffusion process and a physically realistic generative model. The 3DDM denoises along x, y, z axes with different velocities from the non-equilibrium evolution, achieving fewer than 20 diffusion steps and strong generalization to unseen 3D objects and real-world scenes that were not trained.

We propose PFAvatar (Pose-Fusion Avatar), a new method that reconstructs high-quality 3D avatars from Outfit of the Day (OOTD) photos, which exhibit diverse poses, occlusions, and complex backgrounds. Our method consists of two stages: (1) fine-tuning a pose-aware diffusion model from few-shot OOTD examples and (2) distilling a 3D avatar represented by a neural radiance field (NeRF). In the first stage, unlike previous methods that segment images into assets (e.g. garments, accessories) for 3D assembly, which is prone to inconsistency, we avoid decomposition and directly model the full-body appearance. By integrating a pre-trained ControlNet for pose estimation and a novel Condition Prior Preservation Loss (CPPL), our method enables end-to-end learning of fine details while mitigating language drift in few-shot training. Our method completes personalization in just 5 minutes, achieving a 48x speed-up compared to previous approaches. In the second stage, we introduce a NeRF-based avatar representation optimized by canonical SMPL-X space sampling and Multi-Resolution 3D-SDS. Compared to mesh-based representations that suffer from resolution-dependent discretization and erroneous occluded geometry, our continuous radiance field can preserve high-frequency textures (e.g., hair) and handle occlusions correctly through transmittance. Experiments demonstrate that PFAvatar outperforms state-of-the-art methods in terms of reconstruction fidelity, detail preservation, and robustness to occlusions/truncations, advancing practical 3D avatar generation from real-world OOTD albums. In addition, the reconstructed 3D avatars support downstream applications such as virtual try-on, animation, and human video reenactment, further demonstrating the versatility and practical value of our approach.

In this paper, we propose a novel framework for controllable video diffusion, OmniVDiff , aiming to synthesize and comprehend multiple video visual content in a single diffusion model. To achieve this, OmniVDiff treats all video visual modalities in the color space to learn a joint distribution, while employing an adaptive control strategy that dynamically adjusts the role of each visual modality during the diffusion process, either as a generation modality or a conditioning modality. Our framework supports three key capabilities: (1) Text-conditioned video generation, where all modalities are jointly synthesized from a textual prompt; (2) Video understanding, where structural modalities are predicted from rgb inputs in a coherent manner; and (3) X-conditioned video generation, where video synthesis is guided by finegrained inputs such as depth, canny and segmentation. Extensive experiments demonstrate that OmniVDiff achieves state-of-the-art performance in video generation tasks and competitive results in video understanding. Its flexibility and scalability make it well-suited for downstream applications such as video-to-video translation, modality adaptation for visual tasks, and scene reconstruction.

Pre-trained diffusion models have shown great potential in real-world image super-resolution (Real-ISR) tasks by enabling high-resolution reconstructions. While one-step diffusion (OSD) methods significantly improve efficiency compared to traditional multi-step approaches, they still have limitations in balancing fidelity and realism across diverse scenarios. Since the OSDs for SR are usually trained or distilled by a single timestep, they lack flexible control mechanisms to adaptively prioritize these competing objectives, which are inherently manageable in multi-step methods through adjusting sampling steps. To address this challenge, we propose a Realism Controlled One-step Diffusion (RCOD) framework for Real-ISR. RCOD provides a latent domain grouping strategy that enables explicit control over fidelity-realism trade-offs during the noise prediction phase with minimal training paradigm modifications and original training data. A degradation-aware sampling strategy is also introduced to align distillation regularization with the grouping strategy and enhance the controlling of trade-offs. Moreover, a visual prompt injection module is used to replace conventional text prompts with degradation-aware visual tokens, enhancing both restoration accuracy and semantic consistency. Our method achieves superior fidelity and perceptual quality while maintaining computational efficiency. Extensive experiments demonstrate that RCOD outperforms state-of-the-art OSD methods in both quantitative metrics and visual qualities, with flexible realism control capabilities in the inference stage.

Vision-Language Models (VLMs) are widely used in tasks like Open-Vocabulary Object Detection and zero-shot Classification, owing to their powerful generalization. However, recent research reveals that VLMs exhibit significant performance instability when tasked with recognizing concepts at varying granularities (e.g., ``animal'' vs. ``dog''). Prevailing methods inject external knowledge from Large Language Models, but this unconstrained approach distorts the VLM's inherent hierarchical orthogonal geometry, leading to performance collapse on general concepts. To address this, we introduce GeCoin, an innovative Geometrically Constrained framework that safely enhances existing VLMs with external knowledge for improved hierarchical understanding, without additional training. By projecting knowledge into the null-space of a query concept's feature space, GeCoin mathematically guarantees the preservation of general knowledge while integrating specialized information. Extensive experiments across large-scale benchmarks, diverse VLMs, and knowledge from various LLMs (e.g., GPT-3.5, Claude-3, Gemini-Pro) show that GeCoin boosts performance by an average of 3.9% over the strongest baseline—crucially eradicating performance collapse on general concepts.

Dual-lens video inpainting aims to simultaneously restore missing or corrupted contents in videos captured by each lens of binocular systems. Although preliminary explorations have been conducted, existing methods still face two key challenges: limited exploitation of long-range reference information and inadequate modeling of inter-lens consistency in non-standard binocular systems. In this paper, we propose a novel dual-lens video inpainting framework named DLVINet, which addresses these challenges with two core components. Firstly, we develop a sparse spatial-temporal transformer (SSTT) that effectively utilizes the information from distant frames to complete the video contents of each lens individually. By employing sparse spatial-temporal attention with a channel selection mechanism, SSTT not only restores missing regions, but also avoids introducing redundant or irrelevant information. Furthermore, SSTT introduces a multi-scale feed-forward network to enrich the multi-scale representation of completed features. Secondly, we design a cross-lens texture transformer (CLTT) to model inter-lens consistency. By interacting with corresponding features between lenses under the guidance of cross-attention, CLTT captures global inter-lens correspondences. Such a design enables effective cross-view information modeling without being constrained by horizontal parallax, which is particularly critical for non-standard binocular systems. Extensive experiments demonstrate the effectiveness of our DLVINet.

Multi-view diabetic retinopathy (DR) grading has achieved remarkable performance by capturing more comprehensive pathological features than single-view methods. However, complete multi-view fundus images are often difficult to obtain in clinical practice, and the performance degrades significantly when fewer views are available. To overcome this limitation, we propose the first incomplete multi-view DR grading framework, aiming to provide accurate diagnosis regardless of the number of available views. It introduces two novel modules. First, cross-view spatial correlation attention (CSCA) captures region correlations across views, automatically identifying and fusing diagnostically relevant spatial features to improve feature representation. Second, self-supervised mask consistency learning (SMCL) formulates a novel pretext task of missing-view information reconstruction by strategically masking inter- and intra-view regions, enabling the model to infer complete features from incomplete views. Benefiting from CSCA and SMCL, our method enhances structural feature consistency across views and effectively compensates for missing information during DR grading. Extensive experiments demonstrate that our method achieves state-of-the-art grading performance, particularly under realistic conditions where some views are unavailable.

Gait recognition is emerging as a promising technology and an innovative field within computer vision, with a wide range of applications in remote human identification. However, existing methods typically rely on complex architectures to directly extract features from images and apply pooling operations to obtain sequence-level representations. Such designs often lead to overfitting on static noise (e.g., clothing), while failing to effectively capture dynamic motion regions, such as the arms and legs. This bottleneck is particularly challenging in the presence of intra-class variation, where gait features of the same individual under different environmental conditions are significantly distant in the feature space. To address the above challenges, we present a Language-guided and Motion-aware gait recognition framework, named LMGait. To the best of our knowledge, LMGait is the first method to introduce natural language descriptions as explicit semantic priors into the gait recognition task. In particular, we utilize designed gait-related language cues to capture key motion features in gait sequences. To improve cross-modal alignment, we propose the Motion Awareness Module (MAM), which refines the language features by adaptively adjusting various levels of semantic information to ensure better alignment with the visual representations. Furthermore, we introduce the Motion Temporal Capture Module (MTCM) to enhance the discriminative capability of gait features and improve the model’s motion tracking ability. We conducted extensive experiments across multiple datasets, and the results demonstrate the significant advantages of our proposed network. Specifically, our model achieved accuracies of 88.5%, 97.1%, and 97.5% on the CCPG, SUSTech1K, and CASIAB* datasets, respectively, achieving state-of-the-art performance.

We introduce the Hybrid Vector-Occupancy Field (HVOF), a new implicit 3D representation for reconstructing both open and closed surfaces from sparse point clouds. Existing approaches, such as occupancy field and signed distance fields, face severe limitations. They struggle with open surfaces, while unsigned distance field and neural vector field exhibit directional instability in complex topologies and ridge regions. HVOF addresses these challenges by incorporating a smoothly decaying occupancy field around the surface, while capturing precise local geometry using truncated displacement vectors, naturally mitigating direction-field ambiguities near ridge regions. This unified design forms a robust hybrid representation that leverages both occupancy and vector fields. To fulfill it, we design a Hybrid Field variational autoencoder including a hierarchical cross-attention encoder and dual-branch decoder that jointly learn occupancy and vector fields through continuous weighting. Extensive experiments demonstrate that HVOF consistently outperforms state-of-the-art methods across ShapeNet, ABC, and MGN datasets, accurately reconstructing both open and closed surfaces while preserving fine geometric details in complex regions.

We present FoundationSLAM, a learning-based monocular dense SLAM system that addresses the absence of geometric consistency in previous flow-based approaches for accurate and robust tracking and mapping. Our core idea is to bridge flow estimation with geometric reasoning by leveraging the guidance from foundation depth models. To this end, we first develop a Hybrid Flow Network that produces geometry-aware correspondences, enabling consistent depth and pose inference across diverse keyframes. To enforce global consistency, we propose a Bi-Consistent Bundle Adjustment Layer that jointly optimizes keyframe pose and depth under multi-view constraints. Furthermore, we introduce a Reliability-Aware Refinement mechanism that dynamically adapts the flow update process by distinguishing between reliable and uncertain regions, forming a closed feedback loop between matching and optimization. Extensive experiments demonstrate that FoundationSLAM achieves superior trajectory accuracy and dense reconstruction quality across multiple challenging datasets, while running in real-time at 18 FPS, demonstrating strong generalization to various scenarios and practical applicability of our method.

Recent advances in diffusion models have enabled the creation of deceptively real images, posing significant security risks when misused. In this study, we empirically show that different timesteps of DDIM inversion reveal varying subtle distinctions between synthetic and real images that are extractable for detection, taking the forms of such as Fourier power spectrum high-frequency discrepancies and inter-pixel variance distributions. Based on these observations, we propose a novel detection method named ESIDE that directly utilizes features of intermediately noised images by training an ensemble on multiple noised timesteps, circumventing the overtime of conventional reconstruction-based strategies. To enhance human comprehension, we introduce a metric-grounded explanation refinement module to identify and explain AI-generated flaws. Additionally, we present the benchmarks GenHard and GenExplain, offering detection samples of greater difficulty and high-quality rationales for fake images. Extensive experiments show that ESIDE achieves state-of-the-art performance with 98.91% and 95.89% detection accuracy on regular and challenging samples respectively, and demonstrates generalizability and robustness.

Mainstream 3D human pose estimation methods directly predict 3D coordinates of joints from 2D keypoints, suffering from severe depth ambiguity. Pose textual descriptions contain abundant semantic information, which facilitates the model to learn the spatial relationship among different body parts, partially alleviating this issue. Leveraging this insight, we propose a 3D human pose estimation method assisted by textual descriptions. Specifically, we utilize an automatic captioning pipeline to generate textual descriptions of 3D poses based on spatial relations among joints. These descriptions include details regarding angles, distances, relative positions, pitch\&roll and ground-contacts. Subsequently, text features are extracted from these descriptions using a language model, while a 3D human pose estimation model extracts pose features. Aligning the pose features with the text features allows for a more targeted optimization of the estimation model. Therefore, we systematically introduce three alignment approaches to effectively align features extracted by two models operating in entirely different domains. Our method incorporates prior knowledge derived from the textual descriptions into the estimation model and can be seamlessly applied to various existing framework. Experimental results on the Human3.6M and MPI-INF-3DHP datasets demonstrate that our method surpasses state-of-the-art methods.

Domain generalization (DG) and domain adaptation (DA) for 3D semantic segmentation enable the model to maintain high performance while avoiding labor-intensive and time-consuming annotation of target-domain data. However, under adverse weather conditions, the injection of spatial noise will affect the reflectivity of LiDAR point clouds, exacerbate domain distribution discrepancies, and degrade the generalization ability of the model. Current methods mainly rely on sparse convolution-based architecture. Due to its limited receptive field, the model captures varying local geometric information when dealing with point clouds of different sparsities, thereby limiting its transferability. To this end, we propose BeyondSparse, a novel cross-domain 3D semantic segmentation method under adverse weather that incorporates a state-space model into a 3D sparse convolution-based architecture, sequentially modeling all features to learn domain-invariant representations. This method consists of two main components: domain feature decoupling and Mamba-based encoder. The former performs feature disentanglement before sequential modeling, while the latter performs global modeling on voxelized point cloud data. In addition, we introduce a token-style augmentation to capture the intrinsic properties of input data. Extensive experimental results demonstrate that our method outperforms SOTA competitors in both DG and DA tasks, for instance, achieving +4.6% and +0.8% mIoU on ``SynLiDAR to SemanticSTF''.

Out-of-distribution (OOD) detection remains a fundamental challenge for deep neural networks, particularly due to overconfident predictions on unseen OOD samples during testing. We reveal a key insight: OOD samples predicted as the same class, or given high probabilities for it, are visually more similar to each other than to the true in-distribution (ID) samples. Motivated by this class-specific observation, we propose DCAC (Dynamic Class-Aware Cache), a training-free, test-time calibration module that maintains separate caches for each ID class to collect high-entropy samples and calibrate the raw predictions of input samples. DCAC leverages cached visual features and predicted probabilities through a lightweight two-layer module to mitigate overconfident predictions on OOD samples. This module can be seamlessly integrated with various existing OOD detection methods across both unimodal and vision-language models while introducing minimal computational overhead. Extensive experiments on multiple OOD benchmarks demonstrate that DCAC significantly enhances existing methods, achieving substantial improvements, i.e., reducing FPR95 by 6.55% when integrated with ASH-S on ImageNet OOD benchmark.

Deep neural networks are increasingly vulnerable to physically deployable backdoor attacks, which manipulate real-world objects to induce targeted model failures. However, current physical backdoor attacks predominantly rely on perpetually visible triggers appended to target objects. These methods inevitably expose attack traces during the deployment phase, risking human suspicion prior to activation. In this paper, we propose a conditionally-visible physical backdoor attack, which can only be activated under specific optical conditions and thereby overcomes the risk of being detected after deployment and before the attack. Specifically, to ensure robust and reliable activation, we design irregular polygonal pattern as triggers to against across environmental variations. Moreover, we introduce a dual-phase mechanism (dormant and activated) to enable stealthy deployment. Our trigger remains invisible and dormant under non-attack conditions, leaving no physical traces. It activates instantaneously under specific illumination, inducing the target model to perform the desired behavior. We conduct experiments on traffic sign recognition tasks to compare our attack with six digital and seven physical attacks, and assess its performance against potential defenses. Extensive experimental results demonstrate the effectiveness, stealthiness, and robustness of our attack.

Recent advances in deep learning have led to significant improvements in nuclei segmentation from histological images, particularly when labels of all classes are available simultaneously during training. However, in clinical practice, real-world scenarios require a model to perform well in an incremental learning setting, where we anticipate the model to achieve satisfactory performance on previously unseen data while effectively mitigating catastrophic forgetting of old classes. Most previous methods alleviate forgetting by distilling old class knowledge through prototypes; however, they fail to adequately capture fine-grained details to address the challenge of high class similarity, which is particularly severe in histological images. To overcome these limitations, we propose a novel incremental learning method for nuclei segmentation (we call it CiNuSeg), which is composed of two key innovative modules. First, we propose a new Anchor-driven Consistency Learning (ACL) module to construct multi-level class anchors within each sample to effectively capture fine structural and textural details of nuclei, thereby significantly mitigating forgetting. Second, we develop a Dual Region Regularization (DRR) module to suppress new class representations within old class regions while enhancing new class representations within new class regions, strengthening the model's ability to discriminate between different nuclei types and improving inter-class separability. We further introduce an Adaptive Temperature Tuning (ATT) strategy to dynamically balance model stability and plasticity. Extensive experiments conducted on benchmarking MoNuSAC and CoNSeP pathological datasets demonstrate the effectiveness of our method, consistently achieving better performance than SOTAs in different settings. Codes will be available upon publication.

Vision-language fine-tuning has emerged as an efficient paradigm for constructing multimodal foundation models. While textual context often highlights semantic relationships within an image, existing fine-tuning methods typically overlook this information when aligning vision and language, thus leading to suboptimal performance. Toward solving this problem, we propose a method that can improve multimodal alignment and fusion based on both semantics and relationships.Specifically, we first extract multilevel semantic features from different vision encoder to capture more visual cues of the relationships. Then, we learn to project the vision features to group related semantics, among which are more likely to have relationships. Finally, we fuse the visual features with the textual by using inheritable cross-attention, where we globally remove the redundant visual relationships by discarding visual-language feature pairs with low correlation. We evaluate our proposed method on eight foundation models and two downstream tasks, visual question answering and image captioning, and show that it outperforms all existing methods.

Task-oriented dexterous grasping remains challenging in robotic manipulations of open-world objects under severe partial observation, where significant missing data invalidates generic shape completion. In this paper, to overcome this limitation, we study \emph{Task-Oriented Shape Completion}, a new task that focuses on completing the potential contact regions rather than the entire shape. We argue that shape completion for grasping should be explicitly guided by the downstream manipulation task. To achieve this, we first generate multiple task-oriented shape completion candidates by leveraging the zero-shot capabilities of object functional understanding from several pre-trained foundation models. A 3D discriminative autoencoder is then proposed to evaluate the plausibility of each generated candidate and optimize the most plausible one from a global perspective. A conditional flow-matching model named FlowGrasp is developed to generate task-oriented dexterous grasps from the optimized shape. Our method achieves state-of-the-art performance in task-oriented dexterous grasping and task-oriented shape completion, improving the Grasp Displacement and the Chamfer Distance over the state-of-the-art by 16.17% and 55.26%, respectively. In particular, it shows good capabilities in grasping objects with severe missing data. It also demonstrates good generality in handling open-set categories and tasks.