AAAI 2026 论文列表

Developing a multi-modal language model capable of understanding 3D scenes remains challenging due to the limited availability of 3D training data, in contrast to the abundance of 2D datasets used for vision-language models (VLMs). As an alternative, we introduce LLaVA³ (pronounced LLaVA Cube), a novel method that improves the 3D scene understanding capabilities of VLMs using only multi-view 2D images, and without requiring any fine-tuning. Inspired by Cubist painters, who represented multiple viewpoints of a 3D object within a single 2D picture, we propose to describe the 3D scene for the VLM through omnidirectional visual representations of each object. These representations are derived from an intermediate multi-view 3D reconstruction of the scene. Extensive experiments on 3D visual question answering and 3D language grounding show that our approach significantly outperforms previous 2D-based VLM solutions.

Video Anomaly Understanding (VAU) is a novel task focused on describing unusual occurrences in videos. Despite growing interest, the evaluation of VAU remains an open challenge. Existing benchmarks rely on n-gram-based metrics (e.g., BLEU, ROUGE-L) or LLM-based evaluation. The first fails to capture the rich, free-form, and visually grounded nature of LVLM responses, while the latter focuses on assessing language quality over factual relevance, often resulting in subjective judgments that are misaligned with human perception. In this work, we address this issue by proposing FineVAU, a new benchmark for VAU that shifts the focus towards rich, fine-grained and domain-specific understanding of anomalous videos. We formulate VAU as a three-fold problem, with the goal of comprehensively understanding key descriptive elements of anomalies in video: events (What), participating entities (Who) and location (Where). Our benchmark introduces a) FV-Score, a novel, human-aligned evaluation metric that assesses the presence of critical visual elements in LVLM answers, providing interpretable, fine-grained feedback; and b) FineW³, a novel, comprehensive dataset curated through a structured and fully automatic procedure that augments existing human annotations with high quality, fine-grained visual information. Human evaluation reveals that our proposed metric has a superior alignment with human perception of anomalies in comparison to current approaches. Detailed experiments on FineVAU unveil critical limitations in LVLM's ability to perceive anomalous events that require spatial and fine-grained temporal understanding, despite strong performance on coarse grain, static information, and events with strong visual cues.

Open-vocabulary semantic segmentation aims to assign pixel-level labels to images based on textual descriptions, even for categories beyond predefined closed sets. While vision-language foundation models like CLIP are widely used for this task, fine-tuning them for pixel-level predictions often compromises their generalization capabilities. To address this, we propose a novel fine-tuning strategy, CP-CLIP, which generates customized parameters for CLIP without sacrificing its generalization. Our method employs a customized parameter generator that produces newly added parameters based on random noise, using local visual features from CLIP's image encoder as conditions, enabling generalization to new images from unseen scenarios. Additionally, we introduce an orthogonal adaptation technique to ensure the update direction is orthogonal to the pre-trained weights, largely preserving the initial generalization ability. Extensive experiments demonstrate that CP-CLIP achieves state-of-the-art performance across multiple benchmarks in open-vocabulary semantic segmentation.

Personalized text-to-image generation aims to synthesize novel images of a specific subject or style using only a few reference images. Recent methods based on Low-Rank Adaptation (LoRA) enable efficient single-concept customization by injecting lightweight, concept-specific adapters into pre-trained diffusion models. However, combining multiple LoRA modules for multi-concept generation often leads to identity missing and visual feature leakage. In this work, we identify two key issues behind these failures: (1) token-wise interference among different LoRA modules, and (2) spatial misalignment between the attention map of a rare token and its corresponding concept-specific region. To address these issues, we propose Token-Aware LoRA (TARA), which introduces a token mask to explicitly constrain each module to focus on its associated rare token to avoid interference, and a training objective that encourages the spatial attention of a rare token to align with its concept region. Our method enables training-free multi-concept composition by directly injecting multiple independently trained TARA modules at inference time. Experimental results demonstrate that TARA enables efficient multi-concept inference and effectively preserving the visual identity of each concept by avoiding mutual interference between LoRA modules.

3D open-vocabulary scene understanding, which accurately perceives complex semantic properties of objects in space, has gained significant attention in recent years. In this paper, we propose GAGS, a framework that distills 2D CLIP features into 3D Gaussian splatting, enabling open-vocabulary queries for renderings on arbitrary viewpoints. The main challenge of distilling 2D features for 3D fields lies in the multiview inconsistency of extracted 2D features, which provides unstable supervision for the 3D feature field. GAGS addresses this challenge with two novel strategies. First, GAGS associates the prompt point density of SAM with the camera distances to scene objects, which significantly improves the multiview consistency of segmentation results. Second, GAGS further decodes a granularity factor to guide the distillation process and this granularity factor can be learned in a unsupervised manner to only select the multiview consistent 2D features in the distillation process. Experimental results on two datasets show that GAGS improves visual grounding accuracy by an average of 10.9% and semantic segmentation accuracy by an average of 7.0%, with an inference speed 2× faster than baseline methods.

Vision Transformers (ViTs) have achieved strong performance in video action recognition, but their high computational cost limits their practicality. Lightweight CNNs are more efficient but suffer from accuracy gaps. Cross-Architecture Knowledge Distillation (CAKD) addresses this by transferring knowledge from ViTs to CNNs, yet existing methods often struggle with architectural mismatch and overlook the value of stronger homogeneous CNN teachers. To tackle these challenges, we propose a Dual-Teacher Knowledge Distillation framework that leverages both a heterogeneous ViT teacher and a homogeneous CNN teacher to collaboratively guide a lightweight CNN student. We introduce two key components: (1) Discrepancy-Aware Teacher Weighting, which dynamically fuses the predictions from ViT and CNN teachers by assigning adaptive weights based on teacher confidence and prediction discrepancy with the student, enabling more informative and effective supervision; and (2) a Structure Discrepancy-Aware Distillation strategy, where the student learns the residual features between ViT and CNN teachers via a lightweight auxiliary branch, focusing on transferable architectural differences without mimicking all of ViT’s high-dimensional patterns. Extensive experiments on benchmarks including HMDB51, EPIC-KITCHENS-100, and Kinetics-400, demonstrate that our method consistently outperforms state-of-the-art distillation approaches, achieving notable performance improvements with a maximum accuracy gain of 5.95% on HMDB51.

3D Human Pose Estimation (3D HPE) is vital in various applications, from person re-identification and action recognition to virtual reality. However, the reliance on annotated 3D data collected in controlled environments poses challenges for generalization to diverse in-the-wild scenarios. Existing domain adaptation (DA) paradigms like general DA and source-free DA for 3D HPE overlook the issues of non-stationary target pose datasets. To address these challenges, we propose a novel task named lifelong domain adaptive 3D HPE. To our knowledge, we are the first to introduce the lifelong domain adaptation to the 3D HPE task. In this lifelong DA setting, the pose estimator is pretrained on the source domain and subsequently adapted to distinct target domains. Moreover, during adaptation to the current target domain, the pose estimator cannot access the source and all the previous target domains. The lifelong DA for 3D HPE involves overcoming challenges in adapting to current domain poses and preserving knowledge from previous domains, particularly combating catastrophic forgetting. We present an innovative Generative Adversarial Network (GAN) framework, which incorporates 3D pose generators, a 2D pose discriminator, and a 3D pose estimator. This framework effectively mitigates domain shifts and aligns original and augmented poses. Moreover, we construct a novel 3D pose generator paradigm, integrating pose-aware, temporal-aware, and domain-aware knowledge to enhance the current domain's adaptation and alleviate catastrophic forgetting on previous domains. Our method demonstrates superior performance through extensive experiments on diverse domain adaptive 3D HPE datasets.

During the video encoding process, the original spatial domain signal is first transformed into the frequency domain, followed by quantization and compression. As a result, the quality degradation in compressed videos primarily stems from distortions in the frequency domain information. However, existing video enhancement methods typically directly fuse information from adjacent frames in the spatial domain, making it difficult for models to effectively compensate for frequency domain distortions, which leads to suboptimal detail restoration. To address this issue, we propose a Hierarchical Frequency-Guided Alignment Transformer. Additionally, by analyzing the characteristics of the frequency domain, we find that different frequency bands exhibit both correlations and a certain degree of independence. Based on this, we introduce a Frequency-Aware Transformer module that employs a combination of independent and mixed processing to optimize information exchange across different frequency domains, effectively mitigating cross-interference from irrelevant information. Experimental results demonstrate that, compared to existing methods, our approach achieves state-of-the-art performance in objective metrics (PSNR/SSIM), perceptual quality (LPIPS), and subjective visual effects, while reducing model complexity.

Text-to-Video (T2V) generation has advanced greatly, yet maintaining consistency remains challenging, especially for tuning-free long video generation. We attribute the consistency problem to cumulative deviations for long video generation at three levels: the random noise lacking correlation results initial deviation between frames; discrepancy in semantic feature tokens between denoising network blocks gradually accumulates as the frame count grows, leading to greater deviations; attention mechanisms struggle to capture global relationships across distant frames in long videos. To address these, we propose FreeMem, a tuning-free framework leveraging hierarchical memory update and injection: the noise memory stabilizes consistency by manipulating low and high frequency components in the initial noise space; the token memory combats inconsistency through adaptive fusion of historical and current semantic feature tokens between denoising network blocks; and the attention memory establishes persistent cache to model long-range relationships within self attention layers. Evaluated on VBench, FreeMem improves subject and background consistency matrics across various methods, offering a practical solution for low-cost, high-consistency long video generation.

In recent years, neural image compression methods have achieved impressive performance in image compression tasks, most of which are based on variational auto-encoder with hyper-prior and autoregressive Gaussian entropy model. We first demonstrate that the way these end-to-end approaches handle quantization during training leads to a mismatch between the gradients direction of entropy model parameters (i.e., mean and standard deviation) and the direction they should be optimized towards during inference, making neural network difficult to learn accurate estimates of entropy model parameters. To address this issue, we then propose a two-step improvement: in the first step, use straight-through estimator to align the forward propagation during training with inference, thereby correcting the gradients of standard deviation parameters; in the second step, utilize gradients transfer that we propose and MSE-guided gradients to manually compensate for the gradients of mean parameters lost due to straight-through estimator. Finally, we also propose to freeze the auto-encoder and hyper auto-encoder in pre-trained models provided by existing works, and fine-tune only the modules that predict the entropy model parameters, enabling efficient validation of proposed improvements. Experimental results show that our improvements bring appreciable performance gains to state-of-the-art neural image compression models in recent years. Meanwhile, our improvements require no modification to the structure of pre-trained models and only lightweight fine-tuning, which shows strong plug-and-play capability and practical utility.

Recent advances in vision–language models (VLMs) have shed light on human-level embodied intelligence. However, existing benchmarks for VLM-driven embodied agents still rely on high-level commands or discretised action spaces—``non-native'' settings that diverge markedly from the real world. Moreover, current benchmarks focus exclusively on high-level tasks, while lacking joint evaluation and analysis on both low- and high-level. To bridge these gaps, we present \textbf{NativeEmbodied}, a challenging benchmark for VLM-driven embodied agents that adopts a unified, native low-level action space. Built upon diverse simulated scenes, NativeEmbodied first designs three representative high-level tasks in complex scenarios to evaluate overall performance. For more detailed and comprehensive performance analysis, we further decouple the entangled skills behind complex tasks and construct four types of low-level tasks, each corresponding to a key fundamental embodied skill. This joint evaluation across task and skill granularities enables a fine-grained assessment of embodied agent. Comprehensive experiments on the best VLMs reveal pronounced deficiencies in certain fundamental embodied skills. Further analysis shows that these bottlenecks severely constrain performance on high-level tasks. Our NativeEmbodied not only pinpoints the key challenges faced by current VLM-driven embodied agents, but also provides valuable insight for future development of this field.

Few-shot detection-based counters estimate the number of instances in the image specified only by a few test-time exemplars. A common approach to localize objects across multiple sizes is to merge backbone features of different resolutions. Furthermore, to enable small object detection in densely populated regions, the input image is commonly upsampled and tiling is applied to cope with the increased computational and memory requirements. Because of these ad-hoc solutions, existing counters struggle with images containing diverse-sized objects and densely populated regions of small objects. We propose GeCo2, an end-to-end few-shot counting and detection method that explicitly addresses the object scale issues. A new dense query representation gradually aggregates exemplar-specific feature information across scales that leads to high-resolution dense queries that enable detection of large as well as small objects. GeCo2 surpasses state-of-the-art few-shot counters in counting as well as detection accuracy by ~10% while running ~3x faster at smaller GPU memory footprint.

The increasing demand for AR/VR applications has highlighted the need for high-quality content, such as 360° live wallpapers. However, generating high-quality 360° panoramic contents remains a challenging task due to the severe distortions introduced by equirectangular projection (ERP). Existing approaches either fine-tune pretrained diffusion models on limited ERP datasets or adopt tuning-free methods that still rely on ERP latent representations, often resulting in distracting distortions near the poles. In this paper, we introduce SphereDiff, a novel approach for synthesizing 360° static and live wallpaper with state-of-the-art diffusion models without additional tuning. We define a spherical latent representation that ensures consistent quality across all perspectives, including near the poles. Then, we extend MultiDiffusion to spherical latent representation and propose a dynamic spherical latent sampling method to enable direct use of pretrained diffusion models. Moreover, we introduce distortion-aware weighted averaging to further improve the generation quality. Our method outperforms existing approaches in generating 360° static and live wallpaper, making it a robust solution for immersive AR/VR applications.

Change captioning aims to describe changes between a pair of images. However, existing works rely on visual features alone, which often fail to capture subtle but meaningful changes because they lack the ability to represent explicitly structured information such as object relationships and compositional semantics. To alleviate this, we present CORTEX (COmpositional Reasoning-aware TEXt-guided), a novel framework that integrates complementary textual cues to enhance change understanding. In addition to capturing cues from pixel-level differences, CORTEX utilizes scene-level textual knowledge provided by Vision Language Models (VLMs) to extract richer image text signals that reveal underlying compositional reasoning. CORTEX consists of three key modules: (i) an Image-level Change Detector that identifies low-level visual differences between paired images, (ii) a Reasoning-aware Text Extraction (RTE) module that use VLMs to generate compositional reasoning descriptions implicit in visual features, and (iii) an Image-Text Dual Alignment (ITDA) module that aligns visual and textual features for fine-grained relational reasoning. This enables CORTEX to reason over visual and textual features and capture changes that are otherwise ambiguous in visual features alone.

Recent studies have revealed Neural Collapse (NC) in deep classifiers, where last-layer weights and features align into an equiangular tight frame (ETF), concentrating class information along specific embedding directions. However, conventional fine-tuning typically disregards this structure, initializing task-specific classifier heads randomly. To explicitly leverage this phenomenon, we propose a simple yet effective method for metric learning: (1) initializing the classifier head along each class’s NC direction from a pretrained model to preserve the emergent structure, and (2) injecting small isotropic Gaussian noise during finetuning to boost generalization. In addition, we provide a theoretical bound proving that our method explicitly reduces cumulative weight drift from the NC-initialization, compared to standard finetuning. This suggests that our method better preserves the pretrained model’s class-specific structure. Empirically, this structural preservation yields Recall@K gains: reduced weight drift correlates with better performance. Concurrent decreases in the Neural Collapse 1 (NC1) measure confirm that stronger intra‐class cohesion underlies these improvements. Furthermore, we validate the effectiveness of our method on class‐imbalanced benchmarks.

We present Infinite-Story, a training-free framework for consistent text-to-image (T2I) generation tailored for multi-prompt storytelling scenarios. Built upon a scale-wise autoregressive model, our method addresses two key challenges in consistent T2I generation: identity inconsistency and style inconsistency. To overcome these issues, we introduce three complementary techniques: Identity Prompt Replacement, which mitigates context bias in text encoders to align identity attributes across prompts; and a unified attention guidance mechanism comprising Adaptive Style Injection and Synchronized Guidance Adaptation, which jointly enforce global style and identity appearance consistency while preserving prompt fidelity. Unlike prior diffusion-based approaches that require fine-tuning or suffer from slow inference, Infinite-Story operates entirely at test time, delivering high identity and style consistency across diverse prompts. Extensive experiments demonstrate that our method achieves state-of-the-art generation performance, while offering over 6x faster inference (1.72 seconds per image) than the existing fastest consistent T2I models, highlighting its effectiveness and practicality for real-world visual storytelling.

We propose unsupervised multi-scenario (UMS) person re-identification (ReID) as a new task that expands ReID across diverse scenarios (cross-resolution, clothing change, etc.) within a single coherent framework. To tackle UMS-ReID, we introduce image-text knowledge modeling (ITKM) -- a three-stage framework that effectively exploits the representational power of vision-language models. We start with a pre-trained CLIP model with an image encoder and a text encoder. In Stage I, we introduce a scenario embedding in the image encoder and fine-tune the encoder to adaptively leverage knowledge from multiple scenarios. In Stage II, we optimize a set of learned text embeddings to associate with pseudo-labels from Stage I and introduce a multi-scenario separation loss to increase the divergence between inter-scenario text representations. In Stage III, we first introduce cluster-level and instance-level heterogeneous matching modules to obtain reliable heterogeneous positive pairs (e.g., a visible image and an infrared image of the same person) within each scenario. Next, we propose a dynamic text representation update strategy to maintain consistency between text and image supervision signals. Experimental results across multiple scenarios demonstrate the superiority and generalizability of ITKM; it not only outperforms existing scenario-specific methods but also enhances overall performance by integrating knowledge from multiple scenarios.

Autoregressive models, built based on the Next Token Prediction (NTP) paradigm, show great potential in developing a unified framework that integrates both language and vision tasks. Pioneering works introduce NTP to autoregressive visual generation tasks. In this work, we rethink the NTP for autoregressive image generation and extend it to a novel Next Patch Prediction (NPP) paradigm. Our key idea is to group and aggregate image tokens into patch tokens with higher information density. By using patch tokens as a more compact input sequence, the autoregressive model is trained to predict the next patch, significantly reducing computational costs. To further exploit the natural hierarchical structure of image data, we propose a multi-scale coarse-to-fine patch grouping strategy. With this strategy, the training process begins with a large patch size and ends with vanilla NTP where the patch size is 1x1, thus maintaining the original inference process without modifications. Extensive experiments across a diverse range of model sizes demonstrate that NPP could reduce the training cost to around 0.6 times while improving image generation quality by up to 1.0 FID score on the ImageNet 256x256 generation benchmark. Notably, our method retains the original autoregressive model architecture without introducing additional trainable parameters or specifically designing a custom image tokenizer, offering a flexible and plug-and-play solution for enhancing autoregressive visual generation.

Contrastive Language-Image Pretraining (CLIP) has demonstrated impressive generalization on vision-language tasks by aligning images and short texts. However, its inherent 77-token length limits the capacity of capturing complex semantics in long captions. Existing long-text adaptations for CLIP typically rely on either multi-stage training or truncation-based alignment, both inevitably resulting in semantic degradation and cumbersome tuning. Therefore, we propose OneLIP, a unified framework that extends CLIP to understand long captions within a single training stage, eliminating the need for brittle truncation or multi-stage pipelines. OneLIP addresses semantic degradation by introducing two key innovations: (1) Token Refinement and Importance-guided Modeling (TRIM) module, which selects and refines informative tokens via SVD-based contribution scoring and cross-modal relevance modeling; (2) Per-sample Online Hard Negative Mining (PO-HNM) strategy dynamically maintains sample-specific negatives based on dual-consistency difficulty tracking, which is superior in long-text scenarios where key semantics are distributed in scattered positions. Extensive experiments on long-text image retrieval, short-text image retrieval, zero-shot classification, and text-to-image generation demonstrate OneLIP's robustness and versatility across diverse input lengths, offering a faithful solution for long-text representation learning of CLIP.