IJCAI 2026 论文列表

Diffusion models have achieved remarkable success in video generation; however, the high computational cost of the denoising process remains a major bottleneck. Existing approaches have shown promise in reducing the number of diffusion steps, but they often suffer from significant quality degradation when applied to video generation. We propose Guided Progressive Distillation (GPD), a framework that accelerates the diffusion process for fast and high-quality video generation. GPD introduces a novel training strategy in which a teacher model progressively guides a student model to operate with larger step sizes. The framework consists of two key components: (1) an online-generated training target that reduces optimization difficulty while improving computational efficiency, and (2) frequency-domain constraints in the latent space that promote the preservation of fine-grained details and temporal dynamics. Applied to the Wan2.1 model, GPD reduces the number of sampling steps from 48 to 6 while maintaining competitive visual quality on VBench. Compared with existing distillation methods, GPD demonstrates clear advantages in both pipeline simplicity and quality preservation.

Multi-modal point cloud completion aims to recover complete 3D geometric structures from partial observations by integrating auxiliary data. Although image-guided techniques are well-established, the potential of natural language as a source of high-level semantic cues remains under-explored. Therefore, effectively introducing natural language and synergizing it with the image modality to assist 3D structure recovery is key to breaking current performance bottlenecks. To address this, we propose Query-Aware Gating Gramian Alignment (QGGA), a framework for multi-modal completion. QGGA employs a novel query-dependent bidirectional attention mechanism to dynamically filter information flows between modalities and adaptively regulate fusion intensity. To facilitate alignment among the three heterogeneous modalities, we utilize the volume of the parallelotope spanned by multi-modal local features as a consistency metric, constraining geometric alignment by minimizing this volume. Finally, the fused and aligned features are fed into a coarse-to-fine decoder to first predict a coarse point cloud and then progressively recover the complete shape through stepwise upsampling and refinement. Furthermore, we construct ShapeNet-ViPC-Desc, a benchmark dataset enriched with detailed text descriptions. Experimental results demonstrate that QGGA achieves new state-of-the-art (SOTA) performance on this benchmark. The source code is freely accessible at https://github.com/whysq/QGGA.

Existing degradation-resistant infrared-visible image fusion methods struggle to effectively handle composite degradations, where multiple degradation types exhibit intricate coupling and mutual interference. To address this challenge, we propose SpeciFuse, an infrared-visible image fusion network that learns degradation-specific representations. By explicitly modeling the specificity of individual degradations through a siamese architecture trained on single-degradation data, our method effectively addresses the conflicts arising from multiple degradations, thereby achieving robust fusion performance that generalizes to diverse composite degradation scenarios. In SpeciFuse, a degradation-type specificity decoupling module is designed to disentangle the feature representations by attenuating their redundant correlations in a latent space. It minimizes off-diagonal elements to enforce independence among the encodings of distinct degradation types, while preserving dominant diagonal elements to maintain shared content information. Furthermore, a degradation-aware gated fusion module leverages these decoupled features to dynamically modulate cross-modal fusion across both channel and spatial dimensions. This facilitates a degradation-conditional fusion process that adaptively suppresses degradation artifacts while preserving beneficial information across varying conditions. In extensive comparisons with state-of-the-art methods, SpeciFuse demonstrates more robust performance under diverse composite degradations. The code is available at https://github.com/xbsj-cool/SpeciFuse.

When vision contradicts text, multimodal large language models (MLLMs) consistently favor text—even when images provide clear evidence otherwise. This bias poses risks for applications requiring visual grounding, yet its cause remains unclear. In this paper, we uncover a surprising finding: models often get it right initially, forming correct vision-based predictions in their intermediate layers, before changing their minds and favoring text in the final output. We call this "late-layer textual override". The visual information is encoded, it simply does not survive to the output. More intriguingly, we find that how predictions change reveals whether they're correct: 85% of failures shift toward text, while 89% of successes shift toward vision. This directional signature enables a simple but powerful intervention: when we detect a confident visual prediction being suppressed, we restore it. We propose CALRD (Conflict-Aware Layer Reference Decoding), a training-free method that recovers overridden predictions at inference time. Experiments across five MLLMs of varying architectures demonstrate up to 9.4% absolute improvements on conflict benchmarks while largely preserving standard performance, without training or external knowledge. It recovers what the model already knew but failed to preserve.

Bi-temporal Semantic Change Detection (SCD) is a fundamental analytical framework for capturing state transitions across various domains, with remote sensing being a key application area in this work. However, the prevailing paradigm of comparing two discrete temporal snapshots suffers from inherent limitations due to temporal discontinuity. These limitations lead to pronounced radiometric inconsistencies and pseudo-change noise in Bi-temporal SCD. To address these, we propose the Pseudo-Temporal Feature Fusion Network (PTF-Net), which reconceptualizes bi-temporal SCD as continuous pseudo-temporal evolution modeling. This novel paradigm facilitates the recovery of the missing temporal context between observation points. At its core lies an innovative Non-Linear Style Interpolation Mechanism, which projects discrete bi-temporal observations onto a smooth latent manifold to simulate plausible surface evolution dynamics. The mechanism synthesizes a sequence of semantically coherent intermediate representations, which inherently bridge the temporal gap and disentangles style variations from semantic changes. To fully exploit this synthesized sequence, we design a dedicated Temporal Dynamics Perception Branch. This branch employs efficient temporal interactions to robustly discriminate structural mutations from transient noise, effectively acting as a semantic filter. Extensive experiments on three public datasets reveal that PTF-Net consistently outperforms state-of-the-art methods, achieving comprehensive superiority across mIoU, Fscd, and SeK metrics.

Gait recognition has achieved remarkable success in constrained environments, yet its performance often degrades significantly in cross-domain and cross-vertical-view scenarios. This is primarily due to the fact that domain-specific silhouette geometry causes models to overfit to extrinsic geometric characteristics rather than learning generalizable motion patterns. To address this issue, we present ViSA-Gait, which utilizes Vision Foundation Models (VFMs) as a “Semantic Compass” for stable universal human body knowledge guidance, enabling a lightweight backbone to robustly capture fine-grained gait dynamics. Specifically, we introduce a token-based distillation mechanism where a Spatial Token Learner (STL) and a Temporal Token Learner (TTL) filter dense VFM features into motion-consistent descriptors. These descriptors are then adaptively injected into a lightweight 3D-CNN backbone via a Gated Cross-Attention (GCA) mechanism, functioning as “Semantic Anchors” that regularize the feature space and guide the model to focus on intrinsic body motion. Extensive experiments demonstrate that ViSA-Gait achieves SOTA performance on cross-domain and cross-vertical-view benchmarks while remaining competitive within-domain, offering a new perspective on bridging the gap between geometry-based analysis and universal semantic understanding.

Real-world image generation generally requires multi-turn editing, where users iteratively refine a small region while the majority of the image remains stable across turns. Despite this strong region-level stability, existing diffusion transformer (DiT)–based editing pipelines recompute the entire image at each turn, incurring substantial redundant computation on unchanged regions. Moreover, caching methods in existing DiT acceleration frameworks ignore semantic correspondence across prompts, leading to either unnecessary recomputation or unsafe reuse that degrades editing quality. To this end, we propose RegionCache, a semantic-aware reuse framework for multi-turn image editing that selectively reuses diffusion states from unchanged regions. RegionCache identifies reusable regions by detecting semantic overlap between consecutive prompts and localizing their spatial support via cross-attention maps, and employs an adaptive reuse schedule that dynamically determines how long cached regions can be safely reused based on prompt semantic similarity and contextual consistency. Experiments on PixArt-alpha show that RegionCache achieves 1.43x–2.55x end-to-end inference speedup while maintaining comparable image quality.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel attribute–object compositions by leveraging shared knowledge from seen primitives. While recent works have begun to incorporate test-time adaptation to mitigate performance degradation caused by label-space distribution shifts, they suffer from primitive-level knowledge underutilization and cross-modal semantic asymmetry, hindering robust understanding of novel compositions. To address these issues, we introduce MLDA, a test-time adaptation framework for CZSL that jointly derives multi-level prototypes and performs dynamic vision–language alignment. Specifically, we construct and progressively update composition- and primitive-level prototype sets to accumulate structured semantic knowledge. By augmenting test inputs with diverse visual perspectives and enriched compositional descriptions, MLDA leverages optimal transport to achieve consistent matching of semantically salient information across modalities. Furthermore, composition activation scores are dynamically updated over the test stream to emphasize informative compositions and suppress irrelevant ones. Extensive experiments are conducted to verify the effectiveness of the proposed method. The code is available at https://github.com/keepgoingjkg/MLDA.

Cross-view Referring Expression Comprehension (REC) requires a model to accurately localize objects in a spatial representation of one perspective based on descriptions generated from a different perspective. Essentially, it demands maintaining referential consistency under perspective shifts. This problem is prevalent in engineering application scenarios (e.g., Geographic Information Systems (GIS), urban planning, etc), where descriptions are typically generated from a top-down or planning view, while targets must be identified in a spatial representation distinct from the source. However, existing methods often assume shared perspectives or rely on an explicit observer to interpret observer-centric spatial relations (e.g., "left/right" of the observer). Consequently, they struggle to handle cross-view grounding cases commonly arising in engineering application scenarios, where descriptions are defined by the source-view representation itself and do not involve any observing subject. To this end, we propose a novel Subject-Agnostic Cross-View REC task and construct a controlled simulation environment to systematically study referential consistency under perspective changes. Methodologically, we propose a Reward-Driven Consistency Learning framework. By imposing mutually constraining reward signals on the grounding and localization results between the source and target views, we guide the model to maintain referential consistency across different view representations. Experimental results show that our method significantly outperforms baselines in multiple cross-view settings, exhibiting superior generalization and robustness under perspective shifts.

This paper studies cross-domain/species action recognition from low-resolution videos with sharp appearance variations and domain-specific biases. Its attractive viewpoint is mining domain-invariant representations of cross-domain/species features under rough spatial details to enhance recognition and generalization, overcoming the significant decline of existing methods. To address this, we propose a generalized Action recognition framework for Low-resolution conditions with Domain-invariant Representation learning, named ActLDR, designed to learn domain-invariant representations. First, it decomposes video understanding into spatial and temporal pathways for explicitly separating domain-dependent appearance cues from robust motion dynamics; Second, it constructs a Spatial-Temporal Feature Exchange module to enable cross-branch refinement and suppress domain bias; Third, we inject Gaussian feature interference to simulate feature corruption and enforce prediction-level consistency to encourage stable representations. Empirical results demonstrate that our proposal outperforms previous methods, significantly improving robustness across resolutions, domains, and species, and demonstrating outstanding generalization and transferability.

Estimating object mass from visual input is challenging because mass depends jointly on geometric volume and material-dependent density, neither of which is directly observable from RGB appearance. Consequently, mass prediction from pixels is ill-posed and therefore benefits from physically meaningful representations to constrain the space of plausible solutions. We propose a physically structured framework for single-image mass estimation that addresses this ambiguity by aligning visual cues with the physical factors governing mass. From a single RGB image, we recover object-centric three-dimensional geometry via monocular depth estimation to inform volume and extract coarse material semantics using a vision-language model to guide density-related reasoning. These geometry, semantic, and appearance representations are fused through an instance-adaptive gating mechanism, and two physically guided latent factors (volume- and density-related) are predicted through separate regression heads under mass-only supervision. Experiments on image2mass and ABO-500 show that the proposed method consistently outperforms state-of-the-art methods.

Neural Radiance Field (NeRF) has shown remarkable performance in novel view synthesis but requires numerous multiview images, limiting its practicality in few-shot scenarios. Ray augmentation has been proposed to alleviate overfitting caused by sparse training data by generating additional rays. However, existing methods, which generate augmented rays only near the original rays, exhibit pronounced floaters and appearance distortions due to limited viewpoints and inconsistent rays obstructed by nearby obstacles and complex surfaces. To address these problems, we propose DivCon-NeRF, which introduces novel sphere-based ray augmentations to significantly enhance both diversity and consistency. By employing a virtual sphere centered at the predicted surface point, our method generates diverse augmented rays from all 360-degree directions, facilitated by our consistency mask that effectively filters out inconsistent rays. We introduce tailored loss functions that leverage these augmentations, effectively reducing floaters and visual distortions. Consequently, our method outperforms recent few-shot NeRF approaches on the Blender, LLFF, and DTU datasets. Furthermore, DivCon-NeRF demonstrates strong generalizability by effectively integrating with both regularization- and framework-based few-shot NeRFs.

Vision Transformers (ViTs) exhibit notable susceptibility to adversarial attacks, presenting a significant challenge for their deployment in security-sensitive applications. Despite their considerable empirical successes, a rigorous theoretical foundation for ViT's adversarial generalization behavior has not been adequately established. To address this limitation, we leverage empirical Rademacher complexity to analyze the mechanism of perturbation accumulation through deep ViTs layers. We establish a high-probability generalization bound for ViTs in classification tasks under adversarial settings. Our theoretical framework elucidates the roles of several factors in mitigating perturbation effects, norm regularization of weight matrices (in both MLP and attention modules) and depth-wise propagation constraints on layer-wise norms. Extensive experiments on benchmark datasets corroborate our theoretical insights, bridging the gap between ViTs architecture design and adversarial robustness.

The rapid advancement of AIGC video generation calls for evaluation frameworks that move beyond technical fidelity and incorporate human-centered aesthetic assessment. Existing benchmarks often overlook fine-grained perceptual qualities such as visual aesthetics, artistic style, and human preference. To address this limitation, we introduce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy with two primary dimensions—Aesthetic and Generation—and 52 sub-dimensions. Guided by this taxonomy, we curate 1,016 diverse prompts and collect over 60,000 videos generated by 12 mainstream video generation models. More importantly, VGA-BenchV2 substantially expands human-labeled supervision by adding 36,000 task-level annotations, including 16,200 for aesthetic quality, 13,200 for aesthetic tagging, and 6,600 for generation quality, corresponding to 13.46×, 11.15×, and 1.55× scale-ups over VGA-Bench, respectively. Leveraging this enlarged annotation corpus, we develop a hybrid evaluator architecture consisting of VAQA-Net for continuous aesthetic scoring and two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality assessment. Extensive experiments demonstrate strong alignment with human judgments across diverse generation models. Beyond evaluation, VGA-BenchV2 further introduces an evaluation-to-optimization pipeline, where the learned aesthetic evaluator serves as a reward model for reinforcement learning-based generator fine-tuning. This closes the loop from benchmark construction and human supervision to automated evaluation and model optimization, enabling video generators to improve not only in realism but also in aesthetic quality and human preference alignment. Resources are available at https://huggingface.co/datasets/BestiVictoryLab/VGA-Bench.

Referring Multi-Object Tracking (RMOT) faces a fundamental structural contradiction between the high-discriminability demand and the sparse semantic supervision. This mismatch is particularly acute in highly homogeneous scenarios that require fine-grained discrimination over complex compositional semantics. However, under sparse supervision, models overfit to salient yet insufficient cues, leading to shortcut learning and degraded compositional discrimination. To resolve this, we propose COAL (Counterfactual and Observation-enhanced Alignment Learning), a framework that advances RMOT beyond isolated structural optimization through knowledge regularization. First, we introduce Explicit Semantic Injection (ESI) via a VLM to densify the observation space and enhance instance discriminability. Second, leveraging LLM reasoning, we propose Counterfactual Learning (CFL) to augment supervision, enforcing strict attribute verification for robust compositional recognition. These strategies are unified within a Hierarchical Multi-Stream Integration (HMSI) architecture, which distills external knowledge into domain-specific discriminative representations. Experiments on Refer-KITTI and Refer-KITTI-V2 benchmarks validate COAL's efficacy. Notably, it surpasses the state-of-the-art by 7.28% HOTA on the highly challenging Refer-KITTI-V2. These results demonstrate the effectiveness of knowledge regularization for resolving the sparsity–discriminability paradox in RMOT.

Continual test-time adaptation (CTTA) adapts a pre-trained medical segmentation model online to an unlabeled target stream whose distribution changes over time. However, most existing CTTA methods rely on pseudo-labeling and self-supervised objectives, which inevitably yield noisy supervision under domain shifts. To mitigate this limitation, we introduce off-the-shelf Vision Foundation Models (VFMs) as external knowledge sources. Zero-shot VFMs are insufficient for medical segmentation because they lack medical semantics, yet they contain rich and heterogeneous generic knowledge. To exploit such external knowledge, we propose Reciprocal Distillation with a Frozen Foundation Model (ReDi-FM), a novel framework with two core components: using structure-aware prompts from the source model to guide the frozen VFM to generate target-adapted supervision, and distilling the resulting knowledge back into the adapting model. For robust distillation, we introduce two complementary objectives: uncertainty-driven hard distillation for precise guidance in ambiguous regions, and hard class-balanced soft distillation for richer supervision of under-represented and challenging structures. A consensus-aware gate further stabilizes adaptation when the two teachers disagree. Extensive experiments on multi-domain medical segmentation benchmarks demonstrate that ReDi-FM outperforms state-of-the-art CTTA methods. Code is available at https://github.com/M4cheal/ReDi-FM.

While recent video generative models can synthesize high-fidelity videos, they struggle to portray plausible physical interactions and the resulting state transitions, a critical bottleneck for applications in robotics and VR/AR. To address this, we introduce a framework to generate a scalable synthetic dataset of controllable interactions. Our pipeline leverages a structured taxonomy and state-of-the-art image editing models to create explicit `start' and `end' state images, which serve as visual anchors for the interaction. To generate a seamless video utilizing these anchors, we propose State-Guided Sampling (SGS), a novel sampling technique that mitigates artifacts common in naive conditional generation. Furthermore, we develop and validate a new automated evaluation system that aligns with human judgments to ensure data quality. Experiments show that fine-tuning a base model on our dataset significantly enhances its ability to generate plausible interactions.

Skeleton-based human activity recognition (HAR) has achieved strong empirical performance, yet most existing models remain black boxes and difficult to interpret. In this work, we introduce a neurosymbolic formulation of skeleton-based HAR that reframes action recognition as concept-driven first-order logical reasoning over motion primitives. Our framework bridges representation learning and symbolic inference by grounding first-order logic predicates in learnable spatial and temporal motion concepts. Specifically, we employ a standard spatio-temporal skeleton encoder to extract latent motion representations, which are then mapped to interpretable concept predicates via a spatio-temporal concept decoder that explicitly separates pose-centric and dynamics-centric abstractions. These concept predicates are composed through differentiable first-order logic layers, enabling the model to learn human-readable logical rules that govern action semantics. To impose semantic structure on the learned concepts, we align skeleton representations with LLM-derived descriptions of atomic motion primitives, establishing a shared conceptual space for perception and reasoning. Extensive experiments on NTU RGB+D 60/120 and NW-UCLA demonstrate that our approach achieves competitive recognition performance while providing explicit, interpretable explanations grounded in logical structure. Our results highlight neurosymbolic reasoning as an effective paradigm for interpretable spatio-temporal action understanding.

Multi-object tracking (MOT) has long faced challenges with identity switches, especially for targets with low appearance discriminability and complex motion. Existing end-to-end trackers typically enhance robustness by modeling long-term temporal information across target-level representations, yet these representations remain insufficiently discriminative for spatially and visually similar targets. In this paper, we present MoTRa, a Motion-aware Target Representation learning framework for end-to-end MOT that enriches target-level representations with adaptive motion cues. As a result, each representation adaptively balances motion cues and appearance-related content cues under varying conditions, enhancing its discriminability for tracking. To achieve this, we propose a feature fusion and alignment module that extracts motion cues and adaptively fuses them with content features into target representations. To further regularize the dynamically fused representations, we propose a target-specific contrastive learning strategy that promotes intra-trajectory consistency and inter-target separability. Experimental results demonstrate the effectiveness of our method, achieving competitive performance on multiple benchmarks.