AAAI 2026 Papers

Bokeh rendering simulates the shallow depth-of-field effect in photography, enhancing visual aesthetics and guiding viewer attention to regions of interest. Although recent approaches perform well, rendering controllable bokeh without additional depth inputs remains a significant challenge. Existing classical and neural controllable methods rely on accurate depth maps, while generative approaches often struggle with limited controllability and efficiency. In this paper, we propose BokehFlow, a depth-free framework for controllable bokeh rendering based on flow matching. BokehFlow directly synthesizes photorealistic bokeh effects from all-in-focus images, eliminating the need for depth inputs. It employs a cross-attention mechanism to enable semantic control over both focus regions and blur intensity via text prompts. To support training and evaluation, we collect and synthesize four datasets. Extensive experiments demonstrate that BokehFlow achieves visually compelling bokeh effects and offers precise control, outperforming existing depth-dependent and generative methods in both rendering quality and efficiency.

Three-dimensional scene reconstruction from sparse-view satellite images is a long-standing and challenging task. While 3D Gaussian Splatting (3DGS) and its variants have recently attracted attention for its high efficiency, existing methods remain unsuitable for satellite images due to incompatibility with rational polynomial coefficient (RPC) models and limited generalization capability. Recent advances in generalizable 3DGS approaches show potential, but they perform poorly on multi-temporal sparse satellite images due to limited geometric constraints, transient objects, and radiometric inconsistencies. To address these limitations, we propose SkySplat, a novel self-supervised framework that integrates the RPC model into the generalizable 3DGS pipeline, enabling more effective use of sparse geometric cues for improved reconstruction. SkySplat relies only on RGB images and radiometric-robust relative height supervision, thereby eliminating the need for ground-truth height maps. Key components include a Cross-Self Consistency Module (CSCM), which mitigates transient object interference via consistency-based masking, and a multi-view consistency aggregation strategy that refines reconstruction results. Compared to per-scene optimization methods, SkySplat achieves an 86 times speedup over EOGS with higher accuracy. It also outperforms generalizable 3DGS baselines, reducing MAE from 13.18 m to 1.80 m on the DFC19 dataset significantly, and demonstrates strong cross-dataset generalization on the MVS3D benchmark.

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for 3D scenes, widely adopted due to its exceptional efficiency and high-fidelity visual quality. Given the significant value of 3DGS assets, recent works have introduced specialized watermarking schemes to ensure copyright protection and ownership verification. However, can existing 3D Gaussian watermarking approaches genuinely guarantee robust protection of the 3D assets? In this paper, for the first time, we systematically explore and validate possible vulnerabilities of 3DGS watermarking frameworks. We demonstrate that conventional watermark removal techniques designed for 2D images do not effectively generalize to the 3DGS scenario due to the specialized rendering pipeline and unique attributes of each gaussian primitives. Motivated by this insight, we propose GSPure, the first watermark purification framework specifically for 3DGS watermarking representations. By analyzing view-dependent rendering contributions and exploiting geometrically accurate feature clustering, GSPure precisely isolates and effectively removes watermark-related Gaussian primitives while preserving scene integrity. Extensive experiments demonstrate that our GSPure achieves the best watermark purification performance, reducing watermark PSNR by up to 16.34dB while minimizing degradation to original scene fidelity with less than 1dB PSNR loss. Moreover, it consistently outperforms existing methods in both effectiveness and generalization.

Wide-angle videos in few-shot action recognition (FSAR) effectively express actions within specific scenarios. However, without a global understanding of both subjects and background, recognizing actions in such samples remains challenging because of the background distractions. Receptance Weighted Key Value (RWKV), which learns interaction between various dimensions, shows promise for global modeling. While directly applying RWKV to wide-angle FSAR may fail to highlight subjects due to excessive background information. Additionally, temporal relation degraded by frames with similar backgrounds is difficult to reconstruct, further impacting performance. Therefore, we design the CompOund SegmenTation and Temporal REconstructing RWKV (Otter). Specifically, the Compound Segmentation Module (CSM) is devised to segment and emphasize key patches in each frame, effectively highlighting subjects against background information. The Temporal Reconstruction Module (TRM) is incorporated into the temporal-enhanced prototype construction to enable bidirectional scanning, allowing better reconstruct temporal relation. Furthermore, a regular prototype is combined with the temporal-enhanced prototype to simultaneously enhance subject emphasis and temporal modeling, improving wide-angle FSAR performance. Extensive experiments on benchmarks such as SSv2, Kinetics, UCF101, and HMDB51 demonstrate that Otter achieves state-of-the-art performance. Extra evaluation on the VideoBadminton dataset further validates the superiority of Otter in wide-angle FSAR.

CLIP is a seminal multimodal model that maps images and text into a shared representation space by contrastive learning on billions of image–caption pairs. Inspired by the rapid progress of large language models (LLMs), we investigate how the superior linguistic understanding and broad world knowledge of LLMs can further strengthen CLIP—particularly in handling long, complex captions. We introduce an efficient fine-tuning framework that embeds an LLM into a pretrained CLIP while incurring almost the same training cost as regular CLIP fine-tuning. Our method first “embedding-izes” the LLM for the CLIP setting, then couples it to the pretrained CLIP vision encoder through a lightweight adaptor trained on only a few million image–caption pairs. With this strategy we achieve large performance gains—without large-scale retraining—over state-of-the-art CLIP variants such as EVA02 and SigLIP-2. The LLM-enhanced CLIP delivers consistent improvements across a wide spectrum of downstream tasks, including linear-probe classification, zero-shot image–text retrieval with both short and long captions (in English and other languages), zero-shot/supervised image segmentation, object detection, and used as tokenizer for multimodal large-model benchmarks.

Recent generative unlearning models synthesize high quality samples while protecting private information by unlearning the identity. However, existing generative identity unlearning methods face two challenges in multi-identity unlearning: 1) identity conflicts, which cause conflicts of model parameters in the continuous erasure of multiple identities; 2) fragile unlearning, where the model's unlearning ability deteriorates or fails under malicious attacks. In this paper, we introduce a critical yet under-explored task called robust multi-identity unlearning, with the goals of resolving identity conflicts to achieve interference-free unlearning and protecting against malicious attacks to achieve robust unlearning. To satisfy these goals, we propose a novel framework, RObust generatiVE continual identity unlearning against Relearning attacks (ROVER). By filtering unlearning requests with latent similarity, our method effectively isolates benign unlearning from malicious attacks to preserve identity removal integrity. Meanwhile, residual orthogonal resonator resolves identity conflicts in the continuous erasure of multiple identities, preserving stability in benign continual unlearning. Moreover, we introduce the phantom guard network to block malicious attacks by absorbing adversarial gradients, ensuring irreversible identity unlearning. The extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on the task of robust multi-identity unlearning against relearning attacks.

With the development of diffusion models, enhancing spatial controllability in text-to-image generation has become a vital challenge. As a representative task for addressing this challenge, layout-to-image generation aims to generate images that are spatially consistent with the given layout condition. Existing layout-to-image methods typically introduce the layout condition by integrating adapter modules into the base generative model. However, the generated images often exhibit low visual quality and stylistic inconsistency with the base model, indicating a loss of pretrained knowledge. To alleviate this issue, we construct the Layout Synthesis (LaySyn) dataset, which leverages images synthesized by the base model itself to mitigate the distribution shift from the pretraining data. Moreover, we propose the Layout Control (Laytrol) Network, in which parameters are inherited from MM-DiT to preserve the pretrained knowledge of the base model. To effectively activate the copied parameters and avoid disturbance from unstable control conditions, we adopt a dedicated initialization scheme for Laytrol. In this scheme, the layout encoder is initialized as a pure text encoder to ensure that its output tokens remain within the data domain of MM-DiT. Meanwhile, the outputs of the layout control network are initialized to zero. In addition, we apply Object-level Rotary Position Embedding to the layout tokens to provide coarse positional information. Qualitative and quantitative experiments demonstrate the effectiveness of our method.

Cloth-changing person re-identification (CC-ReID) aims to identify individuals across non-overlapping cameras despite clothing variations. Existing methods are often constrained by two primary limitations: approaches using auxiliary modalities typically rely on a single specific cue, limiting their robustness, while feature disentanglement methods struggle with discrete labels that create inconsistencies between ground truth labels and modality semantic similarity. To overcome these limitations, we propose DRDnet, a unified framework that synergistically integrates dual auxiliary cues and advanced relation modeling. Specifically, our Dual-Stream Disentanglement (DSD) module leverages textual descriptions and parsing images to decouple clothing factors through high-level semantic supervision and pixel-level operations, yielding robust clothing-agnostic features. Simultaneously, our Modal Relation Modeling (MRM) module constructs feature memory banks and employs adaptive soft label smoothing, effectively enhancing image-text semantic alignment and reinforcing identity consistency across clothing changes. We evaluate DRDnet on several CC-ReID benchmarks to demonstrate its effectiveness and provide state-of-the-art performance across all benchmarks.

Snapshot compressive imaging (SCI) captures multispectral images (MSIs) using a single coded two-dimensional (2-D) measurement, but reconstructing high-fidelity MSIs from these compressed inputs remains a fundamentally ill-posed challenge. Recent diffusion-based methods improve quality but are limited by scarce MSI training data, domain shifts from RGB-pretrained models, and slow multi-step sampling. These drawbacks restrict their practicality in real-world applications. Unlike prior approaches that rely on expensive iterative refinement or subspace-based diffusion embeddings (e.g., DiffSCI, PSR-SCI)—we introduce a fundamentally different paradigm: a self-supervised One-Step Diffusion (OSD) framework designed specifically for SCI. The key novelty lies in using a single-step diffusion refiner to correct an initial reconstruction, eliminating iterative denoising entirely while preserving generative quality. Moreover, we adopt a self-supervised equivariant learning strategy to train both the predictor and refiner directly from raw 2-D measurements, enabling generalization to unseen domains without ground-truth MSI. To further address limited MSI data, we design a band-selection–driven distillation strategy that transfers core generative priors from large-scale RGB datasets, effectively bridging the domain gap. Extensive experiments confirm that our approach sets a new standard—yielding PSNR gains of 3.44dB, 1.61dB, and 0.28dB on the Harvard, NTIRE, and ICVL datasets respectively, while cutting reconstruction time from 8.9s to just 0.22s per image. These gains in efficiency and adaptability advance SCI reconstruction, enabling accurate and practical real-world deployment.

As digital twins become central to the transformation of modern cities, accurate and structured 3D building models emerge as a key enabler of high-fidelity, updatable urban representations. These models underpin diverse applications including energy modeling, urban planning, autonomous navigation, and real-time reasoning. Despite recent advances in 3D urban modeling, most learning-based models are trained on building datasets with limited architectural diversity, which significantly undermines their generalizability across heterogeneous urban environments. To address this limitation, we present BuildingWorld, a comprehensive and structured 3D building dataset designed to bridge the gap in stylistic diversity. It encompasses buildings from geographically and architecturally diverse regions—including North America, Europe, Asia, Africa, and Oceania—offering a globally representative dataset for urban-scale foundation modeling and analysis. Specifically, BuildingWorld provides about Five million LOD2 building models collected from diverse sources, accompanied by both real and simulated airborne LiDAR point clouds. This enables comprehensive research on 3D reconstruction, building detection and segmentation, as well as roof structure segmentation. Cyber City, a virtual city model, is introduced to enable the generation of unlimited training data with customized and structurally diverse point cloud distributions. Furthermore, we provide standardized evaluation metrics tailored for building reconstruction, aiming to facilitate the training, evaluation, and comparison of large-scale vision models and foundation models in structured 3D urban environments

The spreading of AI-generated images (AIGI), driven by advances in generative AI, poses a significant threat to in- formation security and public trust. Existing AIGI detectors, while effective against images in clean laboratory settings, fail to generalize to in-the-wild scenarios. These real-world images are noisy, varying from “obviously fake” images to realistic ones derived from multiple generative models and further edited for quality control. We address in-the-wild AIGI detection in this paper. We introduce MIRAGE, a challenging benchmark designed to emulate the complexity of in-the-wild AIGI. MIRAGE is constructed from two sources: (1) a large corpus of Internet-sourced AIGI verified by human experts, and (2) a synthesized dataset created through the collaboration between multiple expert generators, closely simulating the realistic AIGI in the wild. Building on this benchmark, we propose MIRAGE-R1, a vision- language model with heuristic-to-analytic reasoning, a reflective reasoning mechanism for AIGI detection. MIRAGE-R1 is trained in two stages: a supervised-fine-tuning cold start, followed by a reinforcement learning stage. By further adopting a inference-time adaptive thinking strategy, MIRAGE-R1 is able to provide either a quick judgment or a more robust and accurate conclusion, effectively balancing inference speed and performance. Extensive experiments show that our model leads state-of-the-art detectors by 5% and 10% on MIRAGE and public benchmark, respectively.

Video-based multimodal large language models (V-MLLMs) have shown vulnerability to adversarial examples in video-text multimodal tasks. However, the transferability of adversarial videos to unseen models—a common and practical real-world scenario—remains unexplored. In this paper, we pioneer an investigation into the transferability of adversarial video samples across V-MLLMs. We find that existing adversarial attack methods face significant limitations when applied in black-box settings for V-MLLMs, which we attribute to the following shortcomings: (1) lacking generalization in perturbing video features, (2) focusing only on sparse key-frames, and (3) failing to integrate multimodal information. To address these limitations and deepen the understanding of V-MLLM vulnerabilities in black-box scenarios, we introduce the Image-to-Video MLLM (I2V-MLLM) attack. In I2V-MLLM, we utilize an image-based multimodal large language model (I-MLLM) as a surrogate model to craft adversarial video samples. Multimodal interactions and spatiotemporal information are integrated to disrupt video representations within the latent space, improving adversarial transferability. Additionally, a perturbation propagation technique is introduced to handle different unknown frame sampling strategies. Experimental results demonstrate that our method can generate adversarial examples that exhibit strong transferability across different V-MLLMs on multiple video-text multimodal tasks. Compared to white-box attacks on these models, our black-box attacks (using BLIP-2 as a surrogate model) achieve competitive performance, with average attack success rate (AASR) of 57.98% on MSVD-QA and 58.26% on MSRVTT-QA for Zero-Shot VideoQA tasks, respectively.

Recent advances in Referring Expression Comprehension (REC) have been largely driven by supervised learning on curated datasets, where each expression is assumed to refer to exactly one known object. However, such assumptions rarely hold in real-world scenarios, where expressions can refer to multiple objects, fail to refer to any, or involve novel categories and complex semantics. These challenges define the task of open-world REC, which demands robust generalization and structured reasoning beyond the scope of traditional REC methods. In this work, we introduce a novel, training-free framework that decouples visual perception from linguistic reasoning to address open-world REC. Our method first transforms the visual scene into a rich textual representation using an open-vocabulary multimodal perception module. It then employs a reasoning language model to interpret the referring expression and perform explicit logical inference over the perceived scene, enabling transparent decision-making and strong generalization in open-world scenarios. Experiments on three standard REC benchmarks as well as two more challenging ones, gRefCOCO and D³, demonstrate that our framework achieves highly competitive zero-shot performance, often surpassing supervised baselines.

Text-to-video generation models have shown significant progress in recent years. However, they still struggle with compositional text prompts, such as attribute binding for multiple objects, temporal dynamics associated with differ- ent objects, and interactions between objects. Inspired by ef- fective human creative workflow, we propose GENMAC, a multi-agent collaboration framework that enables composi- tional text-to-video generation. The framework incorporates a three-stage collaborative workflow: DESIGN, GENERATION, and REDESIGN, with an iterative loop between the latter two stages to progressively verify and refine the generated videos. In the DESIGN stage, a large language model (Design Agent) plans objects with layouts, and then a video gener- ation model synthesizes videos in the GENERATION stage. The REDESIGN stage is the most challenging stage that aims to verify the generated videos, suggest corrections, and re- design the text prompts, frame-wise layouts, and guidance scales for the next iteration of generation. To avoid halluci- nation of single-agent and naive multi-agent frameworks, we apply a division-of-labor strategy in this stage by introducing a sequence of specialized agents, executed by MLLMs (mul- timodal large language models): Verification Agent, Sugges- tion Agent, Correction Agent, and Output Structuring Agent. Furthermore, to tackle diverse scenarios of compositional text-to-video generation, we design a self-routing mechanism to adaptively select the proper correction agent from a suite of correction agents, each specialized for one scenario. Ex- tensive experiments demonstrate the effectiveness of GEN- MAC by generating videos based on long compositional text prompts and achieving state-of-the-art in the compositional text-to-video generation benchmark.

Semi-supervised medical image segmentation is an effective method for addressing scenarios with limited labeled data. Existing methods mainly rely on frameworks such as mean teacher and dual-stream consistency learning. These approaches often face issues like error accumulation and model structural complexity, while also neglecting the interaction between labeled and unlabeled data streams. To overcome these challenges, we propose a Bidirectional Channel-selective Semantic Interaction (BCSI) framework for semi-supervised medical image segmentation. First, we propose a Semantic-Spatial Perturbation (SSP) mechanism, which disturbs the data using two strong augmentation operations and leverages unsupervised learning with pseudo-labels from weak augmentations. Additionally, we employ consistency on the predictions from the two strong augmentations to further improve model stability and robustness. Second, to reduce noise during the interaction between labeled and unlabeled data, we propose a Channel-selective Router (CR) component, which dynamically selects the most relevant channels for information exchange. This mechanism ensures that only highly relevant features are activated, minimizing unnecessary interference. Finally, the Bidirectional Channel-wise Interaction (BCI) strategy is employed to supplement additional semantic information and enhance the representation of important channels. Experimental results on multiple benchmarking 3D medical datasets demonstrate that the proposed method outperforms existing semi-supervised approaches.

Understanding dynamic 4D scenes from an egocentric perspective—modeling changes in 3D spatial structure over time—is crucial for human–machine interaction, autonomous navigation, and embodied intelligence. While existing egocentric datasets contain dynamic scenes, they lack unified 4D annotations and task-driven evaluation protocols for fine-grained spatio-temporal reasoning, especially on motion of objects and human, together with their interactions. To address this gap, we introduce EgoDynamic4D, a novel QA benchmark on highly dynamic scenes, comprising RGB-D video, camera poses, globally unique instance masks, and 4D bounding boxes. We construct 927K QA pairs accompanied by explicit Chain-of-Thought (CoT), enabling verifiable, step-by-step spatio-temporal reasoning. We design 12 dynamic QA tasks covering agent motion, human–object interaction, trajectory prediction, relation understanding, and temporal–causal reasoning, with fine-grained, multidimensional metrics. To tackle these tasks, we propose an end-to-end spatio-temporal reasoning framework that unifies dynamic and static scene information, using instance-aware feature encoding, time and camera encoding, and spatially adaptive down-sampling to compress large 4D scenes into token sequences manageable by LLMs. Experiments on EgoDynamic4D show that our method consistently outperforms baselines, validating the effectiveness of multimodal temporal modeling for egocentric dynamic scene understanding.

Facial Attribute Recognition (FAR) holds significant potential for wide-ranging applications. However, traditionally trained FAR models exhibit unfairness, largely due to data bias—where certain sensitive attributes correlate statistically with target attributes. To address this, we propose a group-attention mechanism: first, each image is categorized into subgroups (e.g., Male/Female&short hair, Male/Female&long hair). Within the attention mechanism, distinct Query parameters are used for each group, with shared Key and Value parameters. As group-specific Query parameters are trained on subgrouped data, the noted bias is effectively mitigated. Consequently, integrating this Group-Attention into Vision Transformer (ViT) yields our novel Group-Decoupled ViT (GD-ViT) model. Moreover, to further attenuate the statistical correlation between sensitive and target attributes, we propose a Mask-Guided Correlation Suppression learning strategy. Specifically, in Stage 1, it first leverages a min-max dual-loss optimization strategy to train GD-ViT in capturing key regions related to sensitive attributes yet irrelevant to target attributes. Then, in Stage 2, it trains another GD-ViT by masking sensitive regions identified in Stage 1, fusing the masked output (as intermediate input) with the model’s intermediate outputs. This weakens regions associated with sensitive attributes while enhancing others, suppressing the learning of key features related to sensitive attributes. Consequently, it encourages the model to focus more on intrinsic target attribute regions and balances the learning process between the sensitive attribute and the target attribute. Extensive experiments demonstrate that our method achieves superior performance across three benchmark datasets for fair facial attribute recognition.

In the domain of moment retrieval, accurately identifying temporal segments within videos based on natural language queries remains challenging. Traditional methods often employ pre-trained models that struggle with fine-grained information and deterministic reasoning, leading to difficulties in aligning with complex or ambiguous moments. To overcome these limitations, we explore Deep Evidential Regression (DER) to construct a vanilla Evidential baseline. However, this approach encounters two major issues: the inability to effectively handle modality imbalance and the structural differences in DER's heuristic uncertainty regularizer, which adversely affect uncertainty estimation. This misalignment results in high uncertainty being incorrectly associated with accurate samples rather than challenging ones. Our observations indicate that existing methods lack the adaptability required for complex video scenarios. In response, we propose Debiased Evidential Learning for Moment Retrieval (DEMR), a novel framework that incorporates a Reflective Flipped Fusion (RFF) block for cross-modal alignment and a query reconstruction task to enhance text sensitivity, thereby reducing bias in uncertainty estimation. Additionally, we introduce a Geom-regularizer to refine uncertainty predictions, enabling adaptive alignment with difficult moments and improving retrieval accuracy. Extensive testing on standard datasets and debiased datasets ActivityNet-CD and Charades-CD demonstrates significant enhancements in effectiveness, robustness, and interpretability, positioning our approach as a promising solution for temporal-semantic robustness in moment retrieval.

In image enhancement tasks, such as low-light and underwater image enhancement, a degraded image can correspond to multiple plausible target images due to dynamic photography conditions. This naturally results in a one-to-many mapping problem. To address this, we propose a Bayesian Enhancement Model (BEM) that incorporates Bayesian Neural Networks (BNNs) to capture data uncertainty and produce diverse outputs. To enable fast inference, we introduce a BNN-DNN framework: a BNN is first employed to model the one-to-many mapping in a low-dimensional space, followed by a Deterministic Neural Network (DNN) that refines fine-grained image details. Extensive experiments on multiple low-light and underwater image enhancement benchmarks demonstrate the effectiveness of our method.