AAAI 2026 Papers

Film set design plays a pivotal role in cinematic storytelling and shaping the visual atmosphere. However, the traditional process depends on expert-driven manual modeling, which is labor-intensive and time-consuming. To address this issue, we introduce FilmSceneDesigner, an automated scene generation system that emulates professional film set design workflow. Given a natural language description, including scene type, historical period, and style, we design an agent-based chaining framework to generate structured parameters aligned with film set design workflow, guided by prompt strategies that ensure parameter accuracy and coherence. On the other hand, we propose a procedural generation pipeline which executes a series of dedicated functions with the structured parameters for floorplan and structure generation, material assignment, door and window placement, and object retrieval and layout, ultimately constructing a complete film scene from scratch. Moreover, to enhance cinematic realism and asset diversity, we construct SetDepot-Pro, a curated dataset of 6,862 film-specific 3D assets and 733 materials. Experimental results and human evaluations demonstrate that our system produces structurally sound scenes with strong cinematic fidelity, supporting downstream tasks such as virtual previs, construction drawing and mood board creation.

Large-scale three dimensional vehicle aerodynamics prediction poses critical computational challenges in modern automotive design, where traditional CFD methods require prohibitive simulation times that conflict with rapid design iteration demands. While recent neural operator approaches show promise, existing methods struggle with computational complexity in dense meshes and fail to preserve essential topological information when processing large-scale point clouds. We propose FCMO, a physics-aware neural operator that integrates fluid mechanics principles with selective state space modeling for efficient large-scale vehicle aerodynamics. FCMO introduces four synergistic components: FlowCurv Anchor Sampling that intelligently selects mesh nodes based on normalized local curvature and windward sensitivity. Additionally, dual-scale physics-aware position encoding with adaptive k-NN construction transforms 3D irregular meshes into causality-preserving sequences through feature-guided serpentine scanning. The model integrates a flow-aware Mamba processor incorporating selective mechanisms that dynamically modulate state transitions based on wall distance and flow characteristics. Finally, a physics-constrained decoder enforces conservation laws through mixed weighted interpolation. Extensive experiments on Ahmed-Body and DrivAerNet benchmarks demonstrate that FCMO achieves consistent state-of-the-art performance with 5.2% improvement in surface pressure prediction, 9.3% enhancement in wall shear stress estimation, and 11.4% boost in drag coefficient accuracy, while maintaining superior computational efficiency with 9.4% fewer FLOPs and 9.9% reduced memory usage compared to existing methods.

We present Autoregressive Representation Alignment (ARRA), a new training framework that unlocks global-coherent text-to-image generation in autoregressive LLMs without architectural modifications. Different from prior works that require complex architectural redesigns, ARRA aligns LLM's hidden states with visual representations from external visual foundational models via a global visual alignment loss and a hybrid token, . This token enforces dual constraints: local next-token prediction and global semantic distillation, enabling LLMs to implicitly learn spatial and contextual coherence while retaining their original autoregressive paradigm. Extensive experiments validate ARRA's plug-and-play versatility. When training T2I LLMs from scratch, ARRA reduces FID by 16.6% (ImageNet), 12.0% (LAION-COCO) for autoregressive LLMs like LlamaGen, without modifying original architecture and inference mechanism. For training from text-generation-only LLMs, ARRA reduces FID by 25.5% (MIMIC-CXR), 8.8% (DeepEyeNet) for advanced LLMs like Chameleon. For domain adaptation, ARRA aligns general-purpose LLMs with specialized models (e.g., BioMedCLIP), achieving an 18.6% FID reduction over direct fine-tuning on medical imaging (MIMIC-CXR). These results demonstrate that training objective redesign, rather than architectural modifications, can resolve cross-modal global coherence challenges. ARRA offers a complementary paradigm for advancing autoregressive models.

Reconstructing dynamic objects from monocular RGB-D video is critical for advancing 3D vision applications and enhancing user experience. However, monocular RGB-D video provides limited 3D observations, making the reconstruction of unobserved regions highly under-constrained. Despite recent advances that combine neural implicit surfaces with diffusion models, the inherent limitations of implicit representations and the lack of effective guidance in diffusion priors lead to blurry appearance and inaccurate geometry in dynamic object reconstruction. To address the issue, we present MGD, which leverages scene-adaptive diffusion priors and Mesh-guided Gaussians for realistic rendering and geometrically accurate reconstruction of dynamic objects, including unobserved regions. The dynamic 3D objects reconstructed by MGD are represented using our proposed Mesh-guided Gaussians, which leverage global and local Gaussians to capture large-scale deformations and fine-grained appearance details, respectively. Additionally, in order to utilize depth information, we integrate a depth ControlNet into the diffusion model and conduct scene-adaptive fine-tuning. We design a self-generated image-pair strategy to produce image pairs used for fine-tuning. Extensive experiments demonstrate that MGD achieves state-of-the-art performance in both high-fidelity reconstruction and structural completeness, while maintaining real-time efficiency during training and rendering.

Recent advancements in 4D generation have demonstrated its remarkable capability in synthesizing photorealistic renderings of dynamic 3D scenes. However, despite achieving impressive visual performance, almost all existing methods overlook the generation of spatial audio aligned with the corresponding 4D scenes, posing a significant limitation to truly immersive audiovisual experiences. To mitigate this issue, we propose Sonic4D, a novel framework that enables spatial audio generation for immersive exploration of 4D scenes. Specifically, our method is composed of three stages: 1) To capture both the dynamic visual content and raw auditory information from a monocular video, we first employ pre-trained expert models to generate the 4D scene and its corresponding monaural audio. 2) Subsequently, to transform the monaural audio into spatial audio, we localize and track the sound sources within the 4D scene, where their 3D spatial coordinates at different timestamps are estimated via a pixel-level visual grounding strategy. 3) Based on the estimated sound source locations, we further synthesize plausible spatial audio that varies across different viewpoints and timestamps using physics-based simulation. Extensive experiments have demonstrated that our proposed method generates realistic spatial audio consistent with the synthesized 4D scene in a training-free manner, significantly enhancing the immersive experience for users.

Distilling knowledge from human demonstrations is a promising way for robots to learn and act. Existing methods, which often rely on coarsely-aligned video pairs, are typically constrained to learning global or task-level features. As a result, they tend to neglect the fine-grained frame-level dynamics required for complex manipulation and generalization to novel tasks. We posit that this limitation stems from a vicious circle of inadequate datasets and the methods they inspire. To break this cycle, we propose a paradigm shift that treats fine-grained human-robot alignment as a conditional video generation problem. To this end, we first introduce H&R, a novel third-person dataset containing 2,600 episodes of precisely synchronized human and robot motions, collected using a VR teleoperation system. We then present Human2Robot, a framework designed to leverage this data. Human2Robot employs a Video Prediction Model to learn a rich and implicit representation of robot dynamics by generating robot videos from human input, which in turn guides a decoupled action decoder. Our real-world experiments demonstrate that this approach not only achieves high performance on seen tasks but also exhibits significant one-shot generalization to novel positions, objects, instances, and even new task categories.

While adversarial attacks can effectively deceive deep neural networks, their real-world applicability is often limited by complex and conspicuous patterns that reveal their attack intent to human observers. To overcome this limitation, we propose UYE, a novel camouflage framework designed to simultaneously mislead DNNs and evade human perception. UYE incorporates two key components: an attention refiner leveraging a pre-trained vision encoder to optimize adversarial patterns for robust attacks across diverse environments, and a perception evaluator trained on a preference dataset curated using tailored prompts from human-aligned large multimodal models to ensure natural and unobtrusive camouflage generation. Extensive experiments demonstrate that UYE outperforms state-of-the-art methods in achieving an optimal balance between human stealth and model deception while maintaining effectiveness in real-world scenarios.

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in visual classification tasks. Existing methods for enhancing VLMs on this task often rely heavily on direct category-to-image matching, which limits generalization and results in suboptimal performance. In addition, these methods provide no understanding of why a specific category is chosen. To address these limitations, we introduce a new deliberative visual classification task that decomposes the classification process into multiple deliberative steps and leverages Large Language Models (LLMs) to perform explicit reasoning before the final decision. Specifically, we propose a Retrieval-driven Reasoning model (RdR) with two components, i.e., retrieval database construction and deliberative category prediction. The first component leverages LLMs to extract category-relevant descriptors and constructs a retrieval database for effective image–descriptor matching. The second component facilitates multiple deliberative steps and performs explicit reasoning based on the retrieved descriptors to augment the category prediction. Extensive experiments on multiple datasets demonstrate that RdR consistently outperforms strong baselines, highlighting its robustness and generalization ability.

Most visual models are designed for sRGB images, yet RAW data offers significant advantages for object detection by preserving sensor information before ISP processing. This enables improved detection accuracy and more efficient hardware designs by bypassing the ISP. However, RAW object detection is challenging due to limited training data, unbalanced pixel distributions, and sensor noise. To address this, we propose SimROD, a lightweight and effective approach for RAW object detection. We introduce a Global Gamma Enhancement (GGE) module, which applies a learnable global gamma transformation with only four parameters, improving feature representation while keeping the model efficient. Additionally, we leverage the green channel's richer signal to enhance local details, aligning with the human eye’s sensitivity and Bayer filter design. Extensive experiments on multiple RAW object detection datasets and detectors demonstrate that SimROD outperforms state-of-the-art methods like RAW-Adapter and DIAP while maintaining efficiency. Our work highlights the potential of RAW data for real-world object detection.

We propose PUNO, a novel deep operator-based framework for point cloud upsampling, addressing the challenge of reconstructing high-resolution geometries from sparse point clouds. PUNO generalizes the neural operators proven effective in image super-resolution to 3D point cloud upsampling. Moreover, it first designs a network for point cloud tasks to achieve vertex displacement and manifold parameterization, thereby forming a coarse geometric representation that is compatible with super-resolution neural operators. This is followed by iterative kernel integral approximations in the function space and backprojection to generate the target coordinates, fully utilizing the high-frequency information in the function space. Unlike prior work, PUNO performs transformations in both the data domain and the function domain, with the solution space containing richer basis functions, yielding finer results that mitigate the ill-posed nature of sparse data. It also benefits global continuity. Extensive experiments demonstrate its superior accuracy, robustness, and generalization ability.

Blurry video super-resolution (BVSR) remains fundamentally ill-posed due to the simultaneous loss of high-frequency spatial details and reliable motion cues in blurry low-resolution frames. While cascade-based and joint BVSR methods struggle under severe blur, existing event-guided VSR approaches largely assume clean inputs and are ineffective against complex motion degradation. These methods fail to model blurry representations or leverage event signals for blur-aware motion cues, leading to sub-optimal performance. We propose BluR-EVSR, a unified framework that implicitly models Blurry Representations and leverages Event cameras to jointly address both blur and resolution degradation for VSR. The framework begins with a self-supervised degradation learning strategy guided by event streams and neighboring frames, enabling adaptive blur representation without requiring explicit supervision. A dynamic routing mechanism encodes spatially varying degradations, while a motion-saliency degradation-aware attention module injects motion saliency priors to facilitate efficient RGB-event fusion. Integrated into a bidirectional recurrent framework, BluR-EVSR enables temporally consistent and detail-preserving restoration with low computational cost. Extensive experiments across multiple benchmarks show that our method significantly outperforms prior BVSR and event-based approaches.

Joint rendering and deformation of mesh and 3D Gaussian Splatting (3DGS) have significant value as both representations offer complementary advantages for graphics applications. However, due to differences in representation and rendering pipelines, existing studies render meshes and 3DGS separately, making it difficult to accurately handle occlusions and transparency. Moreover, the deformed 3DGS still suffers from visual artifacts due to the sensitivity to the topology quality of the proxy mesh. These issues pose serious obstacles to the joint use of 3DGS and meshes, making it difficult to adapt 3DGS to conventional mesh-oriented graphics pipelines. We propose UniMGS, the first unified framework for rasterizing mesh and 3DGS in a single-pass anti-aliased manner, with a novel binding strategy for 3DGS deformation based on proxy mesh. Our key insight is to blend the colors of both triangle and Gaussian fragments by anti-aliased α-blending in a single pass, achieving visually coherent results with precise handling of occlusion and transparency. To improve the visual appearance of the deformed 3DGS, our Gaussian-centric binding strategy employs a proxy mesh and spatially associates Gaussians with the mesh faces, significantly reducing rendering artifacts. With these two components, UniMGS enables the visualization and manipulation of 3D objects represented by mesh or 3DGS within a unified framework, opening up new possibilities in embodied AI, virtual reality, and gaming. We will release our source code to facilitate future research.

With the rapid development of the low-altitude economy, multimodal visual tracking in UAV scenarios has attracted extensive attention. UAVs are typically equipped with independent visible (RGB) and thermal infrared (TIR) sensors, resulting in an inherent spatial misalignment between the two modalities. However, existing RGBT tracking methods generally rely on spatially aligned data inputs, making them unsuitable for unaligned RGBT tracking task in UAV scenarios. In this work, we introduce the new task called unaligned UAV RGBT tracking and construct the first large-scale unaligned RGB and TIR video dataset to promote the research and development of this field. The dataset contains 1,453 pairs of UAV-captured RGBT sequences with precise dual-modal bounding box annotations, and covers 42 object categories, 22 typical challenge attributes, and diverse spatial misalignment scales to better simulate real-world challenging scenarios. To address the limitations of existing methods that fail to handle the spatial misalignment issue in UAV scenarios, we propose the novel RGBT tracking approach. In particular, we design a mixture of shift estimation experts module to adaptively estimate the spatial shifts across two modalities at different scales, and a cross-modal alignment and fusion module to further compensate for nonlinear deformations and integrate multimodal information. Extensive experiments on the created dataset demonstrate that the proposed tracker significantly outperforms existing state-of-the-art tracking methods, validating its practicality and robustness in real-world unaligned UAV tracking scenarios.

Infrared and visible image fusion aims to integrate complementary information, such as thermal saliency from infrared imagery and fine-grained texture details from visible imagery. However, real-world multi-modal misalignment and geometric deformation often introduce severe artifacts. Most existing methods focus on feature extraction within Euclidean space, thereby neglecting the inherent hierarchical structures embedded in multimodal representations. While Euclidean space excels at preserving local structural details and supporting efficient computation, hyperbolic space is naturally suited for modeling hierarchical relationships due to its geometric properties. Building upon these observations, this paper proposes a unified framework that jointly optimizes image registration and fusion through a dual-space architecture. This architecture synergistically combines the local fidelity of Euclidean geometry with the hierarchical modeling capability of hyperbolic geometry to enhance multimodal representation learning. Specifically, this paper introduces Hyperbolic Coupled Contrastive Learning Optimization (HCCLO), which aligns and optimizes the hierarchical structures of infrared and visible embeddings in hyperbolic space. Moreover, this paper designs a task-adaptive dual-space features fusion mechanism, which dynamically balances and fuses Euclidean local features with hyperbolic hierarchical representations, thereby improving adaptability for downstream tasks. Extensive experiments on misaligned multimodal datasets demonstrate that our method achieves state-of-the-art performance, while effectively capturing both spatial dependencies and hierarchical semantics.

Human Novel View Synthesis (HNVS) aims to synthesize photorealistic human images from novel viewpoints given observations from known views. Despite significant advances achieved by existing methods such as NeRF, diffusion models, and 3DGS, they still face substantial challenges in achieving stable modeling from a single image. In this paper, we introduce Dual-Constraint Human Gaussian Splatting (DcSplat), a novel, simple, and efficient 3D Gaussian-based framework for single-view 3D human reconstruction. To address occlusion-induced texture missing and depth ambiguities, we introduce two key components: a Latent Multi-View Consistency Constraint Mechanism and a Geometric Constraint Module. The former employs a Latent-space Appearance Transformer (LatentFormer) to learn semantically coherent, view-consistent appearance priors via SMPL-guided pseudo-view fusion. The latter refines noisy SMPL-based depth through a U-Net-like structure conditioned on latent appearance features. These two modules are jointly optimized to generate high-quality Gaussian parameters in a unified latent space. Extensive experiments demonstrate that DcSplat outperforms existing SOTA methods in both geometry and texture quality, while achieving fast inference and lower computational cost.

Few-shot Semantic Segmentation (FSS) aims to segment the novel target objects with the guidance of minimal annotated reference examples. The affinity-based method has great advantages in the FSS inference stage for both specialist model and foundation model. However, current affinity calculation merely relies on only support-query matching, without considering the query-specific semantic or the semantic correlation among inter-support samples, which limits the representation ability of affinity map. In this paper, we propose the Generalizing Semantic Mining (GSM) that focuses on exploiting generalizing semantic to improve the affinity calculation. Concretely, we first organize the affinity-based inference into three main steps to reveal the crucial role of affinity map. To address the low-data problem, Target Semantic Reusing module considers the query sample as a proxy reference and assigns it with proxy mask identifying its most generalizing semantic regions. Then, to generate the high-fidelity proxy mask, Query-specific Semantic Modeling module pinpoints the most generalizing regions through prior semantic analysis. Finally, Representative Re-weighting module explicitly modulates affinity calculation via generalization-aware weighting. Experiments on FSS benchmarks demonstrate that our GSM can serve as a plug-and-play free lunch for both specialist models and foundation models.

Vision-Language Models (VLMs) like CLIP struggle to understand negation, often embedding affirmatives and negatives similarly (e.g., matching "no dog" with dog images). Existing methods refine negation understanding via fine-tuning CLIP’s text encoder, risking overfitting. In this work, we propose CLIPGlasses, a plug-and-play framework that enhances CLIP’s ability to comprehend negated visual descriptions. CLIPGlasses adapts a dual-stage design: a Lens module disentangles negated semantics from text embeddings, and a Frame module predicts context-aware repulsion strength, which is integrated into the modified similarity computation to penalize alignment with negated semantics, thereby reducing false positive matches. Experiments show that CLIP equipped with CLIPGlasses achieves competitive in-domain performance and outperforms state-of-the-art methods in cross-domain generalization. Its superiority is especially evident under low-resource conditions, indicating stronger robustness across domains.

Reconstructing a faithful geometric surface from sparse images remains a fundamental challenge in 3D computer vision. While recent methods have achieved remarkable progress, they still struggle to recover reliable geometry due to the lack of multi-view geometric cues, particularly in non-overlapping regions. To address this issue, we introduce VGGS, a Gaussian Splatting (GS) method that exploits multi-view geometric priors from VGGT for efficient and high-fidelity sparse-view surface reconstruction. Our primary contribution is an anchor-calibrated depth estimation scheme, which yields accurate depth maps. The insight is to align the VGGT depth prior to the underlying surface with a sparse set of multi-view consistent anchors, then infer depth for unreliable regions by relative depth estimation. Furthermore, to mitigate misalignment in complex scenes, we propose a relative depth consistency loss that penalizes the rendered depth if its relative depth relationship in local regions is inconsistent to the multi-view prior. Extensive experiments on widely-used benchmarks show that VGGS surpasses state-of-the-art methods in both accuracy and efficiency, delivering 4–7× faster optimization while reducing memory consumption compared to previous GS-based approaches.

Although deep learning-based methods have achieved promising performance in Pansharpening, they generally suffer from severe performance degradation when applied to data from unseen sensors. Existing cross-domain strategies, including retraining, fine-tuning, and zero-shot methods, fail to simultaneously preserve model architecture and maintain low adaptation costs. Therefore, we are the first to define and address a novel task in the pansharpening field: enhancing a model's cross-sensor generalization at an extremely low cost while keeping the model architecture invariant. To tackle this task, we propose SWIFT (Sensitive Weight Identification for Fast Transfer), a plug-and-play framework. SWIFT first employs an unsupervised manifold-based sampling strategy to efficiently select a high-fidelity subset the most informative target-domain samples. It then leverages this subset to probe a source-domain pre-trained model, identifying and updating only the weight subset most sensitive to the domain shift by analyzing the gradient behavior of its parameters. Extensive experiments demonstrate that SWIFT can be applied to various deep learning models, boosting adaptation efficiency by up to 30-fold. On a single NVIDIA RTX 4090 GPU, this reduces adaptation time from hours to as little as one minute. The adapted models not only substantially outperform direct-transfer baselines but also achieve performance competitive with, or even superior to full retraining while using only 3% of the target domain dataset and adapting nearly 10% to 30% of the model’s parameters. This establishs a new state-of-the-art on the WorldView-2 and QuickBird datasets.