ICCV 2025 Papers

Generating immersive 360deg indoor panoramas from 2D top-down views has applications in virtual reality, interior design, real estate, and robotics. This task is challenging due to the lack of explicit 3D structure and the need for geometric consistency and photorealism. We propose Top2Pano, an end-to-end model for synthesizing realistic indoor panoramas from top-down views. Our method estimates volumetric occupancy to infer 3D structures, then uses volumetric rendering to generate coarse color and depth panoramas. These guide a diffusion-based refinement stage using ControlNet, enhancing realism and structural fidelity. Evaluations on two datasets show Top2Pano outperforms baselines, effectively reconstructing geometry, occlusions, and spatial arrangements. It also generalizes well, producing high-quality panoramas from schematic floorplans. Our results highlight Top2Pano's potential in bridging top-down views with immersive indoor synthesis.

Integrating external tools into Large Foundation Models (LFMs) has emerged as a promising approach to enhance their problem-solving capabilities. While existing studies have demonstrated strong performance in tool-augmented Visual Question Answering (VQA), recent benchmarks re- veal significant gaps in real-world tool-use proficiency, particularly in functionally diverse multimodal settings re- quiring multi-step reasoning. In this work, we intro- duce ToolVQA, a large-scale multimodal dataset compris- ing 23K samples, designed to bridge this gap. Unlike pre- vious datasets that rely on synthetic scenarios and sim- plified queries, ToolVQA features real-world visual con- texts and challenging implicit multi-step reasoning tasks, better aligning with real user interactions. To construct this dataset, we propose ToolEngine, a novel data genera- tion pipeline that employs image-guided Depth-First Search (DFS) with a Longest Common Subsequence (LCS)-based example matching mechanism to simulate human-like tool- use reasoning. ToolVQA encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. The LLaVA-7B model fine-tuned on ToolVQA not only achieves impressive per- formance on the ToolVQA test set, but also surpasses the large closed-source model GPT-3.5-turbo on five out-of- distribution (OOD) datasets, showing strong generalizabil- ity in real-world tool-use scenarios. Code is available at https://github.com/Fugtemypt123/ToolVQA-release.

Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions.Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations.

Vision-Language Models (VLMs) excel at visual understanding by leveraging pretrained image encoders to generate visual tokens. However, they struggle with high-resolution images and zoomed-in regions due to the computational burden and token redundancy of uniform patch-based processing, often leading to the loss of critical details. To address these challenges, we propose Token-Efficient Vision Language Model (TEVA), a novel framework that detects key regions and applies dynamic patch sampling to efficiently capture fine-grained details while preserving global context. Our approach first identifies subject-oriented regions using an adaptive detection strategy. Then, a dynamic patch sampling mechanism selects and arranges patches at varying scales, ensuring efficient processing without increasing token count. Extensive experiments demonstrate that Token-Efficient Vision Language Model (TEVA) significantly enhances VLM performance in handling visual details, seamlessly integrating with various decoders and LLMs.

Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM.

Time-of-Flight (ToF) sensors provide efficient active depth sensing at relatively low power budgets; among such designs, only very sparse measurements from low-resolution sensors are considered to meet the increasingly limited power constraints of mobile and AR/VR devices. However, such extreme sparsity levels limit the seamless usage of ToF depth in SLAM. In this work, we propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline tailored for using effectively very sparse ToF input data. Our approach improves upon the state of the art by introducing a multi-frame integration module, which produces dense depth maps by merging cues from extremely sparse ToF depth, monocular color, and multi-view geometry. Extensive experiments on both synthetic and real sparse ToF datasets demonstrate the viability of our approach, as it achieves state-of-the-art tracking and mapping performances on reference datasets.

Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy, coverage efficiency, early-stage performance, and scalability. PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM.

Mamba has shown great potential for computer vision due to its linear complexity in modeling the global context with respect to the input length. However, existing lightweight Mamba-based backbones cannot demonstrate performance that matches Convolution or Transformer-based methods. By observing, we find that simply modifying the scanning path in the image domain is not conducive to fully exploiting the potential of vision Mamba. In this paper, we first perform comprehensive spectral and quantitative analyses, and verify that the Mamba block mainly models low-frequency information under Convolution-Mamba hybrid architecture. Based on the analyses, we introduce a novel Laplace mixer to decouple the features in terms of frequency and input only the low-frequency components into the Mamba block. In addition, considering the redundancy of the features and the different requirements for high-frequency details and low-frequency global information at different stages, we introduce a frequency ramp inception, i.e., gradually reduce the input dimensions of the high-frequency branches, so as to efficiently trade-off the high-frequency and low-frequency components at different layers. By integrating mobile-friendly convolution and efficient Laplace mixer, we build a series of tiny hybrid vision Mamba called TinyViM. The proposed TinyViM achieves impressive performance on several downstream tasks including image classification, semantic segmentation, object detection and instance segmentation. In particular, TinyViM outperforms Convolution, Transformer and Mamba-based models with similar scales, and the throughput is about 2-3 times higher than that of other Mamba-based models. The good balance between the efficiency and performance of TinyViM shows that a properly designed and optimized vision mamba can achieve high performance with a small model size. Code is available at https://github.com/xwmaxwma/TinyViM.

Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pre-trained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM.

Dynamic scene reconstruction is a long-term challenge in 3D vision. Recent methods extend 3D Gaussian Splatting to dynamic scenes via additional deformation fields and apply explicit constraints like motion flow to guide the deformation. However, they learn motion changes from individual timestamps independently, making it challenging to reconstruct complex scenes, particularly when dealing with violent movement, extreme-shaped geometries, or reflective surfaces. To address the above issue, we design a plug-and-play module called TimeFormer to enable existing deformable 3D Gaussians reconstruction methods with the ability to implicitly model motion patterns from a learning perspective. Specifically, TimeFormer includes a Cross-Temporal Transformer Encoder, which adaptively learns the temporal relationships of deformable 3D Gaussians. Furthermore, we propose a two-stream optimization strategy that transfers the motion knowledge learned from TimeFormer to the base stream during the training phase. This allows us to remove TimeFormer during inference, thereby preserving the original rendering speed. Extensive experiments in the multi-view and monocular dynamic scenes validate qualitative and quantitative improvement brought by TimeFormer.Project

Video Temporal Grounding (VTG) aims to precisely identify video event segments in response to textual queries. The outputs of VTG tasks manifest as sequences of events, each defined by precise timestamps, saliency scores, and textual descriptions. Despite recent advances, a fundamental limitation persists in existing Video Large Language Models (Video-LLMs): they process all task tokens through identical and static pathways, failing to recognize that temporal localization, saliency assessment, and textual generation represent fundamentally distinct tasks requiring specialized processing. To address this, we introduce TimeExpert, a Mixture-of-Experts (MoE)-based Video-LLM that effectively decomposes VTG tasks by dynamically routing task-specific tokens (e.g., timestamps, saliency scores) to specialized experts, with increased computational efficiency. Our design choices enable precise handling of each subtask, leading to improved event modeling across diverse VTG applications. Extensive experiments demonstrate that TimeExpert consistently achieves state-of-the-art performance on various VTG tasks such as Dense Video Captioning, Moment Retrieval, and Video Highlight Detection.

Face aging is a typical ill-posed problem influenced by various factors such as environment and genetics, leading to highly diverse outcomes. However, existing methods primarily rely on numerical age representations, making it difficult to accurately capture individual or group-level aging patterns. To address this, we introduce a novel disentangled face representation, where age features are modeled in the image modality--referred to as the Age Prompt--providing richer prior age information to constrain the generation results. To this end, we design an ID-age multi-task co-learning framework and propose the Bidirectional Adversarial Disentanglement(BAD) strategy. This strategy maximizes the disentanglement of ID and age representation through bidirectional adversarial learning, extracting their attribute-invariant representations. Based on this representation, we propose TimeBooth, a personalized face aging model capable of generating diverse and individualized aging results. To optimize training, we construct a cross-age hybrid data pipeline and introduce various training strategies. Finally, we propose the R-AgeMAE metric and validate our method through extensive experiments, demonstrating that TimeBooth outperforms existing methods in both diversity and controllability. Our code will be publicly available at https://github.com/szp-aigc/TimeBooth.

Cameras rely on auto white balance (AWB) to correct undesirable color casts caused by scene illumination and the camera's spectral sensitivity. This is typically achieved using an illuminant estimator that determines the global color cast solely from the color information in the camera's raw sensor image. Mobile devices provide valuable additional metadata---such as capture timestamp and geolocation---that offers strong contextual clues to help narrow down the possible illumination solutions. This paper proposes a lightweight illuminant estimation method that incorporates such contextual metadata, along with additional capture information and image colors, into a lightweight model (~5K parameters), achieving promising results, matching or surpassing larger models. To validate our method, we introduce a dataset of 3,224 smartphone images with contextual metadata collected at various times of day and under diverse lighting conditions. The dataset includes ground-truth illuminant colors, determined using a color chart, and user-preferred illuminants validated through a user study, providing a comprehensive benchmark for AWB evaluation.

Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets. The code is available at https://github.com/kreshuklab/no_tiling_artifacts.

3D Gaussian Splatting (3DGS) has drawn significant attention for its advantages in rendering speed and quality. Most existing methods still rely on the image-wise loss and training paradigm because of its intuitive nature in the Splatting algorithm. However, image-wise loss lacks multi-view constraints, which are generally essential for optimizing 3D appearance and geometry. To address this, we propose RT-Loss along with a tile-based training paradigm, which uses randomly sampled tiles to integrate multi-view appearance and structural constraints in 3DGS. Additionally, we introduce an tile-based adaptive densification control strategy tailored for our training paradigm. Extensive experiments show that our approach consistently improves performance metrics while maintaining efficiency across various benchmark datasets.

Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZero, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZero matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.

Contrastive Language-Image Pre-training (CLIP) models exhibit intriguing properties, particularly in their zero-shot classification capability. However, the reliability of CLIP zero-shot classification is severely undermined by spurious correlations. Existing efforts to enhance the robustness of zero-shot CLIP models often rely on prior knowledge or annotations of spurious correlations, limiting real-world applicability due to the unavailability of such information. Alternative methods attempt to detect distribution shift at test time but require training statistics whose access is often restricted or computationally expensive. To address the challenges brought by spurious correlation under zero-shot settings, we propose a novel test-time reasoning approach. Our method, inspired by human recognition, localizes the object and refines the classification accordingly. The inherent capacity of CLIP for semantic understanding allows us to isolate the object of interest without auxiliary models. Zero-shot classification is then performed exclusively on the localized objects, effectively mitigating the influence of spurious correlation. The proposed approach is interpretable and flexible as it requires no spurious annotations or prior knowledge, making it widely applicable. The substantial improvements across multiple benchmark datasets validated the effectiveness of our approach.

This paper introduces a novel method for detailed 3D shape reconstruction utilizing thermal polarization cues. Unlike state-of-the-art methods, the proposed approach is independent of illumination and material properties. In this paper, we formulate a general theory of polarization observation and show that long-wave infrared (LWIR) polarimetric imaging is free from the ambiguities that affect visible polarization analyses. Subsequently, we propose a method for recovering detailed 3D shapes using multi-view thermal polarimetric images. Experimental results demonstrate that our approach effectively reconstructs fine details in transparent, translucent, and heterogeneous objects, outperforming existing techniques.