ICCV 2025 Papers

Multimodal large language models(MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding(SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR

Remote physiological measurement (RPM) based on video and radar has made significant progress in recent years. However, unimodal methods based solely on video or radar sensor have notable limitations due to their measurement principles, and multimodal RPM that combines these modalities has emerged as a promising direction. Despite its potential, the lack of large-scale multimodal data and the significant modality gap between video and radar pose substantial challenges in building robust video-radar RPM models. To handle these problems, we suggest leveraging unimodal pre-training and present the Spatial alignment and Temporal Matching (SATM) Adapter to effectively fine-tune pre-trained unimodal backbones into a multimodal RPM model. Given the distinct measurement principles of video- and radar-based methods, we propose Spatial Alignment to align the spatial distribution of their features. Furthermore, Temporal Matching is applied to mitigate waveform discrepancies between video and radar signals. By integrating these two modules into adapters, the unimodal backbones could retain their modality-specific knowledge while effectively extracting complementary features from each other. Extensive experiments across various challenging scenarios, including low light conditions and head motions, demonstrate that our approach significantly surpasses the state-of-the-art methods.

Adapting Vision-Language Models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for local sparsity and global density, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for local randomness and global importance, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead.

Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM by pruning tokens in the prefill stage while attending to a sparse subset of visual tokens during the decoding phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a 1.4x speedup on image benchmarks. This approach leads to +5.9% average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a 3.6x and 1.7x speedup in prefill and decoding, respectively.

Surface reconstruction from sparse views aims to reconstruct a 3D shape or scene from few RGB images. The latest methods are either generalization-based or overfitting-based. However, the generalization-based methods do not generalize well on views that were unseen during training, while the reconstruction quality of overfitting-based methods is still limited by the limited geometry clues. To address this issue, we propose SparseRecon, a novel neural implicit reconstruction method for sparse views with volume rendering-based feature consistency and uncertainty-guided depth constraint. Firstly, we introduce a feature consistency loss across views to constrain the neural implicit field. This design alleviates the ambiguity caused by insufficient consistency information of views and ensures completeness and smoothness in the reconstruction results. Secondly, we employ an uncertainty-guided depth constraint to back up the feature consistency loss in areas with occlusion and insignificant features, which recovers geometry details for better reconstruction quality. Experimental results demonstrate that our method outperforms the state-of-the-art methods, which can produce high-quality geometry with sparse-view input, especially in the scenarios with small overlapping views. Project page: https://hanl2010.github.io/SparseRecon/.

Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers 1.38x real-time acceleration and 52% memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM.

3D lane detection has emerged as a critical challenge in autonomous driving, encompassing identification and localization of lane markings and the 3D road surface. Conventional 3D methods detect lanes from dense birds-eye-viewed (BEV) features, though erroneous transformations often result in a poor feature representation misaligned with the true 3D road surface. While recent sparse lane detectors have surpassed dense BEV approaches, they completely disregard valuable lane-specific priors. Furthermore, existing methods fail to utilize historic lane observations, which yield the potential to resolve ambiguities in situations of poor visibility. To address these challenges, we present SparseLaneSTP, a novel method that integrates both geometric properties of the lane structure and temporal information into a sparse lane transformer. It introduces a new lane-specific spatio-temporal attention mechanism, a continuous lane representation tailored for sparse architectures as well as temporal regularization. Identifying weaknesses of existing 3D lane datasets, we also introduce a precise and consistent 3D lane dataset using a simple yet effective auto-labeling strategy. Our experimental section proves the benefits of our contributions and demonstrates state-of-the-art performance across all detection and error metrics on existing 3D lane detection benchmarks as well as on our novel dataset.

Creating high-fidelity 3D meshes with arbitrary topology, including open surfaces and complex interiors, remains a significant challenge. Existing implicit field methods often require costly and detail-degrading watertight conversion, while other approaches struggle with high resolutions. This paper introduces SparseFlex, a novel sparse-structured isosurface representation that enables differentiable mesh reconstruction at resolutions up to 1024^3 directly from rendering losses. SparseFlex combines the accuracy of Flexicubes with a sparse voxel structure, focusing computation on surface-adjacent regions and efficiently handling open surfaces. Crucially, we introduce a frustum-aware sectional voxel training strategy that activates only relevant voxels during rendering, dramatically reducing memory consumption and enabling high-resolution training. This also allows, for the first time, the reconstruction of mesh interiors using only rendering supervision. Building upon this, we demonstrate a complete shape modeling pipeline by training a variational autoencoder (VAE) and a rectified flow transformer for high-quality 3D shape generation. Our experiments show state-of-the-art reconstruction accuracy, with a 82% reduction in Chamfer Distance and a 88% increase in F-score compared to previous methods, and demonstrate the generation of high-resolution, detailed 3D shapes with arbitrary topology. By enabling high-resolution, differentiable mesh reconstruction and generation with rendering losses, SparseFlex significantly advances the state-of-the-art in 3D shape representation and modeling.

Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full fine-tuning is often impractical, parameter-efficient fine-tuning --and particularly side-tuning (ST)-- has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention --a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a 73% the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST.

Large pre-trained transformers have revolutionized artificial intelligence across various domains, and fine-tuning remains the dominant approach for adapting these models to downstream tasks due to the cost of training from scratch. However, in existing fine-tuning methods, the updated representations are formed as a dense combination of modified parameters, making it challenging to interpret their contributions and understand how the model adapts to new tasks.In this work, we introduce a fine-tuning framework inspired by sparse coding, where fine-tuned features are represented as a sparse combination of basic elements, i.e., feature dictionary atoms. Sparse coefficients then serve as indicators of atom importance, identifying the contribution of each atom to the updated representation.The feature dictionary atoms function as fundamental building blocks of the representation, and tuning atoms allows for seamless adaptation to downstream tasks.Leveraging the atom selection capability of sparse coefficients, we first demonstrate that our method enhances image editing performance by improving text alignment through the removal of unimportant feature dictionary atoms.Additionally, we validate the effectiveness of our approach in the text-to-image concept customization task, where our method efficiently constructs the target concept using a sparse combination of feature dictionary atoms, outperforming various baseline fine-tuning methods.

We present a method for Sparse view reconstruction with surface element splatting that runs within 2 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate stat-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view Benchmarks based on established multi-view datasets.

Federated Unsupervised Learning (FUL) aims to learn expressive representations in federated and self-supervised settings. The quality of representations learned in FUL is usually determined by uniformity, a measure of how uniformly representations are distributed in the embedding space. However, existing solutions perform well in achieving intra-client (local) uniformity for local models while failing to achieve inter-client (global) uniformity after aggregation due to non-IID data distributions and the decentralized nature of FUL. To address this issue, we propose Soft Separation and Distillation (SSD), a novel approach that preserves inter-client uniformity by encouraging client representations to spread toward different directions. This design reduces interference during client model aggregation, thereby improving global uniformity while preserving local representation expressiveness. We further enhance this effect by introducing a projector distillation module to address the discrepancy between loss optimization and representation quality. We evaluate SSD in both cross-silo and cross-device federated settings, demonstrating consistent improvements in representation quality and task performance across various training scenarios. Our results highlight the importance of inter-client uniformity in FUL and establish SSD as an effective solution to this challenge. Project page: https://ssd-uniformity.github.io/

Completeness is a widely discussed property in explainability research, requiring that the attributions sum to the model's response to the input. While completeness intuitively suggests that the model's prediction is "completely explained" by the attributions, its global formulation alone is insufficient to ensure faithful explanations. We contend that promoting completeness locally within attribution subregions, in a soft manner, can serve as a standalone guiding principle for producing faithful attributions. To this end, we introduce the concept of the completeness gap as a flexible measure of completeness and propose an optimization procedure that minimizes this gap across subregions within the attribution map. Extensive evaluations across various model architectures demonstrate that our method produces state-of-the-art results.

Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.

Video understanding requires the extraction of rich spatio-temporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by 1.6-3.0% and 1.1-1.9% top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models.

We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms -- significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition.Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model weights and datasets are available at: https://huggingface.co/ds4sd

We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines.