ICCV 2025 Papers

The Mixture of Experts (MoE) architecture has excelled in Large Vision-Language Models (LVLMs), yet its potential in real-time open-vocabulary object detectors, which also leverage large-scale vision-language datasets but smaller models, remains unexplored. This work investigates this domain, revealing intriguing insights. In the shallow layers, experts tend to cooperate with diverse peers to expand the search space. While in the deeper layers, fixed collaborative structures emerge, where each expert maintains 2-3 fixed partners and distinct expert combinations are specialized in processing specific patterns. Concretely, we propose Dynamic-DINO, which extends Grounding DINO 1.5 Edge from a dense model to a dynamic inference framework via an efficient MoE-Tuning strategy. Additionally, we design a granularity decomposition mechanism to decompose the Feed-Forward Network (FFN) of base model into multiple smaller expert networks, expanding the subnet search space. To prevent performance degradation at the start of fine-tuning, we further propose a pre-trained weight allocation strategy for the experts, coupled with a specific router initialization. During inference, only the input-relevant experts are activated to form a compact subnet. Experiments show that, pretrained with merely 1.56M open-source data, Dynamic-DINO outperforms Grounding DINO 1.5 Edge, pretrained on the private Grounding20M dataset.

Text animation serves as an expressive medium, transforming static communication into dynamic experiences by infusing words with motion to evoke emotions, emphasize meanings, and construct compelling narratives. Crafting animations that are semantically aware poses significant challenges, demanding expertise in graphic design and animation. We present an automated text animation scheme, termed "Dynamic Typography", which deforms letters to convey semantic meaning and infuses them with vibrant movements based on user prompts. The animation is represented by a canonical field that aggregates the semantic content in a canonical shape and a deformation field that applies per-frame motion to deform the canonical shape. Two fields are jointly optimized by the priors from a large pretrained text-to-video diffusion model using score-distillation loss with designed regularization, encouraging the video coherence with the intended textual concept while maintaining legibility and structural integrity throughout the animation process. We demonstrate the generalizability of our approach across various text-to-video models and highlight the superiority of our methodology over baselines. Through quantitative and qualitative evaluations, we demonstrate the effectiveness of our framework in generating coherent text animations that faithfully interpret user prompts while maintaining readability.

We present ViTaM-D, a novel visual-tactile framework for reconstructing dynamic hand-object interaction with distributed tactile sensing to enhance contact modeling. Existing methods, relying solely on visual inputs, often fail to capture occluded interactions and object deformation. To address this, we introduce DF-Field, a distributed force-aware contact representation leveraging kinetic and potential energy in hand-object interactions. ViTaM-D first reconstructs interactions using a visual network with contact constraint, then refines contact details through force-aware optimization, improving object deformation modeling. To evaluate deformable object reconstruction, we introduce the HOT dataset, featuring 600 hand-object interaction sequences in a high-precision simulation environment. Experiments on DexYCB and HOT datasets show that ViTaM-D outperforms state-of-the-art methods in reconstruction accuracy for both rigid and deformable objects. DF-Field also proves more effective in refining hand poses and enhancing contact modeling than previous refinement methods. The code, models, and datasets are available at https://sites.google.com/view/vitam-d/.

DUSt3R has recently demonstrated that many tasks in multi-view geometry, including estimating camera intrinsics and extrinsics, reconstructing 3D scenes, and establishing image correspondences, can be reduced to predicting a pair of viewpoint-invariant point maps, i.e., pixel-aligned point clouds defined in a common reference frame. While this formulation is elegant and powerful, it is limited to static scenes. To overcome this limitation, we introduce the concept of Dynamic Point Maps (DPM), which extends standard point maps to support 4D tasks such as motion segmentation, scene flow estimation, 3D object tracking, and 2D correspondence. Our key insight is that, when time is introduced, several possible spatial and temporal references can be used to define the point maps. We identify a minimal subset of these combinations that can be regressed by a network to solve the aforementioned tasks. We train a DPM predictor on a mixture of synthetic and real data and evaluate it across diverse benchmarks, including video depth prediction, dynamic point cloud reconstruction, 3D scene flow, and object pose tracking, achieving state-of-the-art performance.

With the increasing attention to pre-trained vision-language models (VLMs), e.g., CLIP, substantial efforts have been devoted to many downstream tasks, especially in test-time adaptation (TTA). However, previous works focus on learning prototypes only in the textual modality while overlooking the ambiguous semantics in class names. These ambiguities lead to textual prototypes that are insufficient to capture visual concepts, resulting in limited performance. To address this issue, we introduce **ProtoMM**, a training-free framework that constructs multimodal prototypes to adapt VLMs during the test time. By viewing the prototype as a discrete distribution over the textual descriptions and visual particles, ProtoMM has the ability to combine the multimodal features for comprehensive prototype learning. More importantly, the visual particles are dynamically updated as the testing stream flows. This allows our multimodal prototypes to continually learn from the data, enhancing their generalizability in unseen scenarios. In addition, we quantify the importance of the prototypes and test images by formulating their semantic distance as an optimal transport problem. Extensive experiments on 15 zero-shot benchmarks demonstrate the effectiveness of our method, achieving a 1.03% average accuracy improvement over state-of-the-art methods on ImageNet and its variant datasets.

Vision-Language Models (VLM) have emerged as a highly promising approach for Continual Learning (CL) due to their powerful generalized features. While adapter-based VLM can exploit both task-specific and task-agnostic features, current CL methods have largely overlooked the distinct and evolving parameter distributions in visual and language modalities, which are found crucial for effectively mitigating catastrophic forgetting.In this study, we find that the visual modality experiences a broader parameter distribution and greater variance during class increments than the textual modality, leading to higher vulnerability to forgetting.Consequently, we handle the branches of the two modalities asymmetrically. Specifically, we propose a Dynamic Multi-layer Null Space Projection (DMNSP) strategy and apply it only to the visual modality branch, while optimizing the language branch according to the original optimizer. DMNSP can restrict the update of visual parameters within the common subspace of multiple null spaces, further limiting the impact of non-zero residual terms. Simultaneously, combined with a dynamic projection coefficient, we can precisely control the magnitude of gradient projection to the null space, endowing the model with good stability and plasticity.Extensive experiments on TinyImageNet, CIFAR100 and ImageNet-R demonstrate that our method outperforms current approaches in accuracy and knowledge retention, setting a new standard for state-of-the-art performance in class incremental learning.

This paper proposes dynamic human group detection in videos. For detecting complex groups, not only the local appearance features of in-group members but also the global context of the scene are important. Such local and global appearance features in each frame are extracted using a Vision-Language Model (VLM) augmented for group detection in our method. For further improvement, the group structure should be consistent over time. While previous methods are stabilized on the assumption that groups are not changed in a video, our method detects dynamically changing groups by global optimization using a graph with all frames' groupness probabilities estimated by our groupness-augmented CLIP features. Our experimental results demonstrate that our method outperforms state-of-the-art group detection methods on public datasets. Code: https://github.com/irajisamurai/VLM-GroupDetection.git

Remote sensing image segmentation faces persistent challenges in distinguishing morphologically similar categories and adapting to diverse scene variations. While existing methods rely on implicit representation learning paradigms, they often fail to dynamically adjust semantic embeddings according to contextual cues, leading to suboptimal performance in fine-grained scenarios such as cloud thickness differentiation. This work introduces a dynamic dictionary learning framework that explicitly models class ID embeddings through iterative refinement. The core contribution lies in a novel dictionary construction mechanism, where class-aware semantic embeddings are progressively updated via multi-stage alternating cross-attention querying between image features and dictionary embeddings. This process enables adaptive representation learning tailored to input-specific characteristics, effectively resolving ambiguities in intra-class heterogeneity and inter-class homogeneity. To further enhance discriminability, a contrastive constraint is applied to the dictionary space, ensuring compact intra-class distributions while maximizing inter-class separability. Extensive experiments across both coarse- and fine-grained datasets demonstrate consistent improvements over state-of-the-art methods, particularly in two online test benchmarks (LoveDA and UAVid). Code is available at https://github.com/XavierJiezou/D2LS.

In recent years, the introduction of Multi-modal Large Language Models (MLLMs) into video understanding tasks has become increasingly prevalent. However, how to effectively integrate temporal information remains a critical research focus. Traditional approaches treat spatial and temporal information separately. Due to issues like motion blur, it is challenging to accurately represent the spatial information of rapidly moving objects. This can lead to temporally important regions being underemphasized during spatial feature extraction, which in turn hinders accurate spatio-temporal interaction and video understanding. To address this limitation, we propose an innovative video representation method called Dynamic-Image (DynImg). Specifically, we introduce a set of non-key frames as temporal prompts to highlight the spatial areas containing fast-moving objects. During the process of visual feature extraction, these prompts guide the model to pay additional attention to the fine-grained spatial features corresponding to these regions. Moreover, to maintain the correct sequence for DynImg, we employ a corresponding 4D video Rotary Position Embedding. This retains both the temporal and spatial adjacency of DynImg, helping MLLM understand the spatio-temporal order within this combined format. Experimental evaluations reveal that DynImg surpasses the state-of-the-art methods by approximately 2% across multiple video understanding benchmarks, proving the effectiveness of our temporal prompts in enhancing video comprehension.

Blind Face Restoration aims to recover high-fidelity, detail-rich facial images from unknown degraded inputs, presenting significant challenges in preserving both identity and detail. Pre-trained diffusion models have been increasingly used as image priors to generate fine details. Still, existing methods often use fixed diffusion sampling timesteps and a global guidance scale, assuming uniform degradation. This limitation and potentially imperfect degradation kernel estimation frequently lead to under- or over-diffusion, resulting in an imbalance between fidelity and quality. We propose DynFaceRestore, a novel blind face restoration approach that learns to map any blindly degraded input to Gaussian blurry images. By leveraging these blurry images and their respective Gaussian kernels, we dynamically select the starting timesteps for each blurry image and apply closed-form guidance during the diffusion sampling process to maintain fidelity. Additionally, we introduce a dynamic guidance scaling adjuster that modulates the guidance strength across local regions, enhancing detail generation in complex areas while preserving structural fidelity in contours. This strategy effectively balances the trade-off between fidelity and quality. DynFaceRestore achieves state-of-the-art performance in both quantitative and qualitative evaluations, demonstrating robustness and effectiveness in blind face restoration.

Non-prehensile manipulation is crucial for handling objects that are too thin, large, or otherwise ungraspable in unstructured environments. While conventional planning-based approaches struggle with complex contact modeling, learning-based methods have recently emerged as a promising alternative. However, existing learning-based approaches face two major limitations: they heavily rely on multi-view cameras and precise pose tracking, and they fail to generalize across varying physical conditions, such as changes in object mass and table friction. To address these challenges, we propose the Dynamics-Adaptive World Action Model (DyWA), a novel framework that enhances action learning by jointly predicting future states while adapting to dynamics variations based on historical trajectories. By unifying the modeling of geometry, state, physics, and robot actions, DyWA enables more robust policy learning under partial observability. Compared to baselines, our method improves the success rate by 31.5% using only single-view point cloud observations in the simulation. Furthermore, DyWA achieves an average success rate of 68% in real-world experiments, demonstrating its ability to generalize across diverse object geometries, adapt to varying table friction, and robustness in challenging scenarios such as half-filled water bottles and slippery surfaces.

In dynamic scenes, achieving accurate camera localization and reconstructing a long-term consistent map containing only the static background are two major challenges faced by Visual Simultaneous Localization and Mapping (VSLAM). In current traditional dynamic VSLAM systems, the methods used to handle dynamic objects are primarily designed for localization; if applied to reconstruction, they are prone to introducing motion artifacts. Meanwhile, mask compensation strategies in NeRF- or 3DGS-based dynamic VSLAM systems also face challenges, such as the inability to completely eliminate dynamic object artifacts and low real-time performance. To address these issues, we leverage object detection to extract semantic information and propose a dynamic feature detection algorithm based on both geometry and appearance. This algorithm accurately identifies known and unknown moving objects and determines their actual motion states. To mitigate the issue of insufficient detection box coverage, we design a dynamic object box correction algorithm based on clustering and Gaussian mixture models to comprehensively identify moving object regions. Furthermore, to overcome the limitations of sparse features in texture-scarce environments, we introduce a feature densification strategy based on image texture complexity, enhancing reconstruction quality while maintaining real-time performance. Extensive experimental evaluations demonstrate that our system achieves state-of-the-art localization and reconstruction performance in dynamic scenes and can run in real time on resource-constrained devices.

We tackle the challenge of jointly personalizing content and style from a few examples. A promising approach is to train separate Low-Rank Adapters (LoRA) and merge them effectively, preserving both content and style. Existing methods, such as ZipLoRA, treat content and style as independent entities, merging them by learning masks in LoRA's output dimensions. However, content and style are intertwined, not independent. To address this, we propose DuoLoRA--a content-style personalization framework featuring three key components: (1) rank-dimension mask learning, (2) effective merging via layer priors, and (3) Constyle loss, which leverages cycle-consistency in the merging process.First, we introduce ZipRank, which performs content-style merging within the rank dimension, offering adaptive rank flexibility and significantly reducing the number of learnable parameters. Additionally, we incorporate SDXL layer priors to apply implicit rank constraints informed by each layer's content-style bias and adaptive merger initialization, enhancing the integration of content and style. To further refine the merging process, we introduce Constyle loss, which leverages the cycle consistency between content and style.Our experimental results demonstrate that DuoLoRA outperforms state-of-the-art content-style merging methods across multiple benchmarks.

In this paper, a contrastive representation learning framework is proposed to enhance human action segmentation via pre-training using trimmed (single action) skeleton sequences. Unlike previous representation learning works that are tailored for action recognition and that build upon isolated sequence-wise representations, the proposed framework focuses on exploiting multi-scale representations in conjunction with cross-sequence variations. More specifically, it proposes a novel data augmentation strategy, "Shuffle and Warp", which exploits diverse multi-action permutations. The latter effectively assists two surrogate tasks that are introduced in contrastive learning: Cross Permutation Contrasting (CPC) and Relative Order Reasoning (ROR). In optimization, CPC learns intra-class similarities by contrasting representations of the same action class across different permutations, while ROR reasons about inter-class contexts by predicting relative mapping between two permutations. Together, these tasks enable a Dual-Surrogate Contrastive Learning (DuoCLR) network to learn multi-scale feature representations optimized for action segmentation. In experiments, DuoCLR is pre-trained on a trimmed skeleton dataset and evaluated on an untrimmed dataset where it demonstrates a significant boost over state-the-art comparatives in both multi-class and multi-label action segmentation tasks. Lastly, ablation studies are conducted to evaluate the effectiveness of each component of the proposed approach.

Customized text-to-video generation with pre-trained large-scale models has recently garnered significant attention by focusing on identity and motion consistency. Existing works typically follow the isolated customized paradigm, where the subject identity or motion dynamics are customized exclusively. However, this paradigm completely ignores the intrinsic mutual constraints and synergistic interdependencies between identity and motion, resulting in identity-motion conflicts throughout the generation process that systematically degrade. To address this, we introduce DualReal, a novel framework that employs adaptive joint training to construct interdependencies between dimensions collaboratively. Specifically, DualReal is composed of two units: (1) Dual-aware Adaptation dynamically switches the training step (i.e., identity or motion), learns the current information guided by the frozen dimension prior, and employs a regularization strategy to avoid knowledge leakage; (2) StageBlender Controller leverages the denoising stages and Diffusion Transformer depths to guide different dimensions with adaptive granularity, avoiding conflicts at various stages and ultimately achieving lossless fusion of identity and motion patterns. We constructed a more comprehensive evaluation benchmark than existing methods. The experimental results show that DualReal improves CLIP-I and DINO-I metrics by 21.7% and 31.8% on average, and achieves top performance on nearly all motion metrics. Page: https://wenc-k.github.io/dualreal-customization

Images captured under severe weather conditions often suffer from complex, composite degradations, varying in intensity. In this paper, we introduce a novel method, Dual-Level Prototype Learning (DPL), to tackle the challenging task of composite degraded image restoration. Unlike previous methods that rely on fixed embeddings to characterize degradation types, DPL maintains a number of degradation-level prototypes to represent the specific degradation scenes dynamically. Furthermore, considering the diverse factors influencing each degradation type, factor-level prototypes are incorporated to capture variations in individual degradation factors. Image features are matched with both degradation-level and factor-level prototypes, producing detailed scene embeddings that enhance the network's understanding of composite degradations. These scene embeddings are then processed through Dual Scene Embedding Transformer Blocks to guide the restoration process. To further refine the prototype distribution, we propose a Prototype Scatter Learning Loss, which enables prototypes within the same degradation to learn more information and push prototypes between different degradations to be separate. Additionally, we introduce a new dataset named Variable Composite Degradation (VCD) dataset which contains images with different intensities of each type of composite degradation to validate the efficacy of our method. Extensive experiments demonstrate that DPL significantly outperforms existing methods in restoring images with composite degradations.

Video semantic segmentation aims to assign a class label for each pixel in every video frame. Existing methods predominantly follow the reference-target interaction paradigm, focusing on extracting local temporal contexts while neglecting the integration of global temporal information. Moreover, complex dynamics and varying lighting conditions introduce inter-frame intra-class discrepancies in feature representations, leading to unstable predictions. In this paper, we propose a novel framework, the Dual-Temporal Exemplar Representation Network (DTERN), which utilizes the strong representational capability of cluster centers, i.e., exemplars, to effectively model both local and global temporal information. DTERN consists of two core modules: 1) the Local Temporal Exemplar Module (LTEM), which constructs local exemplars to capture local temporal contexts, ensuring stable and reliable predictions. 2) the Global Temporal Exemplar Module (GTEM), which introduces learnable global exemplars to dynamically model global temporal information, thereby improving the effective consistency of segmentation. Furthermore, we observe that the existing Video Consistency (VC) metric fails to evaluate segmentation accuracy and lacks sensitivity to small-object segmentation. To this end, we propose Video Effective Consistency (VEC) to comprehensively evaluate temporal consistency and segmentation effectiveness. Experiments on VSPW and Cityscape demonstrate that DTERN outperforms state-of-the-art methods. The code is available at https://github.com/zlxilo/DTERN.

Single-view 3D reconstruction aims to recover the complete 3D geometry and appearance of objects from a single RGB image. Due to incomplete image information and ambiguity, this task remains challenging. Existing methods struggle with the trade-off between local detail and global topology, and with interference from early RGB-depth fusion in signed distance function optimization. To address these challenges, we propose Dual-S3D, a novel framework for single-view 3D reconstruction. Our method employs a hierarchical dual-path feature extraction strategy based on stages that utilize convolutional neural networks to anchor local geometric details. In contrast, subsequent stages leverage a Transformer integrated with selective state-space model to capture global topology, enhancing scene understanding and feature representation. Additionally, we design an auxiliary branch that progressively fuses precomputed depth features with pixel-level features to decouple visual and geometric cues effectively. Extensive experiments on the 3D-FRONT and Pix3D datasets demonstrate that our approach significantly outperforms existing methods--reducing chamfer distance by 51%, increasing F-score by 33.6%, and improving normal consistency by 10.3%--thus achieving state-of-the-art reconstruction quality.