ICCV 2025 Papers

In this paper, we explore the potential of visual in-context learning to enable a single model to handle multiple tasks and adapt to new tasks during test time without re-training. Unlike previous approaches, our focus is on training in-context learners to adapt to sequences of tasks, rather than individual tasks. Our goal is to solve complex tasks that involve multiple intermediate steps using a single model, allowing users to define entire vision pipelines flexibly at test time. To achieve this, we first examine the properties and limitations of visual in-context learning architectures, with a particular focus on the role of codebooks. We then introduce a novel method for training in-context learners using a synthetic compositional task generation engine. This engine bootstraps task sequences from arbitrary segmentation datasets, enabling the training of visual in-context learners for compositional tasks. Additionally, we investigate different masking-based training objectives to gather insights into how to train models better for solving complex, compositional tasks. Our exploration not only provides important insights especially for multi-modal medical task sequences but also highlights challenges that need to be addressed.

Visual object tracking and segmentation are becoming fundamental tasks for understanding human activities in egocentric vision. Recent research has benchmarked state-of-the-art methods and concluded that first person egocentric vision presents challenges compared to previously studied domains. However, these claims are based on evaluations conducted across significantly different scenarios. Many of the challenging characteristics attributed to egocentric vision are also present in third person videos of human-object activities. This raises a critical question: how much of the observed performance drop stems from the unique first person viewpoint inherent to egocentric vision versus the domain of human-object activities? To address this question, we introduce a new benchmark study designed to disentangle such factors. Our evaluation strategy enables a more precise separation of challenges related to the first person perspective from those linked to the broader domain of human-object activity understanding. By doing so, we provide deeper insights into the true sources of difficulty in egocentric tracking and segmentation, facilitating more targeted advancements on this task.

Meta-learning is a powerful paradigm for tackling few-shot tasks. However, recent studies indicate that models trained with the whole-class training strategy can achieve comparable performance to those trained with meta-learning in few-shot classification tasks. To demonstrate the value of meta-learning, we establish an entropy-limited supervised setting for fair comparisons. Through both theoretical analysis and experimental validation, we establish that meta-learning has a tighter generalization bound compared to whole-class training. We unravel that meta-learning is more efficient with limited entropy and is more robust to label noise and heterogeneous tasks, making it well-suited for unsupervised tasks. Based on these insights, We propose MINO, a meta-learning framework designed to enhance unsupervised performance. MINO utilizes the adaptive clustering algorithm DBSCAN with a dynamic head for unsupervised task construction and a stability-based meta-scaler for robustness against label noise. Extensive experiments confirm its effectiveness in multiple unsupervised few-shot and zero-shot tasks.

Contrastive Language-Image Pretraining (CLIP) excels at learning generalizable image representations but often falls short in zero-shot inference on certain downstream datasets. Test-time adaptation (TTA) mitigates this issue by adjusting components like normalization layers or context prompts, yet it typically requires large batch sizes and extensive augmentations, leading to high computational costs. This raises a key question: Can VLMs' performance drop in specific test cases be mitigated through efficient, training-free approaches? To explore the solution, we investigate token condensation (TC) techniques, originally designed to enhance vision transformer efficiency by refining token usage during inference. We observe that informative tokens improve visual-text alignment in VLMs like CLIP on unseen datasets. However, existing TC methods often fail to maintain in-distribution performance when reducing tokens, prompting us to ask: How can we transform TC into an effective "free-lunch" adaptation strategy for VLMs? To address this, we propose Token Condensation as Adaptation (TCA), a training-free adaptation method that takes a step beyond standard TC. Rather than passively discarding tokens, TCA condenses token representation by introducing reservoir-based domain anchor tokens for information-preserving token reduction and logit correction. TCA achieves up to a 21.4% performance improvement over the strongest baseline on cross-dataset benchmark and the CIFAR-100-Corrupted dataset while reducing GFLOPs by 12.2% to 48.9%, with minimal hyperparameter dependency on both CLIP and SigLIP series. Code is available at https://github.com/Jo-wang/TCA.

Contrastive Language-Image Pre-Training (CLIP) is a popular method for learning multimodal latent spaces with well-organized semantics. Despite its wide range of applications, CLIP's latent space is known to fail at handling complex visual-textual interactions. Recent works attempt to address its shortcomings with data-centric or algorithmic approaches. But what if the problem is more fundamental, and lies in the geometry of CLIP? Toward this end, we rigorously analyze CLIP's latent space properties, and prove that no CLIP-like joint embedding space exists which can correctly do any two of the following at the same time: 1. represent basic descriptions and image content, 2. represent attribute binding, 3. represent spatial location and relationships, 4. represent negation. Informed by this analysis, we propose Dense Cosine Similarity Maps (DCSMs) as a principled and interpretable scoring method for CLIP-like models, which solves the fundamental limitations of CLIP by retaining the semantic topology of the image patches and text tokens. This method improves upon the performance of classical CLIP-like joint encoder models on a wide array of benchmarks. We share our code and data here: https://github.com/Raphoo/DCSM_Ideal_CLIP

Digital agents are increasingly employed to automate tasks in interactive digital environments such as web pages, software applications, and operating systems. While text-based agents built on Large Language Models (LLMs) often require frequent updates due to platform-specific APIs, visual agents leveraging Multimodal Large Language Models (MLLMs) offer enhanced adaptability by interacting directly with Graphical User Interfaces (GUIs). However, these agents face significant challenges in visual perception, particularly when handling high-resolution, visually complex digital environments. This paper introduces Iris, a foundational visual agent that addresses these challenges through two key innovations: Information-Sensitive Cropping (ISC) and Self-Refining Dual Learning (SRDL). ISC dynamically identifies and prioritizes visually dense regions using an edge detection algorithm, enabling efficient processing by allocating more computational resources to areas with higher information density. SRDL enhances the agent's ability to handle complex tasks by leveraging a dual-learning loop, where improvements in referring (describing UI elements) reinforce grounding (locating elements) and vice versa, all without requiring additional annotated data. Empirical evaluations demonstrate that Iris achieves state-of-the-art performance across multiple benchmarks with only 850K GUI annotations, outperforming methods using 10x more training data. These improvements further translate to significant gains in both web and OS agent downstream tasks.

With the increasing demand for the right to be forgotten, machine unlearning (MU) has emerged as a vital tool for enhancing trust and regulatory compliance by enabling the removal of sensitive data influences from machine learning (ML) models. However, most MU algorithms primarily rely on in-training methods to adjust model weights, with limited exploration of the benefits that data-level adjustments could bring to the unlearning process. To address this gap, we propose a novel approach that leverages digital watermarking to facilitate MU by strategically modifying data content. By integrating watermarking, we establish a controlled unlearning mechanism that enables precise removal of specified data while maintaining model utility for unrelated tasks. We first examine the impact of watermarked data on MU, finding that MU effectively generalizes to watermarked data. Building on this, we introduce an unlearning-friendly watermarking framework, termed Water4MU, to enhance unlearning effectiveness. The core of Water4MU is a bi-level optimization (BLO) framework: at the upper level, the watermarking network is optimized to minimize unlearning difficulty, while at the lower level, the model itself is trained independently of watermarking. Experimental results demonstrate that Water4MU is effective in MU across both image classification and image generation tasks. Notably, it outperforms existing methods in challenging MU scenarios, known as "challenging forgets".

A concept of light-fields computed from multiple view images on regular grids has proven its benefit for scene representations, and supported realistic renderings of novel views and photographic effects such as refocusing and shallow depth of field. In spite of its effectiveness of light flow computations, obtaining light fields requires either computational costs or specialized devices like a bulky camera setup and a specialized microlens array. In an effort to broaden its benefit and applicability, in this paper, we propose a novel view synthesis method for light field generation from only single images, named inverse image-based rendering. Unlike previous attempts to implicitly rebuild 3D geometry or to explicitly represent objective scenes, our method reconstructs light flows in a space from image pixels, which behaves in the opposite way to image-based rendering. To accomplish this, we design a neural rendering pipeline to render a target ray in an arbitrary viewpoint. Our neural renderer first stores the light flow of source rays from the input image, then computes the relationships among them through cross-attention, and finally predicts the color of the target ray based on these relationships. After the rendering pipeline generates the first novel view from a single input image, the generated out-of-view contents are updated to the set of source rays, and this procedure is iteratively performed while ensuring the consistent generation of occluded contents. We demonstrate that our inverse image-based rendering works well with various challenging datasets without any retraining or finetuning after once trained on synthetic dataset. In addition, our method outperforms relevant state-of-the-art novel view synthesis methods.

Traditional methods for biological shape inference, such as deep learning (DL) and active contour models, face important limitations in 3D. DL approaches require large annotated datasets, which are often impractical to obtain, while active contour methods depend on carefully tuned heuristics for intensity attraction and shape regularization. We introduce deltaMic, a novel differentiable 3D renderer for fluorescence microscopy that formulates shape inference as an inverse problem. By leveraging differentiable convolutions, deltaMic simulates the image formation process, integrating a parameterized point spread function (PSF) with a triangle mesh-based representation of biological structures. Unlike DL- or contour-based segmentation, deltaMic directly optimizes both shape and optical parameters to align synthetic and real microscopy images, removing the need for large datasets or sample-specific fine-tuning. To ensure scalability, we implement a GPU-accelerated Fourier transform for triangle meshes along with narrow-band spectral filtering. We show that deltaMic accurately reconstructs cell geometries from both synthetic and diverse experimental 3D microscopy data, while remaining robust to noise and initialization. This establishes a new physics-informed framework for biophysical image analysis and inverse modeling.

We present InvRGB+L, a novel inverse rendering model that reconstructs large, relightable, and dynamic scenes from a single RGB+LiDAR sequence. Conventional inverse graphics methods rely primarily on RGB observations and use LiDAR mainly for geometric information, often resulting in suboptimal material estimates due to visible light interference. We find that LiDAR's intensity values--captured with active illumination in a different spectral range--offer complementary cues for robust material estimation under variable lighting. Inspired by this, InvRGB+L leverages LiDAR intensity cues to overcome challenges inherent in RGB-centric inverse graphics through two key innovations: (1) a novel physics-based LiDAR shading model and (2) RGB-LiDAR material consistency losses. The model produces novel-view RGB and LiDAR renderings of urban and indoor scenes and supports relighting, night simulations, and dynamic object insertions--achieving results that surpass current state-of-the-art methods in both scene-level urban inverse rendering and LiDAR simulation.

Text-to-image (T2I) models have recently gained widespread adoption. This has spurred concerns about safeguarding intellectual property rights and an increasing demand for mechanisms that prevent the generation of specific artistic styles. Existing methods for style extraction typically necessitate the collection of custom datasets and the training of specialized models. This, however, is resource-intensive, time-consuming, and often impractical for real-time applications. We present a novel, training-free framework to solve the style attribution problem, using the features produced by a diffusion model alone, without any external modules or retraining. This is denoted as Introspective Style attribution(IntroStyle) and is shown to have superior performance to state-of-the-art models for style attribution. We also introduce a synthetic dataset of Artistic Style Split (ArtSplit) to isolate artistic style and evaluate fine-grained style attribution performance. Our experimental results on WikiArt and DomainNet datasets show that \ours is robust to the dynamic nature of artistic styles, outperforming existing methods by a wide margin.

Realistic images are usually produced by simulating light transportation results of 3D scenes using rendering engines. This framework can precisely control the output but is usually weak at producing photo-like images. Alternatively, diffusion models have seen great success in photorealistic image generation by leveraging priors from large datasets of real-world images but lack affordance controls. Promisingly, the recent ControlNet enables flexible control of the diffusion model without degrading its generation quality. In this work, we introduce IntrinsicControlNet, an intrinsically controllable image generation framework that enables easily generating photorealistic images from precise and explicit control, similar to a rendering engine, by using intrinsic images such as material properties, geometric details, and lighting as network inputs. Beyond this, we notice that there is a domain gap between the synthetic and real-world datasets, and therefore, naively blending these datasets yields domain confusion. To address this problem, we present a cross-domain control architecture that extracts control information from synthetic datasets, and control and content information from real-world datasets. This bridges the domain gap between real-world and synthetic datasets, enabling the blending or editing of 3D assets and real-world photos to support various interesting applications. Experiments and user studies demonstrate that our method can generate explicitly controllable and highly photorealistic images based on the input intrinsic images.

In this paper, we address the problem of novel class discovery (NCD), which aims to cluster novel classes by leveraging knowledge from disjoint known classes. While recent advances have made significant progress in this area, existing NCD methods face two major limitations. First, they primarily focus on single-view data (e.g., images), overlooking the increasingly common multi-view data, such as multi-omics datasets used in disease diagnosis. Second, their reliance on pseudo-labels to supervise novel class clustering often results in unstable performance, as pseudo-label quality is highly sensitive to factors such as data noise and feature dimensionality. To address these challenges, we propose a novel framework named Intra-view and Inter-view Correlation Guided Multi-view Novel Class Discovery (IICMVNCD), which is the first attempt to explore NCD in multi-view setting so far. Specifically, at the intra-view level, leveraging the distributional similarity between known and novel classes, we employ matrix factorization to decompose features into view-specific shared base matrices and factor matrices. The base matrices capture distributional consistency among the two datasets, while the factor matrices model pairwise relationships between samples. At the inter-view level, we utilize view relationships among known classes to guide the clustering of novel classes. This includes generating predicted labels through the weighted fusion of factor matrices and dynamically adjusting view weights of known classes based on the supervision loss, which are then transferred to novel class learning. Experimental results validate the effectiveness of our proposed approach.

Recent deepfake detection algorithms focus solely on uni-modal or cross-modal inconsistencies. While the former disregards audio-visual correspondence entirely rendering them less effective against multimodal attacks, the latter overlooks inconsistencies in a particular modality. Moreover, many models are single-stage supervised frameworks, effective on specific training data but less generalizable to new manipulations. To address these gaps, we propose a two-stage multimodal framework that first learns intra-modal and cross-modal temporal synchronization on real videos, capturing audio-visual correspondences crucial for deepfake detection and localization. We introduce a Gaussian-targeted loss in our pretraining model to focus on learning relative synchronization patterns across multimodal pairs. Using pretrained features, our approach not only enables classification on fully manipulated videos but also supports a localization module for partial deepfakes with only specific segments spoofed. Moreover, the pretraining stage does not require fine-tuning, thus reducing complexity. Our model, tested on various benchmark datasets, demonstrates strong generalization and precise temporal localization.

Recent advances in deep learning have led to increasingly complex models with deeper layers and more parameters, reducing interpretability and making their decisions harder to understand. While many methods explain black-box reasoning, most lack effective interventions or only operate at sample-level without modifying the model itself. To address this, we propose the Concept Bottleneck Model for Enhancing Human-Neural Network Mutual Understanding (CBM-HNMU). CBM-HNMU leverages the Concept Bottleneck Model (CBM) as an interpretable framework to approximate black-box reasoning and communicate conceptual understanding. Detrimental concepts are automatically identified and refined (removed/replaced) based on global gradient contributions. The modified CBM then distills corrected knowledge back into the black-box model, enhancing both interpretability and accuracy. We evaluate CBM-HNMU on various CNN and transformer-based models across Flower-102, CIFAR-10, CIFAR-100, FGVC-Aircraft, and CUB-200, achieving a maximum accuracy improvement of 2.64% and a maximum increase in average accuracy across 1.03%. Source code is available at: https://github.com/XiGuaBo/CBM-HNMU.

Understanding 3D cell shape is crucial in biomedical research, where morphology serves as a key indicator of disease, cellular state, and drug response. However, many existing 3D point cloud classification models lack interpretability, limiting their utility for extracting biologically meaningful insights. In this work, we unify standard point cloud backbones and feature aggregation strategies within a Multiple Instance Learning (MIL) framework to enable inherently interpretable classification. Our approach, PointMIL, improves classification performance while providing fine-grained point-level explanations without relying on post hoc analysis. We demonstrate state-of-the-art mACC (97.3%) and F1 (97.5%) in the IntrA biomedical dataset and evaluate the interpretability using quantitative and qualitative metrics. Additionally, we introduce ATLAS-1, a novel dataset of drug-treated 3D cancer cells, and use it to show how PointMIL captures fine-grained morphological effects of chemical treatments. Beyond biomedical applications, PointMIL generalises to standard benchmarks such as ModelNet40 and ScanObjectNN, offering interpretable 3D object recognition across domains

Large-scale vision-language models (VLMs), such as CLIP, have achieved remarkable success in zero-shot learning (ZSL) by leveraging large-scale visual-text pair datasets. However, these methods often lack interpretability, as they compute the similarity between an entire query image and the embedded category words, making it difficult to explain their predictions. One approach to address this issue is to develop interpretable models by integrating language, where classifiers are built using discrete attributes, similar to human perception. This introduces a new challenge: how to effectively align local visual features with corresponding attributes based on pre-trained VLMs. To tackle this, we propose LaZSL, a locally-aligned vision-language model for interpretable ZSL. LaZSL employs local visual-semantic alignment via optimal transport to perform interaction between visual regions and their associated attributes, facilitating effective alignment and providing interpretable similarity without the need for additional training. Extensive experiments demonstrate that our method offers several advantages, including enhanced interpretability, improved accuracy, and strong domain generalization. Codes available at: https://github.com/shiming-chen/LaZSL.

Language-guided Affordance Segmentation (LASO) aims to identify actionable object regions based on text instructions. At the core of its practicality is learning generalizable affordance knowledge that captures functional regions across diverse objects. However, current LASO solutions struggle to extend learned affordances to object categories that are not encountered during training. Scrutinizing these designs, we identify limited generalizability on unseen categories, stemming from (1) underutilized generalizable patterns in the intermediate layers of both 3D and text backbones, which impedes the formation of robust affordance knowledge, and (2) the inability to handle substantial variability in affordance regions across object categories due to a lack of structural knowledge of the target region.Towards this, we introduce a GeneraLized frAmework on uNseen CategoriEs (GLANCE), incorporating two key components: a cross-modal connector that links intermediate stages of the text and 3D backbones to enrich pointwise embeddings with affordance concepts, and a VLM-guided query generator that provides affordance priors by extracting a few 3D key points based on the intra-view reliability and cross-view consistency of their multi-view segmentation masks. Extensive experiments on two benchmark datasets demonstrate that GLANCE outperforms state-of-the-art methods (SoTAs), with notable improvements in generalization to unseen categories. Our code is available at https://anonymous.4open.science/r/GLANCE.

Motion planning is a crucial component of autonomous robot driving. While various trajectory datasets exist, effectively utilizing them for a target domain remains challenging due to differences in agent interactions and environmental characteristics. Conventional approaches, such as domain adaptation or ensemble learning, leverage multiple source datasets but suffer from domain imbalance, catastrophic forgetting, and high computational costs. To address these challenges, we propose Interaction-Merged Motion Planning (IMMP), a novel approach that leverages parameter checkpoints trained on different domains during adaptation to the target domain. IMMP follows a two-step process: pre-merging to capture agent behaviors and interactions, sufficiently extracting diverse information from the source domain, followed by merging to construct an adaptable model that efficiently transfers diverse interactions to the target domain. Our method is evaluated on various planning benchmarks and models, demonstrating superior performance compared to conventional approaches.