ICCV 2025 Papers

In this paper, we challenge the conventional practice in Open-Vocabulary Semantic Segmentation (OVSS) of using averaged class-wise text embeddings, which are typically obtained by encoding each class name with multiple templates (e.g., a photo of <class>, a sketch of a <class>). We investigate the impact of templates for OVSS, and find that for each class, there exist single-template classifiers---which we refer to as class-experts--- that significantly outperform the conventional averaged classifier.First, to identify these class-experts, we introduce a novel approach that estimates them without any labeled data or training. By leveraging the class-wise prediction entropy of single-template classifiers, we select those yielding the lowest entropy as the most reliable class-experts. Second, we combine the outputs of class-experts in a new fusion process. Our plug-and-play method, coined FLOSS, is orthogonal and complementary to existing OVSS methods, offering an improvement without the need for additional labels or training. Extensive experiments show that FLOSS consistently enhances state-of-the-art OVSS models, generalizes well across datasets with different distribution shifts, and delivers substantial improvements in low-data scenarios where only a few unlabeled images are available. Our code is available at https://github.com/yasserben/FLOSS.

With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results. However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature. This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model. Instead of a pixel-based latent space, we take advantage of a learned orthogonal motion latent space, enabling efficient generation and editing of temporally consistent motion. To achieve this, we introduce a transformer-based vector field predictor with an effective frame-wise conditioning mechanism. Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions. Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency.

Multi-modal anomaly detection (MAD) improves industrial inspection by exploiting complementary 2D and 3D data. However, existing methods struggle in few-shot scenarios due to limited data and modality gaps. Current approaches either fuse multi-modal features or align cross-modal representations; however, they often suffer from high false-positive rates and fail to detect subtle defects, especially when training samples are scarce. To address these challenges, we propose the first few-shot MAD method FIND, a novel dual-student framework that integrates intra-modal reverse distillation and cross-modal feature mapping. FIND employs modality-specific teachers and two collaborative students: an intra-modal student for fine-grained anomaly localization via reverse distillation, and a cross-modal student that captures inter-modal correspondences to detect cross-modal inconsistencies. Extensive experiments on MVTec-3D-AD and Eyecandies show that FIND significantly outperforms state-of-the-art methods in both full-shot and few-shot settings. Ablation studies validate the complementary roles of intra- and cross-modal distillation. Our work significantly advances MAD robustness in data-scarce industrial applications.

Layout-to-image (L2I) generation has exhibited promising results in natural domains, but suffers from limited generative fidelity and weak alignment with user-provided layouts when applied to degraded scenes (i.e., low-light, underwater). We primarily attribute these limitations to the "contextual illusion dilemma" in degraded conditions, where foreground instances are overwhelmed by context-dominant frequency distributions. Motivated by this, our paper proposes a new Frequency-Inspired Contextual Disentanglement Generative (FICGen) paradigm, which seeks to transfer frequency knowledge of degraded images into the latent diffusion space, thereby facilitating the rendering of degraded instances and their surroundings via contextual frequency-aware guidance. To be specific, FICGen consists of two major steps. Firstly, we introduce a learnable dual-query mechanism, each paired with a dedicated frequency resampler, to extract contextual frequency prototypes from pre-collected degraded exemplars in the training set. Secondly, a visual-frequency enhanced attention is employed to inject frequency prototypes into the degraded generation process. To alleviate the contextual illusion and attribute leakage, an instance coherence map is developed to regulate latent-space disentanglement between individual instances and their surroundings, coupled with an adaptive spatial-frequency aggregation module to reconstruct spatial-frequency mixed degraded representations. Extensive experiments on 5 benchmarks involving a variety of degraded scenarios--from severe low-light to mild blur--demonstrate that FICGen consistently surpasses existing L2I methods in terms of generative fidelity, alignment and downstream auxiliary trainability.

Incremental unlearning (IU) is critical for pre-trained models to comply with sequential data deletion requests, yet existing methods primarily suppress parameters or confuse knowledge without explicit constraints on both feature and gradient level, resulting in superficial forgetting where residual information remains recoverable. This incomplete forgetting risks security breaches and disrupts retention balance, especially in IU scenarios. We propose FG-OrIU (Feature-Gradient Orthogonality for Incremental Unlearning), the first framework unifying orthogonal constraints on both features and gradients level to achieve deep forgetting, where the forgetting effect is irreversible. FG-OrIU decomposes feature spaces via Singular Value Decomposition (SVD), separating forgetting and remaining class features into distinct subspaces. It then enforces dual constraints: feature orthogonal projection on both forgetting and remaining classes, while gradient orthogonal projection prevents the reintroduction of forgotten knowledge and disruption to remaining classes during updates. Additionally, dynamic subspace adaptation merges newly forgetting subspaces and contracts remaining subspaces, ensuring a stable balance between removal and retention across sequential unlearning tasks. Extensive experiments demonstrate the effectiveness of our method.

Modern machine learning models, that excel on computer vision tasks such as classification and object detection, are often overconfident in their predictions for Out-of-Distribution (OOD) examples, resulting in unpredictable behaviour for open-set environments. Recent works have demonstrated that the free energy score is an effective measure of uncertainty for OOD detection given its close relationship to the data distribution. However, despite that free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation such that in-distribution and OOD instances can have distinct feature representations yet identical free energy scores. This phenomenon occurs when the vector representing the feature space difference between the in-distribution and OOD sample belongs to the null space of the last layer of a neural-based classifier. To mitigate these issues, we explore lower-dimensional feature spaces to reduce the null space footprint and introduce novel regularisation to maximize the least singular value of the final linear layer, hence enhancing inter-sample free energy separation. We refer to these techniques as Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). Extensive experimentation shows that FEVER-OOD improves OOD detection in CIFAR-10 with 28.22% FPR95 and 94.78 AUROC vs 33.66% FPR95 and 92.15 AUROC of the baseline model, CIFAR-100 with 42.77% FPR95 and 89.98 AUROC vs 50.96% FPR95 and 88.06 AUROC of the baseline model, and achieves state of the art OOD detection in Imagenet-100, with average OOD FPR95 of 36.50% and an AUROC of 92.74 when used with the Dream-OOD model, compared with a 39.33% and 91.84 AUROC without FEVER-OOD.

Micro-expressions (MEs) are brief, low-intensity, often localized facial expressions. They could reveal genuine emotions individuals may attempt to conceal, valuable in contexts like criminal interrogation and psychological counseling. However, ME recognition (MER) faces challenges, such as small sample sizes and subtle features, which hinder efficient modeling. Additionally, real-world applications encounter ME data privacy issues, leaving the task of enhancing recognition across settings under privacy constraints largely unexplored. To address these issues, we propose a FED-PsyAU research framework. We begin with a psychological study on the coordination of upper and lower facial action units (AUs) to provide structured prior knowledge of facial muscle dynamics. We then develop a DPK-GAT network that combines these psychological priors with statistical AU patterns, enabling hierarchical learning of facial motion features from regional to global levels, effectively enhancing MER performance. Additionally, our federated learning framework advances MER capabilities across multiple clients without data sharing, preserving privacy and alleviating the limited-sample issue for each client. Extensive experiments on commonly-used ME databases demonstrate the effectiveness of our approach. Our implementation is publicly available at https://github.com/MELABIPCAS/FED-PsyAU.git.

Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is challenging since the models need to generalize to anomalies across different domains. Recently, CLIP-based anomaly detection methods, such as WinCLIP and AnomalyCLIP, have demonstrated superior performance in the ZSAD task, due to the strong zero-shot recognition of the CLIP model. However, they overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, since some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information. To this end, We propose Frequency Enhanced CLIP (FE-CLIP), taking advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction adapter, and (2) Local Frequency Statistics adapter, in the visual encoder of CLIP, to deeply mine frequency information for the ZSAD task. We apply DCT as the frequency-domain transformation. Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, the proposed FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings.

General-purpose Vision-Language Models (VLMs) have driven major advancements in multimodal AI. Fine-tuning these models with task-specific data enhances adaptability to various downstream tasks but suffers from privacy risks. While potential solutions like federated learning can address user data privacy concerns, model protection is also essential. Other methods that rely on black-box VLM APIs usually require the access of prediction logits, leaving them open to inversion attacks. Moreover, addressing the challenges of tuning complexity and data transmission efficiency in federated VLM scenarios is also crucial. To address these challenges, we propose FDPT--a federated discrete prompt tuning method utilizing black-box VLMs. During client optimization stage, FDPT employs an agent-driven framework leveraging large language models (LLMs) with enhanced reasoning capacities to systematically optimize discrete prompt representations, and also utilizes feedback mechanisms and chain of thought to enhance prediction accuracy. Importantly, it performs optimization by relying not on the predicted logic vectors output by LLMs but on textual results, avoiding reverse attack risks. During global aggregation stage, We mimic human electoral activities by employing evolutionary computation methods underpinned by semantic similarity computation to implement enhanced zero-order optimization for acquiring representative global tokens, thereby achieving knowledge aggregation. FDPT significantly outperforms nine state-of-the-art methods in image classification and visual question-answering, reducing communication overhead while generating highly transferable optimized prompts. Additionally, it exhibits improved robustness to data heterogeneity.

The temporal interpolation task for 4D medical imaging, plays a crucial role in clinical practice of respiratory motion modeling. Following the simplified linear-motion hypothesis, existing approaches adopt optical flow-based models to interpolate intermediate frames. However, realistic respiratory motions should be nonlinear and quasi-periodic with specific frequencies. Intuited by this property, we resolve the temporal interpolation task from the frequency perspective, and propose a Fourier Basis-guided Diffusion model, termed FB-Diff. Specifically, due to the regular motion discipline of respiration, physiological motion priors are introduced to describe general characteristics of temporal data distributions. Then a Fourier motion operator is elaborately devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space of Variational Autoencoder. Well-learned Fourier bases can better simulate respiratory motions with motion patterns of specific frequencies. Conditioned on starting and ending frames, the diffusion model further leverages well-learned Fourier bases via the basis interaction operator, which promotes the temporal interpolation task in a generative manner. Extensive results demonstrate that FB-Diff achieves state-of-the-art (SOTA) perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Codes are available at https://github.com/AlexYouXin/FB-Diff.

The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens.

Pre-trained vision-language models (VLMs) have advanced out-of-distribution (OOD) detection recently. However, existing CLIP-based methods often focus on learning OOD-related knowledge to improve OOD detection, showing limited generalization or reliance on external large-scale auxiliary datasets. In this study, instead of delving into the intricate OOD-related knowledge, we propose an innovative CLIP-based framework based on Forced prompt leArning (FA), designed to make full use of the In-Distribution (ID) knowledge and ultimately boost the effectiveness of OOD detection. Our key insight is to learn a prompt (i.e., forced prompt) that contains more diversified and richer descriptions of the ID classes beyond the textual semantics of class labels. Specifically, it promotes better discernment for ID images, by forcing more notable semantic similarity between ID images and the learnable forced prompt. Moreover, we introduce a forced coefficient, encouraging the forced prompt to learn more comprehensive and nuanced descriptions of the ID classes. In this way, FA is capable of achieving notable improvements in OOD detection, even when trained without any external auxiliary datasets, while maintaining an identical number of trainable parameters as CoOp. Extensive empirical evaluations confirm our method consistently outperforms current state-of-the-art methods. Code is available at https://github.com/0xFAFA/FA.

Recent artificial intelligence (AI) generative models have demonstrated remarkable capabilities in image production, and have been widely applied to face image generation, customization, and restoration. However, many AI-generated faces (AIGFs) still suffer from issues such as unique distortions, unrealistic details, and unexpected identity shifts, underscoring the need for a comprehensive quality evaluation method for AIGFs. To this end, we introduce **FaceQ**, the first comprehensive AI-generated Face image database with fine-grained Quality annotations aligned with human preferences, which consists of 12K images and 491K ratings across multiple dimensions. Using the FaceQ database, we establish **F-Bench**, a benchmark for comparing and evaluating face generation, customization, and restoration models, highlighting strengths and weaknesses across various prompts and evaluation dimensions. Additionally, we assess the performance of existing image quality assessment (IQA) methods on FaceQ, and further propose a large multimodal model (LMM) based Face quality Evaluator (**F-Eval**) to accurately assess the multi-dimensional quality of generated faces in a one-for-all manner. Extensive experimental results demonstrate the state-of-the-art performance of our F-Eval.

Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals, multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct both quantitative and qualitative evaluations of state-of-the-art NVS methods across different evaluation settings. Our results show that current NVS methods are prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released the data to help advance self-driving and urban robotics simulation technology.

Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of "cat" can be decomposed into features like tail, fur, and face for recognition.Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation.In this paper, we introduce ExterNal knowledGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features for downstream tasks as data evolves. Extensive experiments demonstrate the state-of-the-art performance of ENGINE. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE

We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: https://github.com/JungHeeKim29/calibration-token.

Building realistic and animatable avatars still requires minutes of multi-view or monocular self-rotating videos, and most methods lack precise control over gestures and expressions. To push this boundary, we address the challenge of constructing a whole-body talking avatar from a single image. We propose a novel pipeline that tackles two critical issues: 1) complex dynamic modeling and 2) generalization to novel gestures and expressions. To achieve seamless generalization, we leverage recent pose-guided image-to-video diffusion models to generate imperfect video frames as pseudo-labels. To overcome the dynamic modeling challenge posed by inconsistent and noisy pseudo-frames, we introduce a tightly coupled 3DGS-mesh hybrid avatar representation and apply several key regularizations to mitigate inconsistencies caused by imperfect labels. Extensive experiments on diverse subjects demonstrate that our method enables the creation of a photorealistic, precisely animatable, and expressive whole-body talking avatar from just a single image.

Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a 100% reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments.