CVPR 2025 Papers

Adversarial attacks threaten the integrity of deep neural networks (DNNs), particularly in high-stakes applications. In this paper, we present a novel black-box adversarial attack that leverages the diverse checkpoints generated during a single model's training trajectory. Unlike conventional ensemble attacks that require multiple surrogate models with diverse architectures, our approach exploits the intrinsic diversity captured over different training stages of a single surrogate model. By decomposing the learned representations into task-intrinsic and task-irrelevant components, we employ an accuracy gap-based selection strategy to identify checkpoints that predominantly capture transferable, task-intrinsic knowledge. Extensive experiments on ImageNet and CIFAR-10 demonstrate that our method consistently outperforms traditional ensemble attacks in terms of transferability, even under resource-constrained and practical settings. This work offers a resource-efficient solution for crafting highly transferable adversarial examples and provides new insights into the dynamics of adversarial vulnerability.

Accurate 3D gaze estimation in unconstrained real-world environments remains a significant challenge due to variations in appearance, head pose, occlusion, and the limited availability of in-the-wild 3D gaze datasets. To address these challenges, we introduce a novel Self-Training Weakly-Supervised Gaze Estimation framework (ST-SWGE). This two-stage learning framework leverages diverse 2D gaze datasets, such as gaze-following data, which offer rich variations in appearances, natural scenes, and gaze distributions, and proposes an approach to generate 3D pseudo-labels and enhance model generalization. Furthermore, traditional modality-specific models, designed separately for images or videos, limit the effective use of available training data. To overcome this, we propose the Gaze Transformer (GaT), a modality-agnostic architecture capable of simultaneously learning static and dynamic gaze information from both image and video datasets. By combining 3D video datasets with 2D gaze target labels from gaze following tasks, our approach achieves the following key contributions: (i) Significant state-of-the-art improvements in within-domain and cross-domain generalization on unconstrained benchmarks like Gaze360 and GFIE, with notable cross-modal gains in video gaze estimation; (ii) Superior cross-domain performance on datasets such as MPIIFaceGaze and Gaze360 compared to frontal face methods. Code and pre-trained models will be released to the community.

Multi-view clustering aims to improve clustering accuracy by effectively integrating complementary information from multiple perspectives. However, existing methods often encounter challenges such as feature conflicts between views and insufficient enhancement of individual view features, which hinder clustering performance. To address these challenges, we propose a novel framework, EPFMVC, which integrates feature enhancement with progressive fusion to more effectively align multi-view data. Specifically, we introduce two key innovations: (1) a Feature Channel Attention Encoder (FCAencoder), which adaptively enhances the most discriminative features in each view, and (2) a View Graph-based Progressive Fusion Mechanism, which constructs a view graph using optimal transport (OT) distance to progressively fuse similar views while minimizing inter-view conflicts. By leveraging multi-head attention, the fusion process gradually integrates complementary information, ensuring more consistent and robust shared representations. These innovations enable superior representation learning and effective fusion across views. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art techniques, achieving notable improvements in multi-view clustering tasks across various datasets and evaluation metrics.

Pedestrian attribute recognition (PAR) seeks to predict multiple semantic attributes associated with a specific pedestrian. There are two types of approaches for PAR: unimodal framework and bimodal framework. The former one is to seek a robust visual feature. However, the lack of exploiting semantic feature of linguistic modality is the main concern. The latter one utilizes prompt learning techniques to integrate linguistic data. However, static prompt templates and simple bimodal concatenation cannot to capture the extensive intra-class attribute variability and support active modalities collaboration. In this paper, we propose an Enhanced Visual-Semantic Interaction with Tailored Prompts (EVSITP) framework for PAR. We present an Image-Conditional Dual-Prompt Initialization Module (IDIM) to adaptively generate context-sensitive prompts from visual inputs. Subsequently, a Prompt Enhanced and Regularization Module (PERM) is proposed to strengthen linguistic information from IDIM. We further design a Bimodal Mutual Interaction Module (BMIM) to ensure bidirectional modalities communication. In addition, existing PAR datasets are collected over a short period in limited scenarios, which do not align with real-world scenarios. Therefore, we annotate a long-term person re-identification dataset to create a new PAR dataset, Celeb-PAR. Experiments on several challenging PAR datasets show that our method outperforms state-of-the-art approaches.

Prior research on out-of-distribution detection (OoDD) has primarily focused on single-modality models. Recently, with the advent of large-scale pretrained vision-language models such as CLIP, OoDD methods utilizing such multi-modal representations through zero-shot and prompt learning strategies have emerged. However, these methods typically involve either freezing the pretrained weights or only partially tuning them, which can be suboptimal for downstream datasets. In this paper, we highlight that multi-modal fine-tuning (MMFT) can achieve notable OoDD performance. Despite some recent works demonstrating the impact of fine-tuning methods for OoDD, there remains significant potential for performance improvement. We investigate the limitation of naive fine-tuning methods, examining why they fail to fully leverage the pretrained knowledge. Our empirical analysis suggests that this issue could stem from the modality gap within in-distribution (ID) embeddings. To address this, we propose a training objective that enhances cross-modal alignment by regularizing the distances between image and text embeddings of ID data. This adjustment helps in better utilizing pretrained textual information by aligning similar semantics from different modalities (i.e., text and image) more closely in the hyperspherical representation space. We theoretically demonstrate that the proposed regularization corresponds to the maximum likelihood estimation of an energy-based model on a hypersphere. Utilizing ImageNet-1k OoD benchmark datasets, we show that our method, combined with post-hoc OoDD approaches leveraging pretrained knowledge (e.g., NegLabel), significantly outperforms existing methods, achieving state-of-the-art OoDD performance and leading ID accuracy.

Automated radiology report generation offers an effective solution to alleviate radiologists' workload. However, most existing methods focus primarily on single or fixed-view images to model current disease conditions, which limits diagnostic accuracy and overlooks disease progression. Although some approaches utilize longitudinal data to track disease progression, they still rely on single images to analyze current visits. To address these issues, we propose enhanced contrastive learning with Multi-view Longitudinal data to facilitate chest X-ray Report Generation, named MLRG. Specifically, we introduce a multi-view longitudinal contrastive learning method that integrates spatial information from current multi-view images and temporal information from longitudinal data. This method also utilizes the inherent spatiotemporal information of radiology reports to supervise the pre-training of visual and textual representations. Subsequently, we present a tokenized absence encoding technique to flexibly handle missing patient-specific prior knowledge, allowing the model to produce more accurate radiology reports based on available prior knowledge. Extensive experiments on MIMIC-CXR, MIMIC-ABN, and Two-view CXR datasets demonstrate that our MLRG outperforms recent state-of-the-art methods, achieving a 2.3% BLEU-4 improvement on MIMIC-CXR, a 5.5% F1 score improvement on MIMIC-ABN, and a 2.7% F1 RadGraph improvement on Two-view CXR.

Trajectory prediction plays a crucial role in autonomous driving systems, and exploring its vulnerability has garnered widespread attention. However, existing trajectory prediction attack methods often rely on single-point attacks to make efficient perturbations. This limits their applications in real-world scenarios due to the transient nature of single-point attacks, their susceptibility to filtration, and the uncertainty regarding the deployment environment. To address these challenges, this paper proposes a novel LiDAR-induced attack framework to impose multi-frame attacks by optimization-driven adversarial location search, achieving endurance, efficiency, and robustness. This framework strategically places objects near the adversarial vehicle to implement an attack and introduces three key innovations. First, successive state perturbations are generated using a multi-frame single-point attack strategy, effectively misleading trajectory predictions over extended time horizons. Second, we efficiently optimize adversarial objects' locations through three specialized loss functions to achieve desired perturbations. Lastly, we improve robustness by treating the adversarial object as a point without size constraints during the location search phase and reduce dependence on both the specific attack point and the adversarial object's properties. Extensive experiments confirm the superior performance and robustness of our framework.

Implicit neural representations (INRs) such as NeRF and SIREN encode a signal in neural network parameters and show excellent results for signal reconstruction. Using INRs for downstream tasks, such as classification, is however not straightforward. Inherent symmetries in the parameters pose challenges and current works primarily focus on designing architectures that are equivariant to these symmetries. However, INR-based classification still significantly under-performs compared to pixel-based methods like CNNs. This work presents an end-to-end strategy for initializing SIRENs together with a learned learning-rate scheme, to yield representations that improve classification accuracy. We show that a simple, straightforward, Transformer model applied to a meta-learned SIREN, without incorporating explicit symmetry equivariances, outperforms the current state-of-the-art. On the CIFAR-10 SIREN classification task, we improve the state-of-the-art without augmentations from 38.8% to 59.6%, and from 63.4% to 64.7% with augmentations. We demonstrate scalability on the high-resolution Imagenette dataset achieving reasonable reconstruction quality with a classification accuracy of 60.8% and are the first to do INR classification on the full ImageNet-1K dataset where we achieve a SIREN classification performance of 23.6%. To the best of our knowledge, no other SIREN classification approach has managed to set a classification baseline for any high-resolution image dataset. Our code is available at https://github.com/SanderGielisse/MWT

Human-object interaction (HOI) reconstruction has garnered significant attention due to its diverse applications and the success of capturing human meshes. Existing HOI reconstruction methods often rely on explicitly modeling interactions between humans and objects. However, such a way leads to a natural conflict between 3D mesh reconstruction, which emphasizes global structure, and fine-grained contact reconstruction, which focuses on local details. To address the limitations of explicit modeling, we propose the End-to-End HOI Reconstruction Transformer with Graph-based Encoding (HOI-TG). It implicitly learns the interaction between humans and objects by leveraging self-attention mechanisms. Within the transformer architecture, we devise graph residual blocks to aggregate the topology among vertices of different spatial structures. This dual focus effectively balances global and local representations. Without bells and whistles, HOI-TG achieves state-of-the-art performance on BEHAVE and InterCap datasets. Particularly on the challenging InterCap dataset, our method improves the reconstruction results for human and object meshes by 8.9% and 8.6%, respectively.

In recent years, large text-to-video (T2V) synthesis models have garnered considerable attention for their abilities to generate videos from textual descriptions. However, achieving both high imaging quality and effective motion representation remains a significant challenge for these T2V models. Existing approaches often adapt pre-trained text-to-image (T2I) models to refine video frames, leading to issues such as flickering and artifacts due to inconsistencies across frames. In this paper, we introduce EVS, a training-free \underline E ncapsulated \underline V ideo \underline S ynthesizer that composes T2I and T2V models to enhance both visual fidelity and motion smoothness of generated videos. Our approach utilizes a well-trained diffusion-based T2I model to refine low-quality video frames by treating them as out-of-distribution samples, effectively optimizing them with noising and denoising steps. Meanwhile, we employ T2V backbones to ensure consistent motion dynamics. By encapsulating the T2V temporal-only prior into the T2I generation process, EVS successfully leverages the strengths of both types of models, resulting in videos of improved imaging and motion quality. Experimental results validate the effectiveness of our approach compared to previous approaches.Our composition process also leads to a significant improvement of 1.6x-4.5x speedup in inference time.

This work presents a novel text-to-vector graphics generation approach, Dream3DVG, allowing for arbitrary viewpoint viewing, progressive detail optimization, and view-dependent occlusion awareness. Our approach is a dual-branch optimization framework, consisting of an auxiliary 3D Gaussian Splatting optimization branch and a 3D vector graphics optimization branch. The introduced 3DGS branch can bridge the domain gaps between text prompts and vector graphics with more consistent guidance. Moreover, 3DGS allows for progressive detail control by scheduling classifier-free guidance, facilitating guiding vector graphics with coarse shapes at the initial stages and finer details at later stages. We also improve the view-dependent occlusions by devising a visibility-awareness rendering module. Extensive results on 3D sketches and 3D iconographies, demonstrate the superiority of the method on different abstraction levels of details, cross-view consistency, and occlusion-aware stroke culling.

Driven by the great success of Large Language Models (LLMs) in the 2D image domain, their applications in 3D scene understanding has emerged as a new trend. A key difference between 3D and 2D is that the situation of an egocentric observer in 3D scenes can change, resulting in different descriptions (e.g., "left" or "right"). However, current LLM-based methods overlook the egocentric perspective and simply use datasets from a global viewpoint. To address this issue, we propose a novel approach to automatically generate a situation-aware dataset by leveraging the scanning trajectory during data collection and utilizing Vision-Language Models (VLMs) to produce high-quality captions and question-answer pairs. Furthermore, we introduce a situation grounding module to explicitly predict the position and orientation of observer's viewpoint, thereby enabling LLMs to ground situation description in 3D scenes. We evaluate our approach on several benchmarks, demonstrating that our method effectively enhances the 3D situational awareness of LLMs while significantly expanding existing datasets and reducing manual effort.

The unprecedented advancements in Large Language Models (LLMs) have profoundly impacted natural language processing but have yet to fully embrace the realm of scalable vector graphics (SVG) generation. While LLMs encode partial knowledge of SVG data from web pages during training, recent findings suggest that semantically ambiguous and tokenized representations within LLMs may result in hallucinations in vector primitive predictions. Additionally, LLM training typically lacks modeling and understanding of the rendering sequence of vector paths, which can lead to occlusion between output vector primitives. In this paper, we present LLM4SVG, an initial yet substantial step toward bridging this gap by enabling LLMs to better understand and generate vector graphics. LLM4SVG facilitates a deeper understanding of SVG components through learnable semantic tokens, which precisely encode these tokens and their corresponding properties to generate semantically aligned SVG outputs. Using a series of learnable semantic tokens, a structured dataset for instruction following is developed to support comprehension and generation across two primary tasks. Our method introduces a modular architecture to existing large language models, integrating semantic tags, vector instruction encoders, fine-tuned commands, and powerful LLMs to tightly combine geometric, appearance, and language information. To overcome the scarcity of SVG-text instruction data, we developed an automated data generation pipeline that collected our SVGX-SFT Dataset, consisting of high-quality human-designed SVGs and 580k SVG instruction following data specifically crafted for LLM training, which facilitated the adoption of the supervised fine-tuning strategy popular in LLM development. By exploring various training strategies, we developed LLM4SVG, which significantly moves beyond optimized rendering-based approaches and language-model-based baselines to achieve remarkable results in human evaluation tasks. Code, model, and data will be released at: https://ximinng.github.io/LLM4SVGProject/

Dataset distillation has demonstrated strong performance on simple datasets like CIFAR, MNIST, and TinyImageNet but struggles to achieve similar results in more complex scenarios. In this paper, we propose EDF (emphasizes the discriminative features), a dataset distillation method that enhances key discriminative regions in synthetic images using Grad-CAM activation maps. Our approach is inspired by a key observation: in simple datasets, high-activation areas typically occupy most of the image, whereas in complex scenarios, the size of these areas is much smaller. Unlike previous methods that treat all pixels equally when synthesizing images, EDF uses Grad-CAM activation maps to enhance high-activation areas. From a supervision perspective, we downplay supervision signals produced by lower trajectory-matching losses, as they contain common patterns. Additionally, to help the DD community better explore complex scenarios, we build the Complex Dataset Distillation (Comp-DD) benchmark by meticulously selecting sixteen subsets, eight easy and eight hard, from ImageNet-1K. In particular, EDF consistently outperforms SOTA results in complex scenarios, such as ImageNet-1K subsets. Hopefully, more researchers will be inspired and encouraged to improve the practicality and efficacy of DD. Our code and benchmark have been made public at NUS-HPC-AI-Lab/EDF.

Diffusion models have revolutionized the field of talking head generation, yet still face challenges in expressiveness, controllability, and stability in long-time generation. In this research, we propose an EmotiveTalk framework to address these issues. Firstly, to realize better control over the generation of lip movement and facial expression, a Vision-guided Audio Information Decoupling (V-AID) approach is designed to generate audio-based decoupled representations aligned with lip movements and expression. Specifically, to achieve alignment between audio and facial expression representation spaces, we present a Diffusion-based Co-speech Temporal Expansion (Di-CTE) module within V-AID to generate expression-related representations under multi-source emotion condition constraints. Then we propose a well-designed Emotional Talking Head Diffusion (ETHD) backbone to efficiently generate highly expressive talking head videos, which contains an Expression Decoupling Injection (EDI) module to automatically decouple the expressions from reference portraits while integrating the target expression information, achieving more expressive generation performance. Experimental results show that EmotiveTalk can generate expressive talking head videos, ensuring the promised controllability of emotions and metric stability during long-time generation, yielding state-of-the-art performance compared to existing methods. The main page of our paper can be found in https://emotivetalk.github.io/.

Affective Image Manipulation (AIM) seeks to modify user-provided images to evoke specific emotions. This task is inherently complex due to its twofold objective: evoking the intended emotion while preserving image composition. Existing AIM methods primarily adjust color and style, often failing to elicit precise, profound emotional shifts. Drawing on psychological insights, we introduce EmoEdit, which extends AIM by incorporating content modifications to enhance emotional impact. Specifically, we construct EmoEditSet, a large-scale AIM dataset of 40,120 paired data through emotion attribution and data construction. To make generative models emotion-aware, we design an Emotion Adapter and train it using EmoEditSet. We further propose an instruction loss to capture semantic variations in each data pair. Our method is evaluated both qualitatively and quantitatively, demonstrating superior performance over state-of-the-art techniques. Additionally, we showcase the portability of our Emotion Adapter to other diffusion-based models, enhancing their emotion knowledge with diverse semantics. Code is available at: https://github.com/JingyuanYY/EmoEdit.

Given a piece of text, a video clip, and a reference audio, the movie dubbing task aims to generate speech that aligns with the video while cloning the desired voice. The existing methods have two primary deficiencies: (1) They struggle to simultaneously hold audio-visual sync and achieve clear pronunciation; (2) They lack the capacity to express user-defined emotions. To address these problems, we propose EmoDubber, an emotion-controllable dubbing architecture that allows users to specify emotion type and emotional intensity while satisfying high-quality lip sync and pronunciation. Specifically, we first design Lip-related Prosody Aligning (LPA), which focuses on learning the inherent consistency between lip motion and prosody variation by duration level contrastive learning to incorporate reasonable alignment. Then, we design Pronunciation Enhancing (PE) strategy to fuse the video-level phoneme sequences by efficient conformer to improve speech intelligibility. Next, the speaker identity adapting module decodes acoustics prior and inject the speaker style embedding. After that, the proposed Flow-based User Emotion Controlling (FUEC) is used to synthesize waveform by flow matching prediction network conditioned on acoustics prior. In this process, the FUEC determines the gradient direction and guidance scale based on the user's emotion instructions by the positive and negative guidance mechanism, which focuses on amplifying the desired emotion while suppressing others. Extensive experimental results demonstrate favorable performance compared to several state-of-the-art methods. The code and trained models will be made available at https://github.com/GalaxyCong/DubFlow.

Visual In-Context Learning (VICL) enables adaptively solving vision tasks by leveraging pixel demonstrations, mimicking human-like task completion through analogy. Prompt selection is critical in VICL, but current methods assume the existence of a single "ideal" prompt in a pool of candidates, which in practice may not hold true. Multiple suitable prompts may exist, but individually they often fall short, leading to difficulties in selection and the exclusion of useful context. To address this, we propose a new perspective: prompt condensation. Rather than relying on a single prompt, candidate prompts collaborate to efficiently integrate informative contexts without sacrificing resolution. We devise Condenser, a lightweight external plugin that compresses relevant fine-grained context across multiple prompts. Optimized end-to-end with the backbone, Condenser ensures accurate integration of contextual cues. Experiments demonstrate Condenser outperforms state-of-the-arts across benchmark tasks, showing superior context compression, scalability with more prompts, and enhanced computational efficiency compared to ensemble methods, positioning it as a highly competitive solution for VICL. Code is open-sourced at https://github.com/gimpong/CVPR25-Condenser.