针对生成模型身份保留评估的挑战，提出Beyond the Pixels层次化框架，通过分解主体特征并引导VLM结构化推理，在与人类判断一致性及诊断性上优于现有方法，并构建了新基准。
摘要：
评估生成模型中的身份保留仍是一项关键但未解决的挑战。现有指标依赖全局嵌入或粗糙的VLM提示，无法捕捉细粒度的身份变化，且诊断性洞察有限。我们引入Beyond the Pixels这一层次化评估框架，将身份评估分解为特征级转换。我们的方法通过（1）将主体分层分解为（类型，风格）→属性→特征决策树，以及（2）提示具体转换而非抽象相似度分数，引导VLMs进行结构化推理。这种分解将VLM分析建立在可验证的视觉证据之上，减少幻觉并提高一致性。我们在四个最先进的生成模型上验证了该框架，证明其在测量身份一致性方面与人类判断高度一致。此外，我们引入了一个专门用于压力测试生成模型的新基准，包含1078个跨不同主体类型的图像-提示对，包括拟人化和动画角色等代表性不足的类别，每个提示平均捕捉6到7个转换轴。
英文摘要：
Evaluating identity preservation in generative models remains a critical yet unresolved challenge. Existing metrics rely on global embeddings or coarse VLM prompting, failing to capture fine-grained identity changes and providing limited diagnostic insight. We introduce Beyond the Pixels, a hierarchical evaluation framework that decomposes identity assessment into feature-level transformations. Our approach guides VLMs through structured reasoning by (1) hierarchically decomposing subjects into (type, style) -> attribute -> feature decision tree, and (2) prompting for concrete transformations rather than abstract similarity scores. This decomposition grounds VLM analysis in verifiable visual evidence, reducing hallucinations and improving consistency. We validate our framework across four state-of-the-art generative models, demonstrating strong alignment with human judgments in measuring identity consistency. Additionally, we introduce a new benchmark specifically designed to stress-test generative models. It comprises 1,078 image-prompt pairs spanning diverse subject types, including underrepresented categories such as anthropomorphic and animated characters, and captures an average of six to seven transformation axes per prompt.
Who Can See Through You? Adversarial Shielding Against VLM-Based Attribute Inference Attacks
East China Normal University（东华大学）；Zhongguancun Academy（中关村学院）；Dept. of Comp. Sci. and Tech., Institute for AI, Tsinghua University（计算机科学与技术系，人工智能研究院，清华大学）
Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset
University of Wisconsin-Madison（威斯康星大学麦迪逊分校）；USC（南加州大学）；University of Michigan-Ann Arbor（密歇根大学安娜堡分校）；University of Washington（华盛顿大学）；University of Maryland（马里兰大学）；Lehigh University（理海大学）；UIUC（伊利诺伊大学厄巴纳-香槟分校）；Peking University（北京大学）；University of Rochester（罗切斯特大学）；University of California, Davis（加州大学戴维斯分校）
GeoWorld-VLM: Geometry from World Models for Vision-Language Models
Harvard AI and Robotics Lab（哈佛人工智能与机器人实验室）；Kempner Institute for the Study of Natural and Artificial Intelligence（凯普纳自然与人工智能研究 institute）；Harvard University（哈佛大学）
DP^2-VL: Private Photo Dataset Protection by Data Poisoning for Vision-Language Models
Shandong University（山东大学）；Shanghai Jiao Tong University（上海交通大学）；Donghua University（东华大学）；Shanghai Normal University（上海师范大学）；East China Normal University（华东师范大学）；Hefei University of Technology（合肥工业大学）
Prompt-OT: An Optimal Transport Regularization Paradigm for Knowledge Preservation in Vision-Language Model Adaptation
Morgan Stanley（摩根士丹利）；Clemson University（克莱姆森大学）；Arizona State University（亚利桑那州立大学）；Washington University in St. Louis（圣路易斯华盛顿大学）；University of Arizona（亚利桑那大学）；University of Notre Dame（圣母大学）
Nüwa: Mending the Spatial Integrity Torn by VLM Token Pruning
Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)（人工智能与数字经济广东实验室）；School of Artificial Intelligence, Xidian University（西安电子科技大学人工智能学院）；The Hong Kong Polytechnic University（香港理工大学）；Great Bay University（大鹏大学）；Shenzhen University（深圳大学）；Huawei（华为）
Large pre-trained vision-language models (VLMs) reduce the time for developing predictive models for various vision-grounded language downstream tasks by providing rich, adaptable image and text representations. However, these models suffer from societal biases owing to the skewed distribution of various identity groups in the training data. These biases manifest as the skewed similarity between the representations for specific text concepts and images of people of different identity groups and, therefore, limit the usefulness of such models in real-world high-stakes applications. In this work…
When the Edit Changes the Patient: Measuring Identity Preservation in Counterfactual Retinal Images
Technical University of Munich（慕尼黑工业大学）；TUM University Hospital（慕尼黑工业大学医院）；Munich Center of Machine Learning (MCML)（慕尼黑机器学习中心）；Munich Data Science Institute (MDSI)（慕尼黑数据科学研究所）；Sapienza University of Rome（罗马第一大学）；Faculty of Medicine, University of Southampton（南安普顿大学医学院）；Moorfields Eye Hospital NHS Foundation Trust（摩尔菲尔兹眼科医院NHS信托基金）；Imperial College London（伦敦帝国学院）
Few-Shot Generative Model Adaption via Identity Injection and Preservation
School of Cyberspace Security, Hangzhou Dianzi University（杭州电子科技大学网络安全学院）；Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；School of Software Engineering, Xi’an Jiaotong University（西安交通大学软件工程学院）；Lishui University and Lishui Institute of Hangzhou Dianzi University（丽水大学和杭州电子科技大学丽水研究所）；School of Faculty of Applied Science, Macao Polytechnic University（澳门理工学院应用科学学院）；School of Communication Engineering, Hangzhou Dianzi University（杭州电子科技大学通信工程学院）
Training-Free Diffusion Framework for Stylized Image Generation with Identity Preservation
Computer Vision Group, Part AI Research Center（部分人工智能研究中心计算机视觉组）；Department of Computer Engineering, Amirkabir University of Technology（阿米卡比尔技术大学计算机工程系）
Nowadays, the wide application of virtual digital human promotes the comprehensive prosperity and development of digital culture supported by digital economy. The personalized portrait automatically generated by AI technology needs both the natural artistic style and human sentiment. In this paper, we propose a novel StyleIdentityGAN model, which can ensure the identity and artistry of the generated portrait at the same time. Specifically, the style-enhanced module focuses on artistic style features decoupling and transferring to improve the artistry of generated virtual face images. Meanwhile…
We present a novel framework to generate images of different age while preserving identity information, which is known as face aging. Different from most recent popular face aging networks utilizing Generative Adversarial Networks(GANs) application, our approach do not simply transfer a young face to an old one. Instead, we employ the edge map as intermediate representations, firstly edge maps of young faces are extracted, a CycleGAN-based network is adopted to transfer them into edge maps of old faces, then another pix2pixHD-based network is adopted to transfer the synthesized edge maps, conc…
University of Pisa（比萨大学）；University of Trento（特伦托大学）；Fondazione Bruno Kessler（布鲁诺·凯塞勒基金会）；ETH Zürich（苏黎世联邦理工学院）
We propose AnonyGAN, a GAN-based solution for face anonymisation which replaces the visual information corresponding to a source identity with a condition identity provided as any single image. With the goal to maintain the geometric attributes of the source face, i.e., the facial pose and expression, and to promote more natural face generation, we propose to exploit a Bipartite Graph to explicitly model the relations between the facial landmarks of the source identity and the ones of the condition identity through a deep model. We further propose a landmark attention model to relax the manual…
Transilvania University of Brasov（布拉索夫特兰西瓦尼亚大学）；Xperi Corporation（Xperi公司）
This paper describes a new model which generates images in novel poses e.g. by altering face expression and orientation, from just a few instances of a human subject. Unlike previous approaches which require large datasets of a specific person for training, our approach may start from a scarce set of images, even from a single image. To this end, we introduce Character Adaptive Identity Normalization GAN (CainGAN) which uses spatial characteristic features extracted by an embedder and combined across source images. The identity information is propagated throughout the network by applying condi…
Realistic age-progressed photos provide invaluable biometric information in a wide range of applications. In recent years, deep learning-based approaches have made remarkable progress in modeling the aging process of the human face. Nevertheless, it remains a challenging task to generate accurate age-progressed faces from infant or toddler photos. In particular, the lack of visually detectable gender characteristics and the drastic appearance changes in early life contribute to the difficulty of the task. We propose a new deep learning method inspired by the successful Conditional Adversarial …
From Cradle to Cane: A Two-Pass Framework for High-Fidelity Lifespan Face Aging
VCIP, College of Computer Science, Nankai University（VCIP，计算机科学学院，南开大学）；Samsung Research China - Beijing (SRC-B)（三星中国北京研究院）；School of Electrical and Information Engineering, Zhengzhou University（电气信息学院，郑州大学）；School of Computer , Zhengzhou University of Aeronautics（计算机学院，郑州航空大学）；Program of Computer Science, City University of Hong Kong (Dongguan)（计算机科学项目，香港城市大学（东莞））；City University of Hong Kong（香港城市大学）；Computer Vision Center, Barcelona（巴塞罗那计算机视觉中心）
