Recent research has shown remarkable success in revealing "steering" directions in the latent spaces of pre-trained GANs. These directions correspond to semantically meaningful image transformations e.g., shift, zoom, color manipulations), and have similar interpretable effects across all categories that the GAN can generate. Some methods focus on user-specified transformations, while others discover transformations in an unsupervised manner. However, all existing techniques rely on an optimization procedure to expose those directions, and offer no control over the degree of allowed interaction between different transformations. In this paper, we show that "steering" trajectories can be computed in closed form directly from the generator's weights without any form of training or optimization. This applies to user-prescribed geometric transformations, as well as to unsupervised discovery of more complex effects. Our approach allows determining both linear and nonlinear trajectories, and has many advantages over previous methods. In particular, we can control whether one transformation is allowed to come on the expense of another (e.g. zoom-in with or without allowing translation to keep the object centered). Moreover, we can determine the natural end-point of the trajectory, which corresponds to the largest extent to which a transformation can be applied without incurring degradation. Finally, we show how transferring attributes between images can be achieved without optimization, even across different categories.
Spatial Steerability of GANs via Self-Supervision from Discriminator
University of Oxford（牛津大学）；University of California, Los Angeles（加州大学洛杉矶分校）；Chinese University of Hong Kong（香港中文大学）；Ant Group（蚂蚁集团）；Australian National University（澳大利亚国立大学）
Reconstruct Face from Features Using GAN Generator as a Distribution Constraint
Yonsei University（延世大学）；Fuzhou University（福州大学）；TCL AI Lab（TCL人工智能实验室）；Anhui University（安徽大学）；Tsinghua Shenzhen International Graduate School（清华深圳国际研究生院）
Make Encoder Great Again in 3D GAN Inversion through Geometry and Occlusion-Aware Encoding
Hong Kong University of Science and Technology（香港科技大学）；Tsinghua Shenzhen International Graduate School（清华大学深圳国际研究生院）；International Digital Economy Academy (IDEA)（国际数字经济研究院（IDEA））
3D GAN反演旨在从单个图像输入同时实现高重建保真度和合理的3D几何形状。然而，现有的3D GAN反演方法依赖于针对每个单独情况的耗时优化。在这项工作中，我们介绍了一种新的基于编码器的反演框架的基础上EG3D，最广泛使用的三维GAN模型之一。我们利用EG3D的潜在空间的固有属性来设计一个正则化和一个背景深度正则化。这使得我们能够训练几何感知编码器能够将输入图像转换为相应的潜在代码。此外，我们探索EG3D的特征空间，并开发了一个自适应细化阶段，提高了EG3D中特征的表示能力，以增强细粒度纹理细节的恢复。最后，我们提出了一个遮挡感知融合操作，以防止在未观察到的区域失真。我们的方法取得了令人印象深刻的结果，可与基于优化的方法相媲美，同时运行速度快达500倍。我们的框架非常适合语义编辑等应用程序。
A Comprehensive Survey on Data-Efficient GANs in Image Generation
University of Science and Technology of China（中国科学技术大学）；Huazhong University of Science and Technology（华中科技大学）；The University of Sydney（悉尼大学）；JD Explore Academy（京东探索研究院）
DisCO: Portrait Distortion Correction with Perspective-Aware 3D GANs
The University of Tokyo（东京大学）；National Institute of Informatics（国立信息学研究所）；National Yang Ming Chiao Tung University（国立阳明交通大学）；University of Maryland, College Park（马里兰大学帕克分校）；Snap Inc.（Snap公司）
针对近距离特写入脸的透视畸变问题，本文提出 DisCO，通过透视感知 3D GAN inversion 联合优化相机参数与潜码，并引入调度和几何正则化，实现自然的人像校正。
CDE-GAN: Cooperative Dual Evolution Based Generative Adversarial Network
School of Electronic Information and Communication, Huazhong University of Science and Technology（华中科技大学电子信息与通信学院）；Shenzhen Research Institute, Huazhong University of Science and Technology（华中科技大学深圳研究院）；Discipline of ICT, University of Tasmania（塔斯马尼亚大学信息与通信技术学科）；School of Information Science and Technology, Nantong University（南通大学信息科学与技术学院）
Image Inversion: A Survey from GANs to Diffusion and Beyond
College of Control Science and Engineering, Zhejiang University（浙江大学控制科学与工程学院）；YouTu Lab, Tencent（腾讯优图实验室）；College of Computer and Information Science, Southwest University（西南大学计算机与信息科学学院）；Department of Computer Science & Engineering, Shanghai Jiao Tong University（上海交通大学计算机科学与工程系）；School of Intelligence Science and Technology, Nanjing University（南京大学智能科学与技术学院）
Revisiting the Privacy Risks of Split Inference: A GAN-Based Data Reconstruction Attack via Progressive Feature Optimization
School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen（哈尔滨工业大学深圳学院）；Tsinghua Shenzhen International Graduate School, Tsinghua University（清华大学深圳国际研究生院）；Department of Computer Science and Technology, Tsinghua University（清华大学计算机科学与技术系）
针对分割推理中间特征泄露问题，提出结合渐进式特征优化与 L1-ball 约束的 GAN 数据重构攻击，显著提升高分辨率、分布外及深层 DNN 场景下的重构效果。
Combating Mode Collapse in GAN training: An Empirical Analysis using Hessian Eigenvalues
Fraunhofer ITWM（弗劳恩霍夫工业数学研究所）；University of Heidelberg（海德堡大学）；TU Kaiserslautern（凯泽斯劳滕工业大学）；Fraunhofer Center Machine Learning（弗劳恩霍夫机器学习中心）；Offenburg University（奥芬堡大学）
Generative Adversarial Networks for Image Super-Resolution: A Survey
School of Engineering, The Hong Kong University of Science and Technology（香港科技大学工程学院）；School of Software, Northwestern Polytechnical University（西北工业大学软件学院）；College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics（南京航空航天大学人工智能学院）；Department of Distributed Systems and IT Devices, Silesian University of Technology（桑特大学技术学院）；School of Computer Science and Technology, Harbin Institute of Technology（哈尔滨工业大学计算机科学与技术学院）
Next-Generation Parallel Decoder for LPDR: Architectural Optimization and Class-Balanced GAN-Augmentation
School of Electrical Engineering；Computer Science National University of Sciences \& Technology Islamabad, Pakistan sobaid.mscs25seecs.edu.pk；Computer Science National University of Sciences \& Technology Islamabad, Pakistan nchandio.mscs25seecs.edu.pk；Computer Science National University of Sciences \& Technology Islamabad, Pakistan njamil.mscs25seecs.edu.pk；Computer Science National University of Sciences \& Technology Islamabad, Pakistan
