China
Africa
Explore chapters and articles related to this topic
Detecting DeepFakes for future robotic systems
Published in Brij B. Gupta, Nadia Nedjah, Safety, Security, and Reliability of Robotic Systems, 2020
Eric Tjon, Melody Moh, Teng-Sheng Moh
StyleGAN demonstrates significant abilities for image synthesis on a dataset of face images (Karras et al. 2019). It uses training data from a high-quality dataset of real faces. StyleGAN utilizes progressive growing, where the model adds more resolution to the input and output after successfully training at a lower resolution. This technique results in large and clear output images. The architecture also learns styles and facial features in an unsupervised manner. Using this feature allows the model to control certain features in the output face, such as glasses, skin tone, and hair color.
Understanding and Building Generative Adversarial Networks
Published in Monika Mangla, Subhash K. Shinde, Vaishali Mehta, Nonita Sharma, Sachi Nandan Mohanty, Handbook of Research on Machine Learning, 2022
StyleGAN is the state-of-the-art for image generation using GANs. The images in Figure 5.24 are from a StyleGAN trained on the Flickr faces HQ (FFHQ) Dataset containing 70000 images (1024×1024) of people from various ethnicities having a diverse set of features.
Manifold-enhanced CycleGAN for facial expression synthesis
Published in The Imaging Science Journal, 2022
In recent years, the StyleGAN has been used as a new generator architecture in the generation of high-resolution and high-realism facial images without landmark extraction. The facial attributes of the generated image can be edited by means of changing the latent code, as shown in Figure 8. When editing a target image not generated by StyleGAN, we would need to invert the image first to its latent code. Abdal et al. [31] proposed two approaches to embed instances from the image space to the latent space. The first one is training an encoder that maps image to the latent space. This method has fast-processing speed, but does not have the ability to process images outside of the training datasets. Another approach is to randomly initialize a latent code and use gradient descent to optimize it. Despite this method improving the generalization ability, the process of mapping images to the latent space of StyleGAN is still complicated and unstable, a problem that needs to be addressed in the future.