FairStyle: Debiasing StyleGAN2 with Style Channel Manipulations

Karakas, Cemre Efe; Dirik, Alara; Yalçınkaya, Eylül; Yanardag, Pinar

doi:10.1007/978-3-031-19778-9_33

Cited by 9 publications

(3 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Most editing techniques are tailored to Style-GAN, the state-of-the-art of 2D GANs [26][27][28][29]. Several editing techniques [10,18,19,25,34] build upon Style-GAN as it uses an intermediate disentangled latent space, usually referred to as w-space. Before editing, a latent space representation of the input image has to be recovered using a process typically referred to as Inversion or Projection [1,2,12,50].…”

Section: Related Workmentioning

confidence: 99%

VIVE3D: Viewpoint-Independent Video Editing using 3D-Aware GANs

Frühstück¹,

Sarafianos²,

Xu³

et al. 2023

Preprint

View full text Add to dashboard Cite

Figure 1. We propose VIVE3D, a novel method that creates a powerful personalized 3D-aware generator using a low number of selected images of a target person. Given a new video of that person, we can faithfully modify several facial attributes as well as the camera viewpoint of the head crop. Finally, we seamlessly composite the edited face with the source frame in a temporally and spatially consistent manner, while retaining a plausible composition with the static components of the frame outside of the generator's region. The dotted squares in the center frame denote the reference regions for the three different camera poses in the column below.

show abstract

Section: Related Workmentioning

confidence: 99%

VIVE3D: Viewpoint-Independent Video Editing using 3D-Aware GANs

Frühstück¹,

Sarafianos²,

Xu³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…The majority of research on the fairness and bias of GANs seeks to either discover and explain the biases or to remove the detrimental impact of using unbalanced data on generation outputs. The three primary areas of bias and fairness research are: discovering and explaining biases; debiasing pre-trained GANs; and improving the training and generation performance of GANs using biased datasets [5].…”

Section: Figure 1 General Architecture Of Generative Adversarial Networkmentioning

confidence: 99%

Real-Time Deepfake Image Generation Based on Stylegan2-ADA

Talib,

Abed

2023

RIA

View full text Add to dashboard Cite

Training Generative Adversarial Networks (GAN) usually leads to hyper-specialization due to few data and this causes training to diverge. This paper proposes a method that significantly stabilizes training without making changes. The method which will be used is stylegan2-ADA method to get fake images, the images will be entered in several steps, where the first step is using 76,400 Flickr-Faces-HQ (FFHQ) images and training them to get fake images. The program will be dividing images inside seven test folders, as the performance rate of 1000 images is 83.3%, which is a very good percentage when compared with the stylegan2 method because our proposed method contains augmentation that generates many images through the use of few images. The second step is represented by using personal images, we used two personal images and made a projection between them. The result of the performance of generating 200 images is 99.9%. Additionally, will be took a direct photo using the computer camera in real-time mode, and i generated 300 images. The generation performance is 99.9%, and our approach outperformed earlier ones in terms of accuracy, the ability to produce images without noise, and ability to create fake images of people who are not actually there.

show abstract

“…Why, or why not, certain portions of latent space respond to concepts in a consistent manner? Answering this question can shed insight on how to better manipulate semantics in GAN latent spaces, while also uncovering potential dataset biases learned by generative models [1,16].…”

Section: Introductionmentioning

confidence: 99%

Concept Lens: Visually Analyzing the Consistency of Semantic Manipulation in GANs

Jeong,

Li,

Berger

et al. 2023

2023 IEEE Visualization and Visual Analytics (VIS)

View full text Add to dashboard Cite

Figure 1: Concept Lens aims to support users in exploring and understanding the semantic manipulations offered by GAN latent spaces. The design of Concept Lens consists of (A) a hierarchical encoding of concepts, (B) a hierarchical encoding of codes, and (C) a grid-based view that shows images generated by combinations of codes (rows) and concepts (columns) selected by user.Nodes in the hierarchies color-encode the consistency and magnitude of edits made by concepts. In this example, conditioned on a group of codes at (F), the blue node (D) indicates consistent edits (e.g. the addition of spots), while the purple node (E) suggests inconsistent edits, shown as variability in species of animals.

show abstract

FairStyle: Debiasing StyleGAN2 with Style Channel Manipulations

Cited by 9 publications

References 20 publications

VIVE3D: Viewpoint-Independent Video Editing using 3D-Aware GANs

VIVE3D: Viewpoint-Independent Video Editing using 3D-Aware GANs

Real-Time Deepfake Image Generation Based on Stylegan2-ADA

Concept Lens: Visually Analyzing the Consistency of Semantic Manipulation in GANs

Contact Info

Product

Resources

About