Generative Adversarial Networks or popularly known as GANs, have been successfully used in various areas such as computer vision, medical imaging, style transfer, natural language generation, to name a few. In one of our articles, we discussed the beginner’s guide to GANs and how it proves to be a front-runner for gaining the ultimate artificial general intelligence (AGI).
In this article, we list down the top 8 GAN-based projects one can try their hands-on.
(The list is in no particular order)
1| Create Anime
GAN Models: For anime creations, you can work with several GAN models such as IllustrationGAN, AnimeGAN, PSGAN.
About: Designing your own anime characters can be time-consuming and needs lots of creative efforts. With GAN, you will be able to automatically generate anime characters without any professional knowledge. In the paper, Towards the Automatic Anime Characters Creation with GANs, the researchers proposed a model that has the capability to create anime faces at high quality with a promising rate of success.
The contribution can be described as three-fold, which are a clean dataset that is collected from Getchu. It is a suitable GAN model, and the approach is to train a GAN from images without tags that can be leveraged as a general approach to training supervised as well as a conditional model without any tag data.
Read the paper here.
2| Face Synthesis
GAN Models: For the face synthesis, you can work with several GAN models such as FaceID-GAN, TP-GAN, GP-GAN.
About: Face synthesis has achieved advanced development by using generative adversarial networks. Synthesising a face image of various viewpoints while preserving its identity is a crucial task. Some of its applications are video surveillance and face analysis.
In the paper, FaceID-GAN: Learning a Symmetry Three-Player GAN for Identity-Preserving Face Synthesis, the researchers proposed FaceID-GAN to generate identity preserving faces, which treats a classifier of face identity as the third player, competing with the generator by distinguishing the identities of the real and synthesised faces.
Read the paper here.
3| Generate Realistic Photographs
GAN Models: For generating realistic photographs, you can work with several GAN models such as ST-GAN.
About: Generating an image based on simple text descriptions or sketch is an extremely challenging problem in computer vision. It has several practical applications such as criminal investigation and game character creation.
In the paper, ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing, the researchers proposed a novel Generative Adversarial Network (GAN) architecture that utilises Spatial Transformer Networks (STNs) as the generator is known as Spatial Transformer GANs (ST-GANs). One of the key advantages of ST-GAN is its applicability to high-resolution images indirectly since the predicted warp parameters are transferable between reference frames.
Read the paper here.
4| Age Synthesis
GAN Models: For age synthesis or face ageing, you can work with several GAN models such as Age-cGAN, Dual cGANs, A3GAN.
About: Generative adversarial networks have the capability to produce synthetic images of exceptional visual fidelity. Face ageing aims at aesthetically rendering a given face to predict its future appearance, has received significant research attention in a few years.
In the paper, Face Ageing with conditional Generative Adversarial Networks, the researchers proposed the GAN-based method for automatic face ageing known as Age-cGAN (Age Conditional Generative Adversarial Network), the first GAN to generate high-quality synthetic images within required age categories.
Read the paper here.
5| Generate New Human Poses
GAN Models: For generating new human poses, you can work with several GAN models such as FD-GAN, Deformable GANs.
About: In the paper, Deformable GANs for Pose-based Human Image Generation, the researchers deal with the problem of generating images, where the foreground object changes because of a viewpoint variation or a deformable motion, such as the articulated human body. This approach can be used with other deformable objects such as human faces or animal bodies, provided that a significant number of key points can be automatically extracted from the object of interest in order to represent its pose.
Read the paper here.
6| Images to Emojis
GAN Models: For Images to Emojis, you can work with several GAN models such as EmojiGAN, EmotiGAN, DC-GAN conditioned on emojis.
About: In the paper, EmojiGAN: learning emojis distributions with a generative model, the researchers proposed an image-to-emoji architecture that is trained on data from social networks and can be used to score a given picture using ideograms. The researchers modelled the distribution of emojis conditioned on an image with a deep generative model. EmojiGAN managed to learn the emoji distribution for a set of given images and generate realistic pictographic representations from a picture.
Read the paper here.
7| Text To Image Synthesis
GAN Models: For Text-To-Image Synthesis, you can work with several GAN models such as StackGAN, DCGAN, GAN-CLS.
About: The text-to-image synthesis is an interesting application of GANs. This aims to learn a mapping from a semantic text space to complex RGB image space and also requires the generated images to be not only realistic but also semantically consistent. Generating photo-realistic images from text is an important problem and has tremendous applications, including photo-editing, computer-aided design, etc.
In the paper, Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks, the researchers proposed Stacked Generative Adversarial Networks (StackGAN) to generate 256×256 photo-realistic images conditioned on text descriptions.
Read the paper here.
8| Image-to-Image Translation
GAN Models: For Image-to-Image translation, you can work with several GAN models such as StarGAN, DualGAN.
About: In the paper, DualGAN: Unsupervised Dual Learning for Image-to-Image Translation, the researchers developed a dual-GAN mechanism, which enables image translators to be trained from two sets of unlabeled images from two domains. According to the researchers, DualGAN has achieved comparable or slightly better results than conditional GAN trained on fully labelled data.
Read the paper here.
