Realistic face generation using a textual description
Anukriti Kumar, Anurag Mudgil, Nakul Dodeja, Dinesh Kumar Vishwakarma · 2021
A lot of research is going on in the field of GANs and their ability to generate photorealistic images from their textual descriptions. Still the images generated by existing tasks focus primarily on either generating birds or flowers from their descriptions. This paper has successfully implemented a sketch-refinement technique for human face creation from their text descriptions. This problem has a wide variety of industrial applications as well such as creating comics automatically, assisting a movie creator to generate frames, art creation assisting an artist to generate sketches and even for educational purposes. Thus, keeping the need for such a system in mind, a two-staged StackGAN architecture obtained from deep convolutional neural networks is proposed in this paper. The features of faces like blonde hair, arched eyebrows are converted into an image of an actual person with these features. Not limiting to this, facial expressions like a wide s mile, happy face are converted from its textual form to the corresponding image of the person with the same expressions. The generation of a high resolution 256 × 256 image using captions provided in CelebA dataset makes it a valuable contribution in the field of research. Further, the proposed research work has successfully obtained an inception score of over 4.04 ± 0.05 10 iterations of evaluation and have shown promising results.