Year: 2025
Type: article
Abstract: In this work, we propose a lightweight and efficient face generation framework that synthesizes realistic human facial images from semantic attribute inputs using Transformer models integrated with StyleGAN2-T. The system takes high-level descriptors such as gender, age, hair colour, eye colour, face shape, hair type, and ethnicity as input, which are processed using Transformer-based encoders to capture contextual relationships and feature dependencies among the attributes. These embeddings are then translated into the latent space of the StyleGAN2-T generator, enabling high-quality facial image synthesis with reduced computational cost and faster inference time. StyleGAN2-T, a distilled variant of StyleGAN2, is employed to maintain image realism while ensuring responsiveness, making the model suitable for real-time applications. The combination of language-based understanding and generative modelling offers a novel pipeline that bridges human-descriptive semantics and machine-driven image synthesis. Experimental results demonstrate the system's ability to generate visually coherent faces across diverse attribute combinations, with potential use cases in digital avatar creation, gaming, virtual reality, and identity reconstruction.
Author: Mrs. B. Haritha
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4409957117