Work

Toward Enhancing LightWeight GAN for Text-Guided Generation of Animated Character Faces

Year: 2025

Type: article

Abstract: Traditional text-guided image generation methods primarily focus on modifying existing images or altering specific elements, which limits their applicability. This paper introduces a significant enhancement to the LightWeight-GAN model, originally designed for image generation from random noise, by transforming it into a text-guided generative system. The proposed approach enables the generation of high-quality animated character faces directly from textual descriptions. To achieve this, we incorporate a mapping network that refines textual inputs before feeding them into the generator, ensuring more precise feature representation. Additionally, we integrate contrastive language-image pretraining (CLIP) to verify the generated images and enforce stronger alignment between textual prompts and visual outputs. The model is trained on a specialized facial dataset, demonstrating its ability to generate semantically accurate and visually compelling character faces. Extensive experiments using the CartoonSet dataset validate the effectiveness of our approach, achieving an FID score of 29.8 across 10 000 generated images. These improvements significantly outperform existing text-to-image generation models, making our system a promising tool for applications in game development, animation, and virtual reality.

Source: IEEE Access

DOI: 10.1109/access.2025.3595928