Text-to-Face Generation Based on Human Preference

Le résumé fourni par la source

Text-to-face image generation technology holds immense application value in fields such as virtual avatar creation, the entertainment industry, and public safety. However, while existing generative models can produce high-resolution images, they often overlook the subjective aesthetic information in text descriptions, leading to a discrepancy between the generated faces and human expectations. To address this, we propose a generative method that incorporates human aesthetic preferences. By constructing an aesthetic scoring model and integrating it into the feedback loop of a diffusion model, we guide and optimize the generation process to better align with human aesthetic standards. Experimental results on our self-built, large-scale, high-quality face dataset demonstrate that this method not only significantly enhances the realism and aesthetic quality of the generated faces but also achieves better consistency with the text descriptions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.