Thai Text-to-Image Prompt Engineering by Pre-trained Large Language with Stable Diffusion Model

Abstract

Text-to-image (T2I) generation is a new area of large language models (LLMs), a type of prompt engineering involving inputting a textual description to generate an image. To shift a new paradigm of Thai natural language processing (Thai-NLP), this paper first presents state-of-the-art Thai Text-to-Image prompt engineering (TH-T2I) to translate Thai text into a semantic image according to the semantic Thai textual description. The pre-trained SCB-MT-EN-TH model is employed for Text-to-Text (T2T) translation. Moreover, the image generation is done according to a semantic text prompt by a stable diffusion model. The T2T is evaluated by Bi-lingual Evaluation Understudy (BLEU), while T2I is done by Inception and Frechet Inception Distance (FID). The images generated by TH-T2I were of high quality, as measured by Inception and FID. TH-T2I contributes to a T2I baseline model in Thai, preserving the Thai cultural language on digital heritage.