Research on Prompt Automation for Improving Image Transformation Performance Based on Generative Model

Recent generative models are widely applied across various industries, with particular attention to text-to-image (T2I) and text guided image-to-image (TI2I) models in the field of image generation. TI2I models aim to preserve the structural characteristics of source images while applying semantic transformation based on prompt. However, existing methods rely on user prompts, leading to inconsistent results depending on the user’s proficiency and writing style. To address this issue, this paper proposes an automated prompt expansion system using Large-Language Model (LLM) to generate structure descriptions of source images. This approach ensures prompt consistency and mitigates structural distortion and semantic bias when performing TI2I using Stable Diffusion-based DDPM inversion.