Parameter-Guided Image Generation with Denoising Diffusion Probabilistic Models

Text-to-image generation models aim to generate visual images by leveraging the rich semantic information inherent in the input text. However, existing models encounter challenges in generating satisfactory images when the input text contains a substantial amount of domain-specific parametric information. To tackle this challenge, we propose a high-detail image generation method for parameterized text inputs, named PIG-DDPM. This method builds upon the Denoising Diffusion Probabilistic Model(DDPM) and utilizes a weighted bag of words model to create a concise and efficient representation of paramet-ric text features. Moreover, the existing DDPM model requires powerful hardware capabilities to generate High-Definition (HD) images, which limits its practical deployment on a large scale. To address this issue, our model incorporates positional information encoding in the textual features and generates HD images by training and stitching together multiple small-sized local images. Through a series of experiments, we substantiate that our method demonstrates noteworthy advantages in terms of both image quality and clarity when compared to the state-of-the-art methods.