MATIC: Multilingual Accurate Textual Image Customization via Joint Generative Artificial Intelligence

The advancements in diffusion model enables the creation of highly detailed images. However, concurrently fusing texts and images poses significant challenges, often struggling with the maintenance of text accuracy across languages, optimal placement, and appropriate typography. To address these challenges, we introduce the Multilingual Accurate Textual Image Customization (MATIC) framework. MATIC employs the Chain-of-Thought (CoT) concept to decompose the textual image generation process into multiple steps, leveraging diverse generative artificial intelligence, including Multimodal Large Language Model (MMLLM) and diffusion model. The framework first generates the desired text and a corresponding prompt for the diffusion model based on user input. The diffusiongenerated image is then examined to remove any undesired text. Meanwhile, the typographic elements are designed to align with the visual content. Finally, the textual image is fused with the aid of a grid coordinate system, evaluated by MMLLM, and further customized by the user through natural language. Experimental results demonstrate that MATIC can produce accurate, high-quality, multilingual textual images that meet user requirements across various domains, including digital marketing, graphic design, and educational content creation.