Expressive Text-to-Image Generation with Rich Text

Plain text has become a prevalent interface for text-based image synthesis and editing. Its limited customization options, however, hinder users from accurately describing desired outputs. For example, plain text makes it hard to specify continuous quantities, such as the precise RGB color value or importance of each word. Creating detailed text prompts for complex scenes is tedious for humans to write and challenging for text encoders to interpret. Furthermore, describing a reference concept or texture in plain text is non-trivial. To address these challenges, we propose using a rich-text editor supporting formats such as font style, size, color, texture fill, footnote, and embedded image. We extract each word's attributes from rich text to enable local style control, explicit token reweighting, precise color rendering, and detailed region synthesis with reference concepts or texture. We achieve these capabilities through a region-based diffusion process. We first obtain each word's mask that characterizes the region guided by the word. For each region, we enforce its text attributes by creating customized prompts, applying guidance within the region, and maintaining its fidelity against plain-text generations or input images through region-based injections. We present various examples of image generation and editing from rich text and demonstrate that our method outperforms strong baselines with quantitative evaluations.