Emergent Text-to-Image Generation Using Short Neologism Prompts and Negative Prompts
Yasusi Kanada, Nicograph International 2024, 2024-6.
[ 日本語ページ ]
[ Paper PDF file ]
[ Poster PDF file ]
[ Slide PDF file ]
[ Full paper PDF file (rejected) ]
[ Full paper (in Japanese) PDF file ]
Abstract: Text-to-image generation models such as Stable Diffusion (SD) and DALL-E can produce a wide variety of images from text prompts. While so-called AI artists often use long prompts to generate desired images, this poster proposes a method for generating diverse painterly images emergently by SD with short prompts of 1-2 words. The reason for such attempts is that the author believes that AI, drawing from learned image data, extracts representations of color, shape, and other aspects, as well as the underlying intentions and motivations of the original images’ authors, and reflects them in the generated images, and that he also believes that the process of the AI user encountering and selecting the generated images is a creative one.
Introduction to this research theme: Bring out the artistic talent of AI and have it create AI artworks
