Published on:
19 October 2023
Primary Category:
Computer Vision and Pattern Recognition
Paper Authors:
Mariia Zameshina,
Olivier Teytaud,
Laurent Najman
Proposes Diverse Diffusion to boost image diversity in text-to-image models
Method selects distant latent vectors to create varied batches
Evaluates color, content, and demographic diversity benefits
Shows potential to improve realism, creativity, and fairness
Applicable as general technique for existing models like Stable Diffusion
Diverse image generation from text
This paper introduces Diverse Diffusion, a method to generate more varied and inclusive images from text prompts using latent diffusion models like Stable Diffusion. The approach focuses on finding distant points in the latent space to produce batches of images with more diversity in color, content, and representation of people. Experiments highlight the benefits for realism, creativity, and fairness.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper