The paper introduces SatisDive, a method for text-to-image generation that ensures every generated image meets a minimum quality threshold (reward floor) while the batch as a whole maintains visual diversity. This satisficing approach avoids the trade-offs of previous methods that either focused solely on reward or diversity, or allowed low-reward images to be compensated by high-reward ones. The authors show that varying the reward floor traces a Pareto frontier between worst-image reward and batch diversity, and their method improves worst-candidate reward by up to 0.43 on FLUX and 0.70 on SANA without sacrificing diversity.
SatisDive method navigates the satisfaction-diversity frontier for text-to-image generation
By imposing a reward floor per image and a diversity cutoff per batch, SatisDive Pareto-dominates FK steering on two base models.
Big Tech
Kevin Zhai · Siva Rajesh Kasa · Soumya Roy · Sumit Negi · Mubarak Shah
University of Central Florida · Amazon
Research Digest··2 min read
Thread:Open-Ended RL Reward Design
In one line
Text-to-image generation can satisfy user preferences while maintaining diversity by ensuring each image meets a minimum reward threshold.
Before you read this
An idea this paper builds onRead it as
The authors formulate text-to-image batch generation as satisficing: each image must meet a reward floor and the batch must meet a diversity cutoff.
The authors introduce SatisDive, a training-free inference-time method for text-to-image .
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.