AI
Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion
Publication: cs.AI updates on arXiv.org
Author: Kevin Zhai, Siva Rajesh Kasa, Soumya Roy, Sumit Negi, Mubarak Shah
Source published: 05/10/2026, 04:00:00 UTC
Briefing created: 05/10/2026, 04:19:24 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 04:03:25 UTC.
Briefings
This article presents a new method called SatisDive for text-to-image generation that optimizes both reward and diversity simultaneously. The authors propose a satisficing approach where every generated image must meet a minimum reward floor, and the entire batch must achieve a diversity cutoff. By varying this floor, they identify a Pareto frontier of optimal trade-offs. SatisDive is a training-free inference-time method that uses batch-relative reward cutoffs to improve worst-candidate rewards and maintain diversity among high-reward candidates. Experiments on the Pick-a-Pic dataset at matched DreamSim show SatisDive improves worst-candidate reward by up to 0.43 with FLUX.1-dev and by up to 0.70 with SANA-1.6B compared to FK steering.
Key points
- SatisDive is a training-free inference-time method that formulates generation as satisficing, requiring every image to satisfy a reward floor and the batch to satisfy a diversity cutoff.
- SatisDive improves worst-candidate reward by up to 0.43 with FLUX.1-dev and by up to 0.70 with SANA-1.6B compared to FK steering on the Pick-a-Pic dataset at matched DreamSim.
Context and limitations
Results are limited to the Pick-a-Pic dataset at matched DreamSim settings. The study focuses on reward and diversity optimization without addressing other potential factors affecting image quality or user satisfaction.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 04:19:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifb74ncowlbnejtsryyrcpe2ymqahersncqsjyxgf3yg5l2p6qse4