RESEARCH · RESEARCH · #1514
SatisDive: inference-time satisficing for the satisfaction–diversity frontier in text-to-image diffusion
The paper (arXiv:2610.02372v1) frames text-to-image generation as a satisficing problem—requiring each candidate to exceed a reward floor while the batch meets a diversity cutoff—and introduces SatisDive, a training-free inference-time method that applies a batch-relative reward cutoff to boost low-reward candidates and encourage diversity among higher-reward ones. On Pick-a-Pic at matched DreamSim ranges, SatisDive improves worst-candidate reward over FK steering by up to 0.43 with FLUX.1-dev (HPSv3 reward) and up to 0.70 with SANA-1.6B (ImageReward), and its satisfaction–diversity curve Pareto-dominates FK steering in reported settings.
KEY POINTS
- The paper (arXiv:2610.02372v1) frames text-to-image generation as a satisficing problem—requiring each candidate to exceed a reward floor while the batch meets a diversity cutoff—and introduces SatisDive, a training-free inference-time method that applies a batch-relative reward cutoff to boost low-reward candidates and encourage diversity among higher-reward ones.
- On Pick-a-Pic at matched DreamSim ranges, SatisDive improves worst-candidate reward over FK steering by up to 0.43 with FLUX.1-dev (HPSv3 reward) and up to 0.70 with SANA-1.6B (ImageReward), and its satisfaction–diversity curve Pareto-dominates FK steering in reported settings.
- Provides a training-free, principled way to traverse and improve the tradeoff between per-image reward and batch diversity, improving worst-case candidate quality in text-to-image generation.
WHY IT MATTERS
Provides a training-free, principled way to traverse and improve the tradeoff between per-image reward and batch diversity, improving worst-case candidate quality in text-to-image generation.