SatisDive
A generation control method applied at the inference stage of text-to-image diffusion models. It reformulates generation as a satisficing task and applies dual constraints—a reward lower bound and diversity truncation—to achieve Pareto-optimal trade-offs between satisfiability and diversity without requiring additional training.
Timeline (last 90 days)
SatisDive is a text-to-image generation method proposed in the arXiv paper "Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion," which reformulates the generation process as a satisficing problem rather than a pure optimization problem.
The approach of combining reward and diversity into a single aggregate score for optimization has a flaw: high diversity can directly offset low reward, meaning the batch may still contain inferior images that clearly deviate from user preferences.
SatisDive sets two hard constraints: a reward lower bound (the minimum reward threshold each candidate image must reach) and a diversity cutoff (the diversity threshold that the entire batch of images as a set must satisfy).
By varying the reward lower bound, a Pareto frontier between satisfaction and diversity can be defined.
SatisDive adopts a training-free strategy that takes effect at inference time, introducing batch-relative reward truncation to distinguish between low-reward and high-reward candidates.
For candidates below the truncation threshold, the algorithm focuses on improving their reward; for candidates above the threshold, it focuses on maintaining diversity among them.
On the Pick-a-Pic dataset using DreamSim as the diversity metric, with FLUX.1-dev as the base model and HPSv3 as the reward model, SatisDive improved the worst candidate reward by up to 0.43 compared to FK steering.
With SANA-1.6B as the base model and ImageReward as the reward model, SatisDive improved the worst candidate reward by up to 0.70 compared to FK steering.
Within the overlapping DreamSim range of the two methods, SatisDive's satisfaction-diversity curve Pareto dominates FK steering's curve in every setting.
SatisDive requires no training and intervenes only at inference time, possessing plug-and-play characteristics; the paper has been validated across multiple combinations including FLUX.1-dev, SANA-1.6B, HPSv3, and ImageReward.
All Facts (10)
SatisDive is a text-to-image generation method proposed in the arXiv paper "Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion," which reformulates the generation process as a satisficing problem rather than a pure optimization problem.
50%UnverifiedThe approach of combining reward and diversity into a single aggregate score for optimization has a flaw: high diversity can directly offset low reward, meaning the batch may still contain inferior images that clearly deviate from user preferences.
50%UnverifiedSatisDive sets two hard constraints: a reward lower bound (the minimum reward threshold each candidate image must reach) and a diversity cutoff (the diversity threshold that the entire batch of images as a set must satisfy).
50%UnverifiedBy varying the reward lower bound, a Pareto frontier between satisfaction and diversity can be defined.
50%UnverifiedSatisDive adopts a training-free strategy that takes effect at inference time, introducing batch-relative reward truncation to distinguish between low-reward and high-reward candidates.
50%UnverifiedFor candidates below the truncation threshold, the algorithm focuses on improving their reward; for candidates above the threshold, it focuses on maintaining diversity among them.
50%UnverifiedOn the Pick-a-Pic dataset using DreamSim as the diversity metric, with FLUX.1-dev as the base model and HPSv3 as the reward model, SatisDive improved the worst candidate reward by up to 0.43 compared to FK steering.
50%UnverifiedWith SANA-1.6B as the base model and ImageReward as the reward model, SatisDive improved the worst candidate reward by up to 0.70 compared to FK steering.
50%UnverifiedWithin the overlapping DreamSim range of the two methods, SatisDive's satisfaction-diversity curve Pareto dominates FK steering's curve in every setting.
50%UnverifiedSatisDive requires no training and intervenes only at inference time, possessing plug-and-play characteristics; the paper has been validated across multiple combinations including FLUX.1-dev, SANA-1.6B, HPSv3, and ImageReward.
50%