The availability of multiple training algorithms and architectures for generative models requires a selection mechanism to form a single model over a group of well-trained generation models. The selection task is commonly addressed by identifying the model that maximizes an evaluation score based on the diversity and quality of the generated data. However, such a best-model identification ap proach overlooks the possibility that a mixture of available models can outper form each individual model. In this work, we numerically show that a mixture of generative models on benchmark image datasets can indeed achieve a better eval uation score (based on FID and KID scores), compared to the individual models. This observation motivates the development of efficient algorithms for selecting the optimal mixture of the models. To address this, we formulate a quadratic optimization problem to find an optimal mixture model achieving the maximum of kernel-based evaluation scores including kernel inception distance (KID) and Renyi kernel entropy (RKE). To identify the optimal mixture of the models us ing the fewest possible sample queries, we view the selection task as a multi armed bandit (MAB) problem and propose the Mixture Upper Confidence Bound (Mixture-UCB) algorithm that provably converges to the optimal mixture of the in volved models. More broadly, the proposed Mixture-UCB can be extended to op timize every convex quadratic function of the mixture weights in a general MAB setting. We prove a regret bound for the Mixture-UCB algorithm and perform several numerical experiments to show the success of Mixture-UCB in finding the optimal mixture of text and image generative models.
The model selection approach by identifying the score-maximizing model overlooks the possibility that a mixture of the generative models, where each sample is generated from a randomly-selected model can outperform every individual model. This motivates the following question: Can there be real-world settings where a non-degenerate mixture of some well-trained generative models obtain a better evaluation score compared to each individual model? In this work, we numerically show that it is possible for a mixture of real-world generative models to improve evaluation scores over the individual models.
Since the evaluation score of a mixture of generative models can improve over the scores of the individual models, a natural question is how to efficiently compute the weights of an optimal mixture of the models using the fewest possible samples from the models. Here, our goal is to minimize the number of sample generation queries from sub-optimal models, which will save the time and monetary costs of identifying the best model. To achieve this, we propose viewing the task as a multi-armed bandit (MAB) problem, in which every generative model represents an arm and our goal is to find the best mixture of the models with the optimal evaluation score.
Be More Diverse than the Most Diverse:
@inproceedings{
rezaei2025be,
title={Be More Diverse than the Most Diverse: Optimal Mixtures of Generative Models via Mixture-{UCB} Bandit Algorithms},
author={Parham Rezaei and Farzan Farnia and Cheuk Ting Li},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=2Chkk5Ye2s}
}