Research map / Bandits and online learning / Multi-armed bandits
Be More Diverse than the Most Diverse: Optimal Mixtures of Generative Models via Mixture-UCB Bandit Algorithms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Mixture-UCB, an online bandit method that selects a mixture of black-box generative models to maximize a kernel-based (KID/RKE-style) diversity-plus-quality objective using as few sample queries as possible.
Problem
Choosing how to combine multiple well-trained generative models: existing work selects a single best model by an evaluation score, while generating samples from many costly models is expensive, so the mixture must be selected with few queries.
Approach
Treat mixture selection as an online learning problem over black-box models: a quadratic kernel-based loss (diversity) plus a linear quality term is minimized via either a UCB-style algorithm (Mixture-UCB-CAB) or online gradient descent (Mixture-UCB-OGD), with a regret-style guarantee for CAB.
Key results
- On FFHQ and LSUN-Bedroom, Mixture-UCB-CAB and OGD yield a lower KID than every individual generator (One-Arm Oracle is the best single model).
- In Table 2 the selected mixture ranks 1st in FID and Coverage and tied 1st in Precision among 5 models.
- Theorem 2 gives an O(sqrt(log T / T)) error bound for Mixture-UCB-CAB, argued to be tight up to logarithmic factors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a Mixture-UCB-CAB or Mixture-UCB-OGD mixture instead of a single generative model | Mixture-UCB · FFHQ and LSUN-Bedroom | KID | Helped |
| Use a Mixture-UCB mixture instead of an individual model | Mixture-UCB · Tables 1 and 2 | FID | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 1 | FID | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 1 | Recall | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 1 | Precision | Hurt |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 1 | Density | Hurt |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 1 | Coverage | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 2 | FID | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 2 | Precision | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 2 | Coverage | Helped |
| Use the selected mixture instead of individual models | Mixture-UCB · Table 2 | Density | Mixed |
| Use a mixture instead of the best individual model | Mixture-UCB · KID-based experiments on LSUN-bedroom and FFHQ | Density | Hurt |
Surprising findings
- Density and Precision are maximized by a mode-collapsed model, so a mixture can lower quality scores relative to the best single model while still improving FID/KID.
Limitations
- No theoretical guarantee for Mixture-UCB-OGD.
- Mixtures may lack coherence and can reduce quality relative to the best single model.
- Requires loading or querying all member models, which may be impractical for memory-constrained deployment.
Experimental setup
Datasets: FFHQ LSUN-Bedroom
Benchmarks: FID KID RKE Precision Density Recall Coverage
Metrics: FID KID RKE Precision Density Recall Coverage
Baselines: One-Arm Oracle offline two-stage mixture best single model by FID
Builds on: UCB KID RKE
Closest papers by meaning
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.88)
- Score-of-Mixture Training: One-Step Generative Model Training Made Simple via Score Estimation of Mixture Distributions (ICML 2025 · similarity 0.88)
- MixMin: Finding Data Mixtures via Convex Minimization (ICML 2025 · similarity 0.85)
- PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation (ICLR 2025 · similarity 0.85)
- An Interpretable Evaluation of Entropy-based Novelty of Generative Models (ICML 2024 · similarity 0.85)
- Towards Theoretical Understandings of Self-Consuming Generative Models (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.