Reproduce “GimmBO: Interactive Generative Image Model Merging via Bayesian Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenxi Liu, Selena Ling, Alec Jacobson
GimmBO searches image-adapter combinations by generating candidates, collecting rankings, and using Bayesian optimization to guide proposals. It constrains total adapter weight, then refines a sparse set. In a 12-participant matching study, median DreamSim similarity was 0.956, versus 0.898 for sliders and 0.926 for Sequential Gallery under an assumed 60-second latency. This controlled result does not establish broader gains.
Key insights
- Effective adapter merges for generative image models often involve a small, sparse subset of available adapters and operate within constrained coefficient magnitude ranges (e.g., sum of coefficients between 1.0 and 3.0).
- Integrating domain-specific knowledge, such as sparsity and magnitude constraints, into a two-stage Bayesian Optimization framework can efficiently navigate high-dimensional, subjective design spaces.
- Preferential feedback (ranking images) combined with intelligent sampling strategies (UCB acquisition function) is effective for user-driven optimization in black-box objective settings, particularly when each evaluation is costly.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.