Conditional Sampling from Frozen Generative Models: From Explicit Rules to Example-Based Guidance

Deep generative models have emerged as both scalable and high-fidelity solutions for generating high-quality synthetic data, effectively capturing the bulk of the training data distribution. However, these models often struggle to adequately generate samples that are rare, underrepresented or that satisfy user-defined conditions or constraints, which are valuable in fields such as finance and healthcare. Retraining generative modes from scratch or using expensive sampling-based methods to capture these targeted outcomes can be computationally prohibitive. To address this challenge, we propose a general framework that enables targeted generation of user-defined conditions from pretrained deep generative models without extensive retraining. Specifically, we address two practical scenarios. In scenarios where explicit rules can evaluate whether generated samples satisfy desired conditions, we propose to use contrastive learning to learn a latent space prior to guide generation towards rule-satisfying outcomes. In settings where only examples of the desired outcomes are provided, we adapt methodologies from the simulation-based inference literature to condition the generation process. Experiments demonstrate that our approach reliably produces condition-satisfying samples, significantly outperforming existing techniques on tabular data in terms of generation quality.