Score-Regularized Joint Sampling with Importance Weights for Flow Matching

Researchers propose a new non-IID sampling framework for flow matching models that improves estimation accuracy by jointly drawing diverse samples and using score-based regularization. The method includes importance weighting techniques to enable unbiased estimation while maintaining sample quality and diversity.

• →New non-IID sampling framework addresses high-variance estimates in flow matching models under limited sampling budgets.
• →Score-based regularization (SR) ensures sample diversity within high-density regions while preventing off-manifold drift.
• →Importance weighting approach uses residual velocity fields to enable unbiased estimation from non-IID samples.
• →Method produces more accurate expectation estimates compared to traditional independent sampling approaches.
• →Research advances reliable characterization of flow matching model outputs with publicly available code.