Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuning

TL;DR

This paper introduces TFFT, a novel fine-tuning method for flow models that controls tail behavior using CVaR, enhancing reliability and discovery in generative tasks without significant computational overhead.

Contribution

The paper proposes a distributional fine-tuning algorithm based on CVaR that efficiently shapes tail behavior in flow models, addressing a gap in existing entropy-regularized methods.

Findings

TFFT effectively controls tail behavior in generative models.

The method achieves tail shaping with computational costs similar to standard fine-tuning.

Demonstrated success in text-to-image and molecular design tasks.

Abstract

Fine-tuning pre-trained diffusion and flow models to optimize downstream utilities is central to real-world deployment. Existing entropy-regularized methods primarily maximize expected reward, providing no mechanism to shape tail behavior. However, tail control is often essential: the lower tail determines reliability by limiting low-reward failures, while the upper tail enables discovery by prioritizing rare, high-reward outcomes. In this work, we present Tail-aware Flow Fine-Tuning (TFFT), a principled and efficient distributional fine-tuning algorithm based on the Conditional Value-at-Risk (CVaR). We address two distinct tail-shaping goals: right-CVaR for seeking novel samples in the high-reward tail and left-CVaR for controlling worst-case samples in the low-reward tail. Unlike prior approaches that rely on non-linear optimization, we leverage the variational dual formulation of CVaR to decompose it into a decoupled two-stage procedure: a lightweight one-dimensional threshold optimization step, and a single entropy-regularized fine-tuning process via a specific pseudo-reward. This decomposition achieves CVaR fine-tuning efficiently with computational cost comparable to standard expected fine-tuning methods. We demonstrate the effectiveness of TFFT across illustrative experiments, high-dimensional text-to-image generation, and molecular design.