Beyond Binary Preference: Aligning Diffusion Models to Fine-grained Criteria by Decoupling Attributes

TL;DR

This paper introduces a hierarchical, fine-grained evaluation framework and a two-stage alignment method for diffusion models, enabling them to better match complex human expertise and detailed criteria in image generation.

Contribution

It proposes a novel hierarchical criteria and a two-stage alignment framework, including Complex Preference Optimization, for aligning diffusion models with fine-grained, non-binary human preferences.

Findings

CPO improves image quality and alignment with expert criteria

Hierarchical criteria enable more nuanced model evaluation

Framework applicable to complex, real-world image generation tasks

Abstract

Post-training alignment of diffusion models relies on simplified signals, such as scalar rewards or binary preferences. This limits alignment with complex human expertise, which is hierarchical and fine-grained. To address this, we first construct a hierarchical, fine-grained evaluation criteria with domain experts, which decomposes image quality into multiple positive and negative attributes organized in a tree structure. Building on this, we propose a two-stage alignment framework. First, we inject domain knowledge to an auxiliary diffusion model via Supervised Fine-Tuning. Second, we introduce Complex Preference Optimization (CPO) that extends DPO to align the target diffusion to our non-binary, hierarchical criteria. Specifically, we reformulate the alignment problem to simultaneously maximize the probability of positive attributes while minimizing the probability of negative attributes with the auxiliary diffusion. We instantiate our approach in the domain of painting generation and conduct CPO training with an annotated dataset of painting with fine-grained attributes based on our criteria. Extensive experiments demonstrate that CPO significantly enhances generation quality and alignment with expertise, opening new avenues for fine-grained criteria alignment.