Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization

2025-02-03

AI Summary

To address pixel-space noise sensitivity, reliance on external vision-language models (VLMs), and low training efficiency in diffusion model preference optimization, this paper proposes Latent-space Step-level Preference Optimization (LPO). LPO innovatively repurposes the diffusion U-Net backbone as a noise-aware latent reward model (LRM), enabling direct preference modeling over noisy latent states and timestep-conditioned reward prediction with latent-space gradient updates. By operating entirely in the latent space, LPO eliminates pixel-space noise interference and VLM dependencies. It achieves state-of-the-art performance across multiple benchmarks in general preference modeling, image aesthetic quality, and image-text alignment. Moreover, LPO improves training efficiency by 2.5×–28× compared to prior methods, demonstrating both effectiveness and scalability.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Learning Preferences or Rankings
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Preference optimization for diffusion models aims to align them with human preferences for images. Previous methods typically leverage Vision-Language Models (VLMs) as pixel-level reward models to approximate human preferences. However, when used for step-level preference optimization, these models face challenges in handling noisy images of different timesteps and require complex transformations into pixel space. In this work, we demonstrate that diffusion models are inherently well-suited for step-level reward modeling in the latent space, as they can naturally extract features from noisy latent images. Accordingly, we propose the Latent Reward Model (LRM), which repurposes components of diffusion models to predict preferences of latent images at various timesteps. Building on LRM, we introduce Latent Preference Optimization (LPO), a method designed for step-level preference optimization directly in the latent space. Experimental results indicate that LPO not only significantly enhances performance in aligning diffusion models with general, aesthetic, and text-image alignment preferences, but also achieves 2.5-28$ imes$ training speedup compared to existing preference optimization methods. Our code will be available at https://github.com/casiatao/LPO.

Problem

• Aesthetic Optimization
• Diffusion Models
• Noise Resilience

Innovation

• Latent Preference Optimization
• Diffusion Models
• Training Efficiency