aiqbench analysis
This paper addresses the practical limitations of reward-based diffusion fine-tuning by introducing DOHF, which leverages Doob's $h$-transform to create an efficient online training algorithm. The method effectively handles rare desirable outcomes and non-differentiable rewards without requiring additional network evaluations, offering a robust framework for improving generative model alignment. By unifying various classifier-free guidance methods under a single probabilistic perspective, it advances the theoretical understanding of diffusion model optimization.
DOHF: Online Diffusion Fine-tuning with Doob's $h$-transform Guidance
arXiv:2609.31882v1 Announce Type: new Abstract: Reward-based diffusion fine-tuning faces practical challenges when desirable outcomes are rare or conditioning corrections are costly to estimate. In this work, we propose Diffusion Online $h$-guidance Fine-tuning (DOHF), which turns Doob's $h$-transform into a practical online training algorithm. DOHF assigns optimality weights to generated samples, estimates the normalized local correction $\nabla\log h$ under the current rollout policy, and distills it directly into the generative model. Theoretically, we characterize the population-optimal DiffusionNFT update as well as the various classfier free guidance methods through a unified $h$-transform perspective. Methodologically, our framework accommodates black-box and non-differentiable rewards without additional network evaluations. We further show improved alignments under three empirical scenarios. Our work demonstrates how adapting probabilistic conditioning through inexpensive estimation and iterative distillation can improve generative learning across statistical sampling and visual generation.
Abstract:Reward-based diffusion fine-tuning faces practical challenges when desirable outcomes are rare or conditioning corrections are costly to estimate. In this work, we propose Diffusion Online $h$-guidance Fine-tuning (DOHF), which turns Doob's $h$-transform into a practical online training algorithm. DOHF assigns optimality weights to generated samples, estimates the normalized local correction $\nabla\log h$ under the current rollout policy, and distills it directly into the generative model. Theoretically, we characterize the population-optimal DiffusionNFT update as well as the various classfier free guidance methods through a unified $h$-transform perspective. Methodologically, our framework accommodates black-box and non-differentiable rewards without additional network evaluations. We further show improved alignments under three empirical scenarios. Our work demonstrates how adapting probabilistic conditioning through inexpensive estimation and iterative distillation can improve generative learning across statistical sampling and visual generation.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.31882 [cs.LG]
(or arXiv:2609.31882v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.31882 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 227 words
Related coverage
Runway’s WorldPrompt and the Engineering of Real-Time Worlds
Latent Space · Sep 25, 2026
MORPH: Generative Retrieval via Diffusion Transformer with Metric-Ordered Sequence Training and Hybrid-Policy Preference Optimization
arXiv · Oct 7, 2026
EWAM: An Enhanced World Action Model for Closed-Loop Online Adaptation in Embodied Intelligence
arXiv · Jun 12, 2026
EntroPack: Fast and Accurate Entropy-Coded Weight Compression at Arbitrary Bitrates
arXiv · Sep 30, 2026