![[2602.16198] Training-Free Adaptation of Diffusion Models via Doob's $h$-Transform](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16198] Training-Free Adaptation of Diffusion Models via Doob's $h$-Transform
Summary
This paper presents a novel training-free adaptation method for diffusion models, leveraging Doob's $h$-transform to enhance sampling efficiency without additional training overhead.
Why It Matters
The research addresses the limitations of existing adaptation methods for diffusion models, which often require extensive training and specific reward assumptions. By proposing a computationally efficient approach, this work has implications for various applications in machine learning, particularly in reinforcement learning and generative modeling.
Key Takeaways
- DOIT (Doob-Oriented Inference-time Transformation) offers a training-free adaptation method for diffusion models.
- The method efficiently transports pre-trained generative distributions to high-reward targets without modifying the original model.
- Theoretical guarantees of convergence to target distributions are established, enhancing the reliability of the approach.
- Empirical results show that DOIT outperforms existing state-of-the-art methods on D4RL offline RL benchmarks.
- This research opens avenues for further exploration in non-differentiable reward scenarios.
Computer Science > Machine Learning arXiv:2602.16198 (cs) [Submitted on 18 Feb 2026] Title:Training-Free Adaptation of Diffusion Models via Doob's $h$-Transform Authors:Qijie Zhu, Zeqi Ye, Han Liu, Zhaoran Wang, Minshuo Chen View a PDF of the paper titled Training-Free Adaptation of Diffusion Models via Doob's $h$-Transform, by Qijie Zhu and 4 other authors View PDF HTML (experimental) Abstract:Adaptation methods have been a workhorse for unlocking the transformative power of pre-trained diffusion models in diverse applications. Existing approaches often abstract adaptation objectives as a reward function and steer diffusion models to generate high-reward samples. However, these approaches can incur high computational overhead due to additional training, or rely on stringent assumptions on the reward such as differentiability. Moreover, despite their empirical success, theoretical justification and guarantees are seldom established. In this paper, we propose DOIT (Doob-Oriented Inference-time Transformation), a training-free and computationally efficient adaptation method that applies to generic, non-differentiable rewards. The key framework underlying our method is a measure transport formulation that seeks to transport the pre-trained generative distribution to a high-reward target distribution. We leverage Doob's $h$-transform to realize this transport, which induces a dynamic correction to the diffusion sampling process and enables efficient simulation-based computatio...
