![[2509.21655] DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2509.21655] DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models
Summary
The paper presents DriftLite, a lightweight approach for inference-time scaling of diffusion models, enhancing adaptation to new distributions without retraining.
Why It Matters
DriftLite addresses the limitations of existing methods in adapting diffusion models, offering a more efficient and stable solution. This is crucial for applications in machine learning where real-time adaptation is necessary, such as in generative AI and data science.
Key Takeaways
- DriftLite provides a training-free method for adapting diffusion models to new distributions.
- The approach utilizes optimal stability control to improve inference dynamics.
- It consistently outperforms existing methods in terms of variance reduction and sample quality.
Computer Science > Machine Learning arXiv:2509.21655 (cs) [Submitted on 25 Sep 2025 (v1), last revised 21 Feb 2026 (this version, v2)] Title:DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models Authors:Yinuo Ren, Wenhao Gao, Lexing Ying, Grant M. Rotskoff, Jiequn Han View a PDF of the paper titled DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models, by Yinuo Ren and 4 other authors View PDF HTML (experimental) Abstract:We study inference-time scaling for diffusion models, where the goal is to adapt a pre-trained model to new target distributions without retraining. Existing guidance-based methods are simple but introduce bias, while particle-based corrections suffer from weight degeneracy and high computational cost. We introduce DriftLite, a lightweight, training-free particle-based approach that steers the inference dynamics on the fly with provably optimal stability control. DriftLite exploits a previously unexplored degree of freedom in the Fokker-Planck equation between the drift and particle potential, and yields two practical instantiations: Variance- and Energy-Controlling Guidance (VCG/ECG) for approximating the optimal drift with minimal overhead. Across Gaussian mixture models, particle systems, and large-scale protein-ligand co-folding problems, DriftLite consistently reduces variance and improves sample quality over pure guidance and sequential Monte Carlo baselines. These results highlight a princ...
