![[2602.20360] Momentum Guidance: Plug-and-Play Guidance for Flow Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20360] Momentum Guidance: Plug-and-Play Guidance for Flow Models
Summary
The paper introduces Momentum Guidance (MG), a novel technique for enhancing flow-based generative models, achieving significant improvements in sample quality without increased computational costs.
Why It Matters
This research addresses the limitations of existing guidance methods in generative modeling, particularly in maintaining sample diversity while improving fidelity. By proposing MG, the authors offer a more efficient alternative that could enhance various applications in machine learning and computer vision.
Key Takeaways
- Momentum Guidance (MG) improves flow-based generative models without increasing inference costs.
- MG utilizes an exponential moving average of past velocities to enhance sample quality.
- Experiments show MG achieves significant improvements in FID scores across benchmarks.
- MG can be combined with classifier-free guidance (CFG) for even better results.
- The method demonstrates consistent quality enhancements in large models like Stable Diffusion 3.
Computer Science > Machine Learning arXiv:2602.20360 (cs) [Submitted on 23 Feb 2026] Title:Momentum Guidance: Plug-and-Play Guidance for Flow Models Authors:Runlong Liao, Jian Yu, Baiyu Su, Chi Zhang, Lizhang Chen, Qiang Liu View a PDF of the paper titled Momentum Guidance: Plug-and-Play Guidance for Flow Models, by Runlong Liao and 5 other authors View PDF HTML (experimental) Abstract:Flow-based generative models have become a strong framework for high-quality generative modeling, yet pretrained models are rarely used in their vanilla conditional form: conditional samples without guidance often appear diffuse and lack fine-grained detail due to the smoothing effects of neural networks. Existing guidance techniques such as classifier-free guidance (CFG) improve fidelity but double the inference cost and typically reduce sample diversity. We introduce Momentum Guidance (MG), a new dimension of guidance that leverages the ODE trajectory itself. MG extrapolates the current velocity using an exponential moving average of past velocities and preserves the standard one-evaluation-per-step cost. It matches the effect of standard guidance without extra computation and can further improve quality when combined with CFG. Experiments demonstrate MG's effectiveness across benchmarks. Specifically, on ImageNet-256, MG achieves average improvements in FID of 36.68% without CFG and 25.52% with CFG across various sampling settings, attaining an FID of 1.597 at 64 sampling steps. Evaluatio...
