![[2602.12932] TFTF: Training-Free Targeted Flow for Conditional Sampling](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.12932] TFTF: Training-Free Targeted Flow for Conditional Sampling
Summary
The paper presents a novel training-free method for conditional sampling in flow matching models, addressing the limitations of importance sampling through a resampling technique and stochastic flow adjustments.
Why It Matters
This research is significant as it introduces a method that enhances conditional sampling without the need for extensive training, potentially improving efficiency in machine learning applications, particularly in high-dimensional and multimodal contexts. Its implications extend to various fields, including image generation and data synthesis.
Key Takeaways
- Introduces a training-free method for conditional sampling.
- Addresses weight degeneracy in high-dimensional settings using resampling.
- Demonstrates significant performance improvements on MNIST and CIFAR-10 datasets.
- Applies to multimodal settings, including text-to-image generation.
- Provides theoretical guarantees of asymptotic accuracy.
Statistics > Machine Learning arXiv:2602.12932 (stat) [Submitted on 13 Feb 2026] Title:TFTF: Training-Free Targeted Flow for Conditional Sampling Authors:Qianqian Qu, Jun S. Liu View a PDF of the paper titled TFTF: Training-Free Targeted Flow for Conditional Sampling, by Qianqian Qu and Jun S. Liu View PDF HTML (experimental) Abstract:We propose a training-free conditional sampling method for flow matching models based on importance sampling. Because a naïve application of importance sampling suffers from weight degeneracy in high-dimensional settings, we modify and incorporate a resampling technique in sequential Monte Carlo (SMC) during intermediate stages of the generation process. To encourage generated samples to diverge along distinct trajectories, we derive a stochastic flow with adjustable noise strength to replace the deterministic flow at the intermediate stage. Our framework requires no additional training, while providing theoretical guarantees of asymptotic accuracy. Experimentally, our method significantly outperforms existing approaches on conditional sampling tasks for MNIST and CIFAR-10. We further demonstrate the applicability of our approach in higher-dimensional, multimodal settings through text-to-image generation experiments on CelebA-HQ. Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG) Cite as: arXiv:2602.12932 [stat.ML] (or arXiv:2602.12932v1 [stat.ML] for this version) https://doi.org/10.48550/arXiv.2602.12932 Focus to learn more ...
