TFTF: Training-Free Targeted Flow for Conditional Sampling
This work addresses conditional sampling for generative models, offering a training-free solution that is incremental but improves performance in specific domains like image generation.
The authors tackled the problem of conditional sampling in flow matching models by proposing a training-free method based on importance sampling and sequential Monte Carlo with stochastic flows, which significantly outperformed existing approaches on MNIST and CIFAR-10 tasks and demonstrated applicability in higher-dimensional settings like CelebA-HQ.
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.