TFTF: Training-Free Targeted Flow for Conditional Sampling

2026-02-16 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences, Data Science & Analytics · Depth: Expert, extended

Summary

A new training-free conditional sampling method, TFTF (Training-Free Targeted Flow), is proposed for flow matching models, addressing the limitations of existing approaches. This method utilizes importance sampling (IS) and incorporates a Sequential Monte Carlo (SMC) resampling technique during intermediate generation stages to mitigate weight degeneracy in high-dimensional settings. To enable diverse trajectories for resampled particles, TFTF introduces a stochastic flow with adjustable noise strength, replacing the deterministic flow. The framework requires no additional training and offers theoretical guarantees of asymptotic accuracy. Experimental results demonstrate that TFTF significantly outperforms existing methods in conditional sampling tasks on MNIST and CIFAR-10, and it is also applicable to higher-dimensional, multimodal text-to-image generation on CelebA-HQ.

Key takeaway

For research scientists and machine learning engineers working on conditional generative models, TFTF offers a robust, training-free approach to achieve high-fidelity, condition-compliant samples. You should consider integrating TFTF into your workflow to bypass the computational and data labeling costs associated with training-based methods, especially when dealing with high-dimensional data or multimodal generation tasks like text-to-image synthesis.

Key insights

TFTF enables training-free conditional sampling in flow matching by combining importance sampling with stochastic flow and SMC resampling.

Principles

• Importance sampling suffers weight degeneracy in high dimensions.
• Stochasticity is crucial for particle diversification during resampling.
• Resampling at intermediate stages improves sample quality and efficiency.

Method

TFTF constructs a conditional velocity field, applies importance weighting, and integrates SMC resampling with a derived stochastic flow to manage weight degeneracy and ensure diverse sample trajectories.

In practice

• Use TFTF for conditional sampling without model retraining.
• Apply stochastic flow to enable particle diversification.
• Implement resampling in intermediate generation stages.

Topics

• Flow Matching Models
• Conditional Sampling
• Sequential Monte Carlo
• Stochastic Flow
• Importance Sampling

Code references

• VSehwag/minimal-diffusion
• Curt-Park/handwritten_digit_recognition
• gnobitab/RectifiedFlow
• huyvnphan/PyTorch_CIFAR10
• BillyXYB/TransEditor

Best for: Computer Vision Engineer, Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows — Training models to utilize their own alignment error as feedback enables self-correction in conditional generative tasks.
• A Time-Reparameterized Cumulative Intensity Extrapolation Sampler for Discrete Flow Matching — TR-CIE improves discrete flow matching sampling quality under limited function evaluations by reparameterizing time and extrapolating cumulative intensities.
• A unified perspective on fine-tuning and sampling with diffusion and flow models — A unified framework clarifies fine-tuning and sampling in diffusion/flow models via exponential tilting.
• Coreset-Induced Conditional Velocity Flow Matching — CCVFM improves generative flow matching by using a data-informed coreset surrogate instead of isotropic noise.
• Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching — Optimizing source distribution in flow matching models significantly improves text-to-image generation performance and convergence speed.
• CrossFlow: One-Step Generation Across Latent and Pixel Spaces — CrossFlow directly maps noisy latents to pixel-space images, combining latent efficiency with direct pixel supervision in one model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.