Velocity Scheduled Flow Matching

2026-07-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

Velocity Scheduled Flow Matching (VSFM) introduces a novel approach to generative modeling by relaxing the constant-speed assumption inherent in traditional Flow Matching (FM). VSFM replaces the standard conditional target x_1 - x_0 with v(t)(x_1 - x_0), where v:[0,1]->R>=0 is a nonnegative velocity profile satisfying integral(v dt) = 1. The method explores six polynomial profiles derived from motion planning. VSFM offers two primary benefits: first, it allows a pretrained linear flow-matching model to be sampled at inference time using any admissible profile via non-uniform τ-schedule integration, achieving up to a 19.8% FID reduction on CIFAR-10 without retraining. Second, training a model from scratch with a braking profile yields an additional 17.4% FID reduction at 4 NFE. These improvements stem from the local truncation error of the Euler integrator on the generated grid.

Key takeaway

For Machine Learning Engineers optimizing generative model performance, Velocity Scheduled Flow Matching offers a direct path to improved sample quality and efficiency. You should consider applying non-uniform τ-schedules to your existing pretrained linear flow-matching models to achieve significant FID reductions without retraining. When developing new models, training with a braking velocity profile can further reduce the number of network evaluations required, enhancing computational efficiency. This method provides tangible gains by leveraging the local truncation error of the Euler integrator.

Key insights

Velocity Scheduled Flow Matching improves sample quality and efficiency by dynamically adjusting the interpolation speed between noise and data.

Principles

• Flow matching's constant speed is a relaxable assumption.
• Non-uniform sampling schedules improve inference.
• Training with braking profiles boosts efficiency.

Method

VSFM replaces the conditional target with v(t)(x_1 - x_0) using a nonnegative profile v where integral(v dt) = 1, integrating the ODE on a non-uniform τ-schedule.

In practice

• Integrate pretrained models with non-uniform τ-schedules.
• Train new models using braking velocity profiles.

Topics

• Velocity Scheduled Flow Matching
• Flow Matching
• Generative Models
• Neural Networks
• ODE Solvers
• Image Generation

Code references

• YangLing0818/consistency_flow_matching

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

Related on AIssential

• ConFlow: Constraints-Guided Learning with Flow Matching for Motion Generation — Integrating constraints directly into training objectives closes the training-inference gap in generative motion models.
• Energy-Weighted Flow Matching: Unlocking Continuous Normalizing Flows for Efficient and Scalable Boltzmann Sampling — Energy-Weighted Flow Matching enables continuous normalizing flows to model Boltzmann distributions using only energy evaluations, not target samples.
• Topological Flow Matching — Topological Flow Matching enhances generative models by integrating domain topology into flow matching for structured data.
• Context-weighted Discrete Flow Matching — Local context density is crucial for reducing uncertainty and improving efficiency in discrete generative modeling.
• Optimal Self-Distillation for Rectified Flow via Linear Probing — Optimal self-distillation provably improves rectified flow models by deriving and applying an optimal mixing coefficient for teacher and true velocities.
• Decision-Weighted Flow Matching for Contextual Stochastic Optimization — Decision-Weighted Flow Matching (DW-FM) aligns generative model training with downstream decision regret by reweighting objectives based on decision sensitivity.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.