Curriculum Sampling: A Two-Phase Curriculum for Efficient Training of Flow Matching

TL;DR

This paper introduces Curriculum Sampling, a two-phase training schedule for flow matching models that improves convergence speed and fidelity by combining middle-biased and uniform timestep sampling.

Contribution

It proposes a novel two-phase curriculum sampling method that adaptively balances sampling strategies to enhance flow matching training efficiency and quality.

Findings

Improved FID from 3.85 to 3.22 on CIFAR-10

Faster peak performance at 100k steps instead of 150k

Highlights the importance of evolving sampling strategies during training

Abstract

Timestep sampling $p(t)$ is a central design choice in Flow Matching models, yet common practice increasingly favors static middle-biased distributions (e.g., Logit-Normal). We show that this choice induces a speed--quality trade-off: middle-biased sampling accelerates early convergence but yields worse asymptotic fidelity than Uniform sampling. By analyzing per-timestep training losses, we identify a U-shaped difficulty profile with persistent errors near the boundary regimes, implying that under-sampling the endpoints leaves fine details unresolved. Guided by this insight, we propose \textbf{Curriculum Sampling}, a two-phase schedule that begins with middle-biased sampling for rapid structure learning and then switches to Uniform sampling for boundary refinement. On CIFAR-10, Curriculum Sampling improves the best FID from $3.85$ (Uniform) to $3.22$ while reaching peak performance at $100$k rather than $150$k training steps. Our results highlight that timestep sampling should be treated as an evolving curriculum rather than a fixed hyperparameter.