PaperlayerView exact revision on arXiv

Source-linked AI summary

Difficulty-Calibrated Interpolation Paths for Conditional Flow Matching

Airin Akter Tania, Md Raihan Khan

arXiv:2608.21286v1cs.CV

TL;DR

Conditional Flow Matching fixes its interpolation schedule in advance even though regression difficulty varies along the path, potentially wasting budget where computation is scarce. The paper proposes Difficulty-Calibrated Flow Matching, which derives a quantile-based schedule from a pilot difficulty profile; it achieves the best CIFAR-10 FID at full sampling budget and gains most in large-batch, few-update settings.

  • Problem

    CFM schedules are fixed independently of the data and model, although regression difficulty varies along the interpolation path and uniform trajectory allocation can waste constrained computation.

  • Method

    DC-FM runs a short linear-path pilot, treats its per-time difficulty profile as a density, and uses its quantile function as the interpolation schedule.

  • Results

    DC-FM achieves the best FID@100 on CIFAR-10 at 5.13, improving over the linear baseline by 0.31 FID, and reaches 8.13 FID@100 versus 10.16 for linear at batch size 512.

  • Takeaways & Limitations

    The calibrated path is competitive across datasets and helps most when update budgets are scarce, while preserving the CFM objective, guidance compatibility, and adding only about 2% training overhead.

Abstract

from arXiv · show

Conditional Flow Matching trains generative models by regressing a network onto the velocity of a prescribed noise-to-data interpolation path. The interpolation schedule that shapes this path is known to affect convergence and sample quality, yet it is invariably fixed in advance, independent of both the data and the model. We show that the regression difficulty of Conditional Flow Matching varies systematically along the path, and we propose Difficulty-Calibrated Flow Matching, which derives the schedule from the model itself: a short pilot run with the linear path records the per-time loss, and the schedule is set to the quantile function of this difficulty profile, so the trajectory lingers where the velocity is hardest to learn. The method has a single hyperparameter, leaves the training objective and its gradient equivalence intact, composes with classifier-free guidance, and adds about two percent training overhead. In controlled experiments on CIFAR-10, MNIST, and Fashion-MNIST with an identical compact U-Net, the calibrated path attains the best FID on CIFAR-10 at full sampling budget and clearly outperforms all fixed schedules in the large-batch, few-update regime, precisely the setting where compute is scarcest.

I. INTRODUCTION

Conditional Flow Matching uses a prescribed interpolation schedule, but regression difficulty varies along the path, making uniform trajectory allocation wasteful under constrained resources. DC-FM measures this difficulty in a pilot run and reshapes the schedule to linger in hard regions while preserving the CFM objective and gradient equivalence.

  • CFM trains a velocity network on closed-form conditional targets for individual noise–data pairs, avoiding direct regression onto an intractable marginal field.
  • Existing interpolation schedules are fixed independently of the data and model, although schedule choice affects convergence and sample quality.
  • Regression difficulty varies along the path, so equal trajectory allocation wastes computation, especially when resources and sampling steps are constrained.
  • DC-FM runs a short linear pilot, treats per-time difficulty as a density, and uses its quantile function to linger where velocity learning is hardest.
  • The method uses one calibration exponent γ, preserves the CFM objective and gradient equivalence, and composes with classifier-free guidance.
  • 5.44 vs. 55.13 FID@100 on CIFAR-10 is reported for DC-FM versus linear, while the method also shows its largest gains in large-batch, few-update settings.

II. RELATED WORK

Related work develops diffusion and flow-based generative models, accelerates few-step generation, and reshapes trajectories through schedules, weighting, or data geometry. DC-FM is positioned among methods that alter path traversal rather than only changing timestep selection.

  • Diffusion and Flow Matching transport a simple prior toward data, with Flow Matching training a velocity field for simulation-free continuous normalizing flows.
  • Acceleration work targets the dominant cost of integrating generative trajectories: the number of function evaluations required for sampling.
  • MeanFlow, SoFlow, AlphaFlow, W-Flow, and Latent Flow Transformer pursue few-step or one-step generation through average velocities, consistency, curricula, or trajectory compression.
  • Closest prior approaches reshape traversal through timestep sampling, noise-level weighting, or data-geometric paths, whereas DC-FM derives its schedule from measured model difficulty.

III. PRELIMINARIES: FLOW MATCHING

Flow Matching prescribes a probability path and trains a network to regress its target velocity, while Conditional Flow Matching makes the target tractable by conditioning on one noise–data pair. The interpolation schedule changes both the path and its velocity field, motivating schedule calibration.

  • Continuous normalizing flows transport a Gaussian source density to the data density by integrating a time-dependent velocity field along an ODE.
  • Flow Matching avoids costly trajectory simulation during training by regressing onto a prescribed target field that generates the chosen probability path.
  • CFM conditions on a single data–noise pair, for which the conditional path and target field are available in closed form.
  • The CFM and FM objectives share the same gradient in θ, making optimization of tractable CFM equivalent to optimization of FM.
  • A monotone schedule α maps noise and data endpoints into an interpolation path, and differentiating that path yields the conditional target velocity.
  • Linear, cosine, and sigmoid schedules produce different paths, and prior work reports that schedule choice affects convergence and sample quality.

IV. DIFFICULTY-CALIBRATED FLOW MATCHING

Fixed schedules allocate trajectory budget uniformly despite unequal regression difficulty along the path. The paper therefore derives the schedule from the model’s measured difficulty to reallocate budget toward difficult regions under constrained computation.

  • Regression difficulty is uneven across t: some path regions dominate the error while others are nearly trivial.
  • Equal trajectory allocation is wasteful under constrained compute, motivating a schedule derived from the model’s measured difficulty.

A. Formulation

The method treats measured per-time learning difficulty as a density over the path and uses its quantile function to construct a valid, difficulty-calibrated schedule. The resulting trajectory slows through hard regions, with γ controlling how strongly difficulty is emphasized.

  • DC-FM defines per-time difficulty as the expected conditional loss and treats it as an unnormalized density over the path coordinate.
  • The schedule is set to the quantile function of the difficulty-derived density, with emphasis exponent γ ≥ 0.
  • If w(s) > 0 on [0, 1], α = R−1 is valid, satisfying α(0) = 0, α(1) = 1, and strict monotonicity.
  • The inverse-function construction provides the schedule velocity in closed form, and α′(t) becomes small where difficulty is large.The trajectory therefore lingers in hard regions and traverses easy regions quickly without changing the CFM formulation.
  • γ = 0 recovers the linear path, γ = 1 makes path-position density proportional to difficulty, and γ > 1 concentrates more aggressively.γ is the method’s single hyperparameter.
  • In practice, pilot losses are averaged into bins, smoothed and floored, cumulatively summed into a piecewise-linear estimate, and inverted by interpolation.The construction adds negligible cost.

B. Learning Objective

DC-FM separates difficulty measurement from standard CFM training: a linear-path pilot estimates difficulty, then a frozen calibrated schedule is used for the main run. Because the schedule is fixed, gradient equivalence remains unchanged, while schedule reshaping changes sampling dynamics rather than merely training-time timestep weights.

  • DC-FM uses two stages so that main training remains standard CFM.
  • The linear-path pilot estimates per-time difficulty because its target norm is independent of t.This makes the measured profile reflect learning difficulty rather than a time-varying target scale introduced by a nonlinear schedule.
  • The calibrated schedule is frozen before training vθ from scratch with the calibrated interpolant.
  • Because α is fixed during Stage 2, CFM gradient equivalence applies unchanged.At inference, samples are generated by integrating the trained velocity field from t = 0 to t = 1.
  • Unlike timestep reweighting, which leaves the inference trajectory unchanged, reshaping α rescales the target velocity and changes sampling dynamics.The paper compares DC-FM with linear training using matched timestep reweighting.

C. Classifier-Free Guidance

Classifier-free guidance is incorporated by combining conditional and unconditional velocity fields at inference. This guidance changes conditional direction, while path calibration changes temporal geometry, so the two mechanisms compose without modification.

  • A single network supports conditional and unconditional fields by randomly replacing the label with a null token during training.The DC-FM objective is unchanged apart from the added conditioning input.
  • At inference, classifier-free guidance forms a guided velocity by extrapolating the conditional field away from the unconditional field.
  • The guidance scale ω satisfies ω = 0 for unconditional sampling and ω > 0 for sharper class alignment.
  • Path calibration reshapes temporal geometry, whereas classifier-free guidance reshapes conditional direction; the two compose without modification.Because the Gaussian interpolant makes the target velocity affine in the endpoints, the guided combination is the flow-matching analogue of score guidance.

V. EXPERIMENTS

The experiments compare interpolation schedules under controlled conditions on CIFAR-10, MNIST, and Fashion-MNIST. Evaluation uses FID at full and low sampling budgets against fixed linear, trigonometric GVP, and logistic schedules.

  • Experiments use CIFAR-10, MNIST, and Fashion-MNIST, with all datasets resized to 32 × 32.Difficulty calibration and ablations use MNIST unless otherwise stated.
  • All methods use an identical compact U-Net with approximately 3.9 M parameters, while only the interpolation schedule differs.Training uses 100 epochs with Adam and otherwise fixed settings, making the comparison controlled.
  • FID is measured against 5,000 real images at NFE ∈ {100, 20}, where lower is better.The two budgets measure quality and low-step efficiency, respectively.
  • DC-FM is compared with linear, trigonometric GVP, and logistic fixed schedules.DC-FM derives α(t) from measured per-time difficulty.

B. Main Results

Across controlled datasets, DC-FM is competitive broadly and strongest on CIFAR-10, while mild calibration outperforms aggressive concentration and gains are largest when updates are scarce.

  • DC-FM achieves the best FID@100 on CIFAR-10 at 5.13, improving over the linear baseline by 0.31 FID.It is within measurement noise of the best method on MNIST and Fashion-MNIST.
  • No fixed schedule dominates across datasets, while the linear path remains the most robust at low NFE.At NFE=20, GVP degrades from 6.84 to 11.06 on MNIST, whereas linear is nearly step-invariant.
  • The main-results comparisons use identical budgets within each table, supporting validity of the within-table method comparisons.
  • Mild calibration is preferable: DC-FM is best near γ = 1 5.81 and degrades for γ ≤0.5.The results indicate that overly aggressive concentration on hard regions harms optimization; as γ →0, the schedule approaches the linear baseline.
  • At batch size 512, DC-FM reaches 8.13 FID@100 versus 10.16 for linear, 10.71 for GVP, and 11.05 for logistic.This is the large-batch, few-update regime targeted by the method.

D. Analysis and Qualitative Results

Pilot difficulty is non-uniform across datasets, and DC-FM converts high-loss regions into slower portions of the interpolation schedule. It produces coherent samples and shows its strongest practical advantage in large-batch, few-update settings, while very low NFE remains a limitation.

  • Difficulty calibration: Pilot loss is high near t=1 and, on CIFAR-10, near t=0, while remaining low in the middle of the path.DC-FM uses this measured profile to slow the schedule through high-loss regions.
  • Compute-constrained evaluation: DC-FM’s advantage grows in the large-batch, few-update regime.The paper identifies this regime as one where fixed schedules waste the most trajectory budget.
  • Qualitative results: DC-FM produces coherent uncurated samples on CIFAR-10 and Fashion-MNIST at NFE=100.The qualitative samples are shown for CIFAR-10 on top and Fashion-MNIST on the bottom.
  • Limitations: At very low NFE, concentrating trajectory time in hard regions can make few-step Euler integration less accurate than the uniform-speed linear path.Closing this low-NFE gap is left to future work.
Loading 2608.21286v1…