Research map / Probabilistic and scientific modeling / Variational and generative sampling

What Time Is It? How Data Geometry Makes Time Conditioning Optional for Flow Matching

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Flow matching and flow sampling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that time-blind flow matching works because high-dimensional low-rank data geometry makes t identifiable from the noisy interpolant, so the time-blindness gap is negligible relative to coupling variance; coupling (OT) is the main lever.

Problem

Why flow matching models trained without explicit time conditioning work at all, given that time is usually thought necessary to disambiguate velocity targets.

Approach

Decomposes the time-blind FM loss into learnable error, coupling variance, and a time-blindness gap; under a spiked-covariance model, the residual-subspace norm of the interpolant gives a closed-form estimator of t at rate O(1/sqrt(d-k)), so the gap is O(1/(d-k)) times the coupling variance.

Key results

• On CIFAR-10 the coupling effect is roughly 25x the time effect; dropping time conditioning costs under 0.7% of total loss.
• Time estimator error on CIFAR-10, CelebA-HQ and FFHQ has mean absolute error 0.006, 0.018 and 0.021 when t is sampled uniformly over [0,1].
• OT coupling improves FID on every dataset whether the model is time-conditioned or time-blind.

Surprising findings

• Time-blind models can recover t from a single noisy interpolant nearly exactly, so explicit time conditioning adds little information.
• Changing the coupling matters far more than removing time conditioning for loss and FID.

Limitations

• Identifiability theory is proven under a spiked-covariance Gaussian model, not for real non-Gaussian data.
• The clock is branchwise: time is identifiable only on intervals where the variance map is one-to-one.
• Does not verify which mechanism trained networks actually use to infer time.

Experimental setup

Datasets: CIFAR-10 CelebA-HQ FFHQ synthetic spiked covariance

Benchmarks: CIFAR-10 CelebA-HQ FFHQ

Metrics: test-set MSE FID-50k estimation error of t

Baselines: time-conditioned flow matching independent coupling

Builds on: flow matching minibatch optimal transport straightness of paths blind denoising diffusion

Closest papers by meaning

• PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation (NeurIPS 2026 · similarity 0.86)
• MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series (NeurIPS 2026 · similarity 0.86)
• Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.86)
• On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.86)
• Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.85)
• On Fitting Flow Models with Large Sinkhorn Couplings (NeurIPS 2026 · similarity 0.85)