Flow matching trains a continuum of velocity regression problems but usually reports a loss averaged over interpolation time, obscuring which regions are limited by intrinsic ambiguity and which remain data-limited. We introduce the Velocity Ambiguity Profile (VAP), $\mathcal{A}_v(t)$, the time-local Bayes-risk floor of velocity prediction, and decompose time-resolved loss as $L_N(t)=\mathcal{A}_v(t)+R(t;N)$, where only the excess term $R(t;N)$ is reducible by more data or a different estimator. For $N$ training samples and a $K$-component mixture in the structured example, we prove two statistical benchmarks. In an intrinsic local-regression benchmark, $\mathcal{A}_v(t)$ enters the time-dependent noise factor of the standard nonparametric rate. In separated isotropic Gaussian mixtures, the Bayes velocity becomes locally affine, and a label-oracle residual-regression benchmark has parametric excess-risk scale $K\mathcal{A}_v(t)/N$, matched within the same separated residual Gaussian-location subproblem. Controlled Gaussian-mixture experiments validate this diagnostic through a dense transition sweep, calibrated oracle-to-blind and empirical-proxy audits, and a local controlled time-sampler stress test in which a frozen reducible-error proxy reduces reducible loss while cross-cell sensitivity delimits aggressive sampling. On CIFAR-100 DINOv2 feature flow, guided time sampling reduces uniformly evaluated velocity loss by 6.78% and global feature Fréchet distance (FD) by 23.6%, while class-conditional kernel inception distance (KID) shows no improvement. VAP turns an averaged flow-matching objective into a time-resolved diagnostic for locating ambiguity-limited and data-limited regions.
Flow Matching and the Averaged Loss Problem
Flow matching has become a central paradigm in generative modeling, offering a flexible alternative to diffusion-based and adversarial approaches. The method trains a vector field that transports a simple base distribution to a data-dependent target distribution by solving a sequence of ordinary differential equations. During training, a continuum of velocity regression problems is posed at different interpolation times between the base sample and the data sample. The standard objective aggregates the regression loss across this entire time interval, typically reporting a single scalar that averages performance over all t. This averaging masks temporal variation: some interpolation intervals may be dominated by the intrinsic structure of the data manifold, while others remain limited by the amount of training data or the capacity of the model. Without a time-resolved diagnostic, researchers cannot distinguish between these two sources of error, and therefore cannot direct capacity or data-collection efforts with precision.
The paper identifies this limitation explicitly. In the prevailing practice, a single loss value such as $\mathbb{E}_{t\sim\mathcal{U}[0,1]}[\|v_\theta(x_t) - \tilde{v}(x_t,t)\|^2]$ is reported, where $v_\theta$ is the predicted velocity and $\tilde{v}$ is the target. The time index t is integrated out, and the resulting number is treated as the overall quality of the model. While convenient for comparison, this scalar hides the fact that early interpolation steps may be data-rich and learnable, whereas later steps near the data manifold may be ambiguous by design. The authors argue that a diagnostic that separates the unavoidable ambiguity from the reducible error is essential for meaningful progress.
Introducing the Velocity Ambiguity Profile
The core contribution of the paper is the Velocity Ambiguity Profile (VAP), denoted $\mathcal{A}_v(t)$. This quantity represents the time-local Bayes-risk floor for velocity prediction: the minimum expected loss that any estimator cannot avoid, regardless of data quantity or architectural choices. The decomposition $L_N(t)=\mathcal{A}_v(t)+R(t;N)$ splits the time-resolved loss into two parts. The first, $\mathcal{A}_v(t)$, is the ambiguity floor and depends only on the underlying data distribution and the geometry of the velocity regression problem at time t. The second, $R(t;N)$, is the excess risk that diminishes as the number of training samples N grows or as the estimator improves. Only $R(t;N)$ is reducible; $\mathcal{A}_v(t)$ is irreducible by definition.
The VAP is defined through the conditional distribution of the velocity given the interpolated point $x_t$ and the time t. Concretely, $\mathcal{A}_v(t)$ is the expected minimum mean-squared error of the true velocity under the optimal Bayesian estimator, averaged over the distribution of $x_t$ at that t. Because this quantity depends on the true data-generating process, it is generally intractable to compute exactly. However, the paper provides theoretical characterizations in stylized settings and empirical proxies that make the concept operational.
The decomposition has a clear interpretation. If at a particular t the VAP is high, it means that the velocity function itself is highly variable or ill-defined in that region of the interpolation space, and no amount of data will make the regression problem easier—the ambiguity is intrinsic. If the VAP is low but the total loss remains high, the excess $R(t;N)$ is large, indicating that more data or a better estimator could substantially improve performance. This separation directs attention to the right knob: reducing ambiguity requires modeling or structural changes, while reducing excess risk requires more data or optimization improvements.
Theoretical Benchmarks and Statistical Characterizations
The paper proves two statistical benchmarks that relate the VAP to known rates in nonparametric regression and mixture models. The first benchmark considers an intrinsic local-regression setting. In this regime, the VAP enters the time-dependent noise factor of the standard nonparametric rate. Specifically, the minimax risk for estimating the velocity at time t scales as $n^{-\frac{2}{2+d_t}}$ where $d_t$ is an effective dimension that depends on the local geometry of the velocity field and is modulated by $\mathcal{A}_v(t)$. The ambiguity profile thus directly inflates the irreducible error term in the convergence rate, making explicit how much of the difficulty is rooted in the data structure rather than insufficient samples.
The second benchmark addresses separated isotropic Gaussian mixtures. Here the authors consider a K-component mixture where each component has an isotropic covariance matrix. Under separation conditions—meaning the components are well-separated relative to their widths—the Bayes velocity becomes locally affine in a neighborhood of each component. This affine structure simplifies the prediction problem considerably. The paper then analyzes a label-oracle residual-regression benchmark, where the oracle provides the true component label and the learner only needs to regress the velocity conditional on that label. The excess risk in this subproblem scales as $K\mathcal{A}_v(t)/N$, showing that the ambiguity profile directly amplifies the sample complexity penalty by a factor proportional to the number of mixture components. This result has a clean intuition: each component contributes its own ambiguity floor, and the total reducible error aggregates linearly with K.
These benchmarks are not just abstract; they are shown to match the empirical behavior in controlled experiments. The parametric scale $K\mathcal{A}_v(t)/N$ is recovered in the separated Gaussian experiments, confirming that the VAP captures a genuinely reducible component of the error that scales with both the number of components and the inverse sample size.
Experimental Validation on Gaussian Mixtures
To validate the diagnostic, the authors conduct a series of controlled experiments on Gaussian-mixture data. They sweep across a dense grid of interpolation times t, computing the VAP empirically via oracle-based estimates and comparing it to the total loss $L_N(t)$. The transition from ambiguity-dominated to data-dominated regimes is mapped out, showing that for early t the loss is largely attributable to $\mathcal{A}_v(t)$, while for later t the excess $R(t;N)$ becomes the primary contributor. This transition sweep provides a visual and quantitative confirmation of the decomposition.
In addition, the paper reports calibrated oracle-to-blind audits. The oracle denotes a benchmark that has access to the true velocity and component labels; the blind refers to a model that must predict the velocity from data alone without oracle assistance. The gap between oracle and blind performance at each t is precisely the reducible excess $R(t;N)$. By tracking this gap across t, the experiments confirm that the VAP predicts where the blind model will struggle and where it has headroom for improvement.
A local controlled time-sampler stress test further demonstrates the diagnostic's utility. In this test, a frozen reducible-error proxy is used to reduce the excess loss $R(t;N)$ while keeping the ambiguity floor fixed. The experiment shows that reducing $R(t;N)$ in selected time cells leads to proportional decreases in the total loss, but that aggressive sampling across cells where the VAP is high yields diminishing returns. Cross-cell sensitivity analysis reveals that sampling strategies must respect the VAP landscape: allocating samples to regions where the ambiguity is low produces the largest gains in overall model performance.
Application to CIFAR-100 DINOv2 Feature Flow
The ultimate testbed for the VAP is a flow-matching model trained on CIFAR-100 features extracted from a DINOv2 vision backbone. The model learns to map noise to data across the interpolation time, effectively learning a generative process over image features. The authors apply guided time sampling, a training strategy that uses the VAP to allocate sampling probability mass across t in proportion to the estimated reducible excess. The idea is simple: if the VAP indicates that a given t is data-limited, the sampler spends more steps there; if the VAP indicates intrinsic ambiguity, the sampler can allocate fewer steps, knowing that additional data will not substantially reduce loss.
The results are striking. Guided time sampling reduces the uniformly evaluated velocity loss by 6.78% compared to uniform time sampling. This reduction demonstrates that the VAP-directed allocation of training effort is effective at lowering the excess risk across the interpolation interval. Global feature Fréchet distance (FD), a metric of distributional similarity between generated and real features, improves by 23.6%. This substantial drop in FD indicates that the generated features are markedly closer to the true data distribution when the sampler respects the VAP landscape.
In contrast, class-conditional kernel inception distance (KID) shows no improvement. KID measures the discrepancy between conditional feature distributions across classes, and its lack of change suggests that the guided sampling primarily affects the overall feature geometry rather than class-specific separation. The authors note that this outcome is consistent with the VAP diagnosis: the ambiguity and excess risk are distributed across the feature space in a way that tightens the global arrangement but does not necessarily sharpen class boundaries. The null result for KID also serves as a useful cautionary tale—diagnostics that work for one metric may not translate to others, and the VAP should be interpreted in conjunction with the specific evaluation criteria of the task.
Interpretation of Results
The 6.78% reduction in velocity loss and the 23.6% drop in Fréchet distance are not merely incremental; they illustrate a principled method for improving flow-matching models without changing the underlying architecture or increasing the total number of training samples. By reweighting the time distribution according to the VAP, the model effectively focuses compute on the times where the excess risk is highest, achieving more with the same budget. This has practical implications for research teams that face fixed compute budgets and need to decide how to allocate training iterations across the interpolation interval.
The absence of KID improvement also highlights an important nuance. The VAP diagnoses ambiguity and excess risk in the velocity regression problem, but the downstream effect on class-conditional metrics depends on how the generated features are used. If the application requires tight class separation, additional techniques such as class-conditional loss terms or feature-level augmentation may be needed alongside VAP-guided sampling. The paper does not claim that the VAP is a panacea for all generative modeling concerns; rather, it provides a targeted diagnostic that surfaces where the flow-matching objective can be improved.
Limitations and Trade-offs
The VAP framework relies on theoretical characterizations that are exact in stylized settings but approximate in high-dimensional, real-world data. The Gaussian-mixture benchmarks, while informative, do not capture the full complexity of natural image manifolds, where the velocity field may have intricate dependencies that are not captured by simple mixture models. The paper acknowledges that the empirical proxies for the VAP, such as oracle estimates and local linear approximations, introduce their own sources of error and should be interpreted as directional guides rather than precise measurements.
Another trade-off arises from the cost of estimating the VAP itself. Computing the Bayes-risk floor requires knowledge of the true velocity distribution, which is generally unavailable. The paper uses oracle-based calculations and local linear approximations, both of which are computationally additional burdens on top of the flow-matching training loop. Researchers must weigh the benefit of the diagnostic against the overhead of its estimation, especially in large-scale experiments where every additional computation can be significant.
A third limitation concerns the generality of the guided-sampling strategy. The experiments show that allocating sampling probability according to the VAP reduces velocity loss and improves Fréchet distance, but the optimal allocation rule may differ across model families, data modalities, and task objectives. The paper provides a general framework, but the specific implementation of time reweighting should be tailored to the particular generative modeling setup. Future work could explore adaptive VAP estimation that updates during training, or integration with other sampling schemes such as curriculum learning or importance sampling.
What It Means in Practice
For practitioners working with flow-matching models, the VAP offers a concrete method for diagnosing and addressing the two sources of error that govern training efficiency. Rather than monitoring a single averaged loss and wondering why performance plateaus, a team can compute or approximate the VAP at several time steps and identify regions where the loss is dominated by ambiguity versus excess risk. If the excess risk is high in certain intervals, guided time sampling or targeted data collection can yield measurable improvements, as the CIFAR-100 experiments demonstrate.
The diagnostic also informs architectural choices. A model that consistently shows high VAP at late interpolation times may benefit from a architecture that better captures the geometry of the data manifold near the target distribution, such as conditional normalization layers or adaptive feature embeddings. Conversely, if the VAP is low across the board but the total loss remains high, the bottleneck is likely optimization or data quantity, and the research effort should focus on stronger optimizers or larger training sets.
Beyond the immediate benefits of improved generative models, the VAP contributes to a broader shift toward time-resolved and capability-aware evaluation in diffusion and flow-based modeling. Just as the PAPAV framework in the embodied-agent paper calls for measuring Perceive, Anticipate, Plan, Act, and Verify separately, the VAP calls for evaluating flow-matching models across the interpolation interval rather than as a single scalar. This shift toward fine-grained diagnostics is likely to become more common as the community moves toward models that are not only high-performing on average but also well-understood in their temporal dynamics.
A third practical takeaway is the importance of reporting multiple metrics, as the paper does with velocity loss, Fréchet distance, and KID. Each metric captures a different aspect of the generated distribution, and a single number can conceal important failure modes. The VAP adds another layer of granularity, enabling researchers to see where in the interpolation space the model struggles and why. This practice of multi-metric reporting, combined with time-resolved diagnostics, will improve the reproducibility and interpretability of future work.