- Rating
- 3
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes PathSpeed, an offline training-time sampler for flow matching that allocates timestep probability according to the estimated motion of the prescribed probability-path marginals.
The method estimates squared sliced-Wasserstein path speed from adjacent marginal samples, smooths and standardizes this profile, and converts it into a fixed sampling distribution used throughout training. The paper provides three main analytical results: an exact finite-difference decomposition of squared path speed into projected mean/scale and standardized-shape components; a continuous interpretation of squared speed as the energy of conditional-average projected motion; and a bound relating integrated path speed to changes in persistence diagrams at a fixed observation scale.
Empirically, the method is evaluated on controlled multimodal/topological distributions, COIL-100, CIFAR-10, and class-conditional ImageNet-256. PathSpeed improves structural recovery on the controlled tasks, gives strong gains over Uniform and several alternative timestep schedules on CIFAR-10, and obtains the best reported mean FID among the tested schedules on ImageNet-256. Ablations examine the shape component, probability placement across time, smoothing, calibration, and compression of the learned distribution.
Critical strengths
-
The paper provides a principled and interpretable signal for timestep allocation. Rather than adapting to the model's current loss, PathSpeed measures a property of the prescribed probability path before training. The moment/shape decomposition and continuous conditional-average-motion interpretation make the construction substantially more informative than a purely heuristic schedule.
-
The controlled experiments and ablations are strong. The structural benchmarks are well matched to the paper's motivation, and the permutation/negation controls are particularly convincing: preserving the same sampling probabilities but assigning them to different times removes the gain. The CIFAR-10 result is also substantial, with PathSpeed clearly outperforming Uniform, fixed schedules, Loss-CDF, and InfoNoise while requiring no additional velocity-network evaluation during training.
Critical weaknesses
-
The comparison to recent adaptive allocation methods is incomplete. The related-work section identifies cDOL as a closely related method that combines an irreducible-loss reference, calibrated weighting, and remaining-gap sampling, but cDOL is not included in the experimental comparison. Since the paper's main empirical claim concerns improved training-time allocation, comparison against this recent method would materially affect my assessment of the practical contribution.
-
The large-scale ImageNet advantage is modest and somewhat representation-dependent. PathSpeed obtains mean FID 27.40 versus 27.95 for InfoNoise with standard deviations 0.65 and 0.68 over three runs. This establishes competitive performance but does not, by itself, provide strong statistical evidence of superiority. Moreover, the ImageNet PathSpeed profile is constructed in decoded Inception-feature space, while FID and several other reported diagnostics also rely on feature-space geometry. I would like to understand how sensitive the result is to this representation choice.
A related conceptual limitation is that the theory explains what squared path speed measures and why structural change requires marginal motion, but it does not establish that sampling proportionally to the proposed calibrated speed profile is optimal for finite-budget optimization. I view the theoretical results as principled motivation rather than a convergence guarantee.
Questions for authors
-
Comparison to cDOL: The paper discusses cDOL as a recent data/model-dependent training allocation method but does not compare against it experimentally. Can the authors provide a matched comparison on CIFAR-10, and ideally ImageNet-256? This would materially affect my assessment of the empirical novelty.
-
Representation dependence: How sensitive is the ImageNet-256 result to the representation used to estimate path speed? In particular, do similar gains hold when the schedule is constructed directly in latent space or using a feature encoder different from the Inception representation used by common image-generation metrics?
-
ImageNet uncertainty: The mean ImageNet FID improvement over InfoNoise is 27.95 to 27.40, while the reported run-to-run standard deviations are approximately 0.68 and 0.65. Are the runs paired by seed, and if so, what is the paired difference or confidence interval? This would help determine whether the large-scale gain is consistent across runs rather than primarily a difference in means.
-
Allocation principle: Proposition 3 gives an entropy-regularized variational interpretation of the softmax calibration, but the objective being optimized is not directly a training-convergence objective. Is there a stronger connection between squared path speed and an optimal finite-budget sampling distribution, for example through gradient variance, approximation error, or convergence rate? I view this as a clarification of the theoretical interpretation rather than a requirement for the method to be valid.
3 comments
2. Representation Independence on ImageNet-256 (W2, Q2)
Replacing Inception with DINOv2 preserves the benefit of path-derived allocation. We recomputed the time distribution using frozen DINOv2 ViT-B/14 features and trained under the same class-conditional SiT-B/2 protocol. All results below use 400k updates, no CFG, and Inception-based FID-50K evaluation (mean SD over three training seeds):
| Method | Features used to construct the schedule | FID-50K |
|---|---|---|
| Uniform | — | 33.20 0.65 |
| InfoNoise | — | 27.95 0.68 |
| PathSpeed | DINOv2 | 27.62 0.63 |
| PathSpeed | Inception | 27.40 0.65 |
The DINOv2 construction reaches FID, retaining 96% of the FID reduction over Uniform ( vs. , a difference of FID) and maintaining a lower mean FID than InfoNoise. This experiment separates the encoder used to construct the time distribution from the encoder used for FID.
The improvement also persists throughout training: at 200k updates, the DINOv2 and Inception constructions obtain and , respectively, compared with for Uniform and for InfoNoise (full 100k–400k progression in our response to Reviewer jcKC, Section 1).
By construction, feature extraction occurs only during the one-time precomputation, and velocity training subsequently uses only the cached 64-bin time distribution. We did not test a latent-identity construction on ImageNet-256 within the discussion period. On CIFAR-10, the default schedule is constructed directly in raw pixels (), with no feature encoder, and reduces FID from (InfoNoise) to .
3. Paired ImageNet-256 Runs and Scope of the Gain (W2, Q3)
Yes, the ImageNet-256 runs are paired by training seed. Their individual 400k-update FIDs and interpolated update counts to reach Uniform's final mean FID threshold (33.20) are:
| Paired run | InfoNoise FID | PathSpeed FID | InfoNoise updates to threshold | PathSpeed updates to threshold |
|---|---|---|---|---|
| A | 28.49 | 27.61 | 223k (0.56) | 193k (0.48) |
| B | 27.19 | 26.67 | 197k (0.49) | 188k (0.47) |
| C | 28.17 | 27.92 | 213k (0.53) | 196k (0.49) |
PathSpeed obtains a lower final FID and reaches the threshold earlier in all three paired runs. The per-seed FID differences (PathSpeed minus InfoNoise) are , , and , with a mean paired difference of FID (SD ; 95% paired- CI , ). Because the three-seed interval includes zero, we describe this as competitive ImageNet performance with a consistent direction across runs rather than a statistically conclusive final-FID separation from InfoNoise. In addition, the paired runs reach the 33.20 threshold 30k, 9k, and 17k updates earlier (193k, 188k, and 196k for PathSpeed vs. 223k, 197k, and 213k for InfoNoise).
ImageNet-256 tests whether the allocation stays competitive when the useful time distribution is smooth; there PathSpeed matches the strongest online schedule without online tracking. The paper's main claim concerns structural recovery, where the effect is large: on controlled distributions, whose structural errors are computed directly in data space, PathSpeed reaches Uniform's final error with 33–49% fewer updates and lowers the 2rings loop error from to (Table 1).
4. Connection to Reducible Training Error and Structural Recovery (W3, Q4)
We agree with your interpretation of Proposition 3 as a principled characterization of the calibrated allocation. Its guarantee concerns the stated entropy-regularized objective; we do not have a result that establishes optimality or a convergence rate for training. The connection to finite-budget learning is supported by the conditional-average-motion interpretation and the following measurements.
On Gaussian25, the known mixture distribution allows us to evaluate the conditional-mean velocity and decompose the regression loss across time bins:
where is learnable approximation error and is irreducible pairing variance. Measured at saved checkpoints from Uniform training:
| Correlation with path speed across time | 1k updates | 3.5k updates | 10k updates |
|---|---|---|---|
| Total velocity MSE | -0.99 | -0.97 | -0.97 |
| Approximation error | 0.79 | 0.59 | 0.25 |
| Pairing variance | -0.99 | -0.99 | -0.99 |
Path speed is positively associated with the model's early approximation error ( at 1k updates). This association weakens as training proceeds (). Total MSE remains strongly anticorrelated with speed ( to ) because it is dominated by the pairing variance , whose profile is fixed by the data and path. These measurements connect the offline signal to the error component that learning can reduce.
The structural controls test how this information affects the trained distribution. With representation and calibration held fixed, adding the shape component reduces CIFAR-10 FID by () and reduces the final 2rings loop error tenfold from to (reaching the threshold in 1,000 updates instead of 1,400), while permuting the same probabilities across time removes the benefit (Table 3). Relative to InfoNoise on CIFAR-10, Recall also increases from to and discrepancy decreases from to (Table 2). We have added these new experimental results and analyses to Appendix I in the revised version.
Thank you for recognizing the interpretability of the path-derived signal and the strength of the structural experiments. We provide new results below on the cDOL comparison, representation dependence, paired ImageNet performance, and the relationship between path speed and approximation error.
1. Comparison with the full method of cDOL and transfer to DDPM++/EDM (W1, Q1)
We independently implemented the full method of cDOL, including optimal-loss estimation, calibrated loss weighting, and adaptive sampling. However, in our original TorchCFM setting, this implementation obtained FID-50K 90.53 at 100k updates. That configuration differs from the published cDOL experiments in architecture, training exposure, and inference protocol. We therefore conducted an additional comparison in the DDPM++/EDM framework used by Xu et al.
We applied PathSpeed's core allocation method to this framework and obtained FID-50K 2.14 after 100k updates. We measure changes between adjacent path marginals and cache a fixed training-time distribution before velocity training.
Both methods use the same DDPM++ network, optimizer configuration, batch size of 512, bf16 precision, and deterministic EDM Heun sampler with 35 network evaluations. Training lasts 100k updates, corresponding to 51.2 million training-image presentations. The cDOL implementation uses its calibrated loss weighting and adaptive sampler.
Fixed-budget quality: FID-50K
The table reports every scheduled FID-50K checkpoint. Each method has one training run, and all entries use the same generation and evaluation protocol.
PathSpeed gives lower FID at every evaluated checkpoint. At 100k updates, FID is 2.14 versus 7.43, a 71% reduction in this matched comparison. At just 25k updates, its FID of 3.92 already improves on our cDOL implementation's 100k result.
Earlier quality attainment: FID-10K monitoring
We additionally evaluated smaller generated sets throughout training to observe the early learning trajectory. The following values are all FID-10K and are kept separate from the FID-50K results above.
On this monitoring metric, PathSpeed reaches FID-10K 9.40 at 14.7k updates, already below the 9.43 obtained by our cDOL implementation at 100k. The advantage appears early and persists as both models continue improving.
Our cDOL implementation has not recovered the published performance. Its final schedule assigns 32.9% of the probability mass to logσ<−4; reproducing the authors' allocation in this low-noise region remains unresolved. We therefore identify the comparison throughout as our independent implementation of cDOL.
Quality and training exposure in the published setting
For context, Xu et al. report the following unconditional CIFAR-10 results using 200 million training-image presentations. Their results use the best-checkpoint protocol; our result below is the fixed 100k-update checkpoint.
Applying PathSpeed's core method reaches the quality range of the published FM results using 25.6% of their training exposure. This additional experiment extends the evidence for offline path-aware allocation to a different architecture and training pipeline, with rapid early improvement and continued gains through the final checkpoint.