- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
RAFM uses an empirical radial source and spherical paths to separate radial estimation from directional transport in flow matching. RAFM-Ang normalizes velocity targets by radius to obtain bounded angular targets, motivated by the radial-moment requirements of conventional velocity regression. The paper provides theoretical analyses and evaluates the method on heavy-tailed synthetic data, real vector datasets, and image and audio generation.
Critical strengths
-
Clear mathematical motivation. The radial-angular decomposition provides a coherent motivation for the method. The analysis makes the radial weighting in velocity regression explicit and establishes bounded angular targets, which is relevant to heavy-tailed settings.
-
Informative controlled comparisons. The matched-source Euclidean FM and fixed-spherical baselines provide useful comparisons for assessing the source and transport design. The Ang-R² ablation directly tests loss weighting while keeping the angular parameterization and sampling field fixed.
-
Clear presentation. The paper is well organized and easy to follow. The progression from radial source matching to spherical transport and angular regression makes the motivation and method clear.
Critical weaknesses
-
Modest methodological contribution. RAFM combines empirical radial matching with spherical flow matching and uses angular regression to avoid the implicit R² weighting of velocity regression. The spherical construction builds on RFM, while norm-preserving generation has a precedent in MSGM. The analysis is useful, but I find the methodological contribution incremental.
-
Unclear quality–cost trade-off. RAFM has substantially lower reported training cost than MSGM, but Table 4 shows that RAFM-Ang requires roughly 1.9–3.5× the training time of Gaussian FM on Student-t/PIV. The comparisons use the same number of updates. It is unclear whether the gains persist when Gaussian FM is given the same training-time budget.
-
Limited image-generation evaluation. Natural-image generation is evaluated only on ImageNette with one DC-AE/SiT configuration. It is unclear whether the gains over Gaussian FM extend to other image datasets or model configurations.
Questions for authors
-
Does RAFM-Ang still outperform Gaussian FM on Student-t/PIV when both methods are given the same wall-clock training budget?
-
Could the authors compare RAFM-Ang and Gaussian FM on another natural-image generation benchmark, such as ImageNet?
We thank the reviewer for the constructive feedback. RAFM primarily targets distributions with challenging radial structure, especially heavy-tailed or strongly non-Gaussian data. Our additional experiments confirm its benefits at equal wall-clock time and demonstrate scalability to ImageNet-256.
W1 (contribution). We acknowledge the precedents of spherical paths (RFM) and norm-preserving generation (MSGM). Our contribution is scale-free angular regression: standard velocity regression implicitly weights the angular loss by R2, whereas RAFM-Ang removes this weighting, has targets bounded by π independently of radial moments, and recovers the FM velocity through rescaling. Training remains simulation-free, unlike MSGM.
A new heavy-tail sweep (Student-t, ν=1.5–5, d=16/32) confirms the mechanism: RAFM-Ang's loss max/median is ≤3.3, versus 34–330 for Gaussian FM, Matched FM, RAFM-Vel, and Student-t FM at ν=1.5. Restoring the R2 weighting (Ang-R2) restores the spikes. RAFM-Ang beats Gaussian FM in all 8 settings and achieves the best Sliced W1 in 5. Further comparisons with spherical controls and MSGM are discussed in our response to reviewer n2NE, with supporting results in its follow-up comment (Tables R1–R3).
W2/Q1 (equal wall-clock). On MI210 GPUs, Gaussian FM was allowed the same training time as RAFM-Ang's 10k updates, completing approximately 11.5k–13.8k updates in the first three settings. Gaussian FM checkpoints were evaluated every 500 updates. Sliced W1 ↓, mean ± SD over 3 seeds:
RAFM-Ang outperforms even Gaussian FM's best checkpoint within the budget in all four settings. Student-t uses freshly drawn samples, explaining differences from the original manuscript results.
The 1.9–3.5× training-time ratios reported in Table 4 of the submitted paper were measured on an RTX 2000 Ada. On MI210, RAFM-Ang's per-update overhead is 15–38%.
W3/Q2 (ImageNet-256). We additionally evaluated class-conditional ImageNet-256 using DC-AE latents and SiT-B/1, training for 400k updates with 3 seeds per method. FID-10k ↓ (mean ± SD):
RAFM-Ang achieves a slightly lower mean FID at every checkpoint, reaching 35.34 versus 36.37 at 400k updates, with comparable training times (8.98 h versus 8.92 h). These results demonstrate scalability to high-dimensional image latents without substantial computational overhead. We emphasize that our primary contribution concerns challenging non-Gaussian radial distributions, where the gains are considerably more pronounced.
Revisions. We will incorporate these experiments, report equal-time budgets and standard deviations, clarify the PIV architectural limitation, and strengthen the discussion of RAFM's scope and relationship to prior work.
We hope these additional experiments and clarifications address the reviewer's concerns and would be happy to address any further questions.