- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Polar Diffusion, which is a geometric reformulation that factorizes the diffusion process over a product space of magnitude (radial) and normalized direction (tangential). And the authors derive the corresponding forward and reverse processes, establish marginal equivalence with the Euclidean VP-SDE, and introduce separate radial and tangential noise-prediction objectives.
Experiments on an embedded spiral, CIFAR-10, and ImageNet , indicate that the polar parameterization, together with moderate radial-loss down-weighting, can improve generation quality.
Critical strengths
-
This paper developed a principled and mathematically coherent reformulation. The polar forward and reverse SDEs are derived explicitly, and the mapped process preserves the marginals of the standard VP-SDE. The score and noise-prediction decomposition also provides a clear connection to the conventional DDPM objectives.
-
The empirical results provide promising initial evidence across both synthetic and image generation tasks. This paper provides the experimental results under 2D spiral embedded, CIFAR-10 and ImageNet. In 2D spiral experiment within Figure 1, Polar diffusion increasingly preserves the underlying structure better than DDPM as the ambient dimension grows, supporting the motivation for separating radial and tangential denoising. On image benchmarks, it slightly improves CIFAR-10 from 3.17 to 3.09, and more substantially improves ImageNet FID from 27.0 to 23.7 under the reported settings.
Critical weaknesses
-
The scaling of the radial and tangential objectives is insufficiently specified. Because the radial residual is scalar while the tangential residual is high-dimensional, their effective balance depends on how the two losses are reduced and normalized. Without this information, it is difficult to determine whether the improvement at reflects the proposed geometric distinction or simply better loss and gradient balancing.
-
The source and robustness of the improvement are not sufficiently isolated. Polar diffusion changes the representation, conditioning, architecture, loss weighting, normalization, and sampling procedure simultaneously. The current ablations do not include an otherwise standard DDPM with the same radial-tangential loss rewighting. Moreover, the CIFAR-10 gain is small and reported without repeated runs, while the ImageNet result is highly sensitive to radial standardization. It therefore remains unclear which component is responsible for the observed improvement.
Questions for authors
-
How stable are the reported FID improvements across random seeds, and how was selected? Reporting repeated runs with uncertainty estimates would be particularly important for the small CIFAR-10 improvement.
-
How are the radial and tangential losses reduced and normalized in implementation? In particular, does exactly reproduce the standard DDPM objective despite their different dimensionalities?
-
Can the authors provide controlled ablations that isolate the contributions of the main components of Polar Diffusion, for example, the polar parameterization, radial standardization, loss reweighting and etc?
3 comments
Thank you for your detailed rebuttal. The authors clearly explained the losses reduction and normalized path, that addressed my pervious question. And, for additional sampling variance and selection, the authors did the additional experiments to prove the stability of the results, which convincing me. But, I still have concerns about the novelty and the source of the empirical improvements. So, I am inclined to keep my initial score, and remain open to further discussion.
Source of the improvement and novelty
Thank you for the follow-up and for staying open to discussion. We address the two remaining concerns below.
Source of the empirical improvement. The controlled runs on CIFAR-10 (50k samples, FID) change one factor at a time:
| Model | Network | Loss () | Sampler | FID |
|---|---|---|---|---|
| DDPM | DDPM UNet | standard | DDPM | 3.14 |
| DDPM | DDPM UNet | standard | Polar | 3.19 |
| DDPM, decomposed loss | DDPM UNet | 1 | DDPM | 3.18 |
| Polar Diffusion | Polar-UNet | 1 | Polar | |
| Polar Diffusion | Polar-UNet | 0.05 | Polar |
(: std over 3 sampling seeds.)
- The polar sampler and the loss decomposition alone do not change FID (rows 1–3).
- The polar parameterization alone does not help: at it is slightly worse than DDPM (row 4).
- The gain comes from down-weighting the radial objective. At a fixed network, improves FID from 3.34 to 3.12, about 10 times the seed std.
- The effect is not specific to -prediction. With x0- and v-prediction, Polar () improves FID over the Euclidean counterpart (x0: ; v: ).
We therefore state the empirical claim precisely in the revision: the improvement comes from the radial/tangential reweighting that the polar formulation exposes, not from the change of coordinates by itself.
Novelty. We see the contribution in three parts:
- An exact polar reformulation of VP diffusion. We derive the forward and reverse polar SDEs (Def. 1, Cor. 1) and the closed-form radial and angular conditional scores (Prop. 2), and we prove that the Euclidean marginals are recovered exactly (Prop. 1). Prior work either replaces the probability path with geodesic (SLERP) paths on a hypersphere (Kumar & Patel, 2026) or analyzes the singularity of the Euclidean score (Liu et al., 2026). Neither gives a marginal-preserving polar SDE for standard diffusion.
- Radial/tangential balance as an explicit, principled control. In the Euclidean objective, the radial and tangential terms are implicitly weighted 1:1 (Cor. 3). Our formulation makes this weight a free parameter that can be applied on top of any prediction target (, x0, v).
- An empirical finding about this control. The 1:1 weighting is not optimal, and down-weighting the radial term improves sample quality across prediction targets, on CIFAR-10 and ImageNet-64.
We hope this clarifies which component drives the gain. We would be glad to discuss further.
Response to Reviewer JDrq: loss normalization, sampling variance, and component ablations
We thank the reviewer for the positive assessment and the precise questions. We have updated the manuscript based on the feedback.
1. How are the losses reduced and normalized? Does wρ=1 reproduce DDPM exactly? Both terms use mean reduction over the n data dimensions:
L=nwx∥ϵ^x−ϵx∥2+nwρ(ϵ^ρ−ϵρ)2.
By Corollary 3, with wx=wρ=1 this equals the standard DDPM loss ∥ϵ^−ϵ∥2/n exactly. We have added this as the first paragraph of Appendix J. We also verified it empirically. We trained the standard DDPM UNet (unchanged architecture, input yt, DDPM sampler) with this decomposed loss at wρ=1, splitting its output into radial and tangential parts. It reaches FID 3.18 on CIFAR-10 (50k samples), matching the DDPM result (3.17 reported by Ho et al.). We have added this experiment to Table 2.
2. Sampling variance and selection of wρ. With wρ=0.05, we generate three independent sampling runs (different random seeds) from the trained model, and the results are stable: FID 3.115±0.019 on CIFAR-10 (50k samples) and 23.39±0.16 on ImageNet-64 (10k samples). We chose wρ by a grid search on CIFAR-10 (Fig. 3) and used the same value on ImageNet-64. We have updated Table 1 to report the mean over these runs. A more principled selection of wρ is left to future work.
3. Ablation studies. We have added the following controlled ablations in Section 5.4 and Appendix K (prediction target: Appendix L):