acceptodds
Under review as a conference paper at ICLR 2027

Polar Diffusion: Decoupling Radial and Tangential Denoising in Diffusion Models

Abstract

Diffusion models typically perform denoising in Euclidean space using a single isotropic prediction objective, implicitly assigning equal importance to errors along all geometric directions. High-dimensional data, however, often concentrate near lower-dimensional structures, where radial and tangential variations can play fundamentally different roles in generation. We introduce Polar Diffusion, a geometric reformulation that factorizes the diffusion process over a product space of magnitude (radial) and normalized direction (tangential). We derive the corresponding product-space score-matching formulation and show that, when mapped back to Euclidean coordinates, the resulting dynamics recover the same marginal distributions as standard variance-preserving diffusion. This factorization exposes a degree of freedom hidden by the Euclidean formulation: radial and tangential denoising objectives can be controlled and weighted independently. Polar Diffusion thus provides a simple geometric extension of conventional diffusion models that enables explicit control over the relative contributions of magnitude and structural errors during denoising. Ablation studies show that moderately down-weighting the radial objective consistently improves generation quality, suggesting that radial and tangential errors need not contribute equally to effective denoising, with greater emphasis on the tangential component providing closer accounting for the distribution of structural information in the data.

Primary area:
generative models
Keywords:
diffusion models, score-based generative modeling, geometric deep learning, riemannian manifold

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JDrqOpenReview ↗
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 64×6464\times64, indicate that the polar parameterization, together with moderate radial-loss down-weighting, can improve generation quality.

Critical strengths
  1. 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.

  2. The empirical results provide promising initial evidence across both synthetic and image generation tasks. This paper provides the experimental results under 2D spiral embedded, 32×3232\times32 CIFAR-10 and 64×6464\times64 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
  1. 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 (wρ=0.05)(w_\rho=0.05) reflects the proposed geometric distinction or simply better loss and gradient balancing.

  2. 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
  1. How stable are the reported FID improvements across random seeds, and how was (wρ=0.05)(w_\rho=0.05) selected? Reporting repeated runs with uncertainty estimates would be particularly important for the small CIFAR-10 improvement.

  2. How are the radial and tangential losses reduced and normalized in implementation? In particular, does (wρ=wx=1)(w_\rho=w_x=1) exactly reproduce the standard DDPM objective despite their different dimensionalities?

  3. 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
AuthorsOpenReview ↗

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=1w_\rho=1 reproduce DDPM exactly? Both terms use mean reduction over the nn data dimensions:

L=wxn∥ϵ^x−ϵx∥2+wρn(ϵ^ρ−ϵρ)2.\mathcal{L} = \frac{w_x}{n}\|\hat\epsilon_x-\epsilon_x\|^2 + \frac{w_\rho}{n}(\hat\epsilon_\rho-\epsilon_\rho)^2 .

By Corollary 3, with wx=wρ=1w_x=w_\rho=1 this equals the standard DDPM loss ∥ϵ^−ϵ∥2/n\|\hat\epsilon-\epsilon\|^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 yty_t, DDPM sampler) with this decomposed loss at wρ=1w_\rho=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 wρw_\rho. With wρ=0.05w_\rho=0.05, we generate three independent sampling runs (different random seeds) from the trained model, and the results are stable: FID 3.115±0.0193.115 \pm 0.019 on CIFAR-10 (50k samples) and 23.39±0.1623.39 \pm 0.16 on ImageNet-64 (10k samples). We chose wρw_\rho 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 wρw_\rho 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):

  • Sampler: sampling a standard DDPM checkpoint with our polar sampler gives FID 3.188, against 3.138 with the DDPM sampler. The sampler is neutral, so the gain comes from training.
  • Loss decomposition alone: training the standard DDPM UNet with the decomposed loss at wρ=1w_\rho=1 gives FID 3.18, matching DDPM (3.17), so splitting the objective by itself does not change results (see point 1).
  • Loss reweighting within the polar parameterization: at a fixed architecture, wρ=1→0.05w_\rho=1 \to 0.05 gives 3.34→3.123.34 \to 3.12 (3 sampling runs). This is the main source of the gain.
  • Inputs and radial standardization: these are ablated in Table 1. We agree that ImageNet-64 is sensitive to radial standardization. Standardizing the radius with dataset statistics is a common input-normalization practice (analogous to normalizing pixel values or latents to zero mean and unit variance); it puts the scalar radius on a scale comparable to the other inputs. We have added this rationale to Sec. 4.1.
  • Prediction target: with x0- and v-prediction, Polar (wρ=0.05w_\rho=0.05) improves over the Euclidean counterpart (x0: 4.69→4.224.69 \to 4.22; v: 4.43→4.194.43 \to 4.19), so the effect is not specific to ϵ\epsilon-prediction.

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.

AuthorsOpenReview ↗

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:

ModelNetworkLoss (wρw_\rho)SamplerFID
DDPMDDPM UNetstandardDDPM3.14
DDPMDDPM UNetstandardPolar3.19
DDPM, decomposed lossDDPM UNet1DDPM3.18
Polar DiffusionPolar-UNet1Polar3.34±0.043.34 \pm 0.04
Polar DiffusionPolar-UNet0.05Polar3.12±0.023.12 \pm 0.02

(±\pm: 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 wρ=1w_\rho=1 it is slightly worse than DDPM (row 4).
  • The gain comes from down-weighting the radial objective. At a fixed network, wρ=1→0.05w_\rho=1 \to 0.05 improves FID from 3.34 to 3.12, about 10 times the seed std.
  • The effect is not specific to ϵ\epsilon-prediction. With x0- and v-prediction, Polar (wρ=0.05w_\rho=0.05) improves FID over the Euclidean counterpart (x0: 4.69→4.224.69 \to 4.22; v: 4.43→4.194.43 \to 4.19).

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:

  1. 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.
  2. 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 (ϵ\epsilon, x0, v).
  3. 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.

Reviewer_zABPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Polar Diffusion, which reformulates the standard diffusion process into polar coordinates: a product space of magnitude (radial) and normalized direction (tangential). They found that the resulting polar diffusion recover the same marginal distributions as standard VP diffusion.

Experiments on CIFAR-10, and ImageNet 64x64 show that down-weighting the radial loss (wρ=0.05w_\rho = 0.05) improves FID over DDPM and iDDPM.

Critical strengths
  1. Clean math and intuitive perspective. The paper provides a clear derivation showing how VP diffusion maps to polar diffusion.

  2. Simple framework, easy to follow.

Critical weaknesses
  1. Marginal empirical gains and unconvincing ablation results.
  • On CIFAR-10, the improvement over the DDPM is very small
  • When using equal weight (w_rho = 1.0), which matches the theoretical decomposition of standard VP DDPM, polar diffusion performs noticeably worse than standard VP DDPM.
  • "Ablation studies show that moderately down-weighting the radial objective consistently improves generation quality" however, they actually seem to decline in Fig3. it is NOT consistently improves generation quality.
  1. Limited evaluation scope and missing comparisons with existing baselines.
  • only CIFAR-10 and ImageNet 64x64, compared only against early baselines(2021 before)
  • The paper lacks comparisons and discussions against recent diffusion parameterization methods (e.g. EDM, x_0- or v-prediction, Flow Matching), making it unclear whether polar diffusion offers advantages over recent training recipes or generalizes to LDM or DiT.
Questions for authors
  1. Performance gap at equal weights: Corollary 3 shows that standard DDPM implicitly uses equal weights(1:1). Why does Polar Diffusion with w_rho =1 perform worse than DDPM on CIFAR-10?
  2. Comparison between Polar Diffusion Sampling and DDPM sampling
  3. Provide more comparisons and discussions on recent diffusion parameterization methods
3 comments
AuthorsOpenReview ↗

Response to Reviewer zABP: equal-weight gap, ablation wording, sampling, and modern parameterizations

We thank the reviewer for the constructive feedback. We have updated the manuscript based on the feedback.

1. Marginal CIFAR-10 gain. We agree that the gain over DDPM is modest on CIFAR-10. We suspect, though we have not verified this, that CIFAR-10's low resolution (32×3232\times32) limits the benefit: there is little fine structure in the images, so improvements to tangential (structural) denoising are hard to see, both in FID and visually. This is consistent with the larger gain on the higher-resolution ImageNet-64 (Table 1: 27.0 →\to 23.39).

2. Why is wρ=1w_\rho=1 worse than DDPM? The objectives match at wρ=1w_\rho=1, but the parameterization does not. The Polar-UNet predicts the radial noise with a separate scalar head (a linear layer on the flattened bottleneck), and its input is xtx_t together with the radius as a condition. We suspect that at equal weight the radial term, which dominates the error at low noise (Liu et al., 2026), takes a large share of the gradient for this head. We have added this discussion as the last paragraph of Appendix K.

3. "Consistently improves" (Fig. 3). Point well taken. Compared with DDPM (FID 3.17), down-weighting the radial objective improves performance for some small values of wρw_\rho (wρ=0.1w_\rho=0.1: 3.11; wρ=0.05w_\rho=0.05: 3.09), but not for every value. We revised the statement to: "appropriately down-weighting the radial objective can improve generation quality over DDPM."

4. Polar sampling vs. DDPM sampling. We sampled a standard DDPM checkpoint with both samplers, using 1000 steps in both cases:

SamplerFIDIS
DDPM3.1389.39
Polar (Alg. 2)3.1889.43

The two samplers are equivalent within noise. The polar sampler therefore does not explain the gain; the gain comes from polar training with reweighting. We have added this comparison to Table 2.

5. Recent parameterizations and baselines. We have trained x0- and v-prediction on CIFAR-10, each in Euclidean form and in Polar form (wρ=0.05w_\rho=0.05):

PredictionEuclidean FIDPolar FID
x04.694.22
v4.434.19

Polar Diffusion improves on both modern prediction targets, which suggests that radial reweighting is complementary to the choice of parameterization. We have added x0- and v-prediction baselines to the toy experiment (Fig. 1) and the CIFAR-10 results to Appendix L; in both, Polar Diffusion outperforms both baselines. Extending polar reweighting to EDM preconditioning, flow matching, and LDM/DiT backbones is left to future work.

Thank you for the response and the new results. The new results on Polar sampling vs. DDPM sampling, Marginal CIFAR-10 gain addresses my concern.

However, my concern about the comparison with baselines stays mostly unresolved. If gain is Marginal on CIFAR-10, I suggest make experiments on ImageNet-64 to better show the effect of Polar Diffusion. If "there is little fine structure in the CIFAR-10 images", the paper should narrow its claims to what the current experiments support. For this reason, I am keeping my score.

note: What is the "(Liu et al., 2026)" here? is it "Improving the euclidean diffusion generation of manifold data by mitigating score function singularity"?

AuthorsOpenReview ↗

ImageNet-64 baselines, narrowed claims, and the Liu et al. reference

Thank you for the follow-up. We are glad that the sampler comparison and the discussion of the CIFAR-10 gain address your concerns.

Baseline comparisons on ImageNet-64. We agree that CIFAR-10 is too small a testbed to show the effect clearly. Repeating the baseline comparisons on ImageNet-64 (a retrained Euclidean baseline in our own evaluation pipeline, Polar at wρ=1w_\rho=1 and 0.050.05, and Euclidean vs. Polar with v-prediction) is beyond our compute budget during the discussion period. We have therefore narrowed the claims as follows, and list these comparisons as future work.

Narrowing the claims. We agree, and we have revised the paper accordingly:

  • We limit the empirical claims to pixel-space VP diffusion with UNet backbones at 32×3232\times32 and 64×6464\times64.
  • We describe the CIFAR-10 gain as modest. The resolution explanation is now stated as an unverified hypothesis, not as a finding.
  • We state the empirical finding precisely: the gain comes from down-weighting the radial objective, while the polar sampler and the loss decomposition alone are neutral (Table 2). The improvement also holds for x0- and v-prediction (Appendix L).
  • We state explicitly that we make no claim about EDM, flow matching, latent diffusion, or DiT/SiT backbones, and list these as future work in Sec. 6.

Liu et al. (2026). Yes. It is Liu, Zhang & Li, "Improving the Euclidean diffusion generation of manifold data by mitigating score function singularity".

Reviewer_wnPZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Standard diffusion models operate in flat Euclidean space and are trained with mean squared error loss which is an isotropic loss as it treats errors in all directions equally. However, real-world images tend to cluster along a lower-dimensional manifold. Here, changes in magnitude and direction play different roles, and therefore this work proposes to decompose every non-zero Euclidean vector into its magnitude and directional components. The paper also derives the resulting forward and reverse SDEs for this polar coordinates, and shows that it recovers the marginal distribution of s variance preserving diffusion process. In addition, the work also derives the closed-form conditional scores. In addition, the loss is decomposed into tangential and radial noise-prediction errors with separate weights. The experiments are done on UNet backbone which is modified to predict these two noise terms. The experiments are shown in CIFAR-10 and ImageNet-64.

Critical strengths
  1. The premise of this work to decomponse diffusion loss into magnitude and direction makes sense. There are some prior works that show advantage of this sort of decomposition post-hoc during inference in the context of Classifier-free guidance.Moving this decomposition to training time therefore seems to make natural sense.
  2. The SDE-based derivations in Section 3 seem correct.
Critical weaknesses
  1. Overlap of core contributions with prior work: In the abstract the paper states “This factorization exposes a degree of freedom hidden by the Euclidean formulation: radial and tangential denoising objectives can be controlled and weighted independently.” However, this also appears in prior work by Kumar and Patel (2026)[1], which the authors have cited in Line 64 but acknowledged in a limited manner by confining the scope to the role of DiNOv2 embeddings in flow matching. However, this paper already wrote down the exact same decomposition of Euclidean objective (See Eq. 4 in the paper) and also identified that optimizing radial loss wastes model capacity and prevents convergence of the angular component (as shown in Figure 4 of the paper).

    1. In addition to above, Line 196 in the paper (exponential map) is same as exponential map integrator (Line 12) in Kumar and Patel (2026). This needs to be properly cited.
    2. This prior work (Kumar and Patel, 2026) also showed that Euclidean linear trajectory which is used in this paper fails on curved manifolds. This happens because high-dimensional data and noise are orthogonal, therefore linear combinations in Euclidean space cut a chord through the low-density interior rather than following the manifold geodesic. They instead proposed SLERP as alternative to mitigate this issue.
 [1] Kumar, Amandeep, and Vishal M. Patel. "Learning on the manifold: Unlocking standard diffusion transformers with representation encoders." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
  2. Limited scale of experiments:

    1. Outdated backbone architecture and parameterization: The paper uses UNet backbone while the community widely uses DiT/SiT backbone. In addition, recently, the use of Flow Matching has become more widespread compared to DDPM.
      1. The paper focuses on ϵ\epsilon-prediction loss but the community has largely moved on to x0x_0-prediction and vv-prediction. These should be included in baselines.
      2. EDM [1] and EDM2 [2] are also competitive baselines here. EDM has a preconditioner that should achieve time-dependent radial normalization.
 EDM2 has magnitude-preserving layers which would force activations to lie on a hypersphere. These comparisons would help establish whether Polar Diffusion helps when proper input/activation normalization is already present.
 [1] Karras, Tero, et al. "Elucidating the design space of diffusion-based generative models." Advances in neural information processing systems 35 (2022): 26565-26577. 
 [2] Karras, Tero, et al. "Analyzing and improving the training dynamics of diffusion models." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.
    2. Datasets: Experiments have been done on CIFAR-10 and ImageNet-64, which are simple but slightly outdated datasets. Empirical gains on CIFAR-10 are small.
  3. Wrong citations: Both citations for CIFAR-10 and ImageNet are incorrect. The paper cites BigGAN paper for CIFAR-10 and VQVAE paper for ImageNet-64.

Questions for authors
  1. I would appreciate if the difference between this work and prior work by Kumar and Patel, 2026 is clearly stated in the paper.
  2. If you compare Polar Diffusion against DDPM with x0x_0-prediction or vv-prediction, the gap in Figure 1 will likely disappear (inspired from results in “Back to basics : Let Denoising Generative models denoise” by Li and He, 2025). This is an important baseline that should be included.
  3. Classifier-free guidance (CFG) scales cause oversaturation, contrast blowout, and burnt-out artifacts. Some recent approaches such as Adaptive Projected Guidance have tried to fix this post-hoc by projecting out the parallel component of the guidance vector at inference time. It would be interesting to see if Polar Diffusion can be better leveraged for this use case.
  4. The paper notes that radial errors dominate tangential errors at small noise levels. Can’t this be mitigated by smarter weighting function (such as min-SNR weighting)?
2 comments
AuthorsOpenReview ↗

Response to Reviewer wnPZ: relation to Kumar & Patel, modern parameterizations, and new results

We thank the reviewer for the careful review. We have updated the manuscript based on the feedback.

1. Relation to Kumar & Patel (2026). We agree that our discussion of this work was too short. We now cite it for the radial/angular split of the Euclidean objective (their Eq. 4) and for the exponential-map update (their Line 12), and we have added a paragraph (Sec. 3.3) on the differences. The main differences are:

  • We do not replace the probability path. Kumar & Patel build geodesic (SLERP) flow-matching paths on a hypersphere. We keep the standard VP-SDE. Polar Diffusion rewrites it in coordinates (ρ,x)∈R×Sn−1(\rho, x) \in \mathbb{R}\times S^{n-1}, and we prove that the Euclidean marginals are recovered exactly (Prop. 1). The radius is a diffusing state variable, not a quantity that is removed or fixed.
  • We derive the full polar dynamics: the forward and reverse polar SDEs (Def. 1, Cor. 1) and closed-form radial and angular conditional scores (Prop. 2). These do not appear in Kumar & Patel.
  • Setting. Their method relies on the hyperspherical geometry of DINOv2 representation encoders. Ours applies directly to pixel space and does not assume that the data lie on a sphere. The sphere serves as a natural space for the structural (directional) information of images; together with the radius, the product R×Sn−1\mathbb{R}\times S^{n-1} represents every nonzero image exactly.

So the loss decomposition itself is not our main claim. Our claim is a marginal-preserving polar reformulation of VP diffusion in which the two terms are separate, controllable objectives.

2. x-prediction / v-prediction baselines (Li & He). We have trained x0-prediction and v-prediction on CIFAR-10, each in Euclidean form and in Polar form (wρ=0.05w_\rho=0.05):

PredictionEuclidean FIDPolar FID
x04.694.22
v4.434.19

The gap does not disappear: Polar improves both x0- and v-prediction, and it closes most of the x0-vs-v gap. We have updated the toy spiral experiment (Fig. 1), where Polar Diffusion also outperforms the x0- and v-prediction baselines, and added the CIFAR-10 results in Appendix L.

3. Min-SNR / time-dependent weighting. This is a good suggestion. A time-dependent wρ(t)w_\rho(t) and its combination with min-SNR weighting are a natural extension of our framework, because the polar decomposition can be applied on top of any per-timestep weighting. We leave a systematic study to future work.

4. EDM / EDM2, DiT/SiT, flow matching, larger datasets. We agree that these comparisons would strengthen the paper. With our compute budget, we focused the rebuttal on isolating the source of the gain (sampling variance, sampler, prediction targets). Combining polar reweighting with EDM preconditioning and with DiT/SiT backbones on latent ImageNet-256 is left to future work. We have stated this limitation explicitly in the Limitations paragraph (Sec. 6).

5. CFG / Adaptive Projected Guidance. We agree this is a promising direction. Polar Diffusion already separates the parallel (radial) and orthogonal (tangential) parts of the prediction, so the guidance could be applied separately to each part. We will study this in future work. Thank you.

6. Citations. Thank you for pointing these out. We have corrected the citations to Krizhevsky (2009) for CIFAR-10 and Deng et al. (2009) / Chrabaszcz et al. (2017) for ImageNet-64.

I thank the authors for providing the x0x_0 and vv-prediction baselines. I understand that the model operates on pixel space and diffuses the radius rather than fixing it but the core conceptual and mathematical claim of this paper which is factorizing the Euclidean objective into controllable radial and tangential loss terms to expose a hidden degree of freedom, is still largely identical to Eq. 4 in Kumar & Patel (2026). Keeping standard VP-SDE instead of geodesic paths does not change the fact that the underlying loss decomposition and the discovery that radial noise optimization wastes model capacity were already published.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.