acceptodds
Under review as a conference paper at ICLR 2027

Compact Representations of Divergence-Free Fields for Self-Supervised Image Denoising

Abstract

We show that divergence-free vector fields admit parsimonious representations, thus extending the classical curl representation beyond three dimensions. Building on this foundation, we propose a novel computationally efficient network architecture that enforces zero divergence by design. Experiments on image denoising—a setting where divergence-free fields are amenable to learning without ground-truth data—demonstrate that our approach sets a new benchmark for constraint-based methods under Gaussian noise, as well as Poisson and real-world noise with minor architectural modifications, bridging the performance gap with leading denoisers.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
divergence, incompressible vector fields, image denoising, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uTRAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce theory for a compact representation of divergence-free vector fields. Prior divergence-free networks require computing increasingly large Jacobian matrices, scaling with dimensionality. This paper therefore proposes a compact representation that avoids computing these large Jacobians and designs a network architecture around this that guarantees zero-divergence. Experiments demonstrate the approach for self-supervised image denoising under Gaussian, Poisson, and real-world noise.

Critical strengths
  • Proofs in Theorems 1 and 2 give a compact, sparse representation where every fif_i is parameterized by two partial derivatives. Also, the proposed expression in Eq. 8 (along with Assumption 1) avoids explicitly computing the Jacobian using a per-pixel MLP.
  • Experimental results in Tables 1 and 2 suggest that the proposed method trains faster and performs better (according to tested datasets and reported PSNRs) than CENSURE, another divergence-free, self-supervised method.
Critical weaknesses
  • Improvement over baseline methods in Table 1 appear to be quite minor (i.e., 0.1-0.2dB between CENSURE and Ours). The larger gains seem to come from reinjection (i.e., Ours$^\dagger$), which seems to be method-agnositic (you can enforce Eq. 19 for other baselines).
Questions for authors
  • Minor request to add confidence intervals around the metrics reported in Table 1 (to give a sense of how reproducible the benchmarked results are)
3 comments

Dear Reviewer uTRA,

Thank you for dedicating your valuable time to the evaluation of our paper and for your positive comments. We have revised the manuscript to address all the comments received (changes are in blue).

  1. Modest improvements without signal-reinjection: Without signal-reinjection, the denoising improvements of our approach are indeed relatively minor compared to CENSURE, in the case of Gaussian noise. However, it achieves this performance gain at a considerably shorter training time, as reported in Table 2. Moreover, unlike CENSURE, our approach generalizes to both Poisson and real-world noise, as demonstrated in Table 1.

  2. Signal-reinjection for other baselines: Regarding the method-agnostic nature of our signal-reinjection technique, we fully agree with your observation. To further investigate this point, we have added Section B.3 and Table 5 to the Appendix, in which we evaluate the potential benefits of applying signal-reinjection to other baseline methods. Our experimental results suggest that signal reinjection consistently improves the performance of constraint-based methods by compensating for their reduced expressivity (except for CENSURE in the case of Poisson and real-world noise), while providing no clear benefit to other types of baselines (see Table 5).

  3. Confidence intervals: Could you please clarify what you mean by confidence intervals? To the best of our knowledge, confidence intervals are not standard in the image denoising literature (see, for instance, [1, 2, 3]). We assume that you may be referring to the variability in the PSNR values reported in Table 1 due to the random seed used to generate synthetic noise. In this regard, we would like to emphasize that our results are highly robust to the choice of random seed. For instance, across 100 randomly selected seeds, the average PSNR values obtained by all methods on BSD68 varied by no more than 0.03 dB. Finally, for the sake of a fair comparison, all methods were evaluated using the same random seed.

  4. Reproducibility of the benchmarked results: With regard to reproducibility, we intend to make the training and evaluation code, along with the corresponding trained model weights, publicly available upon acceptance of the paper.

[1] Zhang, Kai, et al. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE TIP 2017.

[2] Huang, Tao, et al. Neighbor2neighbor: Self-supervised denoising from single noisy images. CVPR 2021.

[3] Zamir, Syed Waqas, et al. Restormer: Efficient transformer for high-resolution image restoration. CVPR 2022.

I thank the authors' for their response, and, after reading through their comments, I am inclined to maintain my current rating.

Thank you for your positive feedback!

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

This paper proposes two theorems for divergence field and compact representations, and designs a neural network structure that natively guarantees zero divergence. The authors evaluate the performance of this method on image denoising tasks for both artificially added noise(Gaussian noise and Poisson noise) and natural noise, and they also compare with a set of supervised and self-supervised baseline.

Critical strengths

The design of the neural network is well-motivated and supported by well-defined theorems. The evaluation is comprehensive on multiple levels and types of synthetic and real world noises. The model shows meaningful improvement compared to other strong baselines(such as neighbor2neighbor) on synthetic noise.

Critical weaknesses

While the proposed model performs well on synthetic noise, it underperforms the baseline methods in some cases on real-world noisy images.

Questions for authors

Following the weakness section, it would be helpful if the authors can provide some explanation/additional experiments regarding the model performance on denoising real-world images.

3 comments

Dear Reviewer XXZp,

Thank you for dedicating your valuable time to the evaluation of our paper and for your positive comments. We have revised the manuscript to address all the comments received (changes are in blue).

Model performance on real-world noise: Although our approach achieves competitive results overall, it does underperform some baseline methods on certain real-world noisy datasets. To understand this behaviour, it is important to recall that our method was primarily designed for Poisson noise, as it relies on the Poisson Unbiased Risk Estimate (PURE). While real-world noise is commonly approximated using a Poisson model [1], the image acquisition process can introduce additional sources of variability, such as those arising from under- or over-exposure [2]. These effects may cause the noise distribution to deviate from the Poisson assumption, as discussed in Section 6.2. Since our approach explicitly relies on this assumption, its performance may deteriorate when such deviations become significant. We believe this could explain the results observed on the FMDD dataset, where the blind-spot network achieves better performance. More generally, this highlights a fundamental challenge in self-supervised denoising of real-world images: the statistical properties of real-world noise are often complex and difficult to characterize, making it challenging to anticipate which method will perform best.

Interestingly, this observation is further supported by an additional experiment conducted in response to Reviewer xmnW, in which we evaluated an alternative blind-spot architecture, PUCA [3], alongside Noise2Self (see Table 6). The results reveal a similarly nuanced picture. On synthetic noise, PUCA consistently underperforms Noise2Self, which can be explained by the structure of its receptive field: a larger proportion of input pixels is excluded from the prediction compared with the Noise2Self architecture (see Fig. 3 of our paper and Fig. 5 of the PUCA paper). However, this performance gap does not carry over consistently to real-world noise. In particular, PUCA outperforms Noise2Self on the SIDD dataset. One possible explanation is that the noise in SIDD exhibits some degree of spatial correlation, which may, under certain circumstances, be better accommodated by PUCA's architecture. Although this interpretation remains speculative, these results further illustrate that the relative performance of denoising methods on real-world data can be difficult to predict based solely on their behaviour under synthetic noise.

Overall, we believe that the ability of our method to remain applicable to real-world noisy images, despite occasional performance gaps with respect to competing approaches, demonstrates its robustness beyond the idealized Poisson noise setting for which it was originally designed.

[1] Hasinoff, Samuel W. Photon, poisson noise. Computer vision: a reference guide. 2014.

[2] Foi, Alessandro, et al. Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data. IEEE TIP 2008.

[3] Jang et al. PUCA: Patch-unshuffle and channel attention for enhanced self-supervised image denoising. NeurIPS 2023.

I thank the authors for the clarification, and my concerns are addressed. I additionally found this relevant work that could be incorporated into the related work session:

Wenquan, Lu et al, Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum, NeurIPS 2025.

I maintain the current rating.

Thank you for your positive feedback! We will add this additional reference to the related work section of the paper.

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

This paper tackles self-supervised image denoising, under divergence-free formula. The paper proves a theorem that every divergence-free field can be built from a cyclic permutation of coordinates, where each output component depends on only two partial derivatives. Based on this, the paper introduces a neural network architecture that restricts the generating function so that each component depends on two pixels and features from a blind-spot network, allowing all required derivatives to be computed with a single backward pass. The paper further extends the method to Poisson noise and proposes an inference-time signal reinjection step.

Critical strengths
  • Huge efficiency gain within the divergence-free, constraint-based approaches.
  • Theorem 2 seems to be a well-established theoretical contribution, which simplifies and generalizes classical curl representation to arbitrary dimensions, and integrates blind-spot and divergence-free approaches.
Critical weaknesses
  • Insufficient comparison against blind-spot approaches: the paper only compares to noise2self for blind-spot baseline, which is known to be a slow blind-spot method. Architecture-based blind-spot networks (e.g., PUCA [1]) are single-pass models, so is expected to have lower cost. As this paper's main goal is on efficiency, I think this comparison is worth included. If these methods achieve comparable performance with lower computational cost, the advantage of this method could be limited.

[1] Jang et al., PUCA: Patch-unshuffle and channel attention for enhanced self-supervised image denoising, NeurIPS 2023

Questions for authors
  • I wonder how the baselines such as Noise2Self and CENSURE could be improved with the signal reinjection method.
3 comments

Dear Reviewer xmnW,

Thank you for dedicating your valuable time to the evaluation of our paper and for your positive comments. We have revised the manuscript to address all the comments received (changes are in blue).

  1. Comparison with PUCA: Following your recommendation, we have included a comparison with an architecture-based blind-spot network. Specifically, we applied the methodology proposed in PUCA to our common backbone architecture: the first convolution is centrally masked, subsequent convolutions are 2-dilated, and patch-unshuffle and patch-shuffle operations are used for downsampling and upsampling, respectively. We have added Section B.4 and Table 6 to the Appendix, where we compare PUCA against both our approach and Noise2Self. On synthetic noise, PUCA yields the weakest performance among the compared methods (see Table 6). We believe that this is due to the structure of PUCA's receptive field, which results in a higher proportion of input pixels being excluded from the prediction (see Fig. 3 of our paper and Fig. 5 of the PUCA paper). The results on real-world noise are less conclusive: Noise2Self performs better on FMDD, whereas PUCA achieves better results on SIDD. One possible explanation is that real-world noise may exhibit some degree of spatial correlation, which could be beneficial to PUCA in some cases. Regarding computational efficiency, although PUCA indeed offers faster inference than Noise2Self, it requires longer training time (see the revised Table 2). Finally, we have also evaluated our approach using PUCA as the π\pi-blind-spot network instead of Noise2Self, demonstrating that our method can accommodate this alternative architecture (see Table 6).

  2. Signal-reinjection for other baselines: We have added Section B.3 and Table 5 to the Appendix, in which we evaluate the potential benefits of applying signal-reinjection to other baseline methods. Our experimental results suggest that signal reinjection consistently improves the performance of constraint-based methods by compensating for their reduced expressivity (except for CENSURE in the case of Poisson and real-world noise), while providing no clear benefit to other types of baselines (see Table 5).

I thank the authors for the rebuttal. I have read their comments, and my main concerns have been addressed. I maintain my original rating.

Thank you for your positive feedback!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.