acceptodds
Under review as a conference paper at ICLR 2027

EMAC-Diffusion: Learning Clean Distributions from Unknown Corruptions via Expectation Maximization

Abstract

We introduce EMAC, an expectation-maximization (EM) algorithm for learning a clean data distribution from a small clean dataset and abundant observations corrupted by an unknown channel. Our method introduces an auxiliary-corruption channel that enables us to train a conditional diffusion model using EM without access to the original corruption channel. We apply this channel to both clean and already-corrupted images in the dataset and impose specific constraints on its design. We theoretically and experimentally identify limitations of prior work in handling low-frequency corruptions and we establish convergence guarantees for our method under appropriate assumptions. Experimentally, we first validate our method in controlled experiments, show how it can be used in the context of a privacy-related application, and finally test it at scale by improving imperfect data that naturally occur in ImageNet at resolution. Qualitative and quantitative analyses show that our method improves over previously proposed baselines on several image-quality metrics.

Primary area:
generative models
Keywords:
Diffusion Models, Generative Models, Inverse Problems, Computer Vision, Expectation Maximization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Ygt8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses a small number of clean images and a large number of images with unknown corruptions to learn the clean image distribution. Specifically, this paper aims to use an auxiliary corruption channel (Q_hat) to make the distribution of clean and degraded images nearly the same after Q_hat processing. Then, through the conditional diffusion model, combined with EM iteration to generate pseudo-clean data, and finally train to obtain the unconditional final generation model. This paper presents theoretical analysis of the method, and verifies the feasibility of the method on CIFAR-10, CelebA and ImageNet.

Critical strengths
  1. The idea is interesting and somewhat novel. Although the idea of reducing distributional differences by adding noise seems to have been explored in Ambient Omni, this paper extends it to more general auxiliary corruption channels, and combines distributional alignment, information preservation and conditional EM into a relatively complete framework.
  2. The paper provides some theoretical analysis, like establishing convergence guarantees under the ideal assumptions. It also analyzes the limitations of ambient omni, which makes the motivation clearer. The experiments also demonstrate improvements over the baselines on CIFAR-10 and CelebA.
Critical weaknesses
  1. This paper assumes that clean image and degraded image should have the same distribution after Q_hat. However, I haven’t found direct verification on how well the distributions are aligned under the selected Q_hat, such as random blur/masking. The influence of the distribution misalignment is also not clear. Thus, although there are some improvements in experiment, it is not clear whether these improvements come from the alignment strategy. There is still a gap between theoretical analysis and practical method.
  2. The selection of Q_hat is mainly empirical and lacks systematic analysis. The paper proposes that the selection of Q_hat should balance the distribution alignment and information preservation. However the paper mainly tests some relatively simple corruption types, and selects random blur/masking based on the FID. There is little direct evidence that these choices achieve a better trade-off between the above distributional alignment and information preservation. It is also unclear whether the selected auxiliary corruptions work well for other types of unknown corruptions, which might limit the generalizability of the paper.
Questions for authors
  1. Could you report some quantitative results to evaluate the distribution alignment under different auxiliary corruption channels, for example, by training a classifier? How does the alignment gap affect the final generation performance?
  2. Why does random gaussian blur outperform other Q_hat? Is it because of better alignment, information preservation or randomization? Could you also consider some other Q_hat, such as downsampling or combined corruptions?
  3. In ImageNet experiments, the final model is based on EDM2 pre-training and fine-tuned on it. Could you add a training-from-scratch comparison, or fine-tune EDM2 using the high quality subset under the same protocol, to verify the independent contribution of EMAC? Besides, could you report more generative metrics like FID or other diversity metrics?
Reviewer_PBiLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies learning a clean image distribution from few clean images and many observations corrupted by an unknown corruption channel. The proposed EMAC introduces a known auxiliary corruption to both clean and corrupted images, alternating conditional diffusion reconstruction with training on clean/reconstructed pairs; the auxiliary channel aims to align observation distributions while preserving information, after which an unconditional generator is trained. The authors lower-bound the Gaussian noise needed for mean-shifted alignment and prove ideal population-EM convergence. Experiments on blurred CIFAR-10, masked CelebA, and mixed-quality ImageNet at 512 resolution are promising.

Critical strengths
  1. The studied problem is relevant, and the idea is useful. In particular, the setting is practically motivated. Introducing a controllable auxiliary channel offers a plausible way to use conditional EM without evaluating or simulating the original corruption operator.

  2. The theory clarifies the method's assumptions. The mean-discrepancy lower bound explains a limitation of Gaussian smoothing for some corruptions. The population analysis also separates convergence of the likelihood objective from recovery of the clean distribution, and correctly identifies the role of the positive clean-data anchor.

Critical weaknesses
  1. About the auxiliary channel: Equation (6) proposes insightful design principles for the auxiliary channel, but it seems that the actual channel is still obtained through heuristic design and parameter search. Although the paper compares multiple channels and their strengths, the existing experiments mainly report the final FID and have not yet directly quantified the distribution discrepancy and degree of information preservation of the auxiliary observations. Therefore, the experiments are insufficient to determine whether the performance improvement indeed comes from a better “distribution alignment–information preservation” trade-off, and how the channel should be selected when facing new, unknown degradations. I recommend supplementing the existing Gaussian noise experiments with different strengths by reporting and analyzing the relationships among distribution consistency, information preservation, and final generation quality.

  2. Error between the ideal and actual conditions: The appendix acknowledges that the convergence analysis relies on the population distribution, strict distribution alignment, and exact conditional fitting and sampling. However, it remains unclear how alignment errors, finite samples, and model approximation errors in practice affect the final restoration results. I suggest adding discussion or analysis.

  3. Evaluation metrics for the ImageNet experiments: The paper reports Coverage, Density, Precision, Recall, and Vendi score metrics on the CIFAR-10 experiments. However, the ImageNet experiments mainly report no-reference image quality metrics and lack evaluation of the coverage and diversity of the generated distribution. Since the method uses quality scores to filter a clean subset and changes the training data through reconstruction, it is unclear whether the improvement in quality scores is accompanied by a decrease in content diversity or distribution coverage. I recommend adding metrics such as FID, recall, coverage, and Vendi score.

  4. Computational overhead and computational complexity: The paper appears not to fully report the computational overhead of the complete training pipeline, nor the model’s computational cost, parameter count, and inference time.

  5. Reporting of experimental and training settings: The paper appears not to fully specify some important experimental settings, such as the initialization and training methods of each method. In addition, the Ambient Omni curves for CelebA are shown up to 8,000 kimg, whereas the clean-only and EMAC conditional models are shown up to more than 16,000 kimg. This difference in training amount should be explained to clarify whether the performance improvement can be attributed to the proposed method.

Questions for authors

Please refer to the Critical Weaknesses.

Reviewer_q5f5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EMAC, an EM-based method for learning a clean distribution from limited clean data and many observations corrupted by an unknown channel. It applies a known auxiliary channel to both datasets and trains a conditional diffusion model through EM, with experiments on CIFAR-10, CelebA, and ImageNet-512.

Critical strengths
  1. The auxiliary-channel idea is simple and well motivated. The alignment and information-preservation requirements give a clear basis for its design.

  2. The analysis gives a concrete explanation of Ambient Omni’s limitations. The noise lower bound, validation-loss curves, and controlled CIFAR-10 results support the analysis.

Critical weaknesses
  1. The auxiliary channel has to be designed by someone who already knows the type of corruption: blur for blurred CIFAR-10, face masks for CelebA (lines 432-434). This weakens the "unknown channel" claim. It would help to explain how the channel could be chosen without this knowledge, and to discuss why the corrupted data helps, since the bound in Theorem 4 uses only the clean data.

  2. On CIFAR-10 the true channel is known, so EM methods that use it (DiffEM, Bai et al. 2024) would show the best result EMAC could reach, but they are not compared. Ambient Omni is run with one fixed noise threshold instead of its original per-image selection, and on CelebA its FID is far above clean-only (15.54 to 81.55 vs. 4.08, Table 6). The baseline may not be at its best.

  3. The computational cost is not reported. Each EM iteration reconstructs every corrupted sample (Algorithm 1, line 3) and then retrains, which is likely much more expensive than Ambient Omni's single run. For ImageNet-512, the number of EM iterations, sampling steps and GPU-hours would be helpful.

  4. The manuscript feels unfinished. A short conclusion and limitations section is missing. The ImageNet resolution in Fig. 21, the metric names in Table 4, and the CelebA dataset size and resolution could also be clarified.

Questions for authors
  1. If the type of corruption were not known in advance, how would the auxiliary channel be chosen? For the channels used in the paper, can you show that clean and corrupted images become indistinguishable after the auxiliary channel?

  2. The bound in Theorem 4 uses only the clean data. What does the corrupted data contribute, in theory or in the finite-sample setting?

  3. On CIFAR-10, can you add DiffEM and Bai et al. (2024), which use the true channel, as upper-bound comparisons?

  4. Can Ambient Omni be run on CelebA with its original per-image threshold selection? Its FID with a fixed threshold is far above clean-only (Table 6), which makes the comparison hard to read.

  5. For ImageNet-512, how many EM iterations and sampling steps were used, and what was the total GPU-hour cost compared with a single Ambient Omni run?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.