acceptodds
Under review as a conference paper at ICLR 2027

Support-guided OOD Removal with Diffusion

Abstract

Many real-world signals are corrupted by localized, structured perturbations whose distribution is unknown: an intermittent background noise in a music recording, a watermark stamped on a photograph, a cast shadow. We study the task of recovering a clean signal x from an observation y containing such a localized out-of-distribution (OOD) corruption. We propose Support-guided OOD Removal with Diffusion, SORD, an iterative algorithm to recover x using only two ingredients: a binary map of the corruption support and a single diffusion model pretrained on clean signals. SORD does not require any paired data or any statistical model of the corruption. We evaluate SORD on natural images and human faces with spatially localized corruptions including handwritten digits, other images, watermarks, and shadows, as well as audio recordings of different musical instruments with temporally localized corruptions including other instruments and human speech. SORD is competitive with, and often outperforms, posterior-sampling inpainting given the same diffusion prior and the same mask, source-separation baselines that require a separate diffusion model for each source, and, on several benchmarks, supervised networks trained on paired clean and corrupted signals.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion model, Noise removal

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_G9wBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes, SORD, a diffusion-based method to remove local OOD perturbation to recover clean data. SORD assumes the setting where the corruption is localized and we have access to this localized corruption support. It uses multiple inpaintings of the support as initial seeds of the estimated clean signals, and refines them iteratively. At each iteration, it first calculates the variance among the residuals between the estimated clean signals and the corruption, and denoises the residuals according to this variance. Then, it subtracts the residuals to update the estimated clean signals, and denoise the estimated signals. It stops when the variance converges. Experiments are performed for both the image and the audio modality, and consider tasks including additive corruption, watermarking and shadows. Results show that SORD generally outperforms or is comparable with prior works.

Critical strengths
  • Generality: SORD does not assume corruption types and is generally applicable across them (as long as they are localized); experiments show applicability to diverse settings including additive, watermark, shadows. It also works across different modalities (images, audios).
  • Performance: On the evaluated settings SORD generally outperforms DPS, the most directly applicable baseline, and often outperforms or is comparable to setting-specific baselines.
Critical weaknesses
  • Efficiency: SORD involves hundreds of network passes per image (Appendix Table 4) and generally requires 32 inpainting examples for best performance (Appendix Table e), which is relatively expensive.
  • Practical utility: performance is somewhat sensitive to accuracy of the corruption support (Appendix Table 5), which could be a concern in practice when perfect masks may not be available. Also the image experiments are performed mostly on distributions with moderate complexity, with traditional pixel-space diffusion models individually trained on the clean distribution; it remains a bit unclear how it performs under more complex data distributions and alternative model variants.
Questions for authors

See Weaknesses above -

  • How would SORD scale when the clean signal distribution becomes more complex (e.g. ImageNet-1000), when it becomes more difficult for the diffusion models to accurately learn and represent the distribution?
  • It would be helpful to have more ablation on the choice of diffusion models, e.g. how would SORD work in combination with general pretrained models not specific to the clean signals and/or with latent diffusion models.
  • For imperfect corruption masks it would also be helpful to have some evaluations beyond simple transformations under more realistic settings, e.g. obtaining the masks from actual segmentation or anomaly detection models.
Reviewer_fE2NOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SORD, a diffusion-based method for removing localized structured corruptions when the corruption support Ω\Omega is known but the corruption distribution itself is unknown. The method assumes only a diffusion prior trained on clean signals. It first obtains an inpainting-based estimate from the trusted observations outside Ω\Omega, and then iteratively reuses the measurements inside the corrupted region, exploiting the fact that yΩ=xΩ+nΩy_\Omega=x_\Omega+n_\Omega may still contain useful information about the underlying clean signal.
The paper evaluates SORD on several image and audio restoration problems, including synthetic localized image corruptions, watermark and shadow removal, and temporally localized audio interference. Its main motivation is to replace explicit modeling of the corruption distribution with structural information about where the corruption occurs.

Critical strengths
  1. Reasonable formulation for corruption-agnostic restoration. The method does not require paired clean-corrupted training data or a corruption-specific model, and instead relies on a diffusion model trained on clean signals as a generic prior. Using such a prior to pull corrupted observations toward the clean-signal distribution is a reasonable design choice in this setting. Moreover, for a fully blind restoration problem with both an unknown corruption type and corruption potentially affecting the entire signal, the problem would be severely underconstrained. Introducing support information is therefore a reasonable way to make corruption-agnostic restoration tractable while still avoiding degradation-specific supervision.

  2. Broad empirical evaluation. The experiments cover both image and audio restoration and multiple types of structured localized corruption, providing reasonably broad empirical evidence for the proposed setting.

Critical weaknesses

1. Strong dependence on oracle support and trusted clean context.

SORD assumes the exact corruption support Ω\Omega is known and treats yΩcy_{\Omega^c} as clean and trusted. Across both image and audio settings, this provides potentially strong contextual information for recovering xΩx_\Omega. The paper does not sufficiently characterize how much of the performance depends on the amount, structure, and informativeness of this clean context, making it difficult to isolate the contribution of the proposed restoration procedure from the support information itself.

2. The empirical and conceptual positioning is incomplete.

Fundamentally, the paper’s OOD framing is not clearly consistent with its own related-work discussion. Lines 077–078 define the localized corruption itself as OOD relative to the clean-signal generative prior, whereas the localized OOD/anomaly-detection literature discussed later concerns identifying observed regions that deviate from normal-data statistics, mainly to obtain the support Ω\Omega. These appear to be different notions of OOD. Once Ω\Omega is given, SORD seems to operate as support-aware restoration with an unknown corruption rather than using any OOD-specific mechanism. This makes both the novelty claim and the choice of comparison methods difficult to assess without a clearer formulation and stronger positioning against degradation-blind restoration. Once the problem is interpreted as recovering a clean signal from an unknown corruption using a learned clean-signal prior, degradation-blind prior-based restoration becomes highly relevant. In particular, for the image experiments, methods such as Deep Image Prior (DIP) [1], DreamClean [2], and Diffusion Image Prior (DIIP) [3] address restoration without requiring a corruption-specific restoration model, with DreamClean and DIIP targeting unknown degradations using diffusion priors. These methods do not receive the oracle corruption support available to SORD, so comparisons on matched image benchmarks would be important for establishing what additional benefit comes from the support information and from SORD itself.

References

[1] Deep Image Prior. CVPR, 2018.

[2] DreamClean: Restoring Clean Image Using Deep Diffusion Prior. ICLR, 2024.

[3] Diffusion Image Prior. ICCV, 2025.

Questions for authors

1. What exactly is OOD in the proposed formulation, and what technical role does it play?

The paper describes the corruption itself as OOD relative to the clean-signal prior, while the related OOD/anomaly-detection literature is mainly used to motivate obtaining the support Ω\Omega. Once Ω\Omega is given, SORD seems to operate as support-aware restoration under an unknown corruption. What is the precise in-distribution reference, and how is this technically different from generic degradation-blind restoration?

2. How much masked-region performance advantage does SORD obtain from its more informative support-aware setting?

SORD is evaluated on the corrupted region Ω\Omega, but it is also given exact support and therefore has access to guaranteed clean observations outside it, yΩc=xΩcy_{\Omega^c}=x_{\Omega^c}. These clean observations can provide strong context for reconstructing xΩx_\Omega, whereas methods operating without oracle corruption support do not receive this information. For the image experiments, comparisons with DIP, DreamClean, and especially Diffusion Image Prior would therefore be important to quantify how much reconstruction advantage SORD actually obtains within the corrupted region under this more informative setting. Since the reported image metrics are computed only over Ω\Omega, such a comparison would directly test whether this information advantage translates into a substantive reconstruction advantage inside the corrupted region.

3. How much of the reported improvement is genuinely SORD-specific rather than standard multi-sample variance reduction?

Averaging multiple stochastic diffusion samples is a standard variance-reduction strategy and is not specific to SORD; stochastic baselines such as DPS could likewise draw KK samples and average them. Since SORD uses KK parallel samples and its performance improves substantially with KK, could the authors provide matched-KK and matched-compute comparisons where the baselines receive the same multi-sample averaging? This would isolate the gain from the proposed refinement and ensemble-spread mechanism from the generic benefit of ensembling.

Reviewer_GPNfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The authors propose SORD, an iterative diffusion-based approach for removing localized, structured corruption. It requires only a binary mask indicating the corruption’s support and a diffusion model trained on clean signals; it does not require paired clean-corrupted data or a model of the corruption distribution.

  • The authors evaluate SORD on image corruptions—including patches, watermarks, and shadows—and localized audio interference from instruments and speech. SORD is competitive with or outperforms several inpainting, source-separation, and supervised restoration baselines across the evaluated benchmarks.

Critical strengths
  • SORD addresses localized, structured corruption with an unknown distribution. It requires only the corruption support and a diffusion prior trained on clean signals, avoiding the need for paired clean-corrupted data or a corruption-specific model.

  • The evaluation covers image restoration and audio interference, including out-of-vocabulary interferers, and compares SORD with inpainting, source-separation, and supervised restoration baselines. For MNIST patches overlaid on CelebA images, SORD achieves 27.16 dB PSNR, compared with 25.78 dB for two-prior SGM and 19.85 dB for DPS inpainting.

Critical weaknesses
  • SORD relies on accurate corruption-support masks, usually known by construction or supplied by the benchmark. The paper does not test human-drawn or coarse masks; small mask errors can substantially reduce performance, such as a drop from 27.16 to 15.20 dB PSNR with a two-pixel shift.

  • Much of the evaluation uses synthetic or controlled corruptions with known masks. The OOV audio experiments are useful, but do not establish broad generalization to naturally occurring, previously unseen corruptions.

  • The baselines are relevant but the gains are inconsistent. SORD improves over DPS inpainting on some tasks, but trails SLBR on CLWD watermark removal (28.06 vs. 31.55 dB) and ShadowDiffusion on ISTD (29.10 vs. 30.60 dB).

Questions for authors
  • How are the corruption-support masks obtained for each benchmark? Are they exact masks derived from the synthetic corruption process or dataset annotations, or are any manually drawn?

  • Are the clean-signal diffusion priors trained on data strictly disjoint from the evaluation examples? Please clarify the train/test splits for the CelebA and audio experiments.

  • Could you evaluate SORD with realistic human-drawn or coarse masks, given its sensitivity to mask errors?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.