Pith. sign in

REVIEW 5 major objections 5 minor 83 references

Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pre-trained image restoration model can be adapted to a new unlabeled degradation domain by distilling quality-gated pseudo-targets from a frozen text-to-image generative oracle, with no architectural changes and no added inference cost.

desk verdict Body is a coherent LEGO recipe for post-training restoration adaptation; the abstract/title describe a different GMD paper — fix that first, then the method is worth taking seriously. read the letter →

arxiv 2512.11121 v2 pith:6B23JNGI submitted 2025-12-11 cs.CV eess.IV

Yuyang Hu , Mojtaba Sahraee-Ardakan , Arpit Bansal , Kangfu Mei , Chenyang Qi , Peyman Milanfar , Mauricio Delbracio This is my paper
classification cs.CVeess.IV
keywords imagerestorationdomainadaptationunsupervisedlearninggenerativepriorspseudo-labelingdiffusionmodelsflowmatchingqualitygating
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes LEGO (framed as generative manifold distillation): a pre-trained image restoration model can be adapted to a new, unlabeled degradation domain by treating a frozen text-to-image model as an offline teacher. The pipeline first lets the original model produce its best guess, refines those guesses into high-quality pseudo-targets through a diffusion/flow ODE inversion-and-regeneration cycle, keeps only pseudo-targets that pass a no-reference quality gate, and finally fine-tunes the original network on a mix of in-distribution pairs and the gated pseudo-pairs. If correct, this closes a practical gap: unsupervised domain adaptation without architectural changes, paired data, or test-time cost, because the large generative model runs once during training. Reported experiments show large perceptual-quality gains on real-world deblurring and super-resolution benchmarks, with human raters preferring LEGO outputs over prior methods.

What carries the argument

The load-bearing object is the frozen generative oracle used as an offline pseudo-labeler: a large text-to-image flow-matching model. It is accessed through forward ODE inversion (to encode the initial restoration as a noise latent) and backward prompt-guided generation (to project it onto the natural-image manifold), with attention injection to preserve structure and classifier-free guidance to balance fidelity and realism. Around it sit the no-reference quality gate, which keeps bad oracle outputs out of the training set, and the mixed-supervision loss, which keeps the fine-tuned model from forgetting its in-distribution behavior while it absorbs the new domain.

What would settle it

Take a small set of out-of-distribution inputs with hidden ground truth, run the LEGO pipeline, and compare gated pseudo-targets against that ground truth. If the gated targets are no closer to the ground truth (by distortion metrics or human fidelity judgment) than the initial restorations, or if the fine-tuned model's gains vanish when the quality gate is replaced by a ground-truth-based filter, the central claim is refuted.

Watch

Extended reading notes

Core claim

LEGO reframes unsupervised domain adaptation as pseudo-supervised distillation. Stage 0 runs the pre-trained model on unlabeled low-quality inputs to get initial estimates. Stage 1 inverts each estimate into the noise-latent space of a frozen flow-matching text-to-image oracle using forward ODE integration, then regenerates a refined image under a sharpness prompt with classifier-free guidance and attention injection; a no-reference quality scorer rejects outputs below a threshold, retaining 69–86% of candidates across the four benchmarks. Stage 2 fine-tunes the original model on a mix of in-distribution pairs and selected pseudo-pairs, so the in-distribution term acts as a regularizer. The

Load-bearing premise

The argument collapses if the oracle's pseudo-targets that pass the no-reference quality gate are not genuinely better supervision than the original model's own restorations—for example, if the gate admits plausible-looking but unfaithful images, the fine-tuned model would learn those hallucinations.

Editorial extensions

If this is right

  • Adapting a deployed restoration model to a new real-world degradation requires only unlabeled low-quality images from that domain; no paired ground truth is needed.
  • The adapted model keeps its original architecture and runs at exactly the same inference cost, since the generative oracle is used only offline during training.
  • Because the in-distribution loss is retained in the fine-tuning objective, the model keeps its original restoration ability while gaining the new domain.
  • Oracle guidance is more effective as a training signal than as a test-time refinement: test-time refinement improves perception but hurts fidelity and adds seconds per image, whereas LEGO improves both fidelity and perception.
  • The same three-stage recipe is applied across deblurring and super-resolution tasks, so the framework's claims are not tied to one degradation type.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the method's ceiling is set by the oracle's pass rate and by the quality gate; because the gate uses the same no-reference metrics that also serve as evaluation metrics, the reported perceptual gains could partly reflect metric alignment rather than true fidelity. A direct test would be to validate gated pseudo-targets against ground truth on a small paired OOD sample.
  • Editorial inference: since pseudo-targets are generated offline and then distilled, the pipeline could be reused to create pseudo-paired datasets for training much smaller student models or for data augmentation in other restoration settings.
  • Editorial inference: the paper fixes the quality threshold across all datasets (pass rates 69–86%); tuning the threshold per domain, or replacing the single scorer with an ensemble, is a natural extension the paper leaves implicit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes LEGO (Learning from a Generative Oracle), a three-stage post-training framework for unsupervised domain adaptation of pre-trained image restoration models. Stage 0 obtains initial restorations from the source-domain model; Stage 1 refines these with an inversion-and-generation procedure using a frozen FLUX oracle, then filters the resulting pseudo-targets by a NIMA quality threshold; Stage 2 fine-tunes the original model on a mix of in-distribution pairs and selected pseudo-pairs. The authors claim state-of-the-art perceptual quality on GoPro-to-REDS and GoPro-to-RealBlur-J deblurring, weak-to-strong synthetic super-resolution, and synthetic-to-DPED real-world super-resolution, with no architectural changes and zero added inference latency. The paper includes ablations of filtering and mixing ratio, a comparison to inference-time oracle guidance, and a human preference study.

Significance. If the claims hold, LEGO is a practical and useful contribution: it converts an unsupervised domain adaptation problem into pseudo-supervised fine-tuning while keeping the restoration architecture and inference cost unchanged. The idea of using a frozen text-to-image model as an offline pseudo-labeler, rather than at test time, is attractive and the experimental design includes a relevant comparison to SDEdit/RF-Solver-style inference-time guidance. The ablations support the importance of quality gating and mixed supervision. However, the validity of the central mechanism rests on the assumption that the oracle's NIMA-gated outputs are better supervision than the initial restorations, and this assumption is not directly validated against ground truth. The paper also contains internal inconsistencies (abstract/body mismatch, NIMA threshold discrepancies) and lacks statistical rigor in several quantitative claims. The contribution is promising but needs additional validation before the reported results can be fully trusted.

major comments (5)
  1. [Abstract vs. §3] The provided abstract describes a method called Generative Manifold Distillation (GMD) with 'quality-gated manifold filter' and 'source-anchored trajectory regularization', but the body presents a different framework, LEGO, with three stages and no manifold-alignment or trajectory-regularization terminology. This is not a minor wording issue: the abstract and body advertise different mechanisms, making it impossible to know what is being evaluated. The authors must reconcile the abstract with the actual method, or provide the missing method if the abstract version is the intended contribution.
  2. [§3.2, Eq. (6); Table 2] The load-bearing assumption is that the FLUX-refined pseudo-targets b x_ood are better supervision than the initial restorations. This is never checked against ground truth. Since REDS and RealBlur-J test sets have paired ground truth, the paper could report PSNR/SSIM/LPIPS of the Stage-1 pseudo-targets on a held-out subset. Without this, the perceptual gains in Table 2 could be driven by hallucinated textures or content drift that pass the NIMA gate. The claim that LEGO 'preserves original robustness' is also only indirectly supported by SSIM changes that are very small (+0.004 on REDS, +0.016 on RealBlur-J). Please add a direct pseudo-target fidelity analysis.
  3. [Eq. (6) vs. Tables 2, 6, 7] NIMA is used both as the quality-gating criterion for selecting pseudo-targets and as one of the main reported perceptual metrics. Therefore, improvements in NIMA after adaptation are partly by construction: the fine-tuned model is trained to reproduce outputs that were explicitly selected to have high NIMA. This circularity also affects the ablation in Table 6 and the ratio sweep in Table 7, where NIMA-like metrics may dominate. The paper should either drop NIMA from the headline metrics or add a non-circular external evaluation (e.g., ground-truth fidelity on paired test splits, or human preference with quantitative win rates).
  4. [Fig. 3 caption; Table 10; Fig. 7 caption] The NIMA threshold is inconsistent across the manuscript: the Fig. 3 caption shows passing at ≥4.0 and failing below 4.0, the supplement Table 10 states α=4.2 for all datasets, and Fig. 7's caption says ≥4.3. This is a key hyperparameter of the quality gate, and the inconsistency undermines the reported pass rates (69.1%–86.4%) and reproducibility. Please correct and report the exact threshold, and ideally a sensitivity analysis over α.
  5. [Tables 2–4; Fig. 6] Quantitative results are reported without error bars or significance tests. Several headline gains are small (e.g., PSNR +0.27 dB on REDS, +0.14 dB on RealBlur-J), and it is unclear whether these are beyond run-to-run variation. The human study is said to report win rates with 95% confidence intervals, but no numerical win rates or p-values are given in the text or supplement. Please provide uncertainty estimates for all main tables and explicit numbers for the human preference results.
minor comments (5)
  1. [Table 2] The asterisk on 'Fully Supervised*' is not defined in the table or the text. Please clarify what the asterisk denotes (likely that it is an upper-bound trained with ground truth on the target domain).
  2. [Table 8] The Synthetic SR row says the test set is 'DIV2K Val (3,000 images)'. DIV2K has 100 validation images; if 3,000 refers to cropped patches, please state this explicitly.
  3. [§4.1] The oracle implementation uses N=50 ODE steps and guidance w=3.5, but no sensitivity analysis is provided for these hyperparameters. A short ablation or reference to a supplement section would improve reproducibility.
  4. [§3.2] The phrase 'attention injection proposed in RF-Solver' is cited but not explained in the main text. Since the supplement compares with/without attention injection, a one-sentence mechanism description in the main text would help readers understand the key design choice.
  5. [References] Several references have inconsistent formatting (e.g., missing author names in [75], [76], and the entry for [26] uses 'DONG Jiangxin' in all caps). Please unify the bibliography style.

Circularity Check

1 steps flagged · score 4.0 of 10

NIMA-gated pseudo-target selection creates partial metric-level circularity; the central adaptation claim still has independent support from FID, LPIPS, and human evaluation.

  1. self definitional [§3.2 (Eq. 6), §4.1 Evaluation metrics, Table 2]
    "we perform quality gating to ensure reliability. We evaluate each generated pseudo-target bx_ood_i using a no-reference image quality assessment (IQA) metric such as NIMA, and retain only those with scores above a threshold α: D_sel_ood = {(y_ood_i, bx_ood_i) | s_IQA(bx_ood_i) ≥ α}. ... Evaluation metrics: ... non-reference metrics (NIMA, MUSIQ, NIQE, BRISQUE, CLIP-IQA, MANIQA)."

    The 'quality' of pseudo-targets is defined by the NIMA score in Eq. 6, and NIMA is then reported as a headline perceptual-quality metric (Table 2: w/ LEGO NIMA 4.460 vs 4.325 baseline). Since Stage 2 fine-tunes L_restore(y_ood, bx_ood) toward these NIMA-selected targets, the reported NIMA improvement is partly enforced by construction rather than by independent validation of pseudo-target fidelity. No comparison of pseudo-targets to ground truth on OOD test splits is provided. The central claim still has non-circular support from FID, LPIPS, PSNR, and human preference, so the circularity is partial and metric-specific.

full rationale

The derivation chain (Stage 0 initial inference, Stage 1 inversion + prompt-guided generation + quality gating, Stage 2 mixed-supervision fine-tuning) is otherwise self-contained: no equation reduces to another by construction, no uniqueness theorem is imported from the authors, and no load-bearing argument rests on an unverified self-citation. Citations to NIMA and MUSIQ are to published, externally used metrics rather than to a self-asserted prior, so they are not circular. The main concern is the overlap between the NIMA-based pseudo-target selection criterion and NIMA as an evaluation metric, which inflates one reported metric by design. Threshold inconsistencies (α=4.0 in Fig. 3, 4.2 in Table 10, 4.3 in Fig. 7) further weaken confidence in the gate. However, because the paper also demonstrates gains on FID, LPIPS, PSNR, and human preference, the central claim does not reduce to the filtering criterion. Score 4 reflects partial metric-level circularity with independent central content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The body introduces no new physical or mathematical entities; it uses FLUX as an existing oracle and defines pseudo-targets as filtered data. The abstract's 'Generative Manifold Distillation' and 'source-anchored trajectory regularization' are named but never defined in the body. The main free parameters are the quality threshold, mixing ratio, guidance scale, and ODE steps, all chosen empirically.

free parameters (4)
  • NIMA threshold alpha = 4.2 (main text), 4.0 (Fig. 3 caption)
    Hand-chosen threshold for quality-gated pseudo-target selection; also an evaluation metric, adding mild circularity. Appears in Eq. (6).
  • Mixed-supervision ratio (in-distribution fraction) = 0.9
    Tuned by ablation on REDS (Table 7); used across all experiments.
  • Oracle guidance scale w = 3.5
    Chosen for FLUX generation; balances fidelity and realism in Eq. (4).
  • ODE steps N = 50
    Fixed Euler solver steps for inversion and generation.
assumptions (4)
  • domain assumption FLUX, a frozen text-to-image rectified-flow model, captures a strong prior over natural images, and inversion+CFG generation can project off-manifold restorations onto the natural image manifold.
    This is the fundamental premise of Stage 1 (§3.2); if the oracle fails to produce better outputs than the initial restoration, the whole pseudo-labeling scheme collapses.
  • domain assumption NIMA no-reference quality scores reliably identify pseudo-targets that are useful supervision.
    Used to construct D_sel in Eq. (6); no independent validation that high-NIMA outputs correspond to higher fidelity.
  • ad hoc to paper Mixing in-distribution data in a 9:1 ratio prevents catastrophic forgetting and overfitting to pseudo-label artifacts.
    The ratio is an empirical choice validated by an ablation on a single dataset (Table 7), not derived from theory.
  • domain assumption Euler discretization with N=50 steps and attention injection preserves semantic structure during inversion and generation.
    The paper relies on RF-Solver with attention injection to keep content identity; no formal guarantee is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior." pith.science (2026). https://pith.science/paper/6B23JNGI

@misc{pith2026251211121,
  author       = {Pith},
  title        = {Pith review of: Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6B23JNGI}},
  note         = {Machine review of arXiv:2512.11121}
}
read the original abstract

Pre-trained image restoration models often fail on out-of-distribution (OOD) real-world degradations. Adapting to these domains is challenging as real-world data lacks paired ground truth, and unsupervised methods often require unstable architectural changes. We propose Generative Manifold Distillation (GMD), which reframes domain adaptation as geometric manifold alignment. GMD operates in a strictly unpaired setting, requiring only low-quality (LQ) target observations. By leveraging the flow-matching dynamics of a frozen text-to-image foundation model, GMD projects off-manifold restorations onto the natural image manifold to generate high-quality pseudo-targets. To ensure stability, a quality-gated manifold filter rejects off-manifold samples, while source-anchored trajectory regularization prevents error accumulation. Ultimately, GMD distills a powerful generative prior into an efficient restoration network. Experiments demonstrate that GMD seamlessly adapts to new distributions using only LQ inputs, drastically improving perceptual quality with zero architectural modifications or added inference latency.

Figures

Figures reproduced from arXiv: 2512.11121 by the authors.

Figure 1
Figure 1. Bridging the Domain Gap with LEGO. (Top) Real￾world deblurring and (Bottom) 4x super-resolution examples. The baseline model (LDM) struggles with unseen, out-of-distribution degradations. Our LEGO adaptation, which requires only unlabeled target images, produces significantly cleaner, sharper, and more faith￾ful results, without any modification to model architecture. Models pre-trained for synthesis act as strong p… view at source ↗
Figure 2
Figure 2. Overview of the LEGO Framework. (a) The Challenge: An in-distribution model trained on Did fails on out-of-distribution Dood data. (b) The LEGO Solution: A post-training adaptation process. Stage 0 gets an initial restoration (x˜ ood) from the in-distribution model. Stage 1 refines x˜ ood into a high-quality pseudo-target using a frozen generative oracle (FLUX). Stage 2 fine-tunes the original model on a mix of in-d… view at source ↗
Figure 3
Figure 3. Selection of pseudo-targets based on image qual￾ity assessment. Top row: A successful selection where the generated target scored above the NIMA threshold (≥ 4.0). Bottom row: A rejection case where the target quality is insuf￾ficient. models, such as diffusion or flow-matching models, describe data generation as a continuous-time process governed by a learned velocity field vθ: d x(τ ) dτ = vθ [PITH_FULL_IMAGE:fig… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the REDS (top) and RealBlur-J (bottom) deblurring dataset. The GoPro-trained LDM baseline (mismatched) and other SOTA methods often leave residual blur or introduce artifacts. LEGO successfully adapts to the out-of-distribution domain (both RE…
Figure 5
Figure 5. Figure 5: Qualitative comparison of real-world SR on the DPED-iPhone dataset. Compared to state-of-the-art baselines, LEGO produces sharper details and more natural textures under real-world degradations. LEGO achieves visually realistic enhancement without introducing artifacts…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Selection of pseudo-targets based on image quality assessment on REDS training set. Top row: A successful selection where the generated target scored above the NIMA threshold (≥ 4.3). Bottom row: A rejection case where the target quality is insufficient [PITH_FULL_IMA…
Figure 8
Figure 8. Figure 8: Visual comparison of pseudo-target generation strategies (Stage 1). We evaluate the intrinsic quality of different generation methods. Standard baselines like SDEdit [41] or the flow matching inversion without Attention Injection help remove artifacts but suffer from s…
Figure 9
Figure 9. Figure 9: Additional Qualitative comparison on the REDS deblurring dataset. The GoPro-trained LDM baseline (mismatched) leaves residual motion blur. LEGO successfully adapts to the out-of-distribution domain, producing sharper, more detailed, and perceptually superior restoratio…
Figure 10
Figure 10. Figure 10: Additional Qualitative comparison on the RealBlur-J deblurring dataset (Set 1). Comparison showing the adaptation to real-world low-light blur. LEGO recovers text and fine structures that are lost by the baseline model. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Additional Qualitative comparison on the RealBlur-J deblurring dataset (Set 2). LEGO successfully adapts to the out-of-distribution domain, producing sharper, more detailed, and perceptually superior restorations. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Additional Qualitative comparison on Synthetic Super-Resolution (Weak → Strong). The baseline model, pre-trained only on weak degradations, fails to generalize to the heavy noise and blur in the target domain. LEGO successfully adapts to the stronger degradation profi…
Figure 13
Figure 13. Figure 13: Additional Qualitative comparison on Synthetic Super-Resolution (Weak → Strong). LEGO demonstrates robust adaptation to severe degradations, effectively removing noise and sharpening details without requiring paired ground truth. 20 [PITH_FULL_IMAGE:figures/full_fig_…
Figure 14
Figure 14. Figure 14: Additional Qualitative comparison on Real-World Super-Resolution (Synthetic → DPED-iPhone). The baseline model, pre-trained on synthetic data, fails to generalize to the complex sensor noise and compression artifacts of the iPhone camera. LEGO successfully adapts to t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

83 extracted references · 2 linked inside Pith

  1. [1]

    Ntire 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. InCVPRW, pages 1122–1131. IEEE Computer Society, 2017. 6

    2017

  2. [2]

    Toward real-world single image super-resolution: A new benchmark and a new model

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. InICCV, pages 3086–3095,

  3. [3]

    Hinet: Half instance normalization network for image restoration

    Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. InCVPRW, pages 182–192, 2021. 12

    2021

  4. [4]

    Unsupervised blind image deblurring based on self- enhancement

    Lufei Chen, Xiangpeng Tian, Shuhua Xiong, Yinjie Lei, and Chao Ren. Unsupervised blind image deblurring based on self- enhancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25691–25700,

  5. [5]

    Hierarchical integration diffusion model for realistic image deblurring.NIPS, 36:29114–29125, 2023

    Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring.NIPS, 36:29114–29125, 2023. 1

    2023

  6. [6]

    Hierarchical integration diffusion model for realistic image deblurring.NIPS, 36, 2024

    Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring.NIPS, 36, 2024. 1, 5, 6

    2024

  7. [7]

    Unpaired deblurring via decoupled diffusion model.arXiv preprint arXiv:2502.01522, 2025

    Junhao Cheng, Wei-Ting Chen, Xi Lu, and Ming-Hsuan Yang. Unpaired deblurring via decoupled diffusion model.arXiv preprint arXiv:2502.01522, 2025. 2

    arXiv 2025

  8. [8]

    Latentpaint: Image inpainting in latent space with diffusion models

    Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334–4343, 2024. 1, 2

    2024

Show all 83 references
  1. [9]

    Adair: Adaptive all-in-one image restoration via frequency mining and modula- tion

    Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modula- tion. InThe Thirteenth International Conference on Learning Representations, 2025. 5, 6

    2025

  2. [10]

    Inversion by di- rect iteration: An alternative to denoising diffusion for image restoration.Transactions on Machine Learning Research, 2023

    Mauricio Delbracio and Peyman Milanfar. Inversion by di- rect iteration: An alternative to denoising diffusion for image restoration.Transactions on Machine Learning Research, 2023. Featured Certification, Outstanding Certification. 1

    2023

  3. [11]

    Diffusion models beat gans on image synthesis.NIPS, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.NIPS, 34:8780–8794, 2021. 1

    2021

  4. [12]

    Learning spatially-variant map models for non-blind image deblurring

    Jiangxin Dong, Stefan Roth, and Bernt Schiele. Learning spatially-variant map models for non-blind image deblurring. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4886–4895, 2021. 2

    2021

  5. [13]

    Ss-jircs: Self- supervised joint image reconstruction and coil sensitivity cali- bration in parallel mri without ground truth

    Weijie Gan, Yuyang Hu, Cihat Eldeniz, Jiaming Liu, Yasheng Chen, Hongyu An, and Ulugbek S Kamilov. Ss-jircs: Self- supervised joint image reconstruction and coil sensitivity cali- bration in parallel mri without ground truth. InProceedings of the IEEE/CVF International Confere...

    2021

  6. [14]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30, 2017. 6

    2017

  7. [15]

    Classifier-free diffusion guid- ance.ARXIV, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guid- ance.ARXIV, 2022. 2, 4

    2022

  8. [16]

    Denoising diffusion probabilistic models.NIPS, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.NIPS, 33:6840–6851, 2020. 1, 4

    2020

  9. [17]

    Cycada: Cycle-consistent adversarial domain adaptation

    Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. InInterna- tional Conference on Machine Learning, pages 1989–1998. Pmlr, 2018. 1

    1989

  10. [18]

    Ref-ldm: A latent diffusion model for reference-based face image restoration.NIPS, 37: 74840–74867, 2024

    Chi-Wei Hsiao, Yu-Lun Liu, Cheng-Kun Yang, Sheng-Po Kuo, Kevin Jou, and Chia-Ping Chen. Ref-ldm: A latent diffusion model for reference-based face image restoration.NIPS, 37: 74840–74867, 2024. 2

    2024

  11. [19]

    Spicer: Self-supervised learning for mri with automatic coil sensitivity estimation and reconstruction.Mag- netic resonance in medicine, 92(3):1048–1063, 2024

    Yuyang Hu, Weijie Gan, Chunwei Ying, Tongyao Wang, Cihat Eldeniz, Jiaming Liu, Yasheng Chen, Hongyu An, and Ulug- bek S Kamilov. Spicer: Self-supervised learning for mri with automatic coil sensitivity estimation and reconstruction.Mag- netic resonance in medicine, 92(3):1048–...

    2024

  12. [20]

    Stochastic deep restora- tion priors for imaging inverse problems.ARXIV, 2024

    Yuyang Hu, Albert Peng, Weijie Gan, Peyman Milanfar, Mauri- cio Delbracio, and Ulugbek S Kamilov. Stochastic deep restora- tion priors for imaging inverse problems.ARXIV, 2024. 2

    2024

  13. [21]

    Multimodal diffusion bridge with attention-based sar fusion for satellite image cloud removal.ARXIV, 2025

    Yuyang Hu, Suhas Lohit, Ulugbek S Kamilov, and Tim K Marks. Multimodal diffusion bridge with attention-based sar fusion for satellite image cloud removal.ARXIV, 2025. 2

    2025

  14. [22]

    Kernel density steering: Inference-time scaling via mode seeking for image restoration

    Yuyang Hu, Kangfu Mei, Mojtaba Sahraee-Ardakan, Ulugbek S Kamilov, Peyman Milanfar, and Mauricio Delbracio. Kernel density steering: Inference-time scaling via mode seeking for image restoration. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 2

    2025

  15. [23]

    Dslr-quality photos on mobile devices with deep convolutional networks

    Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. InProceedings of the IEEE international conference on computer vision, pages 3277–3285, 2017. 4, 5

    2017

  16. [24]

    Eqadap: Equipollent domain adaptation approach to image deblurring.IEEE Access, 10:93203–93211, 2022

    Ibsa Jalata, Naga Venkata Sai Raviteja Chappa, Thanh-Dat Truong, Pierce Helton, Chase Rainwater, and Khoa Luu. Eqadap: Equipollent domain adaptation approach to image deblurring.IEEE Access, 10:93203–93211, 2022. 1

    2022

  17. [25]

    Uncertainty-aware variate de- composition for self-supervised blind image deblurring

    Runhua Jiang and Yahong Han. Uncertainty-aware variate de- composition for self-supervised blind image deblurring. In Proceedings of the 31st ACM International Conference on Mul- timedia, pages 252–260, 2023. 2

    2023

  18. [26]

    Learning spatially-variant map models for non-blind image deblurring

    DONG Jiangxin, S ROTH, and B SCHIELE. Learning spatially-variant map models for non-blind image deblurring. In CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, pages 4884–4893, 2021. 2

    2021

  19. [27]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InICCV, pages 5148–5157, 2021. 6

    2021

  20. [28]

    Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

    Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. InICCV, 2019. 5, 6

    2019

  21. [29]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 1, 2, 3, 5

    2024

  22. [30]

    Srdiff: Single image super-resolution with diffusion probabilistic models.Neuro- computing, 479:47–59, 2022

    Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models.Neuro- computing, 479:47–59, 2022. 1

    2022

  23. [31]

    Diffusion models for image restoration and enhancement–a comprehensive survey

    Xin Li, Yulin Ren, Xin Jin, Cuiling Lan, Xingrui Wang, Wenjun Zeng, Xinchao Wang, and Zhibo Chen. Diffusion models for image restoration and enhancement–a comprehensive survey. ARXIV, 2023. 2

    2023

  24. [32]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InCVPRW, 2017. 6 9

    2017

  25. [33]

    Diffbir: Toward blind image restoration with generative diffusion prior

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. InECCV, pages 430–448. Springer, 2024. 2

    2024

  26. [34]

    Learning deblurring texture prior from unpaired data with diffusion model

    Chengxu Liu, Lu Qi, Jinshan Pan, Xueming Qian, and Ming- Hsuan Yang. Learning deblurring texture prior from unpaired data with diffusion model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14195– 14204, 2025. 2

    2025

  27. [35]

    Unsupervised domain-specific deblurring via disentangled representations

    Boyu Lu, Jun-Cheng Chen, and Rama Chellappa. Unsupervised domain-specific deblurring via disentangled representations. In CVPR, pages 10225–10234, 2019. 1

    2019

  28. [36]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InCVPR, pages 11461–11471, 2022. 1, 2

    2022

  29. [37]

    Controlling vision-language models for multi-task image restoration

    Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sj ¨olund, and Thomas B Sch¨on. Controlling vision-language models for multi-task image restoration. InICLR, 2024. 5, 6, 7

    2024

  30. [38]

    Codi: Conditional diffusion distillation for higher-fidelity and faster image genera- tion

    Kangfu Mei, Mauricio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image genera- tion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ...

    2024

  31. [39]

    Bigger is not always better: Scaling properties of latent diffusion models

    Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M Patel, and Peyman Milanfar. Bigger is not always better: Scaling properties of latent diffusion models. TMLR, 2024. 1

    2024

  32. [40]

    The power of context: How multimodality improves image super-resolution

    Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Pa- tel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image super-resolution. InCVPR, 2025. 1

    2025

  33. [41]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 2, 8, 14, 15

    2022

  34. [42]

    Denoising: a power- ful building block for imaging, inverse problems and machine learning.Philosophical Transactions A, 383(2299):20240326,

    Peyman Milanfar and Mauricio Delbracio. Denoising: a power- ful building block for imaging, inverse problems and machine learning.Philosophical Transactions A, 383(2299):20240326,

  35. [43]

    No-reference image quality assessment in the spatial domain

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12):4695–4708,

  36. [44]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InCVPR, pages 6038–6047,

  37. [45]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InCVPR, 2017. 4, 5, 6

    2017

  38. [46]

    Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study

    Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. InCVPRW, 2019. 4, 5, 6

    2019

  39. [47]

    Multi-adversarial domain adaptation

    Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. InAAAI, 2018. 1

    2018

  40. [48]

    Blur2blur: Blur conversion for un- supervised image deblurring on unknown domains

    Bang-Dang Pham, Phong Tran, Anh Tran, Cuong Pham, Rang Nguyen, and Minh Hoai. Blur2blur: Blur conversion for un- supervised image deblurring on unknown domains. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2804–2813, 2024. 2

    2024

  41. [49]

    Spire: Semantic prompt-driven image restoration

    Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. Spire: Semantic prompt-driven image restoration. InEuropean Con- ference on Computer Vision, pages 446–464. Springer, 2024. 1, 3

    2024

  42. [50]

    Neural blind deconvolution using deep priors

    Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3341–3350, 2020. 2

    2020

  43. [51]

    Multiscale structure guided diffu- sion for image deblurring

    Mengwei Ren, Mauricio Delbracio, Hossein Talebi, Guido Gerig, and Peyman Milanfar. Multiscale structure guided diffu- sion for image deblurring. InICCV, pages 10721–10733, 2023. 1, 2, 12

    2023

  44. [52]

    Real-world blur dataset for learning and benchmarking deblur- ring algorithms

    Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblur- ring algorithms. InECCV, pages 184–201. Springer, 2020. 1, 4, 5, 6

    2020

  45. [53]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, pages 10684–10695, 2022. 1, 2

    2022

  46. [54]

    Semantic image inver- sion and editing using rectified stochastic differential equations

    Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inver- sion and editing using rectified stochastic differential equations. InICLR, 2025. 1, 2

    2025

  47. [55]

    Blind quality assessment of videos using a model of natural scene statistics and motion co- herency

    Michele A Saad and Alan C Bovik. Blind quality assessment of videos using a model of natural scene statistics and motion co- herency. In2012 Conference Record of the Forty Sixth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 332–336. IEEE, 2012. 6

    2012

  48. [56]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 Conference Proceedings, pages 1–10, 2022. 2

    2022

  49. [57]

    Image super-resolution via iterative refinement.PAMI, 45(4):4713–4726, 2022

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement.PAMI, 45(4):4713–4726, 2022. 1, 2

    2022

  50. [58]

    Domain adaptation for image dehazing

    Yuanjie Shao, Lerenhan Li, Wenqi Ren, Changxin Gao, and Nong Sang. Domain adaptation for image dehazing. InCVPR, pages 2808–2817, 2020. 1 10

    2020

  51. [59]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ARXIV, 2020. 1

    2020

  52. [60]

    Nima: Neural image assessment.TIP, 27(8):3998–4011, 2018

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment.TIP, 27(8):3998–4011, 2018. 4, 6

    2018

  53. [61]

    Explor- ing clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Explor- ing clip for assessing the look and feel of images. InAAAI, pages 2555–2563, 2023. 6

    2023

  54. [62]

    Exploiting diffusion prior for real-world image super-resolution.IJCV, 132(12):5929–5949,

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution.IJCV, 132(12):5929–5949,

  55. [63]

    Taming rectified flow for inversion and editing

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. InICML, 2025. 1, 2, 4, 5, 8, 14, 15

    2025

  56. [64]

    Unsupervised real-world super-resolution: A domain adaptation perspective

    Wei Wang, Haochen Zhang, Zehuan Yuan, and Changhu Wang. Unsupervised real-world super-resolution: A domain adaptation perspective. InICCV, pages 4298–4307, 2021. 1

    2021

  57. [65]

    Real- esrgan: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real- esrgan: Training real-world blind super-resolution with pure synthetic data. InICCV, pages 1905–1914, 2021. 1, 4, 6

    1905

  58. [66]

    Image quality assessment: from error visibility to structural similarity.TIP, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.TIP, 13(4):600–612, 2004. 6

    2004

  59. [67]

    Uformer: A general u- shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u- shaped transformer for image restoration. InCVPR, pages 17683–17693, 2022. 5, 6

    2022

  60. [68]

    Deblurring via stochastic refinement

    Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Sa- haria, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. InCVPR, pages 16293–16303, 2022. 1, 2

    2022

  61. [69]

    Deflow: Learning complex image degradations from unpaired data with conditional flows

    Valentin Wolf, Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deflow: Learning complex image degradations from unpaired data with conditional flows. In CVPR, pages 94–103, 2021. 1, 2

    2021

  62. [70]

    Id-blau: Image deblurring by implicit diffusion-based reblurring augmentation

    Jia-Hao Wu, Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Chia-Wen Lin, and Yen-Yu Lin. Id-blau: Image deblurring by implicit diffusion-based reblurring augmentation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25847–25856, 2024. 2

    2024

  63. [71]

    Seesr: Towards semantics-aware real- world image super-resolution

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real- world image super-resolution. InCVPR, pages 25456–25467,

  64. [72]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. InICCV, pages 13095–13105, 2023. 2

    2023

  65. [73]

    Maniqa: Multi- dimension attention network for no-reference image quality assessment

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi- dimension attention network for no-reference image quality assessment. InCVPR, pages 1191–1200, 2022. 6

    2022

  66. [74]

    Dualgan: Unsupervised dual learning for image-to-image translation

    Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849–2857, 2017. 2

    2017

  67. [75]

    Multi- stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Hayat, Fa- had Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi- stage progressive image restoration. InCVPR, 2021. 5, 6

    2021

  68. [76]

    Restormer: Ef- ficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Hayat, Fa- had Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Ef- ficient transformer for high-resolution image restoration. In CVPR, 2022. 5, 6

    2022

  69. [77]

    Deblurring by realis- tic blurring

    Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realis- tic blurring. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2737–2746,

  70. [78]

    The unreasonable effectiveness of deep fea- tures as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep fea- tures as a perceptual metric. InCVPR, pages 586–595, 2018. 6

    2018

  71. [79]

    Neural maximum a posteriori estimation on unpaired data for motion deblurring.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Neural maximum a posteriori estimation on unpaired data for motion deblurring.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2

    2023

  72. [80]

    Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring

    Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, and Meng Wang. Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring. InProceed- ings of the 30th ACM International Conference on Multimedia, pages 6220–6229, 2022. 2

    2022

  73. [81]

    Reference-guided identity preserving face restoration.arXiv preprint arXiv:2505.21905, 2025

    Mo Zhou, Keren Ye, Viraj Shah, Kangfu Mei, Mauricio Del- bracio, Peyman Milanfar, Vishal M Patel, and Hossein Talebi. Reference-guided identity preserving face restoration.arXiv preprint arXiv:2505.21905, 2025. 1 11 8 Supplement This supplementary material provides detailed in...

    2025 arXiv

  74. [82]

    No Attention

    Intrinsic Generation Quality.We first examine the visual fidelity of the pseudo-targets generated in Stage 1. Qualitative comparisons in Figure 8 reveal that methods lacking Attention Injection (both SDEdit and the “No Attention” pipeline) struggle to maintain structural consi...

  75. [83]

    Table 11 presents the quantitative results, comparing our full LEGO pipeline against the SDEdit- based adaptation

    Downstream Adaptation Performance.Based on the visual analysis, we utilize the RF-Solver-generated targets for the downstream adaptation task. Table 11 presents the quantitative results, comparing our full LEGO pipeline against the SDEdit- based adaptation. While SDEdit achiev...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.