Pith. sign in

REVIEW 3 major objections 4 minor 91 references

Residual Rectified Flow adapts pre-trained rectified-flow diffusion models to real-world image restoration by starting sampling from a noisy low-quality state instead of pure noise, achieving four-step inference from an exact acceleration p

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:53 UTC pith:LVRV43PE

load-bearing objection RRF is a clean and useful idea for adapting pretrained rectified-flow backbones to 4-step restoration, but the SOTA claim depends on an unvalidated DMD2-style synthetic-state fix that the paper itself flags as critical. the 3 major comments →

arxiv 2607.25275 v1 pith:LVRV43PE submitted 2026-07-28 cs.CV cs.AI

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field

Zhenning Shi , Chen Xu , Junhao Zhang , Kefei Zhang , Linjie Liu , Zhedong Zheng , Tao Li This is my paper
classification cs.CV cs.AI
keywords real-world image restorationdiffusion modelsrectified flowresidual flowimage super-resolutionLoRA adaptationknowledge distillationfew-step sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a pre-trained rectified-flow diffusion model can be turned into a real-world image restorer by inserting a compact residual term into its transport path. The key move is writing the intermediate state as a linear combination of the high-quality image, the low-quality image, and noise, so that at a specific time the unknown high-quality component cancels out exactly, leaving a known noisy low-quality starting state. Because the learned velocity differs from the standard rectified-flow velocity only by a fixed residual offset, the authors claim that a frozen pre-trained backbone can be adapted with low-rank LoRA updates plus a diffusion-distillation regularizer. If correct, this gives four-step restoration with the generative quality of a billion-scale text-to-image model, while preserving input structure better than Gaussian-noise samplers and more sampling flexibility than one-step mappers.

Core claim

The paper's central discovery is that rectified-flow transport can be redirected through the image residual R = x̂0 − x0 without breaking the linear ODE structure. Writing x_t = (1−t)x0 + tγR + tε and re-expressing it as x_t = [1−(1+γ)t]x0 + γt x̂0 + tε, the coefficient of the unknown high-quality image vanishes at t* = 1/(1+γ). Sampling therefore begins at the exactly known acceleration state x_{t*} = γ/(1+γ) x̂0 + 1/(1+γ) ε and integrates the learned residual velocity resv = γR + ε − x0 back to the high-quality endpoint. Since resv equals the standard rectified-flow target ε − x0 plus the fixed offset γR, the authors argue that a frozen pre-trained rectified-flow backbone needs only LoRA a

What carries the argument

Residual Rectified Flow (RRF): the modified linear path x_t = (1−t)x0 + tγR + tε, with residual R = x̂0 − x0 and residual ratio γ, together with its exact acceleration point t* = 1/(1+γ). This path carries the argument: it makes low-quality-aware initialization possible, yields a closed-form starting distribution that depends only on the observed low-quality image and noise, and guarantees that the learned velocity differs from the standard rectified-flow velocity only by the offset γR, which is why LoRA-only parameter-efficient adaptation is claimed to suffice.

Load-bearing premise

The four-step results stand only if the synthetic states produced by the generator during training successfully close the training–inference distribution gap; if those noisy synthetic states drift from real high-quality states, the accelerated path from the acceleration point will inherit that drift, especially in highly ill-posed tasks like super-resolution.

What would settle it

Train two identical Residual Rectified Flow models on the same data, one with real high-quality images in the training states and one with the synthetic-state replacement described in the paper, then evaluate both under identical 4-step sampling on a real-world super-resolution benchmark such as DRealSR. If the synthetic-replacement model does not match or beat the real-state model, the mitigation is not doing the load-bearing work; alternatively, measure the Wasserstein distance between terminal output distributions when starting states are built with true x0 versus the model's predicted x̂0.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Standard rectified-flow text-to-image backbones can be repurposed for restoration by learning a compact residual offset, removing the need to train a restoration diffusion model from scratch.
  • Starting at the acceleration point replaces 20–50 Gaussian-noise sampling steps with 4 ODE steps, with reported sub-second inference at 512×512 resolution.
  • Because the residual velocity target is a fixed offset from the original rectified-flow target, small LoRA ranks already work (rank 4 is viable; rank 32 is best), supporting parameter-efficient adaptation.
  • Restoration quality improves consistently as the pre-trained backbone grows from 2B to 9B parameters, so the recipe inherits scaling gains from larger generative models.
  • An optional GAN fine-tuning stage trades some PSNR/SSIM for better perceptual and no-reference metrics, giving a fidelity-versus-realism choice for users.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same residual re-parameterization should extend to other inverse problems with a well-defined residual, such as dehazing, deraining, or low-light enhancement; whether LoRA-only adaptation suffices there depends on how compact the residual is, which the paper only tests on super-resolution-like degradations.
  • The paper's stability analysis implies that the residual ratio γ is a fidelity-versus-stochasticity dial, but the paper fixes γ=1; a natural untested extension is to schedule γ per image or per degradation level.
  • The empirical results rest on replacing real training states with synthetic states produced by the generator, following a prior distillation recipe; the paper does not quantify the residual distribution gap this replacement leaves, so a controlled comparison between real-state and synthetic-state training would clarify how much of the four-step success depends on that fix.
  • If each restoration task is captured by a separate low-rank residual adapter on one shared frozen backbone, composability of multiple LoRA adapters on the same rectified-flow model becomes a testable extension the paper does not address.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. ScaleResfusion proposes Residual Rectified Flow (RRF), a modification of standard rectified flow that injects the image residual R into the linear transport path: x_t = (1-t)x_0 + tγR + tε. The paper derives an exact acceleration point t* = 1/(1+γ) at which the unknown clean image coefficient vanishes, allowing sampling to start from a noisy LQ state x_{t*} = γ/(1+γ) x̂_0 + 1/(1+γ) ε. The residual velocity target resv = γR + ε − x_0 differs from the standard RF target by the residual offset γR, which the authors argue enables adaptation of pre-trained rectified-flow backbones via pure LoRA. A DMD-based knowledge-distillation pipeline is introduced for few-step (4-step) sampling. Experiments on DRealSR, RealSR, DIV2K-Val, and LSDIR-Val across SD3, FLUX2, and Z-Image backbones report SOTA or competitive fidelity and perceptual metrics, with user-study support.

Significance. If the reported results hold, the paper makes a useful contribution: it extends the residual/noisy-LQ-initialization idea to modern rectified-flow backbones while preserving the linear transport structure, and it provides a plausible route to scaling restoration models to billions of parameters with low-rank adapters. The algebraic derivation of the acceleration point is correct, and the ablation set is extensive, including backbone scaling, residual ratio, NFE, LoRA rank, and LQ conditioning. The paper also ships code/models and includes a cautionary section on no-reference metrics, which is thoughtful. The central weakness is that the 4-step SOTA claim relies on an unisolated and unquantified training–inference mismatch fix in Appendix C, and the paper's latent-space vs image-space notation leaves a gap in the theoretical foundation.

major comments (3)
  1. [§3.2–§3.4, Eq. (4)–(8), Algorithm 1] The residual R is defined in image space (Eq. 4), but the method trains and samples in the VAE latent space, where x_t is said to denote the latent RRF state (§3.4). The exact acceleration point and the residual velocity target require R = x̂0 − x0 to hold in the same space as the flow. The paper never defines the latent residual explicitly, nor does it justify that the VAE encoder preserves the additive residual relation. If the residual is computed in image space and then added to latent states, the coefficient cancellation in the derivation of t* does not carry over. Please state clearly whether all variables in §3.3 are latent-space variables or redefine the residual accordingly; otherwise the implementation deviates from the theory.
  2. [Appendix C, §3.4, Table 8] The training–inference mismatch is explicitly disclosed in Appendix C as 'critical for highly ill-posed problems such as super-resolution' and is mitigated by replacing noisy real training states with synthetic generator states following DMD2/DynaRS. No quantitative closure evidence is provided: no measurement of the residual distribution gap between training states and inference-time states, and no ablation that isolates the renoising/synthetic-state mechanism. Table 8 ablates the DMD loss as a whole, not the synthetic-state replacement. Given that the headline 4-step SOTA claim rests on this fix, please add a dedicated ablation (with vs. without the synthetic-state replacement) and report a distribution-shift diagnostic (e.g., FID/MMD between the state distributions) or an equivalent demonstration that the mismatch is closed.
  3. [Tables 1–4 and Appendix F] All quantitative metrics are reported on single runs with no error bars, seed variance, or statistical significance testing. Several cross-method differences are small (e.g., LPIPS 0.24 vs. 0.25, DISTS 0.19 vs. 0.20), and the perceptual–distortion trade-off claims are based on point estimates. Please report mean ± std over at least 3 seeds (or bootstrap confidence intervals over test images) for the main tables and key ablations; without this, the SOTA claims are not statistically supported.
minor comments (4)
  1. [Eq. (9)] The SNR expression is stated as proportional to γ without defining the noise scale or whether it refers to latent-space variance. Clarify the constant and the space in which SNR is computed.
  2. [Fig. 4 caption] The caption contains corrupted math/Unicode glyphs (e.g., '~ (0 ),Tx', 'R'); the typesetting should be fixed.
  3. [Algorithm 1] In the inference return statement, the variable x0 is returned, but it is the estimated restored image, not the ground truth. Rename to x̂0^{pred} or 'x0_est' to avoid confusion.
  4. [Appendix A] Equation numbering is reused: Eq. (22) in Appendix A duplicates Eq. (4) in the main text. Renumber the appendix equations.

Circularity Check

0 steps flagged

No significant circularity: the RRF derivation is algebraic and self-contained, and the experimental validation uses external benchmarks.

full rationale

The central derivation (Eqs. 5–8, Theorem 2 in App. G) is a direct algebraic construction: R = x̂0 − x0, x_t = (1−t)x0 + tγR + tε, and resv = γR + ε − x0 = v_RF + γR. The acceleration point t★ = 1/(1+γ) is obtained by solving 1−(1+γ)t★ = 0, which cancels the unknown HQ coefficient; no fitted parameter or renamed empirical pattern is presented as a prediction. The claim that the residual velocity target differs from the standard RF target only by a residual offset is an identity by construction, and the LoRA adaptation claim is an architectural consequence plus an empirical scaling result, not a circular reduction. The citation to the authors' prior Resfusion (Shi et al. 2024) for the geometric 'acceleration point' is not load-bearing because the paper proves the cancellation in Theorem 2. Appendix C explicitly discloses a training–inference mismatch (real x0 in training states vs predicted x_pred0 at inference) and states it is mitigated by replacing noisy real states with noisy synthetic generator states following DMD2. This is a disclosed assumption/limitation with no convergence guarantee or isolated ablation, which is a correctness risk, but it is not circular: the fix is an external distillation procedure, not a parameter fitted to the reported metrics. Benchmarks, ablations, and user studies are all external validation. No circular step could be exhibited with a specific equation reducing the result to its inputs, so the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The paper's load-bearing choices are the tuned hyperparameters (γ, NFE, loss weights, LoRA rank, feature extractor depth) and several domain assumptions: latent-space residual additivity, LoRA expressiveness for the residual offset, and the DMD2-style synthetic-state fix for the train/inference mismatch. These are reasonable but not all rigorously established.

free parameters (5)
  • residual ratio γ = 1 (default; ablated 0.5, 1, 2)
    Controls the SNR of the RRF starting state and is selected via ablation on DRealSR/LSDIR-Val; the final reported numbers use this tuned value.
  • number of function evaluations (NFE) = 4 (default; ablated 1, 2, 4)
    Chosen as the default after NFE ablation; 1- and 2-step variants are also reported and remain competitive.
  • loss weights λ_mse, λ_lpips, λ_dmd, λ_gan = not reported
    The weighted objectives in Eqs. (11)-(12) require hyperparameter choices, but the actual values are not given in the paper.
  • LoRA rank = 32 (ablated 4, 8, 16, 32)
    Higher rank consistently improves metrics; 32 is chosen as default after ablation.
  • feature extractor layers = 23 (ablated 4, 8, 16, 23)
    Deeper LQ feature extraction improves restoration; the full 23-layer extractor is used for SD3.
axioms (5)
  • standard math Rectified Flow linear interpolation x_t = (1−t)x0 + tε and marginal-preserving velocity matching
    Background for §3.1 and Appendix G; taken directly from Liu et al. 2022.
  • ad hoc to paper Residual R can be mapped through the frozen VAE and used additively in latent space
    R is defined in image space in Eq. (4) but the RRF path Eq. (5) and Fig. 5 operate on latent states; the paper never defines a latent residual or proves the encoder behaves linearly.
  • domain assumption Pretrained RF model output on RRF states differs from the desired residual velocity by a residual offset that LoRA can fit
    Central to the pure-LoRA claim in §3.3; only supported empirically by ablations, not by a formal guarantee.
  • ad hoc to paper Residual corrections occupy a low-rank subspace
    Used to justify LoRA sufficiency; the rank ablation supports it empirically, but no first-principles bound is given.
  • domain assumption DMD2-style synthetic-state replacement closes the training/inference mismatch
    Explicitly relied on in Appendix C for few-step SR stability; the paper flags the mismatch but does not prove the mitigation converges or fully closes the gap.

pith-pipeline@v1.3.0-alltime-deepseek · 36768 in / 12981 out tokens · 132101 ms · 2026-08-01T02:53:26.164476+00:00 · methodology

0 comments
read the original abstract

Real-world Image Restoration (Real-IR) aims to recover high-quality (HQ) images from complex and unknown degradations. Although recent diffusion-based methods have substantially improved perceptual quality, their current designs leave two key challenges unresolved. Methods that start from Gaussian noise are slow and often less faithful to the degraded input. Residual-based methods usually train from scratch, which makes it hard to exploit modern pre-trained generative priors. In this paper, we present ScaleResfusion, a scalable diffusion framework for real-world image restoration built on pre-trained text-to-image rectified-flow models. The core of our method is Residual Rectified Flow, which introduces the residual term R into Standard Rectified Flow. Instead of starting from pure noise, it uses a residual transport path that starts from noisy low-quality (LQ) images and admits an exact acceleration point. By learning the residual vector field, Residual Rectified Flow keeps the output distribution and linear diffusion process consistent with the pre-trained rectified-flow models. This makes parameter-efficient fine-tuning possible at scale. We further introduce a knowledge-distillation pipeline to reduce sampling cost while maintaining restoration quality. Extensive experiments on multiple real-world restoration tasks show that ScaleResfusion achieves state-of-the-art performance with much higher efficiency. These results suggest a practical and scalable way to adapt large pre-trained diffusion models to real-world image restoration. Our code and models are available at https://github.com/YukinoshitaLove/ScaleResfusion.

Figures

Figures reproduced from arXiv: 2607.25275 by Chen Xu, Junhao Zhang, Kefei Zhang, Linjie Liu, Tao Li, Zhedong Zheng, Zhenning Shi.

Figure 1
Figure 1. Figure 1: Visual examples of ScaleResfusion for Real-IR. The top row shows LQ inputs and the bottom row shows our restored results with only 4 sampling steps. ScaleResfusion preserves the structure, identity, and content of the input images while synthesizing photo-realistic details. Real-world Image Restoration (Real-IR) aims to recover high-quality (HQ) images from complex and unknown degradations. Although recent… view at source ↗
Figure 2
Figure 2. Figure 2: Initialization comparison from a transport perspective. (a) Gaussian-noise initialization starts far from the restoration manifold and induces unstable [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The residual term 𝑅 exhibits consistent low-energy, edged patterns in both the spatial and frequency domains, suggesting that RRF only needs to learn a compact residual update around the pre-trained transport. This observation suggests that the task-specific transformation for restoration is concentrated in a compact residual subspace, rather than being distributed across the full image-generation process.… view at source ↗
Figure 4
Figure 4. Figure 4: Geometric Interpretation of Residual Rectified Flow (RRF). (a) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of our Training Pipeline, which consists of a generation module and a regularization module. In the generation module, the restored-image [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison with existing diffusion-based restoration methods. The left side shows the result of ScaleResfusion, while the right side compares [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pairwise Arena user study results. Left: preference distribution when [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: More visual comparison with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the right side [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual comparison on WebPhoto-Test [Wang et al [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: No-reference metric comparison on examples without ground truth. Although TSD-SR still obtains better MUSIQ/NIQE scores than ScaleResfusion, its [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Examples where no-reference metrics do not fully reflect visual [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Intermediate restoration results from the same multi-step model. As the number of function evaluations (NFE) increases during sampling, ScaleResfusion [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual ablation on the number of function evaluations. Increasing [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visual ablation on the residual ratio 𝛾. The residual ratio controls the balance between deterministic LQ guidance and stochastic restoration, leading to different fidelity–realism trade-offs. on FLUX2-4B. On LSDIR-Val, it reduces LPIPS from 0.29 to 0.25 and FID from 79.37 to 63.84 on SD3, and reduces LPIPS from 0.39 to 0.23 and FID from 61.09 to 41.69 on FLUX2-4B. These results directly support the motiv… view at source ↗
Figure 17
Figure 17. Figure 17: Visual ablation on GAN fine-tuning. GAN fine-tuning enhances perceptual sharpness and local details, while occasionally reducing content consistency [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗
Figure 19
Figure 19. Figure 19: Visual ablation on DAPE. Replacing DAPE with a fixed prompt [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visual ablation on DMD. DMD reduces the training-inference mis [PITH_FULL_IMAGE:figures/full_fig_p021_18.png] view at source ↗
Figure 20
Figure 20. Figure 20: Visual ablation on LQ feature conditioning. Removing LQ conditioning allows the model to synthesize more diverse details, but it also weakens input [PITH_FULL_IMAGE:figures/full_fig_p022_20.png] view at source ↗
Figure 22
Figure 22. Figure 22: Visual ablation on LoRA rank. Increasing the LoRA rank gives the [PITH_FULL_IMAGE:figures/full_fig_p022_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: More visual comparison on DRealSR with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the [PITH_FULL_IMAGE:figures/full_fig_p026_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: More visual comparison on RealSR with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the right [PITH_FULL_IMAGE:figures/full_fig_p027_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: More visual comparison on LSDIR-Val with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the [PITH_FULL_IMAGE:figures/full_fig_p028_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: More visual comparison on LSDIR-Val with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the [PITH_FULL_IMAGE:figures/full_fig_p029_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: More visual comparison on LSDIR-Val with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the [PITH_FULL_IMAGE:figures/full_fig_p030_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: More visual comparison on LSDIR-Val with existing diffusion-based restoration methods. The left side shows the results of ScaleResfusion, while the [PITH_FULL_IMAGE:figures/full_fig_p031_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: More visual comparison on DIV2K-Val with existing diffusion-based restoration methods. The left side shows the HQ reference images, while the right [PITH_FULL_IMAGE:figures/full_fig_p032_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: More visual comparison on DIV2K-Val with existing diffusion-based restoration methods. The left side shows the HQ reference images, while the right [PITH_FULL_IMAGE:figures/full_fig_p033_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: More visual comparison on WebPhoto-Test [Wang et al [PITH_FULL_IMAGE:figures/full_fig_p034_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: More visual comparison on WebPhoto-Test [Wang et al [PITH_FULL_IMAGE:figures/full_fig_p035_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: More visual comparison on WebPhoto-Test [Wang et al [PITH_FULL_IMAGE:figures/full_fig_p036_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: More visual comparison on WebPhoto-Test [Wang et al [PITH_FULL_IMAGE:figures/full_fig_p037_34.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

91 extracted references · 16 linked inside Pith

  1. [1]

    European conference on computer vision , pages=

    Learning a deep convolutional network for image super-resolution , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  2. [2]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Designing a practical degradation model for deep blind image super-resolution , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  3. [3]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Second-order attention network for single image super-resolution , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  4. [4]

    Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    Enhanced deep residual networks for single image super-resolution , author=. Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

  5. [5]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Image super-resolution using very deep residual channel attention networks , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  6. [6]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Learning deep CNN denoiser prior for image restoration , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  7. [7]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Pre-trained image processing transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  8. [8]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Multi-stage progressive image restoration , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  9. [9]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Swinir: Image restoration using swin transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  10. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Restormer: Efficient transformer for high-resolution image restoration , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  11. [11]

    European conference on computer vision , pages=

    Simple baselines for image restoration , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  12. [12]

    European conference on computer vision , pages=

    Efficient long-range attention network for image super-resolution , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  13. [13]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Activating more pixels in image super-resolution transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [14]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Photo-realistic single image super-resolution using a generative adversarial network , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  15. [15]

    Proceedings of the European conference on computer vision (ECCV) workshops , pages=

    Esrgan: Enhanced super-resolution generative adversarial networks , author=. Proceedings of the European conference on computer vision (ECCV) workshops , pages=

  16. [16]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  17. [17]

    European conference on computer vision , pages=

    Perceptual losses for real-time style transfer and super-resolution , author=. European conference on computer vision , pages=. 2016 , organization=

    2016

  18. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Gan prior embedded network for blind face restoration in the wild , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [19]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Towards real-world blind face restoration with generative facial prior , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  20. [20]

    arXiv preprint arXiv:1409.1556 , year=

    Very deep convolutional networks for large-scale image recognition , author=. arXiv preprint arXiv:1409.1556 , year=

    Pith/arXiv arXiv

  21. [21]

    Advances in neural information processing systems , volume=

    Imagenet classification with deep convolutional neural networks , author=. Advances in neural information processing systems , volume=

  22. [22]

    Advances in neural information processing systems , volume=

    Generative adversarial nets , author=. Advances in neural information processing systems , volume=

  23. [23]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Details or artifacts: A locally discriminative learning approach to realistic image super-resolution , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  24. [24]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  25. [25]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

    Pith/arXiv arXiv 2010

  26. [26]

    Advances in neural information processing systems , volume=

    Generative modeling by estimating gradients of the data distribution , author=. Advances in neural information processing systems , volume=

  27. [27]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    Pith/arXiv arXiv 2011

  28. [28]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  30. [30]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Image super-resolution via iterative refinement , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

    2022

  31. [31]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2023 , publisher=

    2023

  32. [32]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deblurring via stochastic refinement , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  33. [33]

    Pattern Recognition , volume=

    LLDiffusion: Learning degradation representations in diffusion models for low-light image enhancement , author=. Pattern Recognition , volume=. 2025 , publisher=

    2025

  34. [34]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Shadowdiffusion: When degradation prior meets diffusion model for shadow removal , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  35. [35]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Diffusion models in low-level vision: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  36. [36]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  37. [37]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  38. [38]

    arXiv preprint arXiv:2511.22699 , year=

    Z-image: An efficient image generation foundation model with single-stream diffusion transformer , author=. arXiv preprint arXiv:2511.22699 , year=

    Pith/arXiv arXiv

  39. [39]

    European conference on computer vision , pages=

    Diffbir: Toward blind image restoration with generative diffusion prior , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  40. [40]

    International Journal of Computer Vision , volume=

    Exploiting diffusion prior for real-world image super-resolution , author=. International Journal of Computer Vision , volume=. 2024 , publisher=

    2024

  41. [41]

    European conference on computer vision , pages=

    Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  42. [42]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Seesr: Towards semantics-aware real-world image super-resolution , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  43. [43]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  44. [44]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    One-step effective diffusion network for real-world image super-resolution , author=. Advances in Neural Information Processing Systems , volume=

  46. [46]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  47. [47]

    arXiv preprint arXiv:2502.01993 , year=

    One diffusion step to real-world super-resolution via flow trajectory distillation , author=. arXiv preprint arXiv:2502.01993 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2303.11435 , year=

    Inversion by direct iteration: An alternative to denoising diffusion for image restoration , author=. arXiv preprint arXiv:2303.11435 , year=

    Pith/arXiv arXiv

  49. [49]

    Advances in neural information processing systems , volume=

    Denoising diffusion restoration models , author=. Advances in neural information processing systems , volume=

  50. [50]

    arXiv preprint arXiv:2301.11699 , year=

    Image restoration with mean-reverting stochastic differential equations , author=. arXiv preprint arXiv:2301.11699 , year=

    Pith/arXiv arXiv

  51. [51]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Refusion: Enabling large-size realistic image restoration with latent-space diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  52. [52]

    Liu, Guan-Horng and Vahdat, Arash and Huang, De-An and Theodorou, Evangelos A and Nie, Weili and Anandkumar, Anima , journal=. I

  53. [53]

    Advances in neural information processing systems , volume=

    Resshift: Efficient diffusion model for image super-resolution by residual shifting , author=. Advances in neural information processing systems , volume=

  54. [54]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Residual denoising diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  55. [55]

    Advances in Neural Information Processing Systems , volume=

    Resfusion: Denoising diffusion probabilistic models for image restoration based on prior residual noise , author=. Advances in Neural Information Processing Systems , volume=

  56. [56]

    arXiv preprint arXiv:2510.23116 , year=

    Residual Diffusion Bridge Model for Image Restoration , author=. arXiv preprint arXiv:2510.23116 , year=

    Pith/arXiv arXiv

  57. [57]

    2025 , howpublished=

    Black Forest Labs , title=. 2025 , howpublished=

    2025

  58. [58]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  59. [59]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    One-step diffusion with distribution matching distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  60. [60]

    Advances in neural information processing systems , volume=

    Improved distribution matching distillation for fast image synthesis , author=. Advances in neural information processing systems , volume=

  61. [61]

    arXiv preprint arXiv:2508.10779 , year=

    Ultra-High-Definition Reference-Based Landmark Image Super-Resolution with Generative Diffusion Prior , author=. arXiv preprint arXiv:2508.10779 , year=

    Pith/arXiv arXiv

  62. [62]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Animate anyone: Consistent and controllable image-to-video synthesis for character animation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  63. [63]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  64. [64]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Maniqa: Multi-dimension attention network for no-reference image quality assessment , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  65. [65]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Musiq: Multi-scale image quality transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  66. [66]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Exploring clip for assessing the look and feel of images , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  67. [67]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Lsdir: A large scale dataset for image restoration , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  68. [68]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    A style-based generator architecture for generative adversarial networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  69. [69]

    Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    Ntire 2017 challenge on single image super-resolution: Dataset and study , author=. Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    2017

  70. [70]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Toward real-world single image super-resolution: A new benchmark and a new model , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  71. [71]

    Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VIII 16 , pages=

    Component divide-and-conquer for real-world image super-resolution , author=. Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VIII 16 , pages=. 2020 , organization=

    2020

  72. [72]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  73. [73]

    arXiv preprint arXiv:2404.01717 , year=

    AddSR: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation , author=. arXiv preprint arXiv:2404.01717 , year=

    Pith/arXiv arXiv

  74. [74]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The perception-distortion tradeoff , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  75. [75]

    European Conference on Computer Vision , pages=

    Perception-distortion balanced ADMM optimization for single-image super-resolution , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  76. [76]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  77. [77]

    arXiv preprint arXiv:2401.00877 , year=

    Improving the Stability of Diffusion Models for Content Consistent Super-Resolution , author=. arXiv preprint arXiv:2401.00877 , year=

    Pith/arXiv arXiv

  78. [78]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Image quality assessment: Unifying structure and texture similarity , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2020 , publisher=

    2020

  79. [79]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  80. [80]

    IEEE Transactions on Image Processing , volume=

    A feature-enriched completely blind image quality evaluator , author=. IEEE Transactions on Image Processing , volume=. 2015 , publisher=

    2015

Showing first 80 references.