Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

DiffusionReward: Enhancing Blind Face Restoration through Reward Feedback Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Blind face restoration can be improved by reward feedback fine-tuning of diffusion restorers, yielding better perceptual quality, richer detail, and stronger identity consistency than prior state-of-the-art methods.

desk verdict A solid first application of reward feedback learning to blind face restoration, with real gains on external metrics, but the FaceReward evaluation column is circular and should not support the SOTA claim. read the letter →

arxiv 2505.17910 v1 pith:LM6PQ2UI submitted 2025-05-23 cs.CV cs.AI

Bin Wu , Wei Wang , Yahui Liu , Zixiang Li , Yao Zhao This is my paper · ORCID
classification cs.CVcs.AI
keywords blindfacerestorationrewardfeedbacklearningdiffusionmodelsmodelhackingperceptualqualityidentityconsistencyfine-tuning
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

This paper argues that blind face restoration, which must reconstruct a plausible high-quality face from a heavily degraded input, can be improved by fine-tuning the diffusion restoration model itself with reward feedback. The authors propose DiffusionReward, a training loop in which a Face Reward Model scores the restored image against a text description of the original face, and the resulting gradient is pushed through the last denoising step of an off-the-shelf diffusion restorer. Two auxiliary terms keep the fine-tune honest: a structural consistency loss preserves facial fidelity and identity, and a weight regularizer keeps the model close to its pre-trained state to retain generative diversity. During training the reward model is periodically re-trained to prefer the ground-truth face over the current restoration, which the authors argue prevents reward hacking. The reported result is that both a single-step restorer and a multi-step restorer, upgraded this way, beat the prior state of the art on synthetic and real-world face benchmarks.

What carries the argument

The load-bearing mechanism is the parameterized denoising process: a latent diffusion restorer $g_\theta$ is viewed as a generator, and gradients from the reward loss are backpropagated through the final denoising step, with the truncation window set to one step. The guiding signal combines the Face Reward Model score $R(\hat{I}_{HQ},T)$, a structural consistency loss made of LPIPS plus a DWT low-frequency term, and a KL weight regularization toward the initial parameters. The anti-reward-hacking device is the dynamic Face Reward Model update, which after every $n$ generator iterations re-trains the scorer to prefer the ground-truth image over the current restoration using the softmax preference probability in Eqs. (7)-(8).

What would settle it

Run the same DiffusionReward training with the Face Reward Model frozen and every other component identical. If the frozen-reward version matches the dynamic-update version on perceptual quality, identity metrics, and artifact-free appearance on an unseen wild test set, then the dynamic update is not the load-bearing anti-reward-hacking mechanism. A second check: monitor the FRM score of restored images relative to real faces during training; if the score of restored images rises above the score of real faces, the claim that the dynamic update keeps outputs on the real-face manifold fails.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the denoising trajectory of a diffusion-based face restoration model can be treated as a differentiable generator and steered by a face-specific reward model, and that this steering materially improves restoration quality. Trained on pairwise human preferences distilled into a CLIP-style scorer, the Face Reward Model supplies the main gradient; the LPIPS and DWT low-frequency losses anchor the output to the original identity; and the KL weight regularizer prevents the fine-tune from destroying the base model's generative prior. The dynamic update of the reward model, declaring the current restoration the loser against its ground-truth image and re-optimizing the scorer accordingly, is presented as the mechanism that keeps the restored faces on the real-face manifold instead of drifting into high-scoring but unnatural artifacts. With DiffBIR and OSEDiff as base models, the paper reports consistent gains in no-reference quality, aesthetic score, identity distance, and its own FaceReward metric, and a sharp drop in the reward-hacking artifacts shown in its ablation.

Load-bearing premise

The dynamic reward-model update assumes that labeling every current restoration as a loser against its own ground-truth image teaches the reward model to stay on the real-face manifold; if that update only makes the reward model reject the current generator's outputs without guiding it toward better behavior, the paper's main anti-reward-hacking claim collapses.

Editorial extensions

If this is right

  • State-of-the-art diffusion restorers can be upgraded by a post-training reward-feedback pass without changing their inference-time architecture or sampling cost.
  • Reward feedback learning transfers from text-to-image generation to restoration tasks, so any restoration domain with a reliable preference signal becomes a candidate for the same treatment.
  • Dynamic reward updates plus weight regularization give a working countermeasure to reward hacking as it appears in restoration, suppressing stylized painterly outputs and repetitive blemish artifacts.
  • The framework applies to both single-step and multi-step diffusion restoration models, suggesting the mechanism is tied to the denoising parameterization rather than to one specific base model.

Reading between the lines

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

  • If the dynamic Face Reward Model update genuinely prevents reward hacking, the same two-loop scheme could be applied to other conditional diffusion restoration tasks, such as super-resolution of text, hands, or medical images, wherever a domain-specific preference scorer can be trained.
  • The hybrid annotation strategy, human labels plus an SVM trained on six image-quality metrics, means the reward model's ceiling is partly set by those six metrics; a preference scorer trained on richer or more diverse human judgments could plausibly yield larger gains than the ones reported.
  • The weight-regularization result suggests that reward fine-tuning is most effective as a small, local adjustment to a strong pre-trained restorer, rather than as a full retraining recipe.
  • Because the same base model is used before and after fine-tuning, the reported comparisons may understate the practical value: the method converts an existing deployment into a better restorer at no additional inference cost.
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

4 major / 4 minor

Summary. The paper proposes DiffusionReward, a reward-feedback learning framework for blind face restoration. It fine-tunes diffusion-based restorers (OSEDiff and DiffBIR) with a Face Reward Model (FRM) that scores restored images, combined with a structural consistency loss (LPIPS and DWT low-frequency) and a weight regularization term. The FRM is trained on a preference dataset with human and SVM-generated labels, and it is dynamically updated during training so that ground-truth images are always preferred over current restorations, with the stated goal of avoiding reward hacking. Experiments on synthetic and wild face datasets report improvements over the base models and prior SOTA across most of the 11 Table-1 metrics, including a new FaceReward score.

Significance. If the results hold, the paper is one of the first to transfer reward-feedback learning to blind face restoration, and the dynamic FRM update is a potentially instructive way to combat reward hacking. The paper is also transparent in releasing code, data, and models, and its ablation study decomposes the contribution of each loss term. However, the as-stated conclusion that the method 'outperforms existing state-of-the-art methods' is currently overreaching: the headline FaceReward metric is produced by the very model used as the training reward, several quantitative claims are internally inconsistent, and one loss expression is not well defined. These issues are fixable but require a revised evaluation and a careful rewriting of the affected passages.

major comments (4)
  1. [Sec. 4.2 (Table 1), Sec. 3.3 (Eq. 2), Sec. 3.3 (Eqs. 7-8)] The FaceReward column in Table 1 is not an independent evaluation metric. FaceReward is the output of the FRM, and the same FRM provides the training reward in Eq. (2) (L_reward = -R(restored, T)) and is periodically updated to prefer the ground truth over the current generator output (Eqs. 7-8). Therefore, high FaceReward values for the proposed method are expected by construction, and the Table-1 claim that the method achieves top rankings on FaceReward cannot be used as evidence of superiority over prior methods. The external validation of 87.78% vs. 63.05% on 360 pairs is a useful sanity check but is too small to calibrate the metric on the test sets and does not break the circularity. I recommend removing FaceReward from the comparison tables, or supplementing it with a human-preference study on the actual test images and reporting the SVM auto-labeler's accuracy.
  2. [Sec. 3.3, Eq. (5)] The weight regularization term is written as D_KL(theta || theta_base), but KL divergence is defined between probability distributions, not between parameter vectors, and no distribution over theta is specified. As written, Eq. (5) is not a computable loss. If the intent is a Gaussian KL with fixed covariance, it reduces to a scaled L2 distance and should be stated as such; if a distributional interpretation is intended, the distributions need to be defined. This matters because L_reg is one of the three components of the final objective in Eq. (6).
  3. [Sec. 4.3, Table 3] The text states that 'our FRM obviously outperfoms HPS v2' on the basis of Table 3, but the table is mixed: FRM is better on MANIQA (0.6535 vs. 0.6630) and FID (42.59 vs. 48.94), but worse on MUSIQ (69.78 vs. 74.82). The claim of obvious superiority is not supported by the table as presented. Please specify the primary metric, report statistical significance, or qualify the conclusion.
  4. [Sec. 3.3 and Sec. 4.3 (Table 4, Fig. 6(c))] The anti-reward-hacking effect of the dynamic FRM update is central to the method's novelty, but the current evidence is limited to one qualitative comparison and one row in Table 4. The assumption that always preferring the ground-truth image over the current restoration keeps the reward model on the real-face manifold is not demonstrated; an alternative possibility is that the FRM simply learns to reject the current generator's outputs, which could cause optimization to oscillate or drift. I would ask for direct quantitative evidence, such as FID or precision/recall of the restored images during training with and without the dynamic update, or a plot of FRM accuracy on held-out human preference pairs across updates.
minor comments (4)
  1. [Sec. 4.1] The sentence 'we used five reference metrics' is followed by a list of six (SSIM, PSNR, LPIPS, CLIP Score, Deg., and LMD), and the sentence 'four non-reference metrics' is followed by a list of three (MUSIQ, MANIQA, and FID). Please correct the counts.
  2. [Sec. 2 and Sec. 4.1] The GFPGAN method is cited as [1] in the comparison list, but reference [1] is GLEAN; the GFP-GAN citation in Section 2 is [40]. Please fix the citation.
  3. [Sec. 4.2] The sentence that OSEDiff (+ours) and DiffBIR (+ours) 'achieve top rankings in the majority of metrics, such as Deg., LMD, Aesthetic, and FaceReward' is ambiguous: for Deg. and LMD, only DiffBIR (+ours) is the top method, while OSEDiff (+ours) is not. Please state which variant is top for each metric.
  4. [Sec. 4.3] The sentence 'we manually annotate 360 pairs' should clarify that these 360 pairs are a separate evaluation set from the 3,600 training pairs used to train the SVM, since the preceding paragraph describes 3,600 manually annotated pairs.

Circularity Check

2 steps flagged · score 6.0 of 10

Two reported metrics (FaceReward and LPIPS) are exactly the training objectives, so part of the claimed SOTA improvement is forced by construction; other metrics remain independent.

  1. fitted input called prediction [Sec. 3.3 Eq. (2); Sec. 4.1 'Evaluation Metrics'; Sec. 4.2 'Evaluation on Synthetic Dataset'; Table 1]
    "Lreward = −R( ˆIHQ,T). ... In addition, we used our pretrained FRM to score the restored face images, denoting as FaceReward. ... Comparing to state-of-the-art (SOTA) methods, the OSEDiff (+ours) and DiffBIR (+ours) achieve top rankings in the majority of metrics, such as Deg., LMD, Aesthetic, and FaceReward"

    The FaceReward column is computed by the same FRM R that appears in the training loss L_reward = -R. Minimizing L_reward is exactly maximizing FaceReward, so the tuned models' higher FaceReward scores in Table 1 are expected by construction rather than independent evidence. The dynamic update (Eqs. 7-8) re-trains R during optimization, making the metric a moving target aligned with the generator's objective. Although FRM is validated against 360 human pairs (87.78% vs 63.05%), that validates the model's quality assessment; it does not make the FaceReward column an independent evaluation of the method.

  2. self definitional [Sec. 3.3 Eq. (3) and Eq. (6); Sec. 4.1 'Evaluation Metrics'; Table 1]
    "LLPIPS = LPIPS( ˆIHQ, I HQ). ... Ltotal = λ reward Lreward + λ LPIPS LLPIPS + λ DWT LDWT + λ reg Lreg. ... On the Celeba-Test dataset, we used five reference metrics: SSIM [43], PSNR, LPIPS [53], CLIP Score[8], Deg. [29], and LMD [7]"

    LPIPS appears both as a training loss term (with λ_LPIPS weight) and as an evaluation metric in Table 1. Since the restoration network is explicitly optimized to minimize this exact LPIPS function, the reported LPIPS improvement over the base model is a direct consequence of the loss, not an independent measurement. The paper presents the LPIPS column alongside other metrics as evidence of SOTA, so this is a second instance where an evaluation number reproduces the optimization objective.

full rationale

The core ReFL derivation is not circular: the FRM is trained on externally annotated preference data, the generator is optimized with additional structural and regularization losses, and several reported metrics (MUSIQ, MANIQA, Aesthetic, LMD, FID) are independent of the training losses. However, two of the headline evaluation columns are exact or near-exact reproductions of the optimization objectives. FaceReward is the paper's own FRM, which is the reward signal in L_reward, so reporting it as evidence of superiority is circular by construction; the dynamic update makes this worse by re-training the evaluator during optimization. LPIPS is also explicitly a loss term in Eq. (6) and simultaneously a reported metric, so its improvement is partially forced. These circularities are load-bearing because the conclusion explicitly cites FaceReward (and the results section cites LPIPS) among the metrics where the method achieves top rankings. The paper's external human-preference check of the FRM (87.78% vs 63.05%) is genuine evidence for the reward model's validity, but it does not convert the FaceReward column into an independent evaluation of the restoration method. Overall, the central claim retains independent support from other metrics, but the presence of two evaluation numbers that reduce by construction warrants a partial circularity score of 6.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The framework depends on manually tuned loss weights, a dynamic update interval, and several domain assumptions about reward-model reliability. The FRM is a new trained component rather than an independently verified physical entity; its independent evidence is a small human-preference agreement set, but its use as an evaluation metric partially undermines that independence.

free parameters (6)
  • lambda_reward = 0.005 (both bases)
    Weight of the reward loss in Eq. (6); chosen by hand.
  • lambda_LPIPS = 0.02 (OSEDiff), 0.01 (DiffBIR)
    Hand-selected weight of the LPIPS structural loss.
  • lambda_DWT = 0.01
    Hand-selected weight of the DWT low-frequency loss.
  • lambda_reg = 1 (OSEDiff), 1e-4 (DiffBIR)
    Hand-selected strength of the weight regularization term.
  • FRM update interval n = 10 iterations
    Determines how often the reward model is updated; set by hand to balance training stability.
  • Truncated backpropagation steps N = 1
    Limits gradient flow to the last denoising step for efficiency; a design choice.
assumptions (4)
  • domain assumption Diffusion denoising steps can be treated as a differentiable parameterized generator g_theta, so reward gradients can be backpropagated through the last denoising step.
    Central to ReFL; standard in DRaFT and AlignProp but not proven for these restoration models.
  • domain assumption The SVM-auto-labeled preference data (from 3,600 human pairs) is accurate enough to train a reliable FRM.
    The FRM's training labels for 117,540 pairs come mostly from the SVM predictor, whose accuracy is not reported.
  • ad hoc to paper The dynamic FRM update, which always prefers GT over the current restoration, keeps R on the real-face manifold.
    Introduced to prevent reward hacking; no formal guarantee is given.
  • ad hoc to paper D_KL(theta || theta_base) in Eq. (5) is a valid, computable penalty on parameter-space distance.
    No distributional interpretation of model parameters is given; the notation appears to denote an L2-style penalty.
invented entities (1)
  • Face Reward Model (FRM) independent evidence
    purpose: Provides reward feedback to fine-tune the restoration model and is also used as the FaceReward evaluation metric.
    Trained on 117,540 preference pairs and reported to reach 87.78% agreement with human labels on 360 pairs; however, its use as an evaluation metric creates a circular evaluation loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffusionReward: Enhancing Blind Face Restoration through Reward Feedback Learning." pith.science (2026). https://pith.science/paper/LM6PQ2UI

@misc{pith2026250517910,
  author       = {Pith},
  title        = {Pith review of: DiffusionReward: Enhancing Blind Face Restoration through Reward Feedback Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LM6PQ2UI}},
  note         = {Machine review of arXiv:2505.17910}
}
read the original abstract

Reward Feedback Learning (ReFL) has recently shown great potential in aligning model outputs with human preferences across various generative tasks. In this work, we introduce a ReFL framework, named DiffusionReward, to the Blind Face Restoration task for the first time. DiffusionReward effectively overcomes the limitations of diffusion-based methods, which often fail to generate realistic facial details and exhibit poor identity consistency. The core of our framework is the Face Reward Model (FRM), which is trained using carefully annotated data. It provides feedback signals that play a pivotal role in steering the optimization process of the restoration network. In particular, our ReFL framework incorporates a gradient flow into the denoising process of off-the-shelf face restoration methods to guide the update of model parameters. The guiding gradient is collaboratively determined by three aspects: (i) the FRM to ensure the perceptual quality of the restored faces; (ii) a regularization term that functions as a safeguard to preserve generative diversity; and (iii) a structural consistency constraint to maintain facial fidelity. Furthermore, the FRM undergoes dynamic optimization throughout the process. It not only ensures that the restoration network stays precisely aligned with the real face manifold, but also effectively prevents reward hacking. Experiments on synthetic and wild datasets demonstrate that our method outperforms state-of-the-art methods, significantly improving identity consistency and facial details. The source codes, data, and models are available at: https://github.com/01NeuralNinja/DiffusionReward.

Figures

Figures reproduced from arXiv: 2505.17910 by the authors.

Figure 1
Figure 1. An example of issues with diffusion-based face restoration methods. After enhancement [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Training framework of the Face Reward Model. We first train a SVM [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Our ReFL training framework. (Left) We introduce multiple constraints to optimize the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the CelebA-Test. (Zoom in for details) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison between the base model and the our methods on real-world faces. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study visualizations [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Text description example Manual Annotation of Preference Labels. To acquire reliable human preference data, we organized a team of three annotators to manually label image pairs. In total, the annotators provided preference selections for 3,600 image pairs, derived fro…
Figure 8
Figure 8. Figure 8: The brim and eyes of (b) have artifacts, so (a) is a better face image. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: (a) has a more realistic skin texture, (b) has skin that is too smooth and unrealistic, so (a) is [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: The glasses in the text description do not exist in face (b), so face (a) is a better face image. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: More qualitative comparison on the CelebA-Test. (Zoom in for details) [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: More qualitative comparison on the real-world faces. (Zoom in for details) [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Illustration of Reward Hacking. (Left) Examples of facial restoration exhibiting reward [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. GeoMAR: Unleashing Geometrically Aligned Features for Masked Autoregressive Blind Face Restoration

    cs.CV 2026-08 conditional novelty 5.0 of 10

    GeoMAR combines dual-input VLM text priors and MaskGIT-style iterative code prediction for blind face restoration, posting modest perceptual gains on real-world face benchmarks.

  2. Bridging Information Asymmetry: A Hierarchical Framework for Deterministic Blind Face Restoration

    cs.CV 2026-01 conditional novelty 5.0 of 10

    Pref-Restore combines AR semantic tokens, a diffusion generator, and DiffusionNFT-style RL to make blind face restoration more consistent, but its deterministic-identity claim is weakened by self-referential rewards a...

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages · cited by 2 Pith papers

  1. [1]

    Glean: Generative latent bank for large-factor image super-resolution

    Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4, 7

    work page 2021

  2. [2]

    Progressive semantic-aware style transformation for blind face restoration

    Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3

    work page 2021

  3. [3]

    Towards real-world blind face restoration with generative diffusion prior.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    Xiaoxu Chen, Jingfan Tan, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaochun Cao. Towards real-world blind face restoration with generative diffusion prior.IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1

    work page 2024

  4. [4]

    Fsrnet: End-to-end learning face super-resolution with facial priors

    Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3

    work page 2018

  5. [5]

    Directly fine-tuning diffusion models on differentiable rewards.arXiv preprint arXiv:2309.17400, 2023

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards.arXiv preprint arXiv:2309.17400, 2023. 2, 4, 7

    arXiv 2023

  6. [6]

    Support-vector networks.Machine learning, 20:273–297,

    Corinna Cortes and Vladimir Vapnik. Support-vector networks.Machine learning, 20:273–297,

    work page

  7. [7]

    Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder

    Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In European Conference on Computer Vision (ECCV), 2022. 3, 7

    work page 2022

  8. [8]

    Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718, 2021

    Jack Hessel, Ari Holtzman, Maxwell Forbes, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718, 2021. 7

    arXiv 2021

Show all 56 references
  1. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Information Processing Systems (NeurIPS), 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Information Processing Systems (NeurIPS), 2017. 7

    2017

  2. [10]

    Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 3

    2020

  3. [11]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 6

    2022

  4. [12]

    Characterizing photorealism and artifacts in diffusion model-generated images

    Negar Kamali, Karyn Nakamura, Aakriti Kumar, Angelos Chatzimparmpas, Jessica Hullman, and Matthew Groh. Characterizing photorealism and artifacts in diffusion model-generated images. InProceedings of the CHI Conference on Human Factors in Computing Systems, 2025. 2

    2025

  5. [13]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks.IEEE Transactions on Pattern Analysis Machine Intelligence, 43(12):4217–4228, dec 2021. 7, 14

    2021

  6. [14]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3

    2019

  7. [15]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3

    2020

  8. [16]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5, 7

    2021

  9. [17]

    Progressive face super- resolution via attention to facial landmark.arXiv preprint arXiv:1908.08239, 2019

    Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super- resolution via attention to facial landmark.arXiv preprint arXiv:1908.08239, 2019. 3 10

    1908 arXiv

  10. [18]

    Auto-encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 5

    2013

  11. [19]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation.Advances in Neural Information Processing Systems (NeurIPS), 2023

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation.Advances in Neural Information Processing Systems (NeurIPS), 2023. 3

    2023

  12. [20]

    Aesthetic predictor: A linear estimator on top of clip to predict the aesthetic quality of pictures

    LAION-AI. Aesthetic predictor: A linear estimator on top of clip to predict the aesthetic quality of pictures. https://github.com/LAION-AI/aesthetic-predictor , 2022. Accessed: 2025-05-13. 7

    2022

  13. [21]

    Photo-realistic single image super-resolution using a generative adversarial network

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. InProceedings of the IEEE Co...

    2017

  14. [22]

    Blind face restoration via deep multi-scale component dictionaries

    Xiaoming Li, Chaofeng Chen, Shangchen Zhou, Xianhui Lin, Wangmeng Zuo, and Lei Zhang. Blind face restoration via deep multi-scale component dictionaries. InEuropean Conference on Computer Vision (ECCV), 2020. 1

    2020

  15. [23]

    Learning warped guidance for blind face restoration

    Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. InEuropean conference on computer vision (ECCV),

  16. [24]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3

    2021

  17. [25]

    Rich human feedback for text-to-image generation

    Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2024

  18. [26]

    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. InEuropean Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 4, 6, 7

    2024

  19. [27]

    Visual instruction tuning.Advances in Neural Information Processing Systems (NeurIPS), 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in Neural Information Processing Systems (NeurIPS), 2023. 4, 14

    2023

  20. [28]

    Rewards are enough for fast photo-realistic text-to-image generation.arXiv preprint arXiv:2503.13070, 2025

    Yihong Luo, Tianyang Hu, Weijian Luo, Kenji Kawaguchi, and Jing Tang. Rewards are enough for fast photo-realistic text-to-image generation.arXiv preprint arXiv:2503.13070, 2025. 4

    2025 arXiv

  21. [29]

    Pulse: Self- supervised photo upsampling via latent space exploration of generative models

    Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self- supervised photo upsampling via latent space exploration of generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7

    2020

  22. [30]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer.IEEE Signal Processing Letters, 20(3):209–212, 2012. 5

    2012

  23. [31]

    Aligning text-to- image diffusion models with reward backpropagation (2023).arXiv preprint arXiv:2310.03739

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to- image diffusion models with reward backpropagation (2023).arXiv preprint arXiv:2310.03739. 4

    2023 arXiv

  24. [32]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning (ICM...

    2021

  25. [33]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 5

    2022

  26. [34]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Inter- vention (MICCAI), 2015. 6 11

    2015

  27. [35]

    Deep semantic face deblurring

    Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3

    2018

  28. [36]

    Defining and characterizing reward gaming.Advances in Neural Information Processing Systems, 35:9460– 9471, 2022

    Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming.Advances in Neural Information Processing Systems, 35:9460– 9471, 2022. 7

    2022

  29. [37]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InInternational Conference on Machine Learning (ICML), 2015. 3

    2015

  30. [38]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 17

    2010 arXiv

  31. [39]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI),

  32. [40]

    Towards real-world blind face restoration with generative facial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 7

    2021

  33. [41]

    Esrgan: Enhanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops (ECCVW), 2018. 3

    2018

  34. [42]

    Dr2: Diffusion-based robust degradation remover for blind face restoration

    Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

  35. [43]

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

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing (TIP), 13(4):600–612, 2004. 5, 7

    2004

  36. [44]

    Restoreformer: High-quality blind face restoration from undegraded key-value pairs

    Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. 2022. 7

    2022

  37. [45]

    One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3, 6, 7

    2024

  38. [46]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023. 5, 7, 17

    2023 arXiv

  39. [47]

    Lpff: A portrait dataset for face generators across large poses

    Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4, 14

    2023

  40. [48]

    Imagereward: learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 4

    2023

  41. [49]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)...

    2022

  42. [50]

    Face super- resolution guided by facial component heatmaps

    Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super- resolution guided by facial component heatmaps. InEuropean conference on computer vision (ECCV), 2018. 3 12

    2018

  43. [51]

    Difface: Blind face restoration with diffused error contraction.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024

    Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 1, 3, 7

    2024

  44. [52]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 6

    2023

  45. [53]

    The unreason- able effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreason- able effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5, 6, 7

    2018

  46. [54]

    Learning multi-dimensional human preference for text-to-image generation

    Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multi-dimensional human preference for text-to-image generation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  47. [55]

    Td-bfr: Truncated diffusion model for efficient blind face restoration.arXiv preprint arXiv:2503.20537, 2025

    Ziying Zhang, Xiang Gao, Zhixin Wang, Xiaoyun Zhang, et al. Td-bfr: Truncated diffusion model for efficient blind face restoration.arXiv preprint arXiv:2503.20537, 2025. 2

    2025 arXiv

  48. [56]

    loopholes

    Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer.Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 4, 7 13 A Implementation Details of Face Reward Model This section is used...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.