Pith. sign in

REVIEW 3 major objections 6 minor 32 references

ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A latent diffusion model conditioned on multiple high-quality reference images restores low-quality faces while preserving the real person's identity, reaching identity similarity of 0.676, 0.840, and 0.779 on three benchmarks.

desk verdict Useful machinery and a usable dataset, but the headline identity claim rides entirely on the same ArcFace embedding used to build and train on the test set. read the letter →

arxiv 2412.05043 v1 pith:JSJCNETL submitted 2024-12-06 cs.CV cs.LGeess.IV

Chi-Wei Hsiao , Yu-Lun Liu , Cheng-Kun Yang , Sheng-Po Kuo , Kevin Jou , Chia-Ping Chen This is my paper · ORCID
classification cs.CVcs.LGeess.IV
keywords reference-basedfacerestorationlatentdiffusionmodelidentitysimilarityCacheKVtimestep-scaledlossFFHQ-Refblind
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 ReF-LDM, a latent diffusion model that restores a low-quality face image by also taking several high-quality photos of the same person as reference. The central claim is that reference conditioning, implemented by a CacheKV mechanism that runs the denoising network over the reference images only once and reuses their attention keys and values at every denoising step, substantially improves facial identity fidelity. A timestep-scaled identity loss lets the model learn identity without sacrificing image quality. The authors also build FFHQ-Ref, a 20,405-image dataset with identity-disjoint splits, to train and evaluate reference-based face restoration models. If the claim holds, restored faces can look like the actual person rather than a plausible but different face, even under severe degradation.

What carries the argument

The central mechanism is CacheKV: the encoded latent of each reference image is passed through the same denoising U-net once, and the keys and values from each self-attention layer are cached; during each of the T denoising steps, these cached reference keys and values are concatenated with the main tokens so the self-attention layers can attend to reference appearance. The second load-bearing component is the timestep-scaled identity loss, which multiplies the cosine-distance identity loss by $\sqrt{\bar{\alpha}_t}$ so that identity supervision weakens at noisy timesteps and strengthens as the latent approaches a clean face. The FFHQ-Ref dataset, built by grouping FFHQ images with ArcFace cosine distance below 0.4 and splitting by connected components, supplies the training and evaluation pairs.

What would settle it

Take a random sample of FFHQ-Ref test pairs, have human judges decide whether each restoration shows the same person as the reference, and compare human agreement with the ArcFace-based IDS ranking; alternatively, recompute IDS with a face recognition model not used in training and check whether ReF-LDM's advantage over CodeFormer persists.

Watch

Extended reading notes

Core claim

The authors establish that LDM-based face restoration can be made reference-conditioned without the computational cost of passing reference images through the network at every timestep. By extracting keys and values from the denoising U-net once at timestep zero and caching them, then concatenating these cached tokens with the main self-attention keys and values during denoising, the model gains access to reference appearance while avoiding redundant network passes. They further show that scaling an ArcFace identity loss by $\sqrt{\bar{\alpha}_t}$, the diffusion noise schedule value, improves identity similarity without the quality collapse caused by naive identity loss. On the severe FFHQ-Ref benchmark, ReF-LDM reaches an identity similarity score of 0.676, compared with 0.185 for DMDNet and 0.323 for CodeFormer, and it also leads on moderate degradation and on the CelebA-Test-Ref subset.

Load-bearing premise

The identity-fidelity claim rests on automatic grouping of faces by ArcFace embeddings with a cosine-distance threshold of 0.4; if these groupings are wrong, the training references, the evaluation pairs, and the identity metric are all affected in the same direction.

Editorial extensions

If this is right

  • Restoration can use one to several reference images without landmark detection, so severely degraded inputs that break landmark-based reference methods no longer cause failure.
  • The caching design makes multi-reference conditioning practical: about one-fifth the inference time and 39 percent of the GPU memory of spatial concatenation with comparable identity gain.
  • The timestep-scaled identity loss demonstrates that diffusion identity supervision should be down-weighted at noisy timesteps, a principle transferable to other diffusion-based face generation tasks.
  • FFHQ-Ref gives the community an identity-disjoint, licensed benchmark with twice the images and six times the identities of the previous reference-restoration dataset.
  • Identity similarity can be improved substantially on the same image-quality metrics because the identity loss is scheduled to act mainly when the predicted face is already close to natural.

Reading between the lines

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

  • Beyond the paper, the cache-and-reuse KV mechanism could apply to any diffusion task with unaligned conditioning inputs, such as personalized text-to-image generation or video editing where references change pose or viewpoint.
  • Since the identity loss, the dataset grouping, and the evaluation metric all use the same ArcFace embedding space, the reported identity gains could partly reflect the model learning to satisfy ArcFace rather than human identity perception; a human-judged identity test or an evaluation with a different recognition model would clarify this.
  • The dataset construction suggests an inexpensive recipe for retrofitting large existing face datasets with identity groups: threshold a recognition embedding and split by connected components, which could transfer to other domains such as person re-identification or animal face datasets.
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

3 major / 6 minor

Summary. The paper proposes ReF-LDM, a latent diffusion model for reference-based face image restoration. The model conditions a denoising U-Net on a low-quality image and multiple high-quality reference images, using a CacheKV mechanism that extracts keys and values from reference latents in a single U-Net pass and reuses them at every denoising timestep. A timestep-scaled identity loss, proportional to sqrt(alpha_t), is used to avoid the quality degradation caused by applying an ArcFace identity loss at noisy timesteps. The authors also introduce FFHQ-Ref, an identity-grouped dataset derived from FFHQ, with automatic ArcFace-based identity labeling and a manually verified test split. Experiments report IDS, fLPIPS, LPIPS, FID, and NIQE, comparing ReF-LDM with CodeFormer, VQFR, DAEFR, LDM, and DMDNet, and claim significant improvements in identity similarity, especially on severely degraded inputs.

Significance. If the identity-preservation claim is valid, ReF-LDM makes a useful contribution to reference-based face restoration: the CacheKV mechanism is a plausible and efficient way to inject multiple unaligned reference images into a latent diffusion model, and the FFHQ-Ref dataset with identity-based splitting and manual test verification is a valuable resource for the community. The paper includes systematic ablations of the reference-conditioning mechanism, the identity loss scaling, and the number of references, and it reports inference-time and memory numbers that support the efficiency advantage of CacheKV over spatial concatenation. The central limitation is that the headline identity metric (ArcFace IDS) is also the objective used for dataset construction and training, so the current evidence does not independently establish that the restored faces preserve true identity rather than optimize a closed-loop ArcFace score.

major comments (3)
  1. [Sec. 4.1, Eq. (1), Sec. 5.1.2] The identity evaluation is circular. FFHQ-Ref identities and reference pairs are defined by an ArcFace cosine threshold r=0.4 (Sec. 4.1); the training loss minimizes ArcFace cosine distance (Eq. (1), Sec. 3.2.1); and the reported IDS metric is ArcFace cosine similarity (Sec. 5.1.2). Because the model is explicitly trained to maximize the same score on which it is evaluated, and because the test pairs themselves were selected to be close in that embedding, the large IDS gains in Table 7 may reflect optimization of the metric rather than genuine identity preservation. I would like to see IDS computed with a face recognition model that was not used in training or dataset construction (e.g., AdaFace or FaceNet), plus a small human study or manual verification of whether the restored faces are judged to be the same person as the reference.
  2. [Table 7, footnote] The DMDNet comparison is computed only on the images that DMDNet successfully processed: 214/857, 29/857, and 488/2,533 failures on the three benchmarks are excluded. This is especially problematic on FFHQ-Ref-Severe, where DMDNet fails on about a quarter of the test set; the remaining images are likely easier, so the aggregate IDS comparison is biased in ReF-LDM's favor. No confidence intervals or paired statistics are reported. The comparison should be redone on a common subset where all methods succeed, and the failure rate itself should be reported as a separate criterion; ideally the per-image paired IDS differences should be reported with confidence intervals.
  3. [Sec. 4.1 and Sec. 4.3] The FFHQ-Ref training and validation identity labels are produced automatically by ArcFace with a cosine distance threshold of r=0.4, and no manual verification is reported for those splits; only the test split is manually checked. If the automatic labeling is noisy, then the training targets, the reference pairing, and the test-set construction are all affected in the same direction. The paper should quantify the label noise, for example by manually checking a random sample of training pairs and reporting precision, or by comparing against a second face recognition model.
minor comments (6)
  1. [Sec. 2] The sentence 'detect facial components (i.e., eyes, nose, and mouse)' contains a typo: 'mouse' should be 'mouth'.
  2. [Fig. 2 caption] The word 'utlized' should be 'utilized'.
  3. [Table 5] The table title contains the typo 'numbrs'; it should be 'numbers'.
  4. [Sec. 5.3.2] The name 'CodeFomer' should be 'CodeFormer'.
  5. [Table 7] The FID column is present for the two FFHQ-Ref benchmarks but omitted for CelebA-Test-Ref; the paper should state why FID is not reported there, or add it for consistency.
  6. [Table 2] Spatial-concatenation achieves a higher IDS (0.69) than CacheKV (0.65) while CacheKV has better NIQE and much lower compute; the text should frame CacheKV as an efficiency-accuracy trade-off rather than implying it is uniformly superior to spatial concatenation.

Circularity Check

1 steps flagged · score 6.0 of 10

Identity-similarity claim is a closed ArcFace loop: FFHQ-Ref reference pairs, the timestep-scaled identity loss, and the IDS evaluation metric all use the same ArcFace cosine distance, so the reported identity gains are partly forced by construction.

  1. fitted input called prediction [Sec. 4.1 (reference construction), Sec. 3.2.1 Eq. (1) (identity loss), Sec. 5.1.2 (IDS metric)]
    "To determine whether two images belong to the same identity, we utilize the face recognition model ArcFace [3]. ... A distance less than a threshold r = 0.4 indicates that the images are valid references belonging to the same person. ... In our experiments, we use the ArcFace model [3] with cosine distance between the 1D embedding vectors as the identity loss. ... For evaluation metrics, we adopt the identity similarity (IDS) [5, 32], which is the cosine similarity calculated using the face recognition model ArcFace [3]."

    The paper's headline claim of 'significant improvement in face identity similarity' is measured solely by ArcFace IDS. The same ArcFace cosine distance defines (i) which images count as same-identity references in FFHQ-Ref, (ii) the training loss in Eq. (1), and (iii) the evaluation metric. The model is explicitly trained to minimize the ArcFace cosine distance on references selected by that same distance, and then reports high ArcFace cosine similarity as the predicted identity improvement. This is a statistically forced outcome of optimizing the evaluation metric itself, not an independent confirmation of identity fidelity. The manual test-set check and the additional perceptual metrics give partial independent support, but the central identity claim is not self-contained.

full rationale

The central identity-similarity result is partially circular: FFHQ-Ref identities and reference pairs are constructed with ArcFace (Sec. 4.1), the timestep-scaled identity loss minimizes ArcFace cosine distance (Eq. 1, Sec. 3.2.1), and the reported IDS metric is ArcFace cosine similarity (Sec. 5.1.2). A model trained to maximize this exact similarity on ArcFace-selected references will naturally score higher than baselines not trained with this objective, so the Table 7 IDS gains do not independently verify that the restored face preserves the true person's identity. However, the paper also provides qualitative comparisons, LPIPS/fLPIPS/FID/NIQE, and manual verification of test-set identity labels, which give partial independent support. There is no load-bearing self-citation or imported uniqueness theorem; the DMDNet failure footnote is an evaluation-fairness caveat rather than a circular step. Overall circularity score 6 reflects that the main identity claim is substantially defined and driven by the same ArcFace metric without being fully circular.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The model introduces no new physics or new entities. The identity-fidelity claim rests on ArcFace as a proxy for identity and on an automatic threshold for same-person grouping; the image-quality claim rests on the synthetic degradation model. These are domain assumptions rather than derivable facts.

free parameters (5)
  • identity threshold r = 0.4
    ArcFace cosine distance below which two images are treated as the same person (Sec. 4.1); hand-chosen, not validated against human labels, and it defines the training set, test set, and reference count.
  • identity loss weight lambda_time_ID = 0.1
    Loss scale in Eq. 3; chosen by hand (Sec. 5.1.1), ablations do not sweep this value.
  • CFG scale s = 1.5
    Classifier-free guidance scale toward references (Sec. G.1); hand-chosen, no ablation.
  • number of training references = 5
    Fixed to five references during training (Sec. 5.1.1); the model is not trained for the full 2-9 range, creating a train/inference mismatch when fewer or more references are used.
  • test reference minimum distance = 0.1
    Test pairs with ArcFace cosine distance below 0.1 are excluded in Sec. 4.3; post-hoc filtering that makes the test set easier for identity metrics.
assumptions (3)
  • domain assumption ArcFace cosine distance is a valid measure of facial identity similarity
    Used to construct the dataset (Sec. 4.1), as the training loss (Eq. 1), and as the evaluation metric (Sec. 5.1.2); the entire identity-fidelity claim depends on this proxy.
  • domain assumption The synthetic degradation model in Eq. 4 (blur, downsample, noise, JPEG) is representative of real low-quality face inputs
    All training and test LQ images are generated from HQ images with this model; results may not transfer to real degradations.
  • domain assumption The frozen VQGAN autoencoder preserves identity-relevant information in its latent space
    The denoising U-net operates in a learned latent space and both LQ and reference images are encoded with the frozen encoder E; if this space discards fine identity details, the restoration cannot recover them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration." pith.science (2026). https://pith.science/paper/JSJCNETL

@misc{pith2026241205043,
  author       = {Pith},
  title        = {Pith review of: ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSJCNETL}},
  note         = {Machine review of arXiv:2412.05043}
}
read the original abstract

While recent works on blind face image restoration have successfully produced impressive high-quality (HQ) images with abundant details from low-quality (LQ) input images, the generated content may not accurately reflect the real appearance of a person. To address this problem, incorporating well-shot personal images as additional reference inputs could be a promising strategy. Inspired by the recent success of the Latent Diffusion Model (LDM), we propose ReF-LDM, an adaptation of LDM designed to generate HQ face images conditioned on one LQ image and multiple HQ reference images. Our model integrates an effective and efficient mechanism, CacheKV, to leverage the reference images during the generation process. Additionally, we design a timestep-scaled identity loss, enabling our LDM-based model to focus on learning the discriminating features of human faces. Lastly, we construct FFHQ-Ref, a dataset consisting of 20,405 high-quality (HQ) face images with corresponding reference images, which can serve as both training and evaluation data for reference-based face restoration models.

Figures

Figures reproduced from arXiv: 2412.05043 by the authors.

Figure 1
Figure 1. Reference-based face image restoration. Given an input low-quality face image (a), a Latent Diffusion Model (LDM) can reconstruct a high-quality image (b); however, it may not be faithful to the individual’s facial identity. To address this problem, we propose ReF-LDM, which restores a high-quality image with faithful details (c) by utilizing additional reference images (d). With the above components, our ReF-LDM ou… view at source ↗
Figure 2
Figure 2. The proposed ReF-LDM pipeline. Our model accepts a low-quality image and multiple high-quality reference images as input and generates a high-quality image. The blue top panel alone represents a typical LDM [22] denoising process. For an LQ image xLQ, we concatenate its latent zLQ with zt along the channel axis to serve as the input for the denoising U-net. For the reference images {xref}, we design a CacheKV mechan… view at source ↗
Figure 3
Figure 3. Different mechanisms for incorporating reference images into the main denoising process. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Visual ablation results for the timestep [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison. From left to right: input LQ, ground truth, other methods, and our [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Examples of ReF-LDM using reference images from two different individuals. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Example images with mirror padding and watermark artifacts in the CelebRef-HQ dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Distribution of the number of available reference images per image in the FFHQ-Ref dataset [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Race distribution within FFHQ-Ref dataset. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Age and gender distribution within FFHQ-Ref dataset. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Examples where ReF-LDM generates background-region details that differ more from the [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: An example of (d) ReF-LDM demonstrating an illumination change, likely influenced by [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Some failure cases when the input LQ images are occluded. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Some failure cases of side faces when side-face images are absent in the references. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Some successful cases of side faces when side-face images are included in the references. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Data augmentation for reference images. G.3 Training details We trained the VQGAN for 200,000 iterations with batch size 32 on four A6000 GPUs for 7 days. We trained the LDM with only LQ condition for 500,000 iterations with batch size 40 on four A6000 GPUs for 7 days…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages

  1. [1]

    M. Cao, X. Wang, Z. Qi, Y . Shan, X. Qie, and Y . Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023

    work page 2023

  2. [2]

    C. Chen, X. Li, L. Yang, X. Lin, L. Zhang, and K.-Y . K. Wong. Progressive semantic-aware style transformation for blind face restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11896–11905, 2021

    work page 2021

  3. [3]

    J. Deng, J. Guo, X. Niannan, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019

    2019

  4. [4]

    Esser, R

    P. Esser, R. Rombach, and B. Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021

    work page 2021

  5. [5]

    Y . Gu, X. Wang, L. Xie, C. Dong, G. Li, Y . Shan, and M.-M. Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In European Conference on Computer Vision, pages 126–143. Springer, 2022

    work page 2022

  6. [6]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems , 30, 2017

    work page 2017

  7. [7]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    arXiv 2022

  8. [8]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Ad- vances in neural information processing systems , 33:6840–6851, 2020

    work page 2020

Show all 32 references
  1. [9]

    Huang, S

    R. Huang, S. Zhang, T. Li, and R. He. Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis. In Proceedings of the IEEE international conference on computer vision , pages 2439–2448, 2017

    2017

  2. [10]

    Karkkainen and J

    K. Karkkainen and J. Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages 1548–1558, 2021

    2021

  3. [11]

    Karras, T

    T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017

    2017 arXiv

  4. [12]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 4401–4410, 2019. 11

    2019

  5. [13]

    Y .-F. Lau, T. Zhang, Z. Rao, and Q. Chen. Ented: Enhanced neural texture extraction and distribution for reference-based blind face restoration. In Pro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5162–5171, 2024

    2024

  6. [14]

    X. Li, M. Liu, Y . Ye, W. Zuo, L. Lin, and R. Yang. Learning warped guidance for blind face restoration. In Proceedings of the European conference on computer vision (ECCV), pages 272–289, 2018

    2018

  7. [15]

    X. Li, W. Li, D. Ren, H. Zhang, M. Wang, and W. Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2706–2715, 2020

    2020

  8. [16]

    X. Li, S. Zhang, S. Zhou, L. Zhang, and W. Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5904–5917, 2022

    2022

  9. [17]

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Fei, B. Dai, W. Ouyang, Y . Qiao, and C. Dong. Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023

    2023 arXiv

  10. [18]

    completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012

    2012

  11. [19]

    Nitzan, K

    Y . Nitzan, K. Aberman, Q. He, O. Liba, M. Yarom, Y . Gandelsman, I. Mosseri, Y . Pritch, and D. Cohen-Or. Mystyle: A personalized generative prior.ACM Transactions on Graphics (TOG), 41(6):1–10, 2022

    2022

  12. [20]

    Pouyanfar, S

    S. Pouyanfar, S. Sengupta, M. Mohammadi, E. Abraham, B. Bloomquist, L. Dauterman, A. Parikh, S. Lim, and E. Sommerlade. Frr-net: A real-time blind face restoration and relighting network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2023

  13. [21]

    Richardson, Y

    E. Richardson, Y . Alaluf, O. Patashnik, Y . Nitzan, Y . Azar, S. Shapiro, and D. Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021

    2021

  14. [22]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695, 2022

    2022

  15. [23]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi. Image super-resolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713–4726, 2022

    2022

  16. [24]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  17. [25]

    M. Suin, N. G. Nair, C. P. Lau, V . M. Patel, and R. Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6343–6352, 2024. 12

    2024

  18. [26]

    Tsai, Y .-L

    Y .-J. Tsai, Y .-L. Liu, L. Qi, K. C. Chan, and M.-H. Yang. Dual associated encoder for face restoration. In The Twelfth International Conference on Learning Representations, 2024

    2024

  19. [27]

    J. Wang, Z. Yue, S. Zhou, K. C. Chan, and C. C. Loy. Exploiting diffusion prior for real-world image super-resolution. 2024

    2024

  20. [28]

    X. Wang, Y . Li, H. Zhang, and Y . Shan. Towards real-world blind face restora- tion with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9168–9178, 2021

    2021

  21. [29]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text- to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023

    2023

  22. [30]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018

    2018

  23. [31]

    Y . Zhao, T. Hou, Y .-C. Su, X. Jia, Y . Li, and M. Grundmann. Towards authentic face restoration with iterative diffusion models and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7312–7322, 2023

    2023

  24. [32]

    S. Zhou, K. Chan, C. Li, and C. C. Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599–30611, 2022. 13 A Broader Impacts The ReF-LDM has the capability to leverage personal appearances from re...

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.