转眼做人脸复原(face restoration)算法也一段时间了，根据自己的记忆整理一下自己的一些看法，算作个人记录，当然如果有人愿意分享自己的看法也是极好的。先挂下文章链接，下一篇在写总结。
一、前述
人脸修复(face restoration)任务，起源于人脸超分辨率(face super resolution)，可以算是从超分出来的一个分支。作为图像低级任务(low level)中的一个，主要目的就是在低清、受损画质中尽可能的恢复人脸细节。与传统的全图超分的区别在于，人脸是有明确的先验知识在里面的，可以利用这些先验增加模型的信息量，尽可能的恢复人脸的细节。这些先验可以包括面部特征，也称为几何先验（Geometric Prior） ，有基于高清参考人脸的先验（Reference Prior） ，有基于强的生成器先验（Generative Prior） ，也有基于离散的codebook的先验（Vector Quantized Codebook Prior） ，当然，还有近几年爆火的diffusion，作为强的文生图基本模型，他可以作为一种扩散先验（Diffusion Prior） 。
有趣的是，从超分开始，随着不同的先验在人脸修复中被使用，后续的发展中却出现了不同的轨迹 ，例如几何先验、参考先验、生成式先验、codebook先验，这些在人脸中是效果比较突出的，在超分中比较少用到。但是在最近diffusion的发展中，又慢慢的归为了一个轨迹，所谓天下大势，分久必合，合久必分，感觉有这个味道在了。
二、相关论文及代码整理
原始GitHub仓库： https://github.com/qianx77/Face_Restoration_Ref.git
部分实现：https://github.com/qianx77/Face_Restoration
Geometric Prior
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2018 | CVPR | FSRNet: End-to-End Learning Face Super-Resolution with Facial Priors | [paper](https://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_FSRNet_End-to-End_Learning_CVPR_2018_paper.pdf)[code](https://github.com/tyshiwo/FSRNet) | CNN | FSRNet |
| 2021 | TPAMI | Face Restoration via Plug-and-Play 3D Facial Priors | [paper](https://www.zora.uzh.ch/id/eprint/214478/1/ZORA214478.pdf)code | CNN | - |
| 2021 | CVPR | Progressive Semantic-Aware Style Transformation for Blind Face Restoration | [paper](https://openaccess.thecvf.com/content/CVPR2021/papers/Chen_Progressive_Semantic-Aware_Style_Transformation_for_Blind_Face_Restoration_CVPR_2021_paper.pdf)[code](https://github.com/chaofengc/PSFRGAN) | GAN | PSFRGAN |
| 2023 | TNNLS | Multi-prior learning via neural architecture search for blind face restoration | [paper](https://Arxiv.org/pdf/2206.13962.pdf)[code](https://github.com/YYJ1anG/MFPSNet) | CNN | MFPSNet |
| 2022 | CVPR | Blind Face Restoration via Integrating Face Shape and Generative Priors | [paper](https://openaccess.thecvf.com/content/CVPR2022/papers/Zhu_Blind_Face_Restoration_via_Integrating_Face_Shape_and_Generative_Priors_CVPR_2022_paper.pdf)[code](https://github.com/TencentYoutuResearch/BFR-SGPN) | GAN | SGPN |
Reference Prior
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2018 | ECCV | Learning Warped Guidance for Blind Face Restoration | [paper](https://openaccess.thecvf.com/content_ECCV_2018/papers/Xiaoming_Li_Learning_Warped_Guidance_ECCV_2018_paper.pdf)[code](https://github.com/csxmli2016/GFRNet) | CNN | GFRNet |
| 2020 | ECCV | Blind Face Restoration via Deep Multi-scale Component Dictionaries | [paper](https://Arxiv.org/pdf/2008.00418)[code](https://github.com/csxmli2016/DFDNet) | CNN | DFDNet |
| 2020 | CVPR | Enhanced Blind Face Restoration with Multi-Exemplar Images and Adaptive Spatial Feature Fusion | [paper](https://openaccess.thecvf.com/content_CVPR_2020/papers/Li_Enhanced_Blind_Face_Restoration_With_Multi-Exemplar_Images_and_Adaptive_Spatial_CVPR_2020_paper.pdf)[code](https://github.com/csxmli2016/ASFFNet) | CNN | ASFFNet |
| 2022 | TPAMI | Learning Dual Memory Dictionaries for Blind Face Restoration | [paper](https://arxiv.org/pdf/2210.08160.pdf)[code](https://github.com/csxmli2016/DMDNet) | CNN | DMDNet |
| 2022 | ACM | MyStyle: A Personalized Generative Prior | [paper](https://dl.acm.org/doi/pdf/10.1145/3550454.3555436)code | GAN | MyStyle |
| 2023 | arXiv | Personalized Restoration via Dual-Pivot Tuning | [paper](https://arxiv.org/abs/2312.17234)[code](https://github.com/personalized-restoration/personalized-restoration) | Diffusion | Dual-Pivot Tuning |
| 2024 | arXiv | InstantRestore: Single-Step Personalized Face Restoration with Shared-Image Attention | [paper](https://arxiv.org/pdf/2412.06753)[code](https://github.com/snap-research/InstantRestore) | Diffusion | InstantRestore |
| 2024 | CVPR | PFStorer: Personalized Face Restoration and Super-Resolution | [paper](https://openaccess.thecvf.com/content/CVPR2024/papers/Varanka_PFStorer_Personalized_Face_Restoration_and_Super-Resolution_CVPR_2024_paper.pdf)code | Diffusion | PFStorer |
| 2024 | arXiv | RestorerID: Towards Tuning-Free Face Restoration with ID Preservation | [paper](https://arxiv.org/pdf/2411.14125)[code](https://github.com/YingJiacheng/RestorerID) | Diffusion | RestorerID |
| 2024 | arXiv | Overcoming False Illusions in Real-World Face Restoration with Multi-Modal GuidedDiffusion Model | [paper](https://arxiv.org/pdf/2410.04161)code | Diffusion | MGFR |
| 2024 | arXiv | Copy or Not? Reference-Based Face Image Restoration with Fine Details | [paper](https://jianwang-cmu.github.io/25Refine/main.pdf)[code](https://github.com/RefineFIR/RefineFIR) | Diffusion | RefineFIR |
| 2024 | Sensors | PDGrad: Guiding Diffusion Model for Reference-Based Blind Face Restoration with Pivot Direction Gradient Guidance | [paper](https://www.mdpi.com/1424-8220/24/22/7112)code | Diffusion | PDGrad |
| 2024 | CVPR | Restoration by Generation with Constrained Priors | [paper](https://openaccess.thecvf.com/content/CVPR2024/papers/Ding_Restoration_by_Generation_with_Constrained_Priors_CVPR_2024_paper.pdf)[code](https://github.com/adobe-research/gen2res) | Diffusion | gen2res |
| 2024 | WACV | ENTED: Enhanced Neural Texture Extraction and Distribution for Reference-based Blind Face Restoration | [paper](https://openaccess.thecvf.com/content/WACV2024/papers/Lau_ENTED_Enhanced_Neural_Texture_Extraction_and_Distribution_for_Reference-Based_Blind_WACV_2024_paper.pdf)code | Diffusion | ENTED |
| 2025 | AAAI | FaceMe: Robust Blind Face Restoration with Personal Identification | [paper](https://arxiv.org/pdf/2501.05177)[code](https://github.com/modyu-liu/FaceMe) | Diffusion | FaceMe |
Generative Prior (good)
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2020 | MM | HiFaceGAN: Face Renovation via Collaborative Suppression and Replenishment | [paper](https://Arxiv.org/pdf/2005.05005.pdf)[code](https://github.com/Lotayou/Face-Renovation) | GAN | HiFaceGAN |
| 2020 | CVPR | Image Processing Using Multi-Code GAN Prior | [paper](https://openaccess.thecvf.com/content_CVPR_2020/papers/Gu_Image_Processing_Using_Multi-Code_GAN_Prior_CVPR_2020_paper.pdf)[code](https://github.com/genforce/mganprior) | GAN | mGANprior |
| 2021 | CVPR | Towards Real-World Blind Face Restoration with Generative Facial Prior | [paper](https://openaccess.thecvf.com/content/CVPR2021/papers/Wang_Towards_Real-World_Blind_Face_Restoration_With_Generative_Facial_Prior_CVPR_2021_paper.pdf)[code](https://xinntao.github.io/projects/gfpgan) | GAN | GFPGAN |
| 2021 | CVPR | GAN Prior Embedded Network for Blind Face Restoration in the Wild | [paper](https://openaccess.thecvf.com/content/CVPR2021/papers/Yang_GAN_Prior_Embedded_Network_for_Blind_Face_Restoration_in_the_CVPR_2021_paper.pdf)[code](https://github.com/yangxy/GPEN) | GAN | GPEN |
| 2022 | CVPR | Blind Face Restoration via Integrating Face Shape and Generative Priors | [paper](https://openaccess.thecvf.com/content/CVPR2022/papers/Zhu_Blind_Face_Restoration_via_Integrating_Face_Shape_and_Generative_Priors_CVPR_2022_paper.pdf)code | GAN | SGPN |
| 2022 | CVPR | GCFSR: a Generative and Controllable Face Super Resolution Method Without Facial and GAN Priors | [paper](https://arxiv.org/pdf/2203.08444.pdf)[code](https://github.com/hejingwenhejingwen/GCFSR) | GAN | GCFSR |
| 2022 | AAAI | Panini-Net: GAN Prior Based Degradation-Aware Feature Interpolation for Face Restoration | [paper](https://arxiv.org/pdf/2203.08444.pdf)[code](https://github.com/jianzhangcs/panini) | GAN | Panini-Net |
| 2022 | Arxiv | FaceFormer: Scale-aware Blind Face Restoration with Transformers | [paper](https://Arxiv.org/pdf/2207.09790.pdf)code | ViT | FaceFormer |
| 2022 | Arxiv | Blind Face Restoration: Benchmark Datasets and a Baseline Model | [paper](https://arxiv.org/pdf/2206.03697.pdf)[code](https://github.com/bitzpy/blind-face-restoration-benchmark-datasets-and-a-baseline-model) | ViT | STUNet |
| 2023 | CVPR | TFRGAN: Leveraging Text Information for Blind Face Restoration with Extreme Degradation | [paper](https://openaccess.thecvf.com/content/CVPR2023W/MULA/papers/Xie_TFRGAN_Leveraging_Text_Information_for_Blind_Face_Restoration_With_Extreme_CVPRW_2023_paper.pdf)code | GAN | TFRGAN |
| 2023 | TCSVT | DEAR-GAN: Degradation-Aware Face Restoration With GAN Prior | [paper](https://ieeexplore.ieee.org/abstract/document/10044117)code | GAN | DEAR-GAN |
| 2023 | ACMMM | Exploring Correlations in Degraded Spatial Identity Features for Blind Face Restoration | [paper](https://dl.acm.org/doi/abs/10.1145/3581783.3611782)code | GAN | MemGAN |
| 2023 | FSP | Degradation Learning and Skip-Transformer for Blind Face Restoration | [paper](https://www.frontiersin.org/articles/10.3389/frsip.2023.1106465/full)code | GAN | - |
| 2023 | ACMMM | Exploring Correlations in Degraded Spatial Identity Features for Blind Face Restoration | [paper](https://dl.acm.org/doi/abs/10.1145/3581783.3611782)code | GAN | - |
| 2024 | AAAI | Blind Face Restoration under Extreme Conditions: Leveraging 3D-2D Prior Fusion for Superior Structural and Texture Recovery | [paper](https://ojs.aaai.org/index.php/AAAI/article/view/27889/27803)code | GAN | FREx |
Vector Quantized Codebook Prior (good)
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2022 | CVPR | RestoreFormer: High-Quality Blind Face Restoration from Undegraded Key-Value Pairs | [paper](https://Arxiv.org/pdf/2201.06374.pdf)[code](https://github.com/wzhouxiff/RestoreFormer) | ViT | RestoreFormer |
| 2022 | NeurIPS | Towards Robust Blind Face Restoration with Codebook Lookup Transformer | [paper](https://Arxiv.org/pdf/2206.11253.pdf)[code](https://github.com/sczhou/CodeFormer) | ViT | CodeFormer |
| 2022 | ECCV | VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder | [paper](https://Arxiv.org/pdf/2205.06803.pdf)[code](https://github.com/TencentARC/VQFR) | CNN | VQFR |
| 2023 | TPAMI | RestoreFormer++: Towards Real-World Blind Face Restoration from Undegraded Key-Value Pairs | [paper](https://Arxiv.org/pdf/2205.06803.pdf)[code](https://github.com/wzhouxiff/RestoreFormerPlusPlus) | ViT | RestoreFormer++ |
| 2022 | CVPR | Rethinking Deep Face Restoration | [paper](https://openaccess.thecvf.com/content/CVPR2022/papers/Zhao_Rethinking_Deep_Face_Restoration_CVPR_2022_paper.pdf)code | GAN | - |
| 2024 | ICLR | Dual Associated Encoder for Face Restoration | [paper](https://arxiv.org/pdf/2308.07314)[code](https://github.com/LIAGM/DAEFR) | ViT | DAEFR |
Diffusion Prior
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2022 | Arxiv | Difface: Blind Face Restoration with Diffused Error Contraction | [paper](https://arxiv.org/pdf/2212.06512.pdf?trk=public_post_comment-text)[code](https://github.com/zsyOAOA/DifFace) | Diffusion | DifFace |
| 2023 | CVPR | DR2: Diffusion-based Robust Degradation Remover for Blind Face Restoration | [paper](https://arxiv.org/abs/2303.06885)[code](https://github.com/Kaldwin0106/DR2_Drgradation_Remover) | Diffusion | DR2 |
| 2023 | WACV | AT-DDPM: Restoring Faces degraded by Atmospheric Turbulence using Denoising Diffusion Probabilistic Models | [paper](https://openaccess.thecvf.com/content/WACV2023/papers/Nair_AT-DDPM_Restoring_Faces_Degraded_by_Atmospheric_Turbulence_Using_Denoising_Diffusion_WACV_2023_paper.pdf)[code](https://github.com/Nithin-GK/AT-DDPM) | Diffusion | AT-DDPM |
| 2023 | ACMMM | DiffBFR: Bootstrapping Diffusion Model for Blind Face Restoration | [paper](https://dl.acm.org/doi/pdf/10.1145/3581783.3611731)code | Diffusion | DiffBFR |
| 2023 | NeurIPS | PGDiff: Guiding Diffusion Models for Versatile Face Restoration via Partial Guidance | [paper](https://dl.acm.org/doi/pdf/10.1145/3581783.3611731)[code](https://github.com/pq-yang/PGDiff) | Diffusion | PGDiff |
| 2024 | TCSVT | Towards Real-World Blind Face Restoration with Generative Diffusion Prior | [paper](https://arxiv.org/pdf/2312.15736.pdf)[code](https://github.com/chenxx89/BFRffusion) | Diffusion | BFRffusion |
| 2024 | WACV | Diffuse and Restore: A Region-Adaptive Diffusion Model for Identity-Preserving Blind Face Restoration | [paper](https://openaccess.thecvf.com/content/WACV2024/papers/Suin_Diffuse_and_Restore_A_Region-Adaptive_Diffusion_Model_for_Identity-Preserving_Blind_WACV_2024_paper.pdf)code | Diffusion | - |
| 2024 | ECCV | DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior | [paper](https://arxiv.org/pdf/2308.15070)[code](https://github.com/XPixelGroup/DiffBIR) | Diffusion | DiffBIR |
| 2024 | NeurIPS | One-Step Effective Diffusion Network for Real-World Image Super-Resolution | [paper](https://arxiv.org/pdf/2406.08177)[code](https://github.com/cswry/OSEDiff) | Diffusion | OSEDiff |
| 2024 | Arxiv | OSDFace: One-Step Diffusion Model for Face Restoration | [paper](https://arxiv.org/abs/2411.17163)[code](https://github.com/jkwang28/OSDFace) | Diffusion | OSDFace |
Others
| Year | Pub | Title | Links | Arch | Abbreviation |
|---|---|---|---|---|---|
| 2023 | ICCV | Towards Authentic Face Restoration with Iterative Diffusion Models and Beyond | [paper](https://openaccess.thecvf.com/content/ICCV2023/papers/Zhao_Towards_Authentic_Face_Restoration_with_Iterative_Diffusion_Models_and_Beyond_ICCV_2023_paper.pdf)code | Diffusion | IDM |
| 2024 | CVPR | WaveFace: Authentic Face Restoration with Efficient Frequency Recovery | [paper](https://openaccess.thecvf.com/content/CVPR2024/papers/Miao_WaveFace_Authentic_Face_Restoration_with_Efficient_Frequency_Recovery_CVPR_2024_paper.pdf)code | Diffusion | WaveFace |
| 2024 | Arxiv | CLR-Face: Conditional Latent Refinement for Blind Face Restoration Using Score-Based Diffusion Models | [paper](https://arxiv.org/abs/2402.06106)code | Diffusion | CLR-Face |
| 2024 | WACV | Diffuse and Restore: A Region-Adaptive Diffusion Model for Identity-Preserving Blind Face Restoration | [paper](https://openaccess.thecvf.com/content/WACV2024/papers/Suin_Diffuse_and_Restore_A_Region-Adaptive_Diffusion_Model_for_Identity-Preserving_Blind_WACV_2024_paper.pdf)code | Diffusion | Diffuse and Restore |
| 2024 | arXiv | AuthFace: Towards Authentic Blind Face Restoration with Face-oriented Generative Diffusion Prior | [paper](https://arxiv.org/pdf/2410.09864)[code](https://github.com/EthanLiang99/AuthFace) | Diffusion | AuthFace |
| 2025 | Arxiv | InfoBFR: Real-World Blind Face Restoration via Information Bottleneck | [paper](https://arxiv.org/pdf/2501.15443)code | Diffusion | InfoBFR |
| 2025 | WACV | Towards Unsupervised Blind Face Restoration using Diffusion Prior | [paper](https://arxiv.org/pdf/2410.04618)[code](https://github.com/SamsungLabs/DT-BFR) | Diffusion | DT-BFR |