Organizations: School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · vivo BlueImage Lab, vivo Mobile Communication Co., Ltd., Hangzhou, China
Face restoration has achieved significant advancements through the years of development. However, maintaining high fidelity and authenticity while avoiding artifacts remains challenging, especially in extreme degradation scenarios. This highlights the need for models that are more ``honest'' in their reconstruction from low-quality inputs, accurately reflecting original characteristics. In this work, we propose HonestFace, a novel approach for honest face restoration with identity consistency and realistic textures. To achieve this, HonestFace incorporates several key components. First, we propose an identity embedder to effectively capture and preserve crucial identity features from both the low-quality input and multiple reference faces. Second, a masked face alignment method is presented to enhance fine-grained details and textural authenticity. Furthermore, we present a new real-world reference-based face restoration dataset, MultiRefCeleb-Test, with multiple high-quality references and low-quality inputs. Leveraging these contributions within a one-step diffusion model framework, HonestFace delivers excellent restoration results in terms of facial fidelity and realism. Experiments demonstrate that our approach surpasses state-of-the-art methods, achieving superior performance in both visual quality and quantitative assessments. The code and pre-trained models are available at https://github.com/jkwang28/HonestFace .
Figures & tables
Figure 1. Real-world concert scenes captured by smartphones. From left to right: Buyi Mao, Adele, Jason Zhang, Bibi Zhou, Fish Leong, and G.E.M. Concerts are typically held in the evening, and photos are predominantly taken using telephoto lenses. For celebrities like these, who have abundant high-definition images available online, smartphone manufacturers can directly embed high-resolution ID information, enabling automatic matching and background processing when users take photos.
Figure 2. Comparisons of face restoration methods. HonestFace restores a more natural, faithful and ID-consistent face.
Figure 3. Overall training pipeline of HonestFace. First, the LQ input xL is encoded into zL by the VAE encoder Eϕ . Meanwhile, xL and HQ references R={ri} pass through IDE and VRE, then fused into the prompt embedding p . Next, the UNet predicts εθ to estimate z^H . Finally, the VAE decoder Dϕ reconstructs the output x^H . Generator and discriminator are trained alternately.
Figure 4. Key modules in HonestFace pipeline. ( left , a) Facial feature extractor: Generates a feature embedding pF from the most geometrically similar eye patches, selected from the optimal references; ( middle , b) Face identity encoder: Encodes multiple reference faces to produce a combined identity representation pI ; ( right , c) Masked face alignment: A novel perceptual loss that strategically focuses the restoration on perceptually crucial areas, ensuring the fidelity of key textures and natural colors.
Figure 5. Overview of the MultiRefCeleb-Test benchmark. Each identity features multiple high-quality reference images alongside several low-quality inputs. The dataset spans diverse ethnicities and age ranges. It incorporates a wide spectrum of authentic real-world degradations and challenging cases, such as occlusions, varied illumination, and diverse facial expressions.
Figure 6. Visual comparison of synthetic datasets CelebHQRef-Test and Reface-Test. Please zoom in for a better view.
Datasets
Methods
LPIPS ↓
DISTS ↓
C-IQA ↑
M-IQA ↑
MUSIQ ↑
FID ↓
Deg. ↓
LMD ↓
CelebHQRef-Test
ASFFNet ( Li et al., 2020b )
0.1969
0.1411
0.6067
0.5346
71.181
45.618
35.636
3.6047
DMDNet ( Li et al., 2022 )
0.2191
0.1499
0.6417
0.5165
71.119
52.406
34.258
2.6437
PGDiff ( Yang et al., 2023 )
0.2674
0.1559
0.5735
0.4953
70.762
60.693
46.248
3.6478
FaceMe ( Liu et al., 2025 )
0.1945
0.1412
0.6606
0.5758
71.598
46.423
30.578
2.0107
MGFR ( Tao et al., 2025 )
0.2311
0.1520
0.6224
0.6475
70.626
52.103
38.322
2.5274
HonestFace (ours)
0.1809
0.1246
0.6673
0.6858
71.996
41.021
26.106
1.7798
Table 1. Comparison with the reference-based methods. C-IQA stands for CLIP-IQA, and M-IQA stands for MANIQA.
Figure 7. Visual comparison on the real-world MultiRefCeleb-Test dataset. Please zoom in for a better view.
CelebHQRef-Test
Reface-Test
MultiRefCeleb-Test
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
M-IQA ↑
FID ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
M-IQA ↑
FID ↓
M-IQA ↑
FID ↓
CodeFormer ( Zhou et al., 2022 )
25.49
0.7100
0.1791
0.1419
0.5827
48.37
25.65
0.7082
0.1872
0.1497
0.5699
33.57
0.5483
60.61
DAEFR ( Tsai et al., 2024 )
22.85
0.6549
0.2058
0.1439
0.6019
49.35
23.07
0.6574
0.2092
0.1498
0.5842
34.10
0.5541
48.52
OSEDiff ( Wu et al., 2024a )
24.07
0.7105
0.2077
0.1533
0.5229
57.06
24.21
0.7110
0.2076
0.1553
0.5122
46.31
0.7057
84.85
OSDFace ( Wang et al., 2025 )
23.51
0.6742
0.2050
0.1373
0.6316
46.51
23.54
0.6718
0.2156
0.1481
0.6257
32.59
0.7148
63.11
HonestFace (ours)
25.17
0.7147
0.1809
0.1246
0.6858
41.02
25.25
0.7118
0.1784
0.1235
0.6717
25.75
0.6649
57.32
Table 2. Quantitative comparison with the no-reference methods. M-IQA stands for MANIQA.
Table 3. Comprehensive quantitative evaluations. (a) Quantitative comparisons with state-of-the-art reference-based methods on the real-world dataset. (b-e) More experiments for color shift evaluation, ablation studies, and inference complexity comparison. All ablation study experiments are conducted on CelebHQRef-Test. C-IQA stands for CLIP-IQA and M-IQA stands for MANIQA.
Video face restoration aims to enhance degraded face videos into high-quality results with realistic facial details, stable identity, and temporal coherence. Recent diffusion-based methods have brought strong generative priors to restoration and enabled more realistic detail synthesis. However, existing approaches for face videos still rely heavily on generic diffusion priors and multi-step sampling, which limit both facial adaptation and inference efficiency. These limitations motivate the use of one-step diffusion for video face restoration, yet achieving faithful facial recovery alongside temporally stable outputs remains challenging. In this paper, we propose, DVFace, a one-step diffusion framework for real-world video face restoration. Specifically, we introduce a spatio-temporal dual-codebook design to extract complementary spatial and temporal facial priors from degraded videos. We further propose an asymmetric spatio-temporal fusion module to inject these priors into the diffusion backbone according to their distinct roles. Evaluation on various benchmarks shows that DVFace delivers superior restoration quality, temporal consistency, and identity preservation compared to recent methods. Code: https://github.com/zhengchen1999/DVFace.
Face restoration under complex degradations still remains an ill-posed inverse problem due to severe information loss. Although diffusion models benefit from strong generative priors, most methods still condition only on low-quality inputs, making it difficult to recover identity-critical details under heavy degradations. In this work, we propose HDRFace, a High-Dimensional Representation conditioned Face restoration framework that injects semantically rich priors into the conditional flow without modifying the generative backbone. Our pipeline first obtains a structurally reliable intermediate restoration with an off-the-shelf restorer, then uses a pretrained high-dimensional feature encoder to extract fine-grained facial representations from both the low-quality input and the intermediate result, and injects them as additional conditions for generation. We further introduce SDFM, a Structure-Detail aware adaptive Fusion Mechanism that emphasizes global constraints during structure modeling and strengthens representation guidance during detail synthesis, balancing structural consistency and detail fidelity. To validate the generalization ability of our method, we implement the proposed framework on two generative models, SD V2.1-base and Qwen-Image, and consistently observe stable and coherent performance gains across different architectures.
Zirui Wang, Xianhui Lin, Yi Dong +7
City University of Hong Kong · vivo BlueImage Lab, vivo Mobile Communication Co., Ltd
Blind face restoration is highly ill-posed under severe degradation, where identity-critical details may be missing from the degraded input. Same-identity references reduce this ambiguity, but mismatched pose, expression, illumination, age, makeup, or local facial states can lead to overuse of reference appearance. We propose IConFace, a unified reference-aware and no-reference framework with identity--structure asymmetric conditioning. References are distilled into a norm-weighted global AdaFace identity anchor for image-only modulation, while the degraded image is reinforced as the spatial structure anchor through low-rank residuals and block-wise degraded cross-attention with two-route memory. The resulting single checkpoint exploits references when available and falls back to no-reference restoration when absent, improving identity consistency, fine-detail recovery, and degraded-only restoration quality in a unified model.