Source-linked AI summary
The Second Challenge on Real-World Face Restoration at NTIRE 2026: Methods and Results
Jingkai Wang, Jue Gong, Zheng Chen, Kai Liu, Jiatong Li, Yulun Zhang, Radu Timofte, Jiachen Tu, Yaokun Shi, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yingsi Chen, Yijiao Liu, Hui Li, Yu Wang, Congchao Zhu, Alexandru-Gabriel Lefterache, Anamaria Radoi, Chuanyue Yan, Tao Lu, Yanduo Zhang, Kanghui Zhao, Jiaming Wang, Yuqi Li, WenBo Xiong, Yifei Chen, Xian Hu, Wei Deng, Daiguo Zhou, Sujith Roy, Claudia Jesuraj, Vikas B, Spoorthi LC, Nikhil Akalwadi, Ramesh Ashok Tabib, Uma Mudenagudi, Yuxuan Jiang, Chengxi Zeng, Tianhao Peng, Fan Zhang, David Bull Wei Zhou, Linfeng Li, Hongyu Huang, Hoyoung Lee, SangYun Oh, ChangYoung Jeong, Axi Niu, Jinyang Zhang, Zhenguo Wu, Senyan Qing, Jinqiu Sun, Yanning Zhang
TL;DR
Real-world face restoration must recover perceptually high-quality, natural faces from degraded inputs while maintaining identity consistency. This paper reviews an NTIRE 2026 challenge that evaluates competing solutions with identity verification and weighted image-quality metrics, finding strong performance from efficient generative and refinement-based approaches. The challenge results highlight foundation-model adaptation alongside semantic, structural, and identity-preserving constraints.
Problem
Severely degraded face inputs lose visual information, creating an ill-posed restoration problem that requires realistic detail and identity preservation.
Method
The paper reviews an NTIRE 2026 challenge using AdaFace identity verification, weighted IQA evaluation, and unconstrained computational and training-resource rules.
Results
MiPlusCV ranked first, while 9 of 10 valid submissions achieved eligible final rankings after AdaFace verification.
Takeaways & Limitations
Strong challenge results favor efficient generative paradigms, metric-aware refinement, and foundation-model generation constrained by semantic, structural, or identity-preserving signals.
Takeaways & Limitations
Training data must not overlap with the five test datasets, although additional training datasets and simulated degradation pipelines are allowed.
Abstract
from arXiv · showhide
This paper provides a review of the NTIRE 2026 challenge on real-world face restoration, highlighting the proposed solutions and the resulting outcomes. The challenge focuses on generating natural and realistic outputs while maintaining identity consistency. Its goal is to advance state-of-the-art solutions for perceptual quality and realism, without imposing constraints on computational resources or training data. Performance is evaluated using a weighted image quality assessment (IQA) score and employs the AdaFace model as an identity checker. The competition attracted 96 registrants, with 10 teams submitting valid models; ultimately, 9 teams achieved valid scores in the final ranking. This collaborative effort advances the performance of real-world face restoration while offering an in-depth overview of the latest trends in the field.
1. Introduction
Real-world face restoration reconstructs high-quality, natural-looking faces from severely degraded inputs while preserving identity. The challenge advances this goal by encouraging perceptually strong solutions and surveying current restoration trends.
- Face restoration reconstructs high-quality face images from low-quality inputs degraded by blur, noise, compression, and other distortions.
- Severe degradation removes substantial visual information, making face restoration inherently ill-posed.
- Restored faces must combine rich details with clarity, naturalness, realism, and high fidelity.
- CNNs, Transformers, GANs, and diffusion models have substantially advanced face restoration by learning or exploiting face priors.
- The NTIRE 2026 challenge emphasizes richer textures, realistic facial appearances, consistent identity preservation, and strong perceptual performance.
2. NTIRE 2026 Real-world Face Restoration
The challenge evaluates restoration of real-world degraded faces under explicit requirements for perceptual quality and identity consistency. Its protocol combines diverse test data, AdaFace identity verification, and weighted no-reference IQA metrics.
- The task recovers high-quality faces with rich high-frequency details while preserving facial identity to a reasonable degree.
- The challenge imposes no restrictions on model size or FLOPs, prioritizing image quality and identity consistency.
- The test set contains 450 images sampled from CelebChild-Test, LFW-Test, WIDER-Test, CelebA, and WebPhoto-Test.
- Participants must pass dataset-specific AdaFace identity thresholds before their perceptual quality scores can determine ranking.
- Final quality assessment uses NIQE, CLIPIQA, MANIQA, MUSIQ, Q-Align, and FFHQ-referenced FID within a weighted ranking procedure.
3. Challenge Results
MiPlusCV ranked first in the challenge, while the results reveal a shift toward efficient foundation-model generation combined with targeted refinement and identity-preserving constraints. The final ranking followed identity verification and weighted perceptual-quality evaluation.
- 3. Challenge Results: MiPlusCV achieved first place, followed by CEVI-KLETech, HONORAICamera, YuFans, and guaguagua; BVI failed the AdaFace identity test.
- 3.1. Architectures and main ideas: Top-ranked methods largely adapt powerful pre-trained generators, with one-step or distilled diffusion priors dominating the leading approaches.
- 3.1. Architectures and main ideas: Metric-aware refinement, including reward-based post-training and test-time optimization, differentiates several strong solutions.
- 3. Challenge Results: 96 registrations produced 10 valid model submissions, with 9 teams remaining eligible for the final ranking after AdaFace verification.
- 3.4. Conclusions: Strong results combine foundation-model generation with semantic, structural, or identity-preserving constraints rather than relying on unconstrained generation alone.
4. Challenge Methods and Teams
The challenge teams combine generative restoration backbones with targeted structural, semantic, identity, or perceptual refinement strategies. Approaches range from one-step diffusion and test-time optimization to two-stage pipelines and enhanced CodeFormer training.
- MiPlusCV: MiPlusCV uses OSDFace for coarse structure followed by a one-step Z-Image diffusion restorer for perceptual quality.The second stage uses LoRA adapters, direct image supervision, fidelity and perceptual losses, identity supervision, and adversarial learning.
- CEVI-KLETech: CEVI-KLETech inserts SA-FGRC between StyleGAN2-based fidelity, DiffBIR, and DINOv2-guided naturalness stages.The module applies region-specific high-frequency corrections while leaving the low-frequency band unchanged.
- HONORAICamera: HONORAICamera transfers Z-Image-Turbo’s generative prior with fixed timestep 244 for one-step restoration.Training pairs use Real-ESRGAN degradations including blur, noise, and JPEG compression.
- YuFans: YuFans combines one-step SDFace restoration with 10 test-time CLIPIQA gradient-ascent steps under fidelity and total-variation regularization.The base restorer is not further trained; pixel optimization directly modifies its output.
- DeSC-Face: DeSC-Face uses degraded latent tokens as both FLUX.2 backbone conditioning and structured-control input before iterative scene-token restoration.The restored scene-token stream is decoded into the final face image.
- NTR, MaDENN, SN VISION, and BVI: NTR adopts DiffBIR v2.1, while MaDENN, SN VISION, and BVI add identity, semantic, structural, or high-frequency refinement mechanisms.DiffBIR uses SwinIR degradation removal followed by IRControlNet texture synthesis; MaDENN extends CodeFormer with identity and ROI-aware supervision, SN VISION conditions SDXL ControlNet on enhanced RGB and structural maps, and BVI refines one-step diffusion noise.
MiPlusCV
MiPlusCV is affiliated with MiLM Plus, Xiaomi Inc., China.
- MiPlusCV is affiliated with MiLM Plus, Xiaomi Inc., China.
CEVI-KLETech
CEVI-KLETech is affiliated with KLE Technological University, Hubballi, India.
- CEVI-KLETech is affiliated with KLE Technological University, Hubballi, India.
HONORAICamera
HONORAICamera is affiliated with Honor Device Co. Ltd.
- HONORAICamera is affiliated with Honor Device Co. Ltd.
YuFans
YuFans is affiliated with the National University of Singapore and Zhejiang University.
- The listed affiliations are the National University of Singapore and Zhejiang University.
guaguagua
guaguagua is affiliated with Northwestern Polytechnical University in China.
- The listed affiliation is Northwestern Polytechnical University, China.
NTR
NTR includes affiliations spanning universities and a corporation in multiple countries.
- One listed affiliation is the University of Illinois Urbana-Champaign.
- UNSTPB POLITEHNICA Bucharest, Romania, is another listed affiliation.
- SNOW Corporation is also listed as an affiliation.
- Wuhan Institute of Technology and City University of New York are listed together as affiliations.
- The University of Bristol is another listed affiliation.