💻 computer science

The Second Challenge on Real-World Face Restoration at NTIRE 2026: Methods and Results

This paper reviews the NTIRE 2026 challenge on real-world face restoration, detailing the methods and results from 10 participating teams that aimed to generate natural, identity-consistent outputs while advancing state-of-the-art perceptual quality.

Original authors: 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

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: 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 V, 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very old, blurry, and scratched photograph of your favorite celebrity or a family member. You want to fix it, but there's a catch: you don't just want a clear picture; you want it to look like a brand-new, high-definition photo taken today, and it still has to look exactly like that person. If the fix makes them look like a different person, the job is a failure.

This paper is the "report card" for a massive global competition called NTIRE 2026, where the world's smartest computer scientists tried to solve this exact problem: Real-World Face Restoration.

Here is the story of how they did it, explained simply.

🏆 The Big Contest

Think of this like the Olympics for photo-fixing AI.

The Goal: Take a terrible, low-quality face photo and turn it into a stunning, realistic masterpiece.
The Rules: The AI can't just guess random details (like giving someone a different nose). It must keep the person's identity 100% intact.
The Players: 96 teams signed up, but only 10 made it to the final round. 9 of them actually passed the strict "identity check."

🧠 How the AI "Thinks"

In the past, fixing a blurry face was like trying to draw a portrait using only a few blurry dots. It was hard to get the details right.

In 2026, the winners changed the game. Instead of trying to "fix" the pixels one by one, they used AI "Dreamers" (called Diffusion Models).

The Analogy: Imagine you have a muddy window. Old methods tried to scrub the mud off with a rag. The new methods are like a wizard who looks at the mud, remembers what a clean window should look like based on a million other windows they've seen, and essentially "re-imagines" the clean glass right over the mud.

🏅 The Winners and Their Secret Weapons

Here is how the top teams approached the challenge, using simple metaphors:

1. 🥇 MiPlusCV (The Two-Step Chef)

Rank: 1st Place
The Strategy: They didn't try to cook the whole meal in one go.
- Step 1: They used a tool called OSDFace to do the heavy lifting—removing the big blur and getting the basic shape of the face right.
- Step 2: They used a super-fast "one-step" AI (based on Z-Image) to add the delicious garnish: skin texture, pores, and tiny details.
Why they won: They realized that getting the structure right first, then adding the "fancy details" later, produced the most natural-looking results.

2. 🥈 CEVI-KLETech (The Surgeon)

Rank: 2nd Place
The Strategy: They treated the face like a body with different organs.
The Secret: They realized that the skin needs a different kind of "fix" than the eyes or the hair. They used a special tool (Wavelets) to break the image into layers. They kept the "skeleton" (low frequency) alone to protect the identity, but they surgically added high-frequency details only to the skin and hair areas.
Why they won: Precision. They didn't over-fix the whole face; they fixed exactly what needed fixing.

3. 🥉 HONORAICamera (The Speedster)

Rank: 3rd Place
The Strategy: Speed and efficiency.
The Secret: They trained their AI to do the whole job in one single step instead of taking 20 or 30 steps like older models. They used a "Turbo" version of the AI.
Why they won: They proved you don't need to wait hours for a photo to be fixed; you can get high quality instantly.

4. YuFans (The Perfectionist)

Rank: 4th Place
The Strategy: "Good enough isn't good enough."
The Secret: They generated a good photo first, then looked at it with a "robot critic" (an AI that judges photo quality). If the critic said, "The eyes look a bit fake," YuFans tweaked the pixels just for that photo to make the critic happy.
Why they won: They optimized specifically for the judges' scoring system.

🚀 The Big Trends of 2026

The paper highlights three major shifts in how we fix faces:

From "Fixing" to "Re-imagining": We stopped trying to just sharpen blurry pixels. Now, we use powerful AI models that "hallucinate" (create) new, realistic details that were lost, but do it in a way that stays true to the person.
The "One-Step" Revolution: Old methods took many steps to clean up an image. The winners showed that with the right training, AI can do it in a single leap, making it much faster.
Identity is King: No matter how pretty the photo looks, if it doesn't look like the original person, it fails. The best teams built special "guardrails" to ensure the AI never changes the person's face shape or identity.

🏁 The Conclusion

This competition showed us that the future of photo restoration isn't about better cameras; it's about smarter AI. We are moving from simply cleaning up old photos to reconstructing reality with such high fidelity that the restored photo looks like it was taken with a million-dollar camera today.

The winners proved that by combining fast generation, smart structural guidance, and strict identity checks, we can bring our oldest, worst photos back to life better than ever before.

Based on the paper "The Second Challenge on Real-World Face Restoration at NTIRE 2026: Methods and Results," here is a detailed technical summary:

1. Problem Definition

The challenge addresses Real-World Face Restoration, a task aimed at reconstructing high-quality (HQ), natural, and realistic face images from low-quality (LQ) inputs degraded by blur, noise, compression, and other real-world distortions.

Core Challenge: The task is inherently ill-posed because severe degradation removes significant visual information.
Dual Objectives:
1. Perceptual Quality: The output must exhibit rich details, realistic textures, and high fidelity, mimicking high-end imaging devices.
2. Identity Consistency: The restored face must preserve the identity of the original subject.
Constraints: Unlike previous iterations, this challenge imposes no restrictions on computational resources (model size, FLOPs) or training data volume, focusing purely on achieving the best perceptual and identity performance.

2. Methodology and Evaluation Framework

Dataset

Training: Participants were recommended to use FFHQ (70k HQ images) and could utilize other datasets (e.g., FFHQR).
Testing: The evaluation set consisted of 450 images sampled from five diverse real-world datasets: CelebChild-Test, LFW-Test, WIDER-Test, CelebA, and WebPhoto-Test. These cover challenging scenarios like aging, occlusion, extreme poses, and varying lighting.

Evaluation Metrics

The competition employed a two-step evaluation process:

Identity Verification (Gatekeeper): A pre-trained AdaFace model measured cosine similarity between the LQ input and the restored HQ image.
- Thresholds varied by dataset (0.3 for WIDER/WebPhoto, 0.5 for CelebA, 0.6 for LFW/CelebChild).
- Teams failing this threshold (or having >10 failures) were disqualified.
Perceptual Quality Scoring: For qualified teams, a weighted score was calculated based on multiple No-Reference Image Quality Assessment (IQA) metrics:
- Metrics: CLIPIQA, MANIQA, MUSIQ, Q-Align, NIQE, and FID (using FFHQ as reference).
- Formula:
  $\text{Score} = \text{CLIPIQA} + \text{MANIQA} + \frac{\text{MUSIQ}}{100} + \frac{\text{QALIGN}}{5} + \max(0, \frac{10 - \text{NIQE}}{10}) + \max(0, \frac{100 - \text{FID}}{100})$

3. Key Contributions and Trends

The paper analyzes the top-performing methods, identifying five major trends in the field:

Dominance of Efficient Generative Paradigms:
- Top teams moved away from multi-step diffusion models toward one-step or distilled diffusion priors (e.g., OSDFace, Z-Image, SDXL-Turbo).
- This shift demonstrates that high perceptual quality can be achieved with single forward passes, significantly improving inference efficiency.
Metric-Oriented Refinement:
- Leading teams did not rely solely on the backbone generator. They implemented post-training or test-time optimization specifically targeting the challenge metrics (CLIPIQA, MANIQA, MUSIQ).
- Techniques included reward-based post-training and differentiable pixel optimization guided by IQA models.
Semantic and Structural Guidance:
- Despite the power of foundation models, explicit face-aware priors remain crucial for identity safety.
- Methods incorporated semantic parsing (e.g., BiSeNet) and wavelet-domain corrections to ensure different facial regions (eyes, mouth, skin) receive appropriate levels of refinement without distorting anatomy.
Modular Multi-Stage Architectures:
- Instead of monolithic end-to-end training, successful approaches used pipelines separating coarse recovery, perceptual enhancement, and post-processing.
- This modularity allows teams to reuse strong pre-trained generative priors while adding lightweight, task-specific refinement blocks.
Foundation Model Adaptation:
- There is a clear shift from training task-specific restoration networks to adapting large-scale foundation models (e.g., FLUX.2, Stable Diffusion, Z-Image) via parameter-efficient tuning (LoRA) and structured control signals.

4. Results

Participation: 96 teams registered; 10 submitted valid models; 9 teams achieved valid scores after identity verification.
Top Performers:
1. MiPlusCV: Used a two-stage pipeline combining OSDFace (coarse restoration) with a Z-Image-based one-step diffusion restorer, refined via reward-based post-training.
2. CEVI-KLETech: Employed a three-stage baseline with a lightweight Semantic-Aware Frequency-Guided Residual Correction (SA-FGRC) module using wavelet transforms and semantic parsing.
3. HONORAICamera: Fine-tuned Z-Image-Turbo using an OMGSR-style training strategy, fixing inference to a single timestep.
4. YuFans: Combined SDFace (one-step) with test-time CLIPIQA-guided pixel optimization.
5. guaguagua: Utilized DeSC-Face based on FLUX.2 with degradation-aware structured control.
Disqualification: Team BVI failed the AdaFace identity test (82 failed images) and received no final ranking.

5. Significance

State-of-the-Art Advancement: The challenge confirms that foundation models (Diffusion, LCMs) are now the dominant paradigm for real-world face restoration, surpassing traditional CNN/GAN-based approaches in both realism and detail recovery.
Efficiency vs. Quality: The success of one-step and distilled models proves that high-fidelity restoration does not necessarily require computationally expensive multi-step sampling.
Identity Preservation: The results highlight that while generative models are powerful, explicit constraints (semantic guidance, identity loss, wavelet correction) are essential to prevent "hallucination" of incorrect identities or facial structures.
Benchmarking: The paper provides a comprehensive overview of current trends, offering a new benchmark for future research that balances perceptual quality, identity consistency, and inference efficiency.