Source-linked AI summary
NTIRE 2026 Challenge on Short-form UGC Video Restoration in the Wild with Generative Models: Datasets, Methods and Results
Xin Li, Jiachao Gong, Xijun Wang, Shiyao Xiong, Bingchen Li, Suhang Yao, Chao Zhou, Zhibo Chen, Radu Timofte, Yuxiang Chen, Shibo Yin, Yilian Zhong, Yushun Fang, Xilei Zhu, Yahui Wang, Chen Lu, Meisong Zheng, Xiaoxu Chen, Jing Yang, Zhaokun Hu, Jiahui Liu, Ying Chen, Haoran Bai, Sibin Deng, Shengxi Li, Mai Xu, Junyang Chen, Hao Chen, Xinzhe Zhu, Fengkai Zhang, Long Sun, Yixing Yang, Xindong Zhang, Jiangxin Dong, Jinshan Pan, Jiyuan Zhang, Shuai Liu, Yibin Huang, Xiaotao Wang, Lei Lei, Zhirui Liu, Shinan Chen, Shang-Quan Sun, Wenqi Ren, Jingyi Xu, Zihong Chen, Zhuoya Zou, Xiuhao Qiu, Jingyu Ma, Huiyuan Fu, Kun Liu, Huadong Ma, Dehao Feng, Zhijie Ma, Boqi Zhang, Jiawei Shi, Hao Kang, Yixin Yang, Yeying Jin, Xu Cheng, Yuxuan Jiang, Chengxi Zeng, Tianhao Peng, Fan Zhang, David Bull, Yanan Xing, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Wei Zhou, Linfeng Li, Hang Song, Qi Xu, Kun Yuan, Yizhen Shao, Yulin Ren
TL;DR
Short-form UGC restoration must address diverse real-world degradations while preserving fidelity, perceptual quality, and temporal consistency. The paper presents the KwaiVIR benchmark and NTIRE 2026 challenge, combining synthetic and wild data with objective and subjective evaluation. RedMediaTech ranked first in both tracks, while differing rankings across tracks underscored distinct optimization goals.
Problem
Real-world S-UGC videos contain diverse, entangled degradations, but dedicated benchmarks for generative restoration that jointly assess synthetic fidelity and real-world perceptual quality remain limited.
Method
The challenge introduces KwaiVIR, combining synthetic video pairs and wild S-UGC videos with objective metrics and subjective evaluation of fidelity, perception, and temporal consistency.
Results
RedMediaTech ranked first in both tracks, with a subjective score of 3.8525 and an objective score of 61.7395.
Takeaways & Limitations
The results show that comprehensive evaluation is important because perceptual quality, distortion fidelity, and temporal consistency can favor different methods.
Abstract
from arXiv · showhide
This paper presents an overview of the NTIRE 2026 Challenge on Short-form UGC Video Restoration in the Wild with Generative Models. This challenge utilizes a new short-form UGC (S-UGC) video restoration benchmark, termed KwaiVIR, which is contributed by USTC and Kuaishou Technology. It contains both synthetically distorted videos and real-world short-form UGC videos in the wild. For this edition, the released data include 200 synthetic training videos, 48 wild training videos, 11 validation videos, and 20 testing videos. The primary goal of this challenge is to establish a strong and practical benchmark for restoring short-form UGC videos under complex real-world degradations, especially in the emerging paradigm of generative-model-based S-UGC video restoration. This challenge has two tracks: (i) the primary track is a subjective track, where the evaluation is based on a user study; (ii) the second track is an objective track. These two tracks enable a comprehensive assessment of restoration quality. In total, 95 teams have registered for this competition. And 12 teams submitted valid final solutions and fact sheets for the testing phase. The submitted methods achieved strong performance on the KwaiVIR benchmark, demonstrating encouraging progress in short-form UGC video restoration in the wild.
1. Introduction
Short-form UGC videos undergo diverse, entangled real-world degradations that make restoration difficult. The challenge addresses this gap with a benchmark combining synthetic and wild data and evaluating fidelity, perceptual quality, and temporal consistency.
- S-UGC videos captured under unconstrained conditions face acquisition, compression, transcoding, motion, and other real-world distortions.These degradations reduce perceptual quality and challenge practical restoration systems.
- Real-world S-UGC restoration is harder than conventional restoration because degradations are diverse and entangled rather than restricted to controlled categories.Practical methods must also preserve temporal coherence, perceptual realism, and cross-video generalization.
- Generative models can recover plausible details under complex degradations, but dedicated benchmarks for generative S-UGC restoration remain limited.A practical benchmark should evaluate synthetic-data fidelity, real-world perceptual quality, and temporal consistency.
- KwaiVIR combines synthetically distorted video pairs with real-world S-UGC videos and evaluates distortion fidelity, perceptual quality, and temporal consistency.The unified framework is intended to reflect restoration requirements on real short-video platforms.
2. Challenge
The challenge introduces KwaiVIR to represent both synthetic and authentic short-video restoration conditions. Its two-part evaluation separately measures objective quality and subjective fidelity, perception, and temporal consistency.
- KwaiVIR contains synthetic and wild videos, with 200 synthetic and 48 wild training videos, plus 11 validation and 20 test videos.The benchmark combines paired synthetic data with authentic real-world videos.
- The benchmark targets restoration that is faithful to reference content while remaining perceptually convincing in realistic short-video applications.
- The challenge evaluates synthetic videos for restoration accuracy and perceptual quality, while evaluating real-world videos for perceptual quality.Separate objective and subjective tracks accommodate different optimization goals.
- The objective protocol uses PSNR, SSIM, LPIPS, MUSIQ, and WarpError, while the subjective study assesses content fidelity, perceptual quality, and temporal consistency.Subjective evaluation covers 10 synthetic and 10 wild videos; the objective test evaluated 11 videos because of runtime limits.
3. Challenge Results
RedMediaTech ranked first in both subjective and objective evaluations, although the two rankings are not strictly aligned because teams could submit different results for each track. The results support comprehensive evaluation across perceptual quality, fidelity, and temporal consistency.
- RedMediaTech ranked first in the subjective evaluation with a score of 3.8525.The subjective score combines fidelity, perception, and temporal consistency.
- RedMediaTech ranked first in objective evaluation with a final score of 61.7395 and the best PSNR (30.7610), SSIM (0.8504), and LPIPS (0.1910).
- Subjective and objective rankings are not strictly aligned because participants could submit different restoration results for the two tracks.The rankings therefore reflect different optimization goals rather than directly comparable performance.
- The challenge results show that methods can have distinct advantages in perceptual quality, distortion fidelity, and temporal consistency.
4. Teams and Methods
The teams use diverse restoration strategies, including diffusion transformers, generative pipelines, lightweight adversarial models, and detail-aware latent diffusion. Their methods target perceptual quality, fidelity, temporal consistency, efficiency, and robustness to real-world degradations.
- RedMediaTech: RedMediaTech uses a two-stage single-step diffusion framework built on Wan 2.1’s DiT to balance perceptual quality and distortion-oriented metrics.The first stage uses Wan 2.1’s VAE and DiT with MSE and LPIPS losses.
- Generative restoration pipeline: Dual-branch generative restoration combines general restoration, pre-cleaning, DOVE, and RRDB-based fusion across two stages.The fusion network combines degraded inputs with intermediate outputs from the generative branches.
- Anchor-guided diffusion restoration: Anchor-guided restoration enhances sparse frames with ODTSR before STCDiT restores the full video using motion-aware segmented encoding.Adaptive segment boundaries address frames with severely degraded or suddenly lost local structures.
- MiAlgo LM: Mi-GenVR builds on Wan2.1-14B and injects low-quality priors through Self-Attention-Control and Cross-Attention-Control.The framework is designed to bridge pretrained video generation with high-fidelity restoration.
- BVI: BVI modifies TADSR with residual-noise refinement, detail-aware high-frequency and gradient losses, and ratio-capped residual regularization.The losses use a progressively activated warm-up schedule, while ratio capping stabilizes residual correction.
- xingyanan: FlashVSR-UGC-Causal compresses multi-step diffusion into a real-time one-step mapping through distribution matching distillation.The framework targets compression artifacts, low-light noise, and motion blur in short-form UGC videos.
- weichow: NAFNet-GAN is a 12.5M-parameter, single-stage residual U-Net using adversarial training, achieving PSNR 28.42 and LPIPS 0.279.It is trained exclusively on the KwaiVIR synthetic training data.
TaoMC2
The section title identifies a method named “The Trade-off between Texture Preservation and Denoising: A Robust Dual-Branch Diffusion Pipeline for Challenging Short-Form Videos.”
- The method is titled “The Trade-off between Texture Preservation and Denoising: A Robust Dual-Branch Diffusion Pipeline for Challenging Short-Form Videos.”
Video-Restorer
The section title identifies a method named “CoDiVSR: Rethinking What to Condition and What to Disentangle in Diffusion-based Video Super-Resolution.”
- The method is titled “CoDiVSR: Rethinking What to Condition and What to Disentangle in Diffusion-based Video Super-Resolution.”