CVPR NTIRE 2026 UGC Short‑Video Restoration Challenge: Winning Solutions and Highlights

The short‑video user‑generated content (UGC) restoration problem differs fundamentally from traditional video repair because real‑world UGC suffers from uncontrolled capture conditions, motion blur, low lighting, compression artifacts, and temporal inconsistencies.

The NTIRE 2026 challenge built the KwaiVIR benchmark, contributed jointly by the University of Science and Technology of China and Kuaishou. The dataset contains 200 synthetic degraded videos and 48 real‑world UGC videos for training, 11 validation videos, and 20 test videos.

Two evaluation tracks were defined: a subjective track scored by professional judges on fidelity, perceptual quality, and temporal consistency, and an objective track measured by PSNR, SSIM, LPIPS, MUSIQ (no‑reference quality) and WarpError (temporal error). Both tracks allowed participants to submit results for the same videos.

In total, 95 teams registered from 14 domestic and overseas institutions and five technology companies. Only 12 teams submitted valid test results.

Overall, RedMediaTech (from Xiaohongshu) achieved the highest subjective score (3.8525) and the best objective metrics (PSNR 30.7610, SSIM 0.8504, LPIPS 0.1910). The ranking discrepancy between the two tracks highlights the importance of combined subjective‑objective evaluation for generative video restoration.

Top‑3 method details :

RedMediaTech used a single‑step diffusion framework based on Wan 2.1 DiT. Training proceeded in two stages: the first stage initialized with Wan 2.1 VAE + DiT and optimized with MSE + LPIPS; the second stage replaced the VAE with a stronger Qwen‑Image VAE and added a shortcut connection to preserve spatial details. DiT employed 3‑D RoPE for temporal encoding, and data augmentation such as frame skipping and random cropping improved robustness. Training used 8×H20 GPUs for about five days (AdamW, lr 5e‑5) followed by a second stage at lr 2e‑5 for one day, plus ~10 000 high‑resolution internal video clips.

TaoMC2 (Alibaba‑TAO & Beihang) built a two‑stage generative repair pipeline on a text‑to‑video diffusion model. The first stage featured a dual‑branch repair module: a general real‑world repair branch and a pre‑cleaning branch (with an optional open‑source DOVE model for the objective track). The second stage fused the original degraded input with the two branch outputs using an RRDB network and an anchor‑fusion strategy to balance degradation removal and detail preservation.

STCVSR (Nanjing University, Hunan University & OPPO) combined pretrained STCDiT and ODTSR models. ODTSR enhanced sparse anchor frames (one every 25 frames) to provide structural guidance, while STCDiT performed full‑video restoration using motion‑aware VAE latent variables. The pipeline dynamically adjusted segment boundaries for severely degraded frames and required no extra training, relying solely on pretrained weights.

Additional notable entries included Gen‑VSR (Video‑Restorer) (Sun Yat‑sen, Xi’an Jiaotong & Nanyang Technological) with a two‑stage pipeline (temporal alignment + ONNX up‑sampling followed by DOVE super‑resolution) and Lucky one (Beihang & Tsinghua) which fine‑tuned CogVideoX for single‑step diffusion video repair, introduced latent‑pixel supervision, achieved up to 28× inference speedup, and maintained temporal consistency through latent‑space training.

The organizers acknowledge support from the National Natural Science Foundation of China, China Postdoctoral Science Foundation, Anhui Postdoctoral Research Fund, basic research funds of central universities, and the Humboldt Foundation.