Skip to main content
Back

Kuaishou's KwaiVIR Benchmark Pushes AI Video Restoration Into Real-World Chaos

Kuaishou and USTC drop KwaiVIR, a benchmark for restoring gritty short-form videos, with 12 teams pushing single-step diffusion into the wild.

Pro

  • NTIRE 2026 launched KwaiVIR, a new benchmark for restoring real-world short-form UGC videos.
  • Dataset: 200 synthetic training clips, 48 wild clips, 11 validation, and 20 test videos from USTC and Kuaishou.
  • 95 teams registered, 12 submitted final solutions across subjective user-study and objective metric tracks.
  • RedMediaTech won both tracks using a single-step Wan 2.1 DiT with a Qwen-Image VAE swap.
  • Nearly all top methods fine-tune large video diffusion backbones (Wan, CogVideoX, SeedVR2, FlashVSR) with LoRA.
  • Code, data, and fact sheets available on GitHub.

Short-form videos on platforms like Kwai and TikTok look terrible in ways that classic super-resolution models were never trained for. Between shaky phone cameras, aggressive platform transcoding, and layer after layer of re-encoding, the degradations arrive tangled together rather than as clean synthetic noise. The NTIRE 2026 challenge, hosted at CVPR, pushed generative restoration models into exactly that mess, and the results report is now out.

The competition introduced KwaiVIR, a short-form UGC (S-UGC) video restoration benchmark contributed by USTC and Kuaishou Technology, containing both synthetically distorted clips and real-world footage in the wild. 95 teams registered and 12 submitted valid final solutions and fact sheets for the testing phase. Data splits, evaluation code, and team fact sheets are available on GitHub.

Why in-the-wild UGC breaks normal restoration

S-UGC clips collected from real short-video platforms usually suffer from diverse and entangled degradations introduced by consumer-level acquisition, compression, transcoding, and platform-side processing pipelines. A model trained on the usual bicubic-downsample-plus-JPEG recipe simply does not generalize to that. Restoration in the wild becomes particularly hard once both perceptual realism and temporal coherence enter the picture.

Temporal coherence is the awkward part. Image models can be pushed to hallucinate believable detail with diffusion priors, but the moment you run them frame by frame on video, textures start flickering. That is the tension the challenge was designed around: generative priors strong enough to invent missing detail, controlled enough not to shimmer.

Inside the KwaiVIR benchmark

The released data include 200 synthetic training videos, 48 wild training videos, 11 validation videos, and 20 testing videos. The synthetic pairs give teams clean ground truth to train against, while the wild footage has no reference and forces models to prove they work when nobody is holding a pristine copy for comparison.