VRT
A transformer that restores whole video clips in parallel — super-resolution, deblurring, denoising, interpolation, and space-time upscaling in one model family.
About
VRT (arXiv 2201.12288, revised June 2022) is the Video Restoration Transformer from Jingyun Liang's team at ETH Zurich with Meta collaborators. Where recurrent models restore frame by frame, VRT predicts multiple frames in parallel using temporal mutual self-attention: the sequence is split into small clips, mutual attention jointly handles motion estimation, alignment, and fusion inside each clip, and the clip grid is shifted every other layer so information crosses clip boundaries. A parallel warping module then fuses neighboring-frame features.
The authors report gains of up to 2.16 dB over prior methods across up to fourteen benchmark datasets (the v2 revision expands the task set), spanning video super-resolution, deblurring, denoising, frame interpolation, and space-time super-resolution. The code and pretrained checkpoints are on GitHub. The honest caveats: transformers of this size are heavy to train and run, and every number is author-reported — but as a unified restoration backbone it has been widely cited and re-implemented.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (arXiv 2201.12288v2): up to +2.16 dB over prior SOTA across up to 14 datasets covering VSR, deblurring, denoising, interpolation, and space-time VSR
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: paper arXiv 2201.12288 (Liang et al., ETH Zurich/Meta), official repo github.com/JingyunLiang/VRT with code and checkpoints. Strengths: unified multi-task restoration transformer, strong author-reported margins, widely cited. Docked for: (1) all headline numbers are author-reported with no independent audit; (2) heavy compute requirements limit practical adoption; (3) repo activity has slowed and there is no hosted demo or API. License: research code; repo license terms unverified.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv 2201.12288v2 (Liang et al., ETH Zurich/Meta); 5 restoration tasks
- · Author-reported up to +2.16 dB across up to 14 datasets
- · Official repo JingyunLiang/VRT: code and checkpoints
- · Heavy compute; no hosted demo or API; repo activity slowed