Lightweight VAE
Efficient encoding and LQ-conditioned decoding reduce autoencoding overhead, with cached features for incremental video processing.
FastVREfficient Streaming Video Restoration
with One-Step Diffusion
FastVR is a one-step diffusion framework for video restoration, supporting arbitrary-scale super-resolution, and streaming inference for long videos.
1 Alibaba Group2 Xidian University
11 FPS1080p on NVIDIA H20
21.96 GBPeak GPU memory
1 stepFrom LQ to restored video
StreamingBounded temporal memory
Reported model runtime on a single H20 GPU; excludes file I/O and model loading.
01 / LONG-VIDEO RESULTS
Long-video restoration.
AIGC and Real-World videos.
Drag to compare input and FastVR enhancement.
Input
FastVR
0:00
0:00
SYNCED PLAYBACK
Choose a long video
01 / 08
Select a video, then press play to load its input and FastVR result.
If playback is unavailable, access the video files on GitHub.
Diffusion-based video restoration recovers realistic details, but its large-scale use is limited by the cost of VAE encoding and decoding and the quadratic complexity of full attention.
FastVR combines a lightweight VAE with chunk-wise causal attention to enable efficient, one-step streaming restoration. Continuous trajectory learning and velocity consistency regularization improve restoration quality during training.
Experiments on synthetic and real-world benchmarks show strong perceptual quality and temporal consistency, with 11 FPS at 1080p on a single NVIDIA H20 GPU.
02 / UNDER THE HOOD
Designed for one-step diffusion.
Efficient inference. Restoration-oriented training.
Chunk-wise causal attention
Bidirectional interaction within each chunk; bounded preceding context across chunks. KV Cache reuses history during streaming.
Latent-to-pixel training
Latent trajectory learning and velocity consistency are followed by pixel-space L1 and DISTS supervision.
03 / LESS WAITING. MORE VIDEO.
Quality meets efficiency.
1080p video · single NVIDIA H20 GPU
Inference speed
FPS · higher is better ↑Peak GPU memory
GB · lower is better ↓Single H20 GPU · 1080p. Runtime excludes file I/O and model loading.
04 / MEASURED, NOT JUST SEEN
Across five benchmarks.
FastVR achieves the highest MUSIQ and CLIP-IQA scores across all five reported datasets using one-step diffusion. Full-reference and no-reference metrics capture different aspects of restoration quality.
SPMCS · SyntheticBestSecond best
↑ Higher is better. ↓ Lower is better. Ties share the same rank at the reported precision. Values are reproduced from the paper.
A CLOSER LOOK
Visual comparisons
Original PNG comparisons. Click to inspect fine details.
Video comparisons
01 / 06
BibTeX
@misc{chen2026fastvrefficientstreamingvideo,
title = {FastVR: Efficient Streaming Video Restoration with One-Step Diffusion},
author = {Xiaoxu Chen and Qin Yang and Haoran Bai and Sibin Deng and Ying Chen},
year = {2026},
eprint = {2609.36757},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2609.36757}
}