SwiftVR

SwiftVR is a real-time, one-step streaming video-restoration model. It combines mask-free shifted-window attention with a restoration-aware autoencoder for causal chunk-wise inference. FlashDreams provides 2x and 4x post-processing presets and the standalone v2v-swiftvr application.

Quick Start

V2V

uv sync --package flashdreams-swiftvr --inexact
uv run --no-sync flashdreams-run-v2 v2v-swiftvr \
  --mode mp4 --output-path artifacts/swiftvr-2x.mp4 --timeout unbound -- \
  --video-path docs/source/_static/model_clips/wan22/wan22-ti2v-5b.avif

Demo Presets

V2V

Developer Details

• GPU: One CUDA-capable NVIDIA GPU.
• Python: 3.10 or newer.
• PyTorch: 2.9 or newer.
• FFmpeg: Required for MP4 input and output.

Post-processing presets

Cam2V and Interactive Drive commands select these presets with --postprocess-preset PRESET in their application arguments, after --:

Citation

If you use SwiftVR, please cite the original work:

@article{yan2026swiftvr,
  title={SwiftVR: Real-Time One-Step Generative Video Restoration},
  author={Yan, Jiaqi and Chen, Xiangyu and Zhong, Xinlin and Huang, Haibin and Zhang, Chi and Liu, Jie and Zhou, Jiantao and Li, Xuelong},
  journal={arXiv preprint arXiv:2606.09516},
  year={2026}
}