Skip to content

README.md

Latest commit

 

History

History
115 lines (78 loc) · 4.25 KB

README.md

File metadata and controls

115 lines (78 loc) · 4.25 KB
Raw

SwiftVR Deployment Guide

English | 简体中文

Tech Blog

SwiftVR supports image and video restoration and super-resolution. Both use --model_cls swiftvr --task sr.

1. Model Download and Conversion

After setting up the environment, download and convert the official weights from H-oliday/SwiftVR, running these commands from the LightX2V repository root:

hf download H-oliday/SwiftVR --local-dir /path/to/SwiftVR
python tools/convert/examples/convert_swiftvr.py \
  --source /path/to/SwiftVR \
  --output /path/to/SwiftVR_lightx2v

2. Environment Setup

Choose the LightX2V Docker image for your GPU. Pull the corresponding image in a working directory on the host:

# H100, A100, etc.
docker pull lightx2v/lightx2v:26062001

# RTX 5090
docker pull lightx2v/lightx2v:26062001-cu130-5090-fix-260921

Start a container using the selected image:

docker run --gpus all -itd --ipc=host --name [container_name] -v [mount_settings] --entrypoint /bin/bash [image_id]

Inside the container, clone the source and install the dependencies:

git clone https://github.com/ModelTC/LightX2V.git
cd LightX2V
python -m pip install "av>=17" decord imageio-ffmpeg huggingface_hub

To set up your own environment, see the installation guide.

3. Offline Inference

Before running, set lightx2v_path (repository path), model_path (converted model path), config_path (configuration path), and CUDA_VISIBLE_DEVICES (GPU IDs) in the corresponding script, along with the input path and output_path.

Select a configuration under configs/swiftvr/ for your GPU:

GPU Configuration
H100 h100/swiftvr.json
A800 a800/swiftvr.json
RTX 5090 5090/swiftvr.json

To enable compilation, add "use_compile": true to the configuration file or use the corresponding swiftvr_compile.json.

# Image super-resolution
bash scripts/swiftvr/inference/run_swiftvr_image_sr.sh

# Video super-resolution
bash scripts/swiftvr/inference/run_swiftvr_video_sr.sh

Provide exactly one of --image_path or --video_path, with a .png or .mp4 output file, respectively. The default scale factor is 2 (--sr_ratio 2); change it as needed or use --size 1440 2560 to specify the output dimensions (height, width). If both are set, size takes precedence.

For multi-GPU video inference, use h100/swiftvr_parallel.json as a reference. For four GPUs, set parallel.chunk_p_size=4, CUDA_VISIBLE_DEVICES=0,1,2,3, and --nproc_per_node 4, then run:

bash scripts/swiftvr/inference/run_swiftvr_video_sr_4gpu.sh

4. Service Deployment and API Usage

Set the repository path, model path, configuration path, and GPU IDs in the corresponding server script, then choose a single-GPU or four-GPU launch:

# Single GPU; uses the H100 compile configuration by default
bash scripts/swiftvr/server/start_server.sh

# Four GPUs; match the parallel settings as described above
bash scripts/swiftvr/server/start_server_4gpu.sh

The default port is 8000. One sr service supports both image and video super-resolution.

Before sending requests, edit url and message in the client scripts: use absolute input paths accessible to the server and set save_result_path to a server-side output file. Use sr_ratio to set the scale factor or size: [height, width] to specify output dimensions, following the same rules as offline inference.

Once startup and warmup finish, run the corresponding request script from another terminal:

# Image super-resolution
python scripts/swiftvr/server/post_image.py

# Video super-resolution
python scripts/swiftvr/server/post_video.py

📜 Citation

If you use SwiftVR in your research, please cite the SwiftVR paper:

@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}
}