Complete the environment setup and download SeedVR2-3B. The local model directory must contain seedvr2_ema_3b.pth, ema_vae.pth, and pos_emb.pt.
| Task | Shell entry | Config |
|---|---|---|
| 3B video super-resolution | run_seedvr2_3b_sr.sh | seedvr2_3b.json |
| 3B image super-resolution | run_seedvr2_3b_image_sr.sh | seedvr2_3b.json |
| 7B video super-resolution | run_seedvr2_7b_sr.sh | seedvr2_7b.json |
| 7B video super-resolution, eight GPUs | run_seedvr2_7b_sp.sh | seedvr2_7b_sp.json |
For 7B, also download SeedVR2-7B and set dit_original_ckpt in the selected JSON to seedvr2_ema_7b.pth. Keep model_path pointing to the 3B directory for ema_vae.pth and pos_emb.pt.
Set lightx2v_path, model_path, video_path (or image_path), and CUDA_VISIBLE_DEVICES in the selected Shell script. Run from the repository root:
bash scripts/seedvr2/run_seedvr2_3b_sr.shThe 3B video example uses --sr_ratio 2.0 and saves to save_results/output_lightx2v_seedvr2_sr.mp4. Edit --sr_ratio and --save_result_path to change the scale and output location. Image inputs use --image_path and a .png output.
Edit model_path, input_video_path, and save_result_path in seedvr_sr_inference.py, then run:
python scripts/seedvr2/inference/seedvr_sr_inference.pyThe Python example uses configs/seedvr/seedvr2_3b.json; its relative input and output paths are resolved from the current working directory. See the Python API guide for shared usage.