InfVSR

Inference checkpoint for InfVSR: Toward Consistency-Driven Streaming Generative Video Super-Resolution (ICML 2026).

Project · Paper

Checkpoint

InfVSR.ckpt contains the InfVSR model weights in BF16.

File	Description
InfVSR.ckpt	InfVSR model weights, approximately 2.63 GiB
config.json	InfVSR architecture configuration
external_weights.json	Pinned download URLs for external dependencies

External Dependencies

The VAE and visual encoder weights are downloaded separately.

File	Purpose	Download
Wan2.1_VAE.pth	Wan2.1 video VAE	Download
DAPE.pth	DAPE visual conditioning adapter (SeeSR; Hugging Face mirror)	Download
ram_swin_large_14m.pth	RAM visual encoder	Download

Download into the models/ directory of the InfVSR inference code:

hf download Wan-AI/Wan2.1-T2V-1.3B Wan2.1_VAE.pth \
  --revision 37ec512624d61f7aa208f7ea8140a131f93afc9a --local-dir models
hf download alexnasa/SEESR DAPE.pth \
  --revision 4059ae0246bf52f5ac5017821bb6e59a6849a758 --local-dir models
hf download xinyu1205/recognize-anything ram_swin_large_14m.pth --repo-type space \
  --revision 7c76cec1377de0d46df8091a032c4b667aea4459 --local-dir models

Loading

Use the InfVSR implementation from the project repository. Load InfVSR.ckpt directly as the DiT weights.

Load the checkpoint with the provided architecture configuration:

import json
from pathlib import Path
import torch
from diffsynth_wan.wan_video_dit import WanModel

checkpoint_dir = Path("models/InfVSR")
config = json.loads((checkpoint_dir / "config.json").read_text())
weights = torch.load(
    checkpoint_dir / "InfVSR.ckpt",
    map_location="cpu",
    weights_only=True,
)
dit = WanModel(**config["model_config"])
dit.load_state_dict(weights, strict=True, assign=True)

For end-to-end inference, use the InfVSR pipeline together with the three external dependencies above.

Citation

@inproceedings{zhang2026infvsr,
  title={InfVSR: Toward Consistency-Driven Streaming Generative Video Super-Resolution},
  author={Zhang, Ziqing and Liu, Kai and Chen, Zheng and Li, Xi and Chen, Yucong and Duan, Bingnan and Kong, Linghe and Zhang, Yulun},
  booktitle={ICML},
  year={2026},
}

Acknowledgements

InfVSR builds on Wan2.1, DiffSynth-Studio, SeeSR, and Recognize Anything. External assets retain their respective upstream licenses.