In this paper, we propose OSDEnhancer, the first framework that achieves real-world STVSR in one-step diffusion. Given a low-resolution and low-frame-rate video as input, OSDEnhancer generates a high-resolution and high-frame-rate video.
OSDEnhancer begins with a linear initialization to establish essential spatiotemporal structures and adapt the model for one-step reconstruction. It then applies a divide-and-conquer strategy, introducing the temporal coherence (TC) and texture enrichment (TE) LoRAs that progressively specialize in inter-frame dynamics modeling and fine-grained texture recovery, respectively, while collaborating during inference for enhanced overall performance. A bidirectional VAE decoder employs deformable recurrent blocks to leverage the multi-scale structure of the vanilla VAE, enhancing latent-to-pixel reconstruction through joint multi-scale deformable aggregation and inter-frame feature propagation.
🔈News
📌 [Jun 2026] We made an important checkpoint fix on Hugging Face to correct the VAE config and Transformer LoRA key names. To reproduce the results reported in our paper, please use the latest checkpoint files 👉
Important Note: On June 25, 2026, we updated the Hugging Face checkpoint files to correct the VAE config and Transformer LoRA key names. To reproduce the results reported in our paper, please use the latest checkpoint files from Hugging Face.
By default, the inference script automatically loads the checkpoint from Hugging Face. For local checkpoint loading, the checkpoint directory should be organized as follows:
For stable inference, we recommend using a GPU with not less than 80GB of VRAM. We recommend setting spatial_scale = 4 and temporal_scale = 2. To use a local checkpoint, specify --ckpt_path. For long videos or high-resolution inputs, enable chunk-based inference by additionally setting --chunk_length and --overlap, where --chunk_length should satisfy the form of 8N+1.
If you find this work useful for your research, please consider citing our paper 😊
shell
1@article{wei2026osdenhancer,
2title={Taming Real-World Space-Time Video Super-Resolution with One-Step Diffusion},
3author={Wei, Shuoyan and Li, Feng and Zhou, Chen and Cong, Runmin and Zhao, Yao and Bai, Huihui},
4journal={arXiv preprint arXiv:2601.20308},
5year={2026}6}
📕 License & Acknowledgement
This project is released under the Apache License 2.0. OSDEnhancer is built upon CogVideoX. We also sincerely thank the authors of DOVE, EvEnhancer, and RealBasicVSR for their excellent open-source implementations, which provided valuable references for this project.
