SparkVSR: Interactive Video Super-Resolution via Sparse Keyframe Propagation

Jiongze Yu¹, Xiangbo Gao¹, Pooja Verlani², Akshay Gadde², Yilin Wang², Balu Adsumilli², Zhengzhong Tu^†,1

¹Texas A&M University ²YouTube, Google
^†Corresponding author

> 💡 **Your ⭐ star means a lot to us and helps support the continuous development of this project!** # ComfyUI_SparkVSR_SM ----- [SparkVSR](https://github.com/taco-group/SparkVSR): Interactive Video Super-Resolution via Sparse Keyframe Propagation Update ----- * 修复一些错误，超大视频请按工作流的note来调整tile数值 * Preprocess the image node to connect to the pisa_sr node to load the model, or use other methods to upscale a single-frame image, that is, enable reference image mode; otherwise, it is non-reference image mode. ref_indices are the sequence frame numbers of the reference image in the source video. For example,' 0, 'is the first frame. You can also input 0 to only reference the first frame to save resources and achieve better results. * 预处理图片节点接入pisa_sr 节点加载的模型或者你用其他方法放大的单帧图片，即开启图模式，否则是非垫图模式，ref_indices是垫图在源视频的序列帧序号，比如0就是首帧，你也可以输入'0,'(注意有逗号) 只参考首帧以节省资源获得更好的效果. 1.Installation ----- In the ./ComfyUI/custom_nodes directory, run the following: ``` git clone https://github.com/smthemex/ComfyUI_SparkVSR_SM ``` 2.requirements ---- ``` pip install -r requirements.txt ``` 3.checkpoints ---- * cog unet merged or gguf [links](https://huggingface.co/smthem/SparkVSR-GGUF/tree/main) #f32 or bf16 or gguf * cog vae [links](https://huggingface.co/JiongzeYu/SparkVSR/tree/main) * if use pisa_sr need [stable-diffusion-2-1-base](https://www.modelscope.cn/models/stabilityai/stable-diffusion-2-1-base) vae and unet,and pisa_sr.pkl[google](https://drive.google.com/drive/folders/1oLetijWNd59xwJE5oU-eXylQBifxWdss) ``` ├── ComfyUI/models/ | ├── diffusion_models/ | ├──sd21base-f32.safetensors # 3.22G or fb16 if use pisa_sr to refer image optional 可选，使用 pisa_sr超分参考图用 | ├──SparkVSR-S2-F32.safetensors # 20.7G or fb16 | ├── gguf/ | ├──SparkVSR-S2-Q8_0.gguf optional 可选， | ├── vae/ | ├──CogVideoX1.5-5B-I2V-VAE.safetensors #822M | ├──sd21vae.safetensors # 319M if use pisa_sr to refer image optional 可选，使用 pisa_sr超分参考图用 | ├── loras | ├──pisa_sr.pkl # 32M if use pisa_sr to refer image optional 可选,使用 pisa_sr超分参考图用 ``` 4.Example ---- ![](https://github.com/smthemex/ComfyUI_SparkVSR_SM/blob/main/example_workflows/exampleir.png) ![](https://github.com/smthemex/ComfyUI_SparkVSR_SM/blob/main/example_workflows/exampS2.png) 5.Citation ---- ``` @misc{yu2026sparkvsrinteractivevideosuperresolution, title={SparkVSR: Interactive Video Super-Resolution via Sparse Keyframe Propagation}, author={Jiongze Yu and Xiangbo Gao and Pooja Verlani and Akshay Gadde and Yilin Wang and Balu Adsumilli and Zhengzhong Tu}, year={2026}, eprint={2603.16864}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2603.16864}, } ```

Name	Type	Description
unet	COMBO	1 options: none
vae	COMBO	1 options: none
pkl	COMBO	1 options: none
dtype	COMBO	3 options: bfloat16, float16, float32

SparkVSR_SM_SRModel

Inputs (4)

Outputs (1)