---
license: other
license_name: bsd-3-clause-pytorch-tutorials
license_link: https://github.com/pytorch/tutorials/blob/main/LICENSE
tags:
  - kornia
  - super-resolution
---

# kornia/small_sr

Pretrained weights for the small sub-pixel super-resolution network used by `kornia.models.small_sr.SmallSRNet` (`upscale_factor=3`).

The network upscales a single-channel (luminance) image with an efficient sub-pixel convolution layer (Shi et al.,
CVPR 2016). It is the model from PyTorch's
[super-resolution tutorial](https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html).

## Weights

| File | Description |
|------|-------------|
| `superres_epoch100-44c6958e.safetensors` | 3x upscaling, the tutorial's checkpoint |

## Provenance

`superres_epoch100-44c6958e.safetensors` holds the state dict of PyTorch's
[`superres_epoch100-44c6958e.pth`](https://s3.amazonaws.com/pytorch/test_data/export/superres_epoch100-44c6958e.pth)
(sha256 `44c6958ea9df3886ef120bad1fc827ffb913e4e6f667b24dc7aea9b032b6df75`), unchanged: the same keys, dtypes,
shapes and bits. The kornia maintainers converted it (converter revision `0c3db846b2`) and checked that a model
loaded from either file gives bitwise-identical outputs. The sha256 of the converted file is
`84d6ae56ac55d03df3089b42bfc0dc6786b5ebb48a44d522bdc68d66d85cde8c`.

## License

The weights come from the PyTorch tutorials, whose repository is licensed under BSD-3-Clause; see
[`LICENSE`](LICENSE), copied from [pytorch/tutorials](https://github.com/pytorch/tutorials). No separate licence is
stated for the weights file itself.

## Citation

```bibtex
@inproceedings{shi2016real,
    author    = {Shi, Wenzhe and Caballero, Jose and Husz{\'a}r, Ferenc and Totz, Johannes and Aitken, Andrew P.
                 and Bishop, Rob and Rueckert, Daniel and Wang, Zehan},
    title     = {Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional
                 Neural Network},
    booktitle = {CVPR},
    year      = {2016}
}
```