MahoCommerce/Real_HAT_GAN_sharper-ONNX

Real_HAT_GAN_sharper (ONNX)

An ONNX export of Real_HAT_GAN_SRx4_sharper.pth from XPixelGroup/HAT, for the image editor of Maho. HAT is by Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao and Chao Dong, under the Apache-2.0 license.

• Scale: 4×. Input pixel_values, a float32 tensor of shape [1, 3, 192, 192] with values from 0 to 1. Output reconstruction, shape [1, 3, 768, 768].
• The input shape is fixed. Cut a larger image into tiles of 192 × 192 pixels.
• fp32 only. The model gives NaN values in fp16.
• config.json sets model_type to swin2sr, so that transformers.js loads the file with AutoModelForImageToImage. The model is HAT, not Swin2SR.

Exported with spandrel 0.4.2 and torch.onnx.export (opset 17).

@InProceedings{chen2023activating,
  author    = {Chen, Xiangyu and Wang, Xintao and Zhou, Jiantao and Qiao, Yu and Dong, Chao},
  title     = {Activating More Pixels in Image Super-Resolution Transformer},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2023},
}