HAT Super Resolution

HAT Super Resolution is a local AI image upscaling project for developers, researchers, and people working with low-resolution images. It provides pretrained Hybrid Attention Transformer models and code for training your own.

• 4x image upscaling: Pretrained models enlarge images to four times their original dimensions, including GAN-based models for real-world images.
• Fidelity or sharpness: Real_HAT_GAN_SRx4 favors fidelity to the source, while Real_HAT_GAN_SRx4_sharper favors sharper-looking results.
• Smaller model: HAT-S provides a smaller alternative to the full HAT model.
• Limited GPU memory: Tile mode processes images in sections to reduce GPU memory demands during upscaling.
• Custom training: Training code and configurations let researchers and developers train super-resolution models on their own datasets.

The Python project is open source under Apache 2.0 and runs locally using PyTorch and BasicSR. You can also run its models through chaiNNer. Pretrained weights are available through Google Drive or Baidu Netdisk; local processing saves output images on your machine. A separate hosted Replicate demo provides 4x upscaling.

Training needs substantially more GPU memory than inference. The supplied default training setup uses about 20 GB of memory per GPU.