cszn/BSRGAN — reverse-engineered prompt

Build me a simple Python tool for improving low quality photos with BSRGAN.

I want to drop one or more blurry, compressed, or old looking images into an input folder, run a command, and get sharper higher resolution versions in an output folder. It should use the pretrained BSRGAN model for real image super resolution, and include an easy way to download the model if it isn’t already there.

Please make the setup friendly for someone who isn’t a machine learning expert. Include clear install steps, a simple test command, and sensible defaults for scale 4 enhancement. Keep the existing research code intact where possible, but clean up the workflow so it feels like a usable photo upscaler rather than a paper demo.

If CUDA is available, use the GPU, otherwise fall back to CPU with a clear message. Add basic error handling for missing files, unsupported image formats, and missing model weights. Include a small README section explaining how to run it on my own images.