Spaces:

adamelliotfields

/

diffusion

like 9

Running on Zero

In #20, looked at using AuraSR for upscaling. It was too heavy and took 20 seconds to upscale an image (with very high quality). Discovered the ai-forever/Real-ESRGAN PyTorch implementation, and it is very lightweight, both code and weights. Loading the model does take a few seconds, but generation is fast. To get it to work with Diffusers, the pipeline output_type had to be set to "np" to output to a NumPy array instead of a PIL image.