How to keep model weights loaded in GPU?

Hello, I'm using gen_img_diffusers.py and it's pretty cool to generate great images with that script, But when i want to generate a few images it re-loads the models (tokenizer, unet, etc) and lates too much between image generations.

Isn't there a way to keep all those information in GPU just like a Pipeline?

Start by reading gen_img_diffusers.py and trace how the tokenizer, UNet, and other models are initialized for each image generation. Check whether the script already exposes a reusable pipeline or generation loop. Done means multiple images can be generated without reloading the models into GPU memory between generations.

Written by the indexing model from the issue text.