lucidrains/denoising-diffusion-pytorch ? reverse-engineered prompt

Build me a Python package for training and sampling diffusion models on images, with a simple way to use a U Net model and a diffusion wrapper to train on batches of normalized images and then generate new samples later.

I want it to support the easy path where I can point it at a folder of images, set the image size, batch size, learning rate, number of steps, and have it train, save checkpoints, and log sample images along the way. It should also support faster sampling, mixed precision, and multi GPU training if available.

If it makes sense, include the 1D version too, so I can use the same style of API for sequence data as well as images. Keep the code clean and easy to import, with a small example in the README that shows training and sampling from scratch. If you need to look up current docs online for anything, go ahead.