Code for training the model about pytorch-stable-diffusion

[hkproj]

Hi! I don't have any immediate plan to write the training code, especially because this code is meant for learning the model and not to be used in production. If you'd like to train the model for fine-tuning or for production, I highly recommend using Hugging Face.

If you want to learn how training works for Stable Diffusion, then, given that we are using the DDPM formulation in my video, you only need to train the using the Algorithm described in my video on Diffusion Models, in which I also provide the code for the training algorithm. The implementation of that video can be found in my repository about the DDPM paper.

The only difference is that you also need to provide the time step and the caption of the image you're training the model upon, since they're both inputs for the UNet.

But as I've said before, for training a model from scratch, I highly recommend using open source libraries like Hugging Face, because it also takes into consideration all the data augmentation needed for the model. Be aware that to train a working Stable Diffusion from scratch that can generate meaningful outputs, it needs millions of images, which may take forever if you're using your personal computer/laptop.

[TWang1017]

yesss, appreciate that if anyone can give a simple training code

[Inkyl]

I'm also looking forward to finding a training code from scratch @hkproj