Write a doc for centralizing our LoRA efforts

[sayakpaul]

Currently, we have incoming support for LoRA to fine-tune with DreamBooth and Text2Image.

These things enable the community to run fine-tuning on consumer GPUs which is a big enabling factor for many use cases.

So with that, I think it's important to put together a doc highlighting these efforts. This doc could live under https://github.com/huggingface/diffusers/tree/main/docs/source/en/training (as suggested by @patil-suraj).

I will start working on this doc next week. If anyone has any suggestions, feel free to comment. Opening this issue thread so that things don't get slipped through the cracks.

[brucethemoose]

While its for a completely different implementation, I noticed some helpful bits in this: https://rentry.org/lora_train

For instance:

• A note that network_dim has a significant influence on model quality
• A simple image showing the directory structure for training data: https://i.imgur.io/zpbs8FB_d.webp?maxwidth=640&shape=thumb&fidelity=medium
• A minimal vram consumption example config.
• Sane recommendations for number of epochs, LR and such.

[sayakpaul]

Thanks so much for the suggestions! I will definitely keep all these in mind :)