Low-rank Adaptation for Fast Text-to-Image Diffusion Fine-tuning

[specblades]

Low-rank Adaptation for Fast Text-to-Image Diffusion Fine-tuning
https://github.com/cloneofsimo/lora

Main Features:

1. Fine-tune Stable diffusion models twice as faster than dreambooth method, by Low-rank Adaptation
   

2. Get insanely small end result, easy to share and download.
   

3. Easy to use, compatible with diffusers
   

4. Sometimes even better performance than full fine-tuning (but left as future work for extensive comparisons)
   

5. Merge checkpoints by merging LORA
   

It produces 3.5mb file, that can be merged with main model

"style of sks, baby lion", with disney-style LORA model.

"style of sks, superman", with pop-art style LORA model.

[ghost]

It's very interesting, but is there a description for toddlers somewhere? I donbt understand half of text in that github xD