I just extracted a base dimension rank 192 & alpha 192 rank LoRA from my Stable Diffusion XL (SDXL) U-NET + Text Encoder DreamBooth trained model (with my newest very best Text Encoder config) and the result is really really good.
Of course not good as full model :)
LoRA is 1.4 GB, DreamBooth is 6.5 GB
LoRA size totally depends on LoRA rank
Same Seed — Not cherry picked
After Detailer (ADetailer) is on 0.4 denoise
I haven’t tested different Rank effect of LoRA extraction yet
Used Kohya GUI LoRA extract
Probably with extraction rank testing better results can be obtained
You can never get such good LoRA with LoRA training
Check right top of grids to see whether extracted LoRA or Full model
1st img : full
2nd : lora
3rd : full
4th : lora
You can download config from here : https://patreon.com/posts/89213064
You can see quick how to use tutorial here : https://youtu.be/EEV8RPohsbw
