FLUX: My take with another technique on Emma Watson with FLUX.
Can probably lower the weight a bit to like 0.8
Prompt her with just "woman" or "Jenny".
Enjoy!
My take on the FooBar of generative AI: Emma Watson
Hope you enjoy :)
Description
FAQ
Comments (12)
Correction, all of your newest DoRAs have the "LoRA" label in the "type" category.
Wait, there is a DoRA category? I'm gonna change that. Thank you!
@CrunchyBanana Yup, there is! No problem! 👍
Hello, I'm curious abot which model are you using for tranining, Your lora's look amazing.
I'm using the same train of thought I used with my SD1.5 Loras:
Train on the model, that probably most models have merged into them. In my case: Ponyrealism
but to clarify a bit further:
It's not the model that determines the quality that much. Sometimes I make some small merges of different models for fun and training.
99% of the quality is to spend some time with the source images. I'd say about 50% of the time for my Loras I spend on preparing the dataset. The results are always nearly the same, no matter the model or optimizer you use.
@CrunchyBanana Thank you, For me training pony realism (any model) It's becoming a pain in the ass. For some reasonn Euler A it's the sampler tha works better for my loras. But Euler A looks soft and not sharp.
@Hikarias ye Euler is kinda smoothing everything. DPM++ and Restart are nice with realistic Pony. Also Karras and Align your Steps as Scheduler work well.
@CrunchyBanana Yeah I know, But for some reason DPM++ Karras, It's not working well with my lora's I'm not sure why, I have tried using onetrainer and kohya with not good results. Using prodigy optimizer and cosine or constant. 16 rank 16 alpha.
@Hikarias Use Prodigy with cosine. iirc it doesn't adapt when set on constant. But whatever you pick: Make sure your source data is excellent :)
Try Restart/Automatic as Sampler. Its slow, but kinda foolproof once you found a fitting cfg
@CrunchyBanana Yeah I was using Prodigy with cosine, 1 LR, But with the same problem, Probably it's something else. Anyway thanks for your advices! I don't want to waste more your time :)
@Hikarias no problem. Shoot me a message if you have further questions. I'll try to help if I can :)