<@205854764540362752> I saw you mentioned you were doing research on the bleed problem, I was wonde
@Furkan Gözükara SECourses I saw you mentioned you were doing research on the bleed problem, I was wondering if there was any new information?
YI have a Flux fine-tuned on a person, and I'm trying to pair with a LoRA trained on clothing. The face is getting really messed up unless I turn off the last lora block, and then the clothing lora isn't nearly as good, but the face works.
YI've now cropped out the faces in my clothing LoRA dataset, hoping that'll help
Fthey didnt work
YDo you think only training specific LoRA blocks would help me, just cropping the faces out?
Fi tested it
Fit doesnt help
Fcertain blocks
Feven single layer training still turns entire model into your face :d
ZCould someone tell me if stylistic Loras are applied correctly to your fine-tuned models? Because on my own model, Loras don't work with the same settings that work on the shared flux.dev model. ... If it turns out that Loras don't work with fine-tuned models, that would be a failure...))
ZI understand this is a real problem with fine-tuned models?? Why was this not mentioned anywhere? Creating custom Loras and applying them to my characters was the main plan... Why did no one say a word about this? This is a mega failure...
Z
Y@Zhigarev yeah this has been a real issue for me. The only thing that helps is disabling the last double block of the LoRA but it also really hurts the style
ZUnfortunately, this destroys the entire pipeline I had planned for future potential clients, as I won’t even be able to apply something like a branded armored t-shirt from a specific company to them... let alone a vintage photography style, like in this example...
ZWhat are the options for this mistake@Furkan Gözükara SECourses
Floras overwriting your character
Fbecause they are overfit
Zokey whats the solution ?
ZYou do understand that people mostly use models with Loras, right?
Fthey use to generate random stuff
Fnot certain people
Fthat is the whole world difference
ZWell, to create serious works, the consistency of the character and the uniformity of the style are crucial.
ZLet’s focus on how this issue can be resolved?
ZIs it possible, for example, to blend a style or object Lora into your custom model? (I’m not talking about another character or a man.)
Fnormally a good trained style should work
Fbut civitai loras or models are all overfit
Fhow can you solve is a good research
Fmost likely training style and person together
ZHow to reduce the size of an extracted LoRA from 11GB to normal size around 600MB + - more less ?
Fextracted lora loses quality when you reduce the lora network rank
Fonly way is that
Falso save as fp8
ZI performed the extraction in FP8, and the size remained the same at 12GB.
ZThat is, from the FP8 model, which became 12GB, I performed LoRA extraction, and the size did not change at all.
Zsame 12 gb FP8 to 12 gb Lora...
Fwell that is just impossib le you made mistake somewhere
Fbest lora with fp8 is like 2-3 gb
Zwhats wrong here? in the step 2 i choose my finetuned FP8 model
Fsave precision
Fmake it fp8
ZAlright will try, it's just that this isn't mentioned anywhere, or did I miss it??
Fit is for saving space not mandatory
Ffloat is best quality
ZThere is no option to select FP8... only FP16, BF16, Float, and None.
ZThe 11GB size of LoRA… that’s why it’s better in quality. There’s no magic here. If LoRA weighed 1GB while maintaining quality, that would be a real breakthrough. But models with a size of 23GB, where LoRA doesn’t work properly, and extracted LoRAs at 11GB, are a significant negative effect that needs to be known in advance.
Fok then you need 2 steps
ZWhich steps?
Fconvert to fp8
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
ZIt reduced the size, but the image is completely black.
