CSomething like "a woman with sots running down the sides of her face". The caption shoukd tell the AI exactly what its looking at so it understands what it is.
Dhm, wouldn't mentioning the spots make them a variable? so if I afterwards prompt for her with name, the markings wouldn't be there unless mentioned?
CYoure training it, not prompting it.
CYou are creating an object that will update its weights for specific information when used. Put that specific information in you dataset lables
CAnd only use that object if you want it to know that specific information
Fit is easy
Fyou are very likely using shared vram
Fotherwise speed cant be that bad
Fi dont work with NSFW so dont know
Fi dont think needed
Fi am working on a new rare token
Faccording to t5 XXL
Fdumped its token
Diiiinteresting. 25 pics is enough to train a style for Flux.
Fit sounds low but you can try
Dtried it via the civitai trainer, it looks promising actually. trainable artstyles with only a few dozen pic could be amazing
Cthere are some options that allow you to just use one image, actually
LEy Furkan, I'm trying to Extract Flux Lora, but I only got this progress bar to 0, it's frozen
I don't know how to deal with the warning about torch, could you please help me?
I don't know how to deal with the warning about torch, could you please help me?
Lwhat " memory fallback enabled" is? and why is relevant to the training?
Lhow much time did it take?
Fit is around 30 second per it when batch size is 7
Fso 100 steps would be 3000 seconds
JOkay, please forget about the NSFW. It is not important.
Do you think it is worth to try a community Flux Checkpoint for training a Dreambooth Model? If yes, which would you recommend ? Or would you always recommend to use the base version? Best regards and thanks a lot
Do you think it is worth to try a community Flux Checkpoint for training a Dreambooth Model? If yes, which would you recommend ? Or would you always recommend to use the base version? Best regards and thanks a lot
D@Furkan Gözükara SECourses Flux anti-bleeding protection. look at right-most column. No more Furkan Obama 
Lthe issue was I put a "space" in the name of the output model, it wasn't a good idea
LI don't get it, could you tell more please? and a picture of the original woman would be great
Dleft is without lora, just Flux. the middle is regular lora training with the bleeding effect. you can see that not only the prompt "ohwx woman" was changed, but also "woman" by a lot. on the right is the lora with prior regularization: the prompt "woman" looks almost exactly like in the unchanged model with no lora on the left, only "ohwx woman" is changed
Mhow many regularzation images did you use? did you caption your training images?
DI did not use any reg images. it's a code change of the trainer.
Mwhat code did you change?
SVery interesting since prior reg means using reg images
S
Dprior just means original model. I guess the name of "prior preservation" parameter is named that because originally people used images generated by the prior model
SSo what did you change in order to avoid bleeding? Would love to experiment with this aswell
Fcan you eloborate more? last time i tried it wasnt have any improvement :d
Fyes i was gonna ask this
Fbecause it was not working at all last time i tested. which trainer code change fixed?
DOT, but it's experimental.
Dyou might be able to reach a good result with reg images too, because it does something very similar. not exactly though
Cyou do realize that because of what was done to flux - it really is not trainable, right?
Dit's not a secret but this is not a development channel. I've shared on the OT discord
Fso you modified OT code? can you give full details please we can forward to kohya
Malready done
Fcan you link
Fyou forwarded ? i dont see :d
FGitHub
Replace SD3Tokenizer with the original CLIP-L/G/T5 tokenizers.
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
Mdm on twitter i sent him
Fhe doesnt read there :d
Mhe does