but the issue is flux itself has inner captioning
but the issue is flux itself has inner captioning
Fso it still caption me as man :d
Nif it's only size issue, and speed boosts almost twice (6,3s to 3,7s/it) I will just need bigger ssd)))
Nthx for explaining
D@Dr. Furkan Gözükara After seeing your tutorial I am confused. We dont need to caption our training images for flux lora, right? (To achieve better quality)
SI've done a few character loras now and just had to put in the trigger name then ,woman. And I don't think you even needed 'woman'
DSo no point in using something like joycaption to caption all images meant for training?
SI did that with the first couple, then just did the trigger word with the next few. No appeciable difference imo
DOk thanks for confirming
Ffor character yes just ohwx man or woman
Ffor style i will test
Fyes not at this moment
Fwe will see in fine tuning
DThe vae path, T5-XXL and Clip-L Paths options are not showing in my latest installed Kohya GUI
DThe options are visible in your tutorial of v19
Fselect model first
Fit will become available
Fselect dev fp16 flux model
DReally???? ohh that's cool!! but a huge problem at the same time, must to be be a way to override this, its ok if the model recognize you as a man but not any man only "ohwx man"
Fi dont know any overwrite way atm
Fit is how it is designed
Fbut fine tuning may overcome this issue
DI don't understand how it works , to auto caption in have to analyze the images first caption them and then train, but I don't think is doing that, may be is the way that organize its weights, and it auto detect that the patterns relates to the weights that overlaps with the concept of man, but really idk
Fit uses attention of both image itself (this is where auto captioning happens in neural network) and the prompts we provide
DI edited my message to make it more clear, hehe
Dok I see your point
Dthis is documented somewhere or everybody is making assumptions?
Fwell i asked this to top developers and 2 said similar stuff
Fi think people knows it from sd3 architecture
Fmy experiment also proves this
Fi didnt use man class
Fit perfectly learnt
Fand man prompt is still fully impacted
Fbut car prompt is not
DI'm training same parameters no class and regularization images, in the first samples 5 epochs nobody looks like me, this is good I think, the resemblance must apear in the next epochs
Dyes but that happens with sd and sdxl too at a some extent, it generalizes if you train with your name only it bleeds to people even to other genders if you overtrain enough
Fsd and sdxl doesnt have exact same mecahnism though
Fso if you dont caption it will make huge impact
Ffor example even though i fully captioned, it reduced the likelineess very small. impact on sdxl way bigger
Dthat's true
DLets see, it needs more research, It will be cool to know exactly how the base model was trained
Fi am training turbo model
Flets see what comes out of it
Nwhat differences are you seeing\expecting? starting with same configs used on dev? how does speed compare?
Fstill training. i dont expect it to work :d
MHello, trying to train with 3080 10Go on Linux but even with the config Rank_9_7514MB, my GPU is always OOM at the "move text encoders to gpu" step
Many idea ?
Fyou loaded config into dreambooth tab at least once
Fso extract a new config
