this is not a finetune, this is lora training, i went on his discord its just 512x512 lora training
this is not a finetune, this is lora training, i went on his discord its just 512x512 lora training with fal
Plol, I figured it might be something shady
Mhis discord is quite shady too i left
Fbalance matters
Fit doesnt work
Fcurrently loras are overfitting
Fmaybe due to flux nature
Fit needs more research
F512 is way lower quality people really doesnt test
Fyou have to test with exact same configuration and compare
Fi will prepare full fine tuning guide hopefully with Kohya
Falso for flux you can have very different angles in train dataset
Flike looking down
Flooking up
Fcompletey side view
Fit really does amazing job
Dmay be, sdxl was mastered to its limits, some things about flux are still a mystery
Fyep so true
Dtraning people/characters seem to be figured out mostly now, I'm curious, whats next? using the thousands of reg images to further finetune Flux dev?
Fi will do fine tuning with my reg images
Falso a style training
Fwill share style dataset too
Fand also then there is single image training challange :d
Fbut also there is fine tuning config / workflow
Fi bet it will be way better than lora
Ddid you try training with regularization but using no class? all your trainings looked amazing but all man end up looking like you
Dmay be doing that you maintain resemblance linking you to the token ohwx only and not to the token man
DI will watch your video now!!!!
Nis Q8 gguf flux model worse than Q5? becuse I have speed boost on Q8 and don't understand why
Fawesome
Fhere comparison :
Fthis is made for LLMs but same logic
Fso how am i gonna this ? so it will be ohwx token of me, and man token for reg images right?
Fgood idea i can test
Fbut 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
