i am gonna do with kohya
i am gonna do with kohya
�Oh ok it's becasue I haven't really seen anyone yet with an actual working model. I look forward to your research in that area though 
Fye it is next research
Fwith 8 gpus :d
Fi expect to go as low as 12 gb
Fbut we will see
Fthere are some serious optimizations
Dyes I did, no class, now I balanced better the dataset, and I will make a test with regularization images from the class person because I have different ages and genders, and let's see, my dataset was a little unbalanced all man looked like me even I didn't used any class
PDude on reddit suposedly fine tuned it. Link is a discord link that doesn't work :/
https://www.reddit.com/r/StableDiffusion/comments/1f40dvc/finetunning_the_flux1_for_your_own_data/
https://www.reddit.com/r/StableDiffusion/comments/1f40dvc/finetunning_the_flux1_for_your_own_data/
Reddit
Explore this post and more from the StableDiffusion community
M@Dr. Furkan Gözükara have you tried this prompt Anime illustration portrait of ohwx man, forest with lake in the background, high quality, beautiful face, action pose, clear focus, detailed and fantasy setting
D@Dr. Furkan Gözükara but I have to say same dataset same caption including class on sdxl no bleeding at all, I trained 20 or more people at the same time on a single sdxl lora and no bleeding at all, something is missing, idk
Mthis 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
