i think i could now recognise you blindfolded
i think i could now recognise you blindfolded
HI had previously tried manual captioning with GPT Vision too, but it was really tedious and torturous because my internet was very limited and I couldn't download a BLIP model, especially with large datasets 
VIt doesn't do so well vs certain art styles like kawaii if it has lots of busy things happening, but for paintings its 95% great.
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=16)
Fbatch size 7 fails
Fmeanwhile fine tuning works :d
PThanks a lot, I am so stupid! 
Fno issues
Ftesting gradient accumulation steps 7
Finstead of batch size 7
Fto make comparison more fair
Fgradient accumulation steps 7 works
Fbut sloww
OWhat, exactly, are you doing with these images of Dr. Gözükara that you can feel his likeness? 
Whi @Dr. Furkan Gözükara I wonder if I train lora using flux dev but gen image with Flux Snell with trained lora, the quality is still good enough ?
Ghi! I am following the lora tutorial to do a dreambooth flux tune on massed compute - I am loading the 48gb config for my A6000, but when the training starts it says it will be something like 880,000 steps and will take 2500 hours. I have 125 images and am using the default of 40 repeats with everything else left as-is in the config (best configs v5) - is that accurate or am I missing something?
Wchange repeat to 1, the config file does not change this that why you need 880,000 steps
Wa few thousands step is enough
Gthank you! so 22000 steps and ~40 hours is normal?
Git still has 200 epochs maybe that's also unnecessary?
Wyes due to step we only use 1, 200 epochs is a must
Galright, guess I just set an alarm on my phone to check in 2 days lol. thank you!
Wdepend on how do you want to train: checkpoint or lora. style lora or person lora
Gperson lora but I read the post about how finetunes are much better at keeping it flexible
Wfor a person lora more than 6000 steps I think a bit much
Ggot it - even going the dreambooth route?
WI haven't trained checkpoint yet but it will require more hardware power and times since many people have already done it.
GI guess I will see how it pans out in ~12ish hours!
GI've done a regular lora at 6000 but it isn't getting tattoos right. it gets the placement and size right but random designs. maybe it's just a limitation of flux training right now
Fi never tried
Frepeat will be 1
F22k steps looks normal. each config has step speed
Falso i sugest fine tuning batch size 7
F3000 steps takes 3 hours
Fthis is rtx 4090 speed for fine tuning on windows
Fif you are getting like 10 15 second / it something wrong :d
CHello all. I'm new here.
Can I use a trained model of myself (FLUX) in l'ora to generate images with two people? So far, my attempts result in both people having my face. A guide on this would be very helpful
Can I use a trained model of myself (FLUX) in l'ora to generate images with two people? So far, my attempts result in both people having my face. A guide on this would be very helpful
Fwe are researching that
Fcurrently it bleeds
Fthe distilled models of FLUX claimed to have fixed this issue
Fhopefully i will fully investigate
DI confirm sub 7s/it on a 4090
D
Dand it only running at 80% speed
Fawesome
F10 epoch vs 100 epoch
Fhard to decide best epoch it is like 60 or 70
F256 images batch size 1
Fi think i will pick 70 epoch
D@Dr. Furkan Gözükara what is the best speed I can get full finetuning with my 3090ti?
EThanks this is really encouraging to see. People have been telling me you only need 2000 steps or it is overtrained, but most of my attempts look like your examples. 2000 steps PS3 style and closer to 6000+ start to look more photographic. BTW could you make a guide on how to tell if you have over or undertrained if you haven't already?