and how many epochs is another question
and how many epochs is another question
Fsadly i dont know yet for flux
Fbut go with 50 and see if trained good
Fcompare checkpoints
GI remember doing it on SDXL for 3000 images but dont remember steps none of that
Fmore images better for style
Gstill dont have my 48GB A6000 running yet, just using my 3090ti for now
Fye it will take time on that gpu especially if you are not using linux
Gif it took 14hrs on my 3090ti on windows, how much faster would the A6000 do it?
Fsame speed 
Fyou need 4090 for speed up
For L40S
Gwtf
Funless you make multiple batch size
Fwith fp8 on A6000
Fwhich i plan to do
Gmultiple?
Fwhen training 20800 images
Fye batch size like 2 3 i havet tested yet
Gwhy is 4090 faster than a 48GB card?
Fit has more cuda and power
G:1222pepeblunt:
Gso the only good thing is the VRAM i have
Fyep
Gtell me the setting for batch, I will test it myself,
Gcan do 3? 4 batches?
GI will get that system up tonight or tomorrow
Fcaptioning 10400 images started
Gthen I can tell you
Fi will test later
GYeah but the setting, ill test it, where is it
Gi can try it once i reinstall windows
Fwhich settings?
Fbatch size is on gradio
Fi got to sleep
Gill check it and increase it to 4
Gand see if it works
ADoes anyone know if ai toolkit you need to be connected to the internet for it to work?
BThe final slate of characters for my animated pilot Very Special. LoRAs trained using datasets built with the Everly Heights Character Maker.
Pi am rejected acces to Lama on HF. if i use your 1 click installer will i still need to acces HF ? or will it just install?
Fit will install you dont need
Fmy installers handle that 
Fwow new training parameter made huge difference
resemblance reduced but clothing much better
looks like i have to investigate
resemblance reduced but clothing much better
looks like i have to investigate
Fcheck this messagfe
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...
Ffor stylized outputs it has a huge impact
Fi will do 8 more trainings with it
F
Mtop left looks accurate
F
Nis there explanation for what these flags do in the JoyCaption app? none of them are checked in any of the sample screenshots?
