FI did 10 epoke and 10 steps, it's normal that's training during 4hours with 62 images?
Fyep expanding testing prompts will update patreon
Fyes
Fflux is slow
Fi am going to add more detailed configs
Flike x GB y speed z tier
Nwhat did you set for saving? every x epoch or steps?
FI set 0 for saving
Fyou should save checkpoints
Flike every 10 epoch
Fand compare them
Fnewest trainings to compare 
Nwow!!
Nyour speed is good, at least that's what i get with a 3090 on windows 11. but math not adding up for me? by "10 steps" do you mean "10 repeats"? 1650 steps total at 10 epochs means 165 steps per epoch... you can't have repeats set at 10 because that would mean 16.5 images lol. my guess, not all 62 of your images were recognized, or you have batch set > 1?
Ffirst testing lora rank impact with 38 prompts
Nthat grid is going to be a monster 
�nvm just realized electricity was back on lol
Fye
Flora rank really overfits model
Fbut also improves resembalcen
F128 vs 4
Flook clothing difference
Fso fine tuning may firs this issue
Fmaybe rank 32 good
F256 truly best resemblance
�Wow it's so much more incredibly detailed, the sword and the castle in the background. It looks like it kind of removes the background blur that Flux is known for, too
F
Fye it overfits model
Fok time to test higher resolution
F256 is very impressive quality
Fyep
Fmore learns better
Ffine tuning will be much better probably
FFtesting higher res training impact right now
F1280 and 1536
Gwhats the difference between fine tuning(ive never used it) and upscaling?
Fcan i get this quality via massed compute or runpod? if i can create a lora that i could use all the time, i am ready to use several powerful gpu. i have to wait 2h to see the result of my training i am impatient
Fit trains all layers of model
Finstead of a lora
Fyes ytou can use both
Fif you want speed rent L40S
FFi will make a video
Fbut it is same as in these 2 tutorials
Gi can do this on my hardware?
Fdo what?
Gso it makes a checkpoint or a lora? Fine tune
Fcheckpoint will be better
