I'm going to start training with one of your configs, what config do you recommend me to start for a
I'm going to start training with one of your configs, what config do you recommend me to start for a rtx 3090?
Fi havent tested new ones still testing currently among those best one is 1e_04_fp8_bf16_accelerate_full_bf16_32_rank
Fmake sure that you have like 500 mb vram usage before starting training
Dthanks
Fthe new tests may have better config
Fi am downloading them to 8x a6000 machine to test
Fthey are completed but not evaluated yet
DI'm going to test your configs but in my previous tests I hade the same bleeding problem I hade with sd3
Dyou should try to train two people of the same class at the same time to test this, this wasn't a problem at all with sdxl using your configs.
Dmay be this will not be fixed until text encoder training is implementes.
DIf the problem persist we will be able to train just one person at a time or a style but not multiple specific persons.
DI will start a training right now with your recomended setting.
Fi am waiting clip l training to test
Fmy settings are not complete yet 
still in research but give it a try
still in research but give it a tryFthere are so many diff parameters i am yet to test
DThanks, I will test and report if the bleeding persists
NI put everything as you said, thanks!
the link seems to be closed download, found on Civitai model dev from darkwod haginface removed it
the link seems to be closed download, found on Civitai model dev from darkwod haginface removed it
Dother problem I'm seeing is speed to replicate one of my sdxl trainings it will increase from a few days to a couple of months, heheh
NAnother question, is there a list of art styles in flux, tokens to get them
Fcurently it is like 8.3 / 1.8 = 4.6 times slower :d
Fnot that i know
Nwhat are the best ones for Flux right now?
DI was able to train at 1.2 it/s with gradient checkpoint off and tweaking some other options, I thing full bf16 training experimental and some other things I dont remember right now
Dwith sdxl off course
Fhavent tested yet just using default :d
F1.2 it is good speed for RTX 4090 and SDXL
Dmay be It was 1.4/it but oround that speed
Dmy card is a 3090
DI hade slightly better speed on windows 10 0.1 it/s better
Fah you must be using xformers
Fi think it reduces quality slightly
Dyes I think you right
NWhat's the new parameter
Fprobably for turbo model
D
D1e_04_fp8_bf16_accelerate_full_bf16_32_rank config
Dsomething is not right
Dis consuming too much vram
D
Done thing I saw yesterday was that no matter the lora rank I use the vram usage was the same
Dlora rank 32
Dlora rank 128 is consuming less vram??? what??
Dmust be a bug
Flora rank impact on VRAM is very minimal i tested
Fhow much vram usage you have before starting the training?
D0.6
DI can turn off the monitors and get 0.4 gb
F0.6 should work but still using some extra. i wonder if due to operating system
Fi tested vram usages on ubuntu
Dbut yesterday worked
Finteresting