U3090
Dwith T5 attention mask off I get 4.5 it/s in my 3090 ti
Dwith T5 attention mask enabled 6.5 it/s
DI suggest 1 repetition 200 epochs
Uwould you suggest 1 repeat / 200 epochs? 2 repeats / 100 epochs?
Ugreat.
Uthanks!
D5000 steps
Uhave u seen quality differences with / without?
Dno in ma case but is better if you have captions
Dtheoretically
Uso if i'm using captions, i should aswell use t5 attention
Dyes, you can try with and without
Dfor a firts run you can try the fast way 1 repeat 150 or 200 epochs and see the results
Dthen try with attention mask and compare
Djust a suggestion, there is no one solution for every case
Dat this point is trial an error
AAny way to improve these settings for ai toolkit?
message.txt2.23KB
A102 images so far has been going 12 hours
DI only used kohya
Dbut you are using adamw8bit
Dit consumes more vram
DI use adafactor
Dwhat graphic card do you have?
MAnyone use Prodigy for Flux traning?
Dnot me, it too memory hungry
Sare there any differences i should be aware of when making a gguf of a flux.1-dev fp32 safetensor model, compared to the way i would do it for an llm? just looking to make a q8 for people with less vram.
Jusing only original ones from Black forrest labs (Flux1-dev),
there are too many versions out for me to keep track of and whether they are all
compatible with my LoRAs or not.
there are too many versions out for me to keep track of and whether they are all
compatible with my LoRAs or not.
TYeah they are original, you can read the repo page to have the info
N@Dr. Furkan Gözükara 
Congrats on the review, will peruse and need to put likes everywhere
Congrats on the review, will peruse and need to put likes everywhere
Fprobably you loaded into dreambooth tab at least once and corrupted config
Fyes this is a bug
Fyou need to use command
Flike cd checkpoints
Fi didnt try lion
Fyou have to use cfg 1
Fflux has different guidance
FWell 102 is decent number so even if you go 100 epoch it makes 10000 steps 
FI need to test it and see if works
FRegularazation I tested it. It will make your face mixed
F1 repeat 200 epochs
FI trained turbo schenell model
FResults are bad as expected. It trained though
FCurrently training on 4x gpu to see its quality - practical batch size being 4
DI saw it in your video, you tested again without class with the same results?
Fi trained only ohwx but didnt use reg that time
Fcan you explain me exact setup you meaning
Idoes your tutorial give better results than training lora on site like replicate?