if we're doing a full finetune (then extract) instead of regular lora training, is the ~2000-2500 st
if we're doing a full finetune (then extract) instead of regular lora training, is the ~2000-2500 steps still valid?
ATried searching but can't find it (I assume it's in the video tutorial but I can't watch video from work) Can any SwarmUI user tell me where I need to put the T5xxl model so SwarmUI will detect it? Trying to figure out how to use Swarm for Flux and this is the only bit I can't see anywhere
Dmodels\clip
Fit should load your last saved preset
Fif i remember correctly
Fi never compared
Fnot that i know sadly
Fit should scale very well with H100. i would say speed is even faster than lora best settings if you have 48 gb gpu. if you have 24 gb little bit slower
Fsadly looks like min 48 gb is needed for 12 gb gpus
Fi think not. not even fp8 works
Ffp8 training works for lora but not for fine tuning
Fsadly fp 8 fine tuning not supported yet. i asked kohya and he said he doesnt know how to
Fbatch size 7 fine tunng take 3 hours for 3000 steps
Fso 1$ 
Fhopefully will fully research this
Fnot that great. L40 is not great like L40S
V@Dr. Furkan Gözükara What gpu is that for $1?
Fsomething so wrong
FA6000
F3 hours
Fi will show durations at next tutorial hopefully
Fand will post entire info on patreon soon
Fpreparing
Fi will also compare 15 img vs 256, fine tune vs lora, batch size 1 vs 7
Vdo vs dora since ontrainer can do doras for flux, maybe a finetune lora vs lora vs dora?
Fnice
Fbut this depends on dataset size
Fi do more steps with 256 images
Fit needs a special name
For you need to manually select
Fi shared in #tips-info channel
Falso here a tips for you
Ffine tuned model can do much higher high res upscale compared to lora
Fwithout stripe lines
V@Dr. Furkan Gözükara I had wanted to ask, I read some documentation that training captions requires commas to separate the prompt, does that still hold true for flux? I have my prompts structured as sentences and there are no commas only peroids at the end of the sentence.
Fi think not
Fbut you dont need captions for flux
Fat least for distilled dev
Vtoo late
V
VIt just finished and I am happy with the results.
Fnice
VFully captioned, including colors, with activation sentence and word.
V"painted of the style of "word"".
Fyes it works
Fbut a little bit subpar
Fi compared :d
Fbut depends on your dataset
Fit must be consisten