<@205854764540362752> congrats on the latest training article. Can you add something about captionin
@Furkan Gözükara SECourses congrats on the latest training article. Can you add something about captioning (if you used it, how you made it…)?
MI mean, how did the model learn what each expression was? What expression does it inherently know, as far as your dataset goes?
Fi added it actually :d
FI didn't caption them only ohwx man since captioning reduces likeliness and doesn't bring improvements.
Fi updated article added more details
MThe Patreon one?
Fyep
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
MI haven't looked at the training dataset--as far as you can tell, does it look like Flux understand 100% of facial expressions? 50%?
Fit understand alls
Fi shared dataset too
Fbut lower res
MEverything like "anxious"? Like, if you put in a picture of yourself looking worried about something (dunno if you're that good of an actor 
), do you think the model would pick it up and be able to reproduce it when the prompt asks for it?
), do you think the model would pick it up and be able to reproduce it when the prompt asks for it?Fmodel has great internal captioning system
Xyeah yeah, anime style lora
Fi dont have a very good config for sdxl lora atm
Fbut training sdxl dreambooth / fine tuning and extracting lora works great
Dwhat are they offering here?
Fsomething very inferior i guess
Fprobably high batch size, high LR low step
Deven if all that and even if they train only some layers of the model, 5 min and only 1 epoch seems extreme. They must have found something new?
Fi doubt that
Dwe'll see if civitai gets flooded with sh** flux loras soon
Fhigh batch, high LR, single layer makes sense
Fyep
Dthe fine print says 1 epoch. so it isn't high batch or massive parallelization, it's just 1 epoch.
DI don't have any of their tokens to try, but would be interested if someone her has
D*here
F1 epoch but how many repeats
Drepeats are a kohya thing, 1 epoch in the general context of ML usually means going once through your dataset. And look here, 2,5 minutes. Even if you used multiple GPUs, it cannot mean much more than going through your dataset once with the s/it I have seen so far on flux
Dthat being said, I share your doubts
Fi know but civitai already uses kohya repeating 
when i first trained they were at least
when i first trained they were at leastFtheir description i think means that you dont get checkpoints not necessarly 1 repeat
YI have question about training and captions please : How many images you think, for a Pose lora including 2 or multiple characters should be in the folder, and repeats ? And also my captionning, can I send one here with the related image ? It is Sfw but may contain violence, soo i prefer to ask ^^
Fyou can dm me
Nsooo...I've tried some lora trainings recently. What can I say. Sometimes results are pretty good on 70-80 epochs (once on 50 epochs) and sometimes 200 epochs are not good enough. I guess it all depends on dataset.
Nand training on realistic dataset will not give you any cartoonish (comics) style output. Because it's too overtrained on realistic images. I did sample check every 40 epochs (5 checks during training) and after 80th epoch cartoon style becomes realistic
WHi @Furkan Gözükara SECourses can you create another train preset using 24gb Vram to train fp16 Lora. I trained using Rank_3_18950MB. The lora is fp8 compare with fp16 training in your Rank 1 or Rank 2 presets, the prompt understanding of the fp8 lora is reduce and the result of fp8 is so far not satisfy my need
WThank you in advance
WMaybe you can create new preset taking longer time to train but still use only 24gb Vram
DI have just started with Flux, but using onetrainer and bf16 I am below 12 GB vram... 
Dor do you mean the base model in fp16, not the Lora weights?
Dif so, I doubt that makes a difference. even during inference there is not much different between fp8 and NF4
Ffor fp16 you have to reduce resolution
Fno other way
Fhe means training
D? I also mean training
DI'm just saying that even during inference a low-precision base model doesn't make a difference, so how would it during training. The Lora weights are still bf16+
Dfascinating how Flux learns without any distortion. it just slowly adjusts
Fit makes some
