BS 7 is way better
BS 7 is way better
Fi will compare it with bs 1
Fboth 15 images
Fboth 256 images
Fby the way i researched a unique LR for bs7 :d
KLooking at the right column it's way better than the left column
KI need to get into the "Kaggle" thing I guess. I can't keep locking up my computer
Fthese are all same bs7
Fdifferent checkpoints
Ffinding best epopch
KRan 2 trainings without a problem with AI-Toolkit over night though.. Unfortunately bad ones (As expected)
KSorry read it wrong! Read BS 1 for the left column
F15 images bs7 really good
Fnot even cherry pick
F
EJust out of curiosity, approximately how many steps was the checkpoint used for this image trained for? I'm just curious because for me even at around 2000 steps I'm still getting PlayStation2 CGI type looks even though I'm using real photos as a dataset (50 images). It starts to improve at much higher steps.
Flet me check
Fit is 140 epoch but since batch size 7 it is not divisible
Fso it is actually
F420 steps
Fwith batch size 7
Fso 2920 steps with batch size 1
BAny advice for training loras on flux at 512? its struggling to capture likeness
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Ddont
MI'm looking at the Flux licenses, and as far as I can see, only the schnell version is free for commercial use. The dev version can only be used for non-commercial purposes, or must be agreed separately with BFL. I wonder about that, the SDXL license is a bit more flexible since the change.
CDo you mean 512x512 input images? I've done 2 for curiosities sake and both worked fine using the rank 4 config. I stopped them both around 80 epoch and they were adequate by that stage. One was 9 images, the other 16 or so. Are your input images any good?
BI get insane results at 1024x1024 but the training time is slow and the file size is large. I'm trying to find the lowest I can push the Lora training parameters until details and likeness begin to drop off. As I want to build an app and do this at scale.
Ci redid those experiments at 1024 as the 512 was just for the much faster speed and my curiosity, the likeness was fine just not as consistently accurate as the 1024. I didn't tinker with them long or do any overly useful comparisons, I purged the 512 loras and redid them. I did find they seemed more consistent when generating images at 512 also, so if your use case is aiming for that maybe its worth the effort to investigate. Rank 4 config was what I used with some barrel scraping input images id prepped for 1024 and resized to 512 and I was blown away with how well it did, better than any SDXL training and the input images were questionable at best
KWTAF... XlabsSampler only generates black images suddenly
KThis is just a quick workflow to try it out, but it doesn't work 
KYou see anything wrong here?
F@Dr. Furkan Gözükara hellow, i saw there was an update on patreon about fine tune/dreambooth, so can we work on it already?
PMight be the prompt as well? I did the mistake adding man into my prompt when i wanted to replicate myself.. so instead of "ohwx man" just use your token.. in my case "ohwx" ..
VAnyone fine tune an fp8 version of flux with 24gb and no shared gpu?
D@Dr. Furkan Gözükara My first test was a total success, I trained many people at the same time without bleeding between each other, It works perfect and the lora can be used on regular flux dev that is much faster on inference, I get a little bleeding with two subjects with a similar name "Diego man" "Dani man" is minimal so it can be fixed changing the name to "Daniel man" I saw a little class bleeding but I think it can be fixed using regularization images. the model behaves very similar to regular SDXL, Training many subjects at the same time was imposible with regular flux-dev. The model is still a little undertrained so I will continue and try with regularization of class people because my dataset contains people from different genders and ages. my captions are very simple, "name class"
This model is awesome and very promising so far.
https://huggingface.co/nyanko7/flux-dev-de-distill
This model is awesome and very promising so far.
https://huggingface.co/nyanko7/flux-dev-de-distill
Dnote: for this first test I used 512x512 resolution to speed up the training and I used a higher lr for the same reason 0.0001 for unet and TE, may be with lower lr and of couse at 1024x1024 will get better results
F512 reduces likeliness
BYeah I’ve figured
FSDXL is free to use afaik
Ffp8 fine tuning not supported
Fi asked kohya he doesnt know how to
Fthey werent in the same picture right?
Fseperate images
F10 epoch 256 images batch size 7
Ftakes 2:58:25 on single RTX A6000
F30 epoch
Fi like the shading / shadow on head from moonlight epic
Fand this is 100 epoch clearly overfit
Fhere a good example of undertrained to overfit for you . i think epoch 40 is best
