15 images bs7 really good
15 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
Dyes different images, but sdxl has the same problem to generate two different people in the same image, the good thing is that now I can train multiple subjects in one lora, I didn't try to generate two people in the same frame, I will try tomorrow
OAre you doing a fine-tune then extracting a LORA? or is this a straight LORA training?
I've been training finetunes/loras on family members, so having a bunch of subjects together with lower bleeding is a great use case for me. Even if there is still some bleeding, then I can segment or inpaint the faces if it can get the broad features correct.
I've been training finetunes/loras on family members, so having a bunch of subjects together with lower bleeding is a great use case for me. Even if there is still some bleeding, then I can segment or inpaint the faces if it can get the broad features correct.
K@Dr. Furkan Gözükara did you have any success with the one trainer anti bleed thing ?
VI bought 2x used 3090 on the cheap. But I keep getting blue screens If I want to use both of them optimally. Please don't tell me that I have to upgrade my PSU as well
PHow do I use the new koya dreambooth training files in swamp?
PI get an error when loading it backend #0 failed to load model test_1-000150.safetensors
CNote sure if this applies but if youre on massedcompute you need to copy model/clip/vae files into relevant swarm folders. Your checkpoint will go in diffussion_models, or the lora folder if that is what you trained in which case flux1_dev needs to be in diffusion_models
Fawesome
Fit bled i didnt get success
Fye you need like 900 watt only for gpus 
Fput into diffusion_models
