Hi, I'm trying to do sdxl dreambooth with the patreon kaggle notebook and i get 'out of memory' erro
Hi, I'm trying to do sdxl dreambooth with the patreon kaggle notebook and i get 'out of memory' error after the 4th iteration. Any idea?
F4 th iteration?
EI have to commented on it
Fyou mean steps?
MYes
FWhich base model you trying
MRealVis
Fdownloaded from civitai?
Fwhat is your resolution 1024x1024?
Myes 1024,1024 and the training images are 1280x1536
now it works with the
now it works with the
stabilityai/stable-diffusion-xl-base-1.0F1280x1536 will use more vram
Fresize them to 1024x1024 yourself before start training
Fbecause it will downscale only 1 side to 1024
VFurkan, do you offer new checkpoint settings? and do you mean if i train checkpoint, but can finally extract lora from it, too?
Fnew checkpoint settings are shared on patreon
Fand you can extract lora from them
Fworking nice
BDo regularisation images work better when training on dreambooth when they have caption files with them?
Fi never tried them with caption
Msomeone i know tried your dreambooth notebook and got this error. do you know what could be wrong?
Fyou forgatten to enable gpu
SFurkan where you going for vacation?
SCan I come with you Haha
Fi dont have any vacation haha 
not even weekends
not even weekendsFcheckout the newest article : https://medium.com/@furkangozukara/comparison-between-sdxl-full-dreambooth-training-vs-lora-training-vs-lora-extraction-44ada854f1b9
Medium
Hello everyone. I have trained myself recently by using Kohya GUI latest version. Used official SDXL 1.0 base version.
Dhe's always working working working
SAnyone know why I always get bad results from Adetailer? Does it only work well on small further faces? On my images where my face is pretty large (think portrait) Adetailer always makes my face look worse, like its set to a higher CFG Scale (its not) I followed the same settings @Dr. Furkan Gözükara used in the SDXL Dreambooth training video. For some reason this has always been my results with Adetailer and I'm not sure what I could be doing wrong
Fcan you show previous after and your settings?
Fi am using adetaler all the time
JYou are in that same position?
JAI is hard work, thank you for all these tutorials, I would need 48 hours a day to get everything done.
JBTW, do you think teaching SDXL style on a realvisXL model would be better? It seems the custom SDXL models are still flexible with prompts more than the 1.5 were.
Fyes D
Fsadly didnt have chance to test yet
JI'll test it now. Same everything only difference is base vs realvis SDXL
JThe settings already helped me increase the quality by a lot. I just need more steps the next time.
SAwesome keep the work going and don't forget not to burn out 
I know it myself
I know it myselfSCan someone share or show us what a high-quality or ultra-high-quality dataset for training looks like for SDXL Dreambooth? For example how many are enough and what ratio between close-up face, upper body, and full body shots is needed? If a photographer takes shots of the subjects what to take care of etc.
DHello, new to the patreon, looking at the config files and i notice i can't find the network dim or alpha. Is 128/1 still the default we are using here, or is it 128/128?
DHi @Dr. Furkan Gözükara and everyone. My 10 year old nephew wants a cool illustration of himself for xmas so I wan to to do a lora training of him. Can I use the images from the download_man_reg_imgs.sh file? Or is there a child regimages dataset somewhere? Lastly, will the class prompt "man" work in this case or does it have to be "child"? Thanks!
Oguys whats the best XL-based model for landscape?
JThe checkpoint that I have thought a style with, now produces that style even without a triggerword. Is that normal? Or is it overtrained?
MI used your crop & resizer tool and made them 896x1152, It's still using 14.9 GB VRAM so I switched to P100 GPU, it slow but works just fine.
JIs it just me or does that 6s/it for a p100 seem slow?
MIdk how fast it should be. For me 4.5h training time for an sdxl dreambooth for free is good enough.
JAh, for free. Then it's fine
JI get 1.2s/it on my 4090, that is why I was wondering why it's slower on a p100
JOh wait, p100 isn't that new. Nevermind me.
JI was thinking about the A100
MWith kaggle's double gpu it's around 1.3s/it but it has only 14GB Vram
L@Dr. Furkan Gözükara have you tried multi gpu lora training with kohya?
ddp_bucket_view and the other parameter dont seem to exist even on dev branch