Is RTX 4090 24GB good enough for Dreambooth SDXL or should I go with better like RTX 6000 48GB?
Is RTX 4090 24GB good enough for Dreambooth SDXL or should I go with better like RTX 6000 48GB?
BA5000 or 3090 are sufficient for Dreambooth SDXL, works fine with his 24gb settings file
Frunpod
Fbest
Feven rtx 3090 good
GThanks @Furkan Gözükara SECourses
GWhat's the difference between the kohya gui for training Dreambooth and the A1111 Dreambooth extension?
GI am confused as to which tutorial to follow
Fa1111 dreambooth is good for sd 1.5
Fbut only older version working
Ffor sdxl kohya is the one
GHas anyone made a graph of the number of high quality source images for a model and the quality of results?
GI understand that quality is more important than quantity. But as an example let's say... is there any reason to go above a certain number of source images assuming they're all high quality?
GWhat's the diminishing returns like?
Fnot that i know such graph
Fit is mostly based on experience
GGIs there something wrong with the settings in these JSON files?
GI've done training with 24GB_Best and 48GB_Best, and it only creates a single model file, and the model is way worse than when I did the LoRA
G
GThis is for 18 source images btw
GI think I confused this with LoRA. Is the cause of my issue because I put max trained steps in the Regularization Repeats?
GOkay, I rewatched the original LoRA tutorial I got this idea from. When you explained how to calculate how many regularization images you need based on steps and number of source images, I for some reason thought that I needed to put that number in Regularization images Repeats
F5760 repeat in reg images
F
Fit will be 1
MI have the exact same issue. @Furkan Gözükara SECourses are you saying i should not use LoRA to train my face in SDXL? (4060ti16GB)
GI'm re-running my training now after I fixed my issue I had in my settings. If that issue is solved for me I'll let you know.
FI would do dreambooth in your case
FWe have 16gb config
_Is there any difference between sdxl finetuning and kohya's dreambooth? Since both methods use sdxl_train.py?
(except for the captioning method..)
(except for the captioning method..)
Fjust dont use reg images
Fwhen no reg images used dreambooth = fine tuning
_thanks a lot!! now it's little less confusing 
_oh, I have one more questions:
does it make a difference if i train image sets sequential (for instance 3000 steps, save checkpoint and use it as new base model to continue another 3000 steps) VS train all images in one run (eg 6000 steps)
does it make a difference if i train image sets sequential (for instance 3000 steps, save checkpoint and use it as new base model to continue another 3000 steps) VS train all images in one run (eg 6000 steps)
Fin theory yes
Fbut it depends on whether you are using dynamic LR
Fand if scripts are properly working
_that makes sense 
thank you!!!
thank you!!!_https://github.com/kohya-ss/sd-scripts/issues/890
--train_text_encoder in sdxl_train.py does literally NOTHING
super exciting, curious how the training outcome will be after this is fixed.
--train_text_encoder in sdxl_train.py does literally NOTHING
super exciting, curious how the training outcome will be after this is fixed.
GitHub
Hi! I've been trying to perform Dreambooth training of the SDXL text encoders without affecting the unet at all. There doesn't seem to be an option in sdxl_train.py to specifically target o...
Fyep i am waiting too
Fposted it on twitter
MHi all, i have been using runpod for A1111, since im using 4090 i want to fully maximize the gpu to generate image. can i know if i can divide the GPU process so that two people can generate the image at the same time without having to queue up with each other?
Mor perhaps just need to tweak the command line args in webui-user.sh?
Fstart 2 instance of auto1111.
Mhad not found that yet, thanks 
_@Furkan Gözükara SECourses debugged the text encoder bug myself and reproduced the answer from arazela.
If you are curious too, here the debug output.
If you are curious too, here the debug output.
_after implementing the fix from koyha (only local because it's not pushed yet):