can i use 512x512 images that have been upscaled using 4x ultrasharp to 1024x1024 for training in dr
can i use 512x512 images that have been upscaled using 4x ultrasharp to 1024x1024 for training in dreambooth?
Fwhy not i think you can
ZI'll give it a shot and we shall see
Fi have updated both kohya and roop readme files
Fboth working
FGitHub
Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks - Stable-Diffusion/1-Click-DeepFake-Tutorial.md at main · FurkanGozukara/Stable-Diffusion
FGitHub
Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks - Stable-Diffusion/Generate-Studio-Quality-Realistic-Photos-By-Kohya-LoRA-Stable-Diffusion-Training-Full-Tutorial.md at main · ...
Nhas anyone tried training multiple concept? I noticed that it basically doubled the steps, so it seems like it's simply just training one concept and than continuing to the second concept. Is it the same as just doing two training?
Zwhere does dreambooth generate the sample images?
Fyou specify everything in the config
Fpath, number of samples, etc
Zoh yeah i totally forgot about that, thanks for reminding me
F@Dr. Furkan Gözükara did you find if those high quality reg images improve the training results? or not yet
ZHey @Dr. Furkan Gözükara any updates onn how to use the new optimizers in dreambooth?
F@Dr. Furkan Gözükara roop on a qemu VM works wonderful, because you can assign half of your cpus and use script with max-cores, so your system doesn't collapse. everything's fast and smooth
ZI have a question that's been bugging me for a while. How do we know what to set our lora strengths to?
HHi everyone 
HJust joined the server 
Zwelcome
P@Dr. Furkan Gözükara How can one have both kohya and automatic1111 without them interfering each other?
A1) Create e different venv for each
2) Run them on different ports
2) Run them on different ports
PDoes kohya not automatically make a venv folder for windows installation?
AI think so, what is your issue you're having with them interfering with each other?
Ppython version installed on my machine is 3.10.6 and kohya is requiring 3.10.9
AThats strange Kohya works on Linux with 3.10.6
Pwindows !
P10
AYes I know I was just saying
Pah
ACan't you just use 3.10.9 for both?
PI can try but this should not be needed because of venv folder
RThere's a workaround m. A weird one Sr Furkan helped me with. 1) setup 3.10.9, set env path. Install Kohya. 2) make copy of 3.10.9, uninstall original one. The problem is when you install Python system wide it basically wipes parts of the precious phthon install. Has worked for me. Have auto1111 running with 3.10.6 set as system wide Python. Then give 3.10.9 in env path for Kohya startups
RI recommend you become familiar with environment paths. The doctor discusses it on one of his videos about python
RAuto1111 is pretty strict about 3.10.6
AAh, wonder why A1111 and Kohya are like that on Windows but not Linux, strange
RI think I saw 3.20.6 as requirement somewhere
RYea strange
AThink you mean 3.10.6. I was able to run A1111 on 3.10.10 on RunPod, haven't tried Kohya... but clearly different requirements for Linux and Windows 
RYea sorry, 3.10.6
P3.10.6 works for both a1111 and kohya
Pbut for some reason all training is now worse quality compared to a couple of days ago
Pboth runpod and local machine gave same results
ZDoes lion not work with lora in dreambooth? cause the results I got look nothing like my model
AWhat did you set the learning rate to?
Z5e-7
Zand 3e-7 for tenc
Zi'm doing lora extended
AWas it overtrained? Usually the LR for Lion is 1e-7
Zi don't think so, typically overtrained it looks burnt I thought.
Zi've been trying to train a model of my wife for ages now, and I get close, but never really there. I've followed every video I've come across, but still can't get it perfect