i used bad-hands5 negative prompt and kept turning her african
i used bad-hands5 negative prompt and kept turning her african
Fhaha :d ye sometimes simpler prompts much better
Zcan you explain a little bit how the tokenizer works, and how i can use it to help my prompting?
Zor does it not really matter cause they all get combined anyways
AThis looks amazing 
Zdo you recommend kohya or automatic1111 dreambooth for lora generation? or does it matter?
A@Furkan Gözükara SECourses After a week of struggling to understand and to get things to run and not crash in the middle of training I have finally managed to get some acceptable results. Thanks to your tutorials and advice. Now on to style training...
Zif i write multiple words in my prompt, should i surround them with parenthesis?
Zlike i'm doing an indian, and i wrote "face with intricate warpaint"
JInteresting, just watched the LoRA training guide, there's a lot of talk about Network Rank, it seems that lower network ranks seem to work well for some people, whereas the higher ranks seem preferable in other instances
JI experimented a bit on my face, and I feel the higher network rate is much better, especially for finer details and 'ideas'
WI just watched the tutorial about Lora, if I want to train an illustration style, how should I set the Dimension and Alpha?
Iwhere is the tutorial for making a lora?
Ido i have to use dreambooth?
Ii want to be able to use any model, so I dont want to make a dreambooth model
D@Furkan Gözükara SECourses Doctor G, I want to follow along with your latest video with the Kohya LoRA Stable Diffusion training. I'm on GNU/Linux, will that make a difference? I don't think I'd use Visual Studio for example. Thank you for all of the work you put into every one of your videos!!!
EHey! Watching your new Lora video now. Although, I thought using dream booth extension for Lora training was better? Should I continue using dream booth, or Kohya gui?
PWhat version of the LDSR model are you using ? Is xformers required in the command line arggs for LDSR?
ZI totally forgot i had a new installation and forgot to redownload VAE
Zno wonder my stuff ain't been looking top notch
Tthe only decent images i get are from img2img, never gotten any decent results from training. GTX 1070 8GB
D@Furkan Gözükara SECourses Odd, I'm following along with the Kohya LoRA video and I don't seem to have the same interface? I don't see the Network Rank and Network Alpha and I don't see anything about LORA on the interface? I don't know what I could have done wrong.
I\stable-diffusion-webui\venv\lib\site-packages\torchvision\transforms\functional_tensor.py:5: UserWarning: The torchvision.transforms.functional_tensor module is deprecated in 0.15 and will be removed in 0.17. Please don't rely on it. You probably just need to use APIs in torchvision.transforms.functional or in torchvision.transforms.v2.functional.
?
?
DNever mind me. I am so dumb I was in the "DREAM BOOTH" tab NOT the Dreambooth LORA" tab, I guess I missed that part.
ZGreat Kohya video! Didn't realize that you didn't need captions unless you were training multiple styles
Gis it possible to train Kandinsky w/ DB? not super clear to me on how much is model, how much is code 
Iin the new lora tutorial you say to install the old torch i just uninstalled for 2
Iso i used torch 2 in the install for kohys
L@Furkan Gözükara SECourses thank you for latest video about Lora. I am trying to generate reg images, but they all look like sketches, not realistic portraits. I am using same model and VAE in video. Any advice?
Z@Lacerda613 what model are you using?
LRealistic vision 2.0, like in the tutorial
DKeep failing. I think my installations are borked, not right for my system/set-up.
Tsame thing happens to me
JAny opinions on captioning for face TI training? Do you structure your captions like "Photograph of a man, wearing a red shirt, white background" - Or do you tend to just have a single string without commas?
JIt occurred to me a while ago that commas carry their own weights, and do have somewhat of an influence on generation
JI guess the logic also applies to training
Idoesnt seem to matter for me
Ithe BLIP uses one long sentence seems to be fine
Ii just confirm they are correct
JSeems like it, I might just go through and delete all of my commas and run the training again, see if I get wildly different results
JThere's some people out there who place a lot of emphasis on captioning, a lot of people say it doesn't matter
Iand its correct 95%+ pretty cool
Ii bet in a couple years youll just be able to give it a few pictures of people and just tell it to make the best movie ever with them
Tyeah, not very useful until then. I have a feeling its gonna have that same vibe that vhs tapes give when you watch them. Instantly recognizable that its stable diffusion.
EI have a quick question. For Lora training. I’m wondering why ohwx has to be used as instance token? If we’re training a person, can’t we just use their first and last name as the instance token, followed by the class token? Or does that not work as well ? (For instance prompt and class prompt I use file words)
Zanyone seen the ai made pizza commercial yet?
MThanks for the kohya video @Furkan Gözükara SECourses you absolute legend. 
FSorry guys I have been away will read messages
FTy so much for the feedback
FYou don't need. () will make prompt strength 1.1
FYe I also found 128 performed best with alpha value 1. I plan to test even higher values