Some kind of bug in Kohya, I also got this and couldn't fix it, I think restarting the pod resolved
Some kind of bug in Kohya, I also got this and couldn't fix it, I think restarting the pod resolved it.
Fsorry not that one
FThis one:
FIt says it uses 18 images to train on a model of a person, I dont know how it could work on a logo
FI think you can generate a small dataset with the logo in different situations and test
FFull disclosure, I'm trying to train my first TI model, so mine is all theory 
NYou are right, restart worked but then I encounter this
ADid you go to tools, configure the stuff, click "Prepare training data", and then "Copy into Folders Tab" before clicking "Train model" ?
Fhello welcome
Nthat did it 
Fhow many. and why small number?
ADid you also disable xformers?
Nno 
do I have to? I need to do inference after
do I have to? I need to do inference afterAInference is in A1111, you need to disable xformers for training to get better quality
Nok how do I do that?
AUnder "Training parameters" expand "Advanced Configuration" and UNCHECK "Use xformers"
AIt also defaults to "cosine" as "LR scheduler" but I seem to get better results using "constant_with_warmup"
FHi @Dr. Furkan Gözükara, how are you? What do you recommend for doing some Textual Inversion training experiments?
Colab (I'm paying 10 USD plan) or Runpod?
Locally I've a a Apple M2 8-cores 16GB, which I think is super slow for training.
Colab (I'm paying 10 USD plan) or Runpod?
Locally I've a a Apple M2 8-cores 16GB, which I think is super slow for training.
Fdo you have any particular reason for textual inversion
FNo, just that I've seen great models in Civitai, low size, and wanted to do some exploration
N@Dr. Furkan Gözükara do you have any advise on which training method to use if I'm looking to get a sketch style like this:
Ehey!! for anyone doing lora training with 1111 dreambooth extension, how are you getting txt2img to recognize it? for me now, it wont pickup on any Lora unless i use the "sd-webui-additional-networks" extension. although i believe the previous method with "a1111-sd-webui-locon" extension was more convenient (red icon at the top under generate)? (side note, when i train with dreambooth extension, it saves the lora as .pt files not safetensors. it only saves as safetensors if i check the option "generate lora weights for extra networks" , but then i cant use those in the prompt (i receive error)
Ldreambooth is installed but it is not visible as a tab.
Qreload ui
LI did. Closed terminals, changed port. Nothing.
Ldoes anyone know what to do here?
Land to be honest, I have dreambooth installed on my webui in my pc, and I dont see it in the tab either
Lis it a python problem?
Q@Le open extentions tab
Ebut if i train a lora in Kohya SS gui, then bring it to 1111 it works just fine without using the additional networks extension (extra network) should i jsut stick with kohya instead of dreambooth? (prob is right now my kohya has a problem with xformers so traingiing is 8 hours instead of the normal 20 minutes)
Lthis is from runpod
Lno dreambooth on both
L@Dr. Furkan Gözükara have you seen anything like this?
LTried reseting my pod. still nothing
Fdreambooth and lora are better than textual inversion
Fwell if you can get good training dataset dreambooth best
Fin the options there is some compability
Fi havent tested yet
F
Fthat means install failed
Fif you open any desk i can quickly install
LI didn't understand it
Fi will dm you
Zshould we be using xformers for lora training or no?
EDoes really not one person know? Everyone’s Lora implementation must have broken then 
EIt makes the training go way faster (xformers and gradient checkpointing)
Zright but i keep seeing people say not to use it, and i've heard in the past that it was botching peoples trainings
EIt can also depend on what version of xformers you have. Like, using 0.0.19 will give different results to 0.0.20 , or so I’ve read; so you want to find the right xformers . I think it’s mostly to reduce vram usage too. Side note, Like my problem right now is dreambooth extension is giving me issues as it always does but training takes like 20 min with xformers but kohya is more reliable but xformers isn’t working for me and same training takes 8 hours 
ESo I’d say use xformers, unless it’s bitching your photos
