Dit does that, just press train again and it will go
Mim new to making lora thing, and when i create lora on kohya, from 20 img and 100 img source, my file size only 9mb, so what happend with that? i see ppl make lora above 100mb
Mi mean i try using 20 img as source first and 100 img latter, but size still 9mb
VDo you guys know any information on how to utilize ControlNet to indicate the location of eyes and mouth in a sketch when working with Stable Diffusion?
We have the openpose, but as far I know, it works only in humanoid images.
We have the openpose, but as far I know, it works only in humanoid images.
Snetwork rank decides the size
Sless rank, more compression I guess
Mdefault one is 8, lest say i change 128, so 128:8=16x9=144 yeah maybe right
TI always do that every time it gets me that problem but Solved it now by changing how a folder of images should path. Fix it by getting inside the folder and copy directory. weird
THe only teach how to run the lora in khoya but didn't get the main goal of producing the precise training out put from the image subject. He cut the the video and overlay a picture of him not knowing how everything has done. I tried hus settings but none works for me, so i moved back to dreambooth training
TThis regulation images tab is giving is time consuming but it seems no effect all out gpu is stressed. Do you thin reg images do work? in training faces with dream booth?
TOne thing i loved with lora in khoya is it train fast like 10-15mins and ur done. Unlike dreambooth takes 1-2hrs depend on ur data and settings. I hope Lora in khoya will dev better training settings. I cant produce precise face from my subject. Annoying
JAh, finally got Dreambooth working in the Colab.
TWas it Smooth there in colab?
JWell, despite having spent my whole career (I am now retired) as a Systems and Network Engineer I have almost no coding ability, so it could have gone smoother.
JHere was a pretty good output. Took about 15 years off my age, so I approve.
Pat 1.8 im getting "looks like" results. but I'm going for a 1:1
PKohya feels like textual inversion with extra steps and HD usage
PI got a lot of "looks like" with textual inversion. not until I switched to dreambooth did I start getting 1:1
Swhat are your parameters?
Pexactly the same as the video
Sthe video was only for installation
Sif you used 512 photos you failed
Phe ran through a trial run as well
Sshow me your settings
Salso what base model
Pclosed now. I'm going through my dataset.
SI said not to listen to tutorials, the link was only for installing kohya
Plet's move this to DM so we don't clog up the main chat?
Sif your training takes 15 your dataset is too small
Sok
TInteresting
TOn going training on dreambooth. I wanted to cancel my training progress since im seeing better result of my subject in preview image. However, I forgot to check "Generate ckpt file when training is cancelled" but "Genenrate ckpt file when saving during training". Was checked. Im a good to stop the training now? 
Pcheck your stable diffusion models folder
Pyou should see a folder in there named your training model
Pin there you should have a list of ckp files that correspond with the training steps
Pif you have that then you should be good to stop the training and test using those checkpoints
TOh! i forgot lol! Thank for saving my time. Let me check
AHi I have got some strange problems.
AThis time it is the multiple controlnet
AI can't use use it for some reasons
AI can only use it for only 1 controlnet per generation.
AIf I tried to use 2 controlnet in the same time only the first controlnet will be applied.
PI take back what I said about Kohya Lora training. @Steel has shown me the light. Thanks mate, incredible results.
AWhat is your training setting sir? Could you share it here?
THow many images and Learning rate you used? loooks good tho.
P77 images. not cropped. full resolution
Plearning rate 0.0002
Pbut there were a lot of custom settings he walked me through.
PI'll try to compile a .txt file
Pjust testing the Lora now tho