WHi all. Today I have a good feeling to be the day I could generate my first cool IA image of me . Everything updated to last version thanks to Dr Furkan last tutorial.
Dlooking forward to your style training video, i'm trying to train a style, but the result is not that good atm
Fi got everything ready just recording video. i hope it gets recommended by youtube
WOne question do the class images race for a man can interfere in the training for a man photo?. Because it generates me 90% of them black men (i'm Caucasian).
Fyou generated man from the base model?
Fthen it shouldnt
Wyes is it is "photo of man" in class prompt, all black random guys.
Fyour class photos like that?
Fthat is unsual
Fi have publicly shared dataset
FW
Wnice thank you
Finteresting
Wmay be I set something wrong. I will check
WCan I delete a model created?
Fsure
WJust deleting the C:\stable-diffusion-webui-master\models\dreambooth\model folder?
Fyes but if you have checkpoints you also need to delete them
Fto save space
Wok thanks
SUmm, site is working, but with page errors and big time spend, did you have, like, boosty or smth?
Smaybe, i can transfer this money to you by some another way, what is not blocking in Russia
Fyes i can also receive bitcoin or bnb crypto
WAny advise to avoid overtraining? I only got 2 or 3 samples correctly, but the 3 o 4 is over-training, this means, the Tomer Hanuka sample is not Tomer Hanuka stile anymore stile anymore. The problem happen before the chekpoint is saved, so I'm not able to get not over-training. Checkpoint, and it is very difficult to make stiles with the first checkpoint.
Fhow many per epoch do you save?
Fyou can reduce number of epochs
Fthen use x/y/z to compare
Fpost processing style teaching video
WI'm using images, no video yet
Fno i mean i am post processing new tutorial video
Fi recorded it
Wha ok, thanks. I will check the epochs
WTraining steps per image 150, save model frequency 25, save preview frequency 5
F25 could be big
Fi saw 5 epoch made overtraining in my recent tutorial
Fstill processing it not available yet on youtube
Whaha, sure, I don't know where I took that number haha
Dquick question, if I'd like to maintain a look of photo I trained I'm getting good results, for example it's a certain type of "ohwx City", should I train another model on top of a previous checkpoint or should I return to stable diffusion 1.5? For example if I'd like to train a certain type of "elos Bus" to show up in the photos of the "ohwx City" should I continue from the ohwx city checkpoint or train it on the s.d. 1.5 checkpoint?
Fit depends on purpose
Ffor example in upcoming video
Fi trained myself then trained style on top of that
WOk, you create two different models right?
Fit works like this
F1st train model and you get model check point
Fthen create a new model from that checkpoint
Fand train it
Wmmmm interesting
Fwait new video explains a lot
