since runpod is my main setup i'd be very interested to see how I can improve it...
since runpod is my main setup i'd be very interested to see how I can improve it...
Fon runpod i think biggest issue is correct xformers from what i seen. but when i get sufficient donation i will hire myself a runpod to prepare a video hopefully
Iwhat xformer version do you use in your setup?
Idid anyone try doing a dreambooth training with protogen as the basis? and if so what settings?>
HI don’t think that’s a good base model for training.
Maybe try Dreamlike Photoreal.
Maybe try Dreamlike Photoreal.
Aoof.... need some advice on TI embeds. i went through the tutorial on SECourses... even using the sample images y'all had... and i got GREAT results....
Abut when i try with my own source images .... horrible. i'm not sure what i'm doing wrong
Ai'm trying to train an embed on images for my wife
Abut the results are coming out very poor
Iand what kind of settings would you recommend for training?
HInteresting. My 2.1 TI training was subpar
Abut the results come out like this.
A
Cremove bottom right
Athe full body image ?
Athe others ok?
Shave you tried fixing it with negative prompts as well?
Cyes precisely
Ai tried using the embed for "bad-prompt"
Athe images come out more "appealing"... but still totally different person
Cthese could also be optimized by being replaced by something full frame
Awhat's driving me nuts is when i do same thing with the sample images from SECourses... spot on perfect
Ayou mean full face? like up close?
C
Aah ok so all of them up close
Fthat is good
Fthat means your training dataset is not good enough
Ahow's this?
A
Aremoved the bottom right from before and added more close ups
Fi think should learn very good
Awill try. i'm trying to keep it to 10 like your video
Aand samples
Ato match as close as possible and then i'll play with it tweaking as i go
Fi am using this one i explained in tutorial : xformers: 0.0.14.dev
Iso if my training works (meaning I can complete it without errors) and I am on xformers: 0.0.15.dev403 - so would changing it to 14 improve training results?
Aany issue with using xformers 0.0.16? i'm on whatever build comes through auto1111 when you do git pull
Ican't say anything about local installs of a1111 but on runpod xformers 0.16 do not work...
Ialso on the DB github wiki they say only 14 are supposed to work
IGitHub
Contribute to d8ahazard/sd_dreambooth_extension development by creating an account on GitHub.
Aoh for dreambooth
Fye that is why. dreambooth is not compatiable with above 14 i think.not sure for 15 though
F
Fnow i will try different prompts to beautify
F
Fcoma difference
F
Fwith comma
F
Cwell, I am on TI aswell... today I have tried the newest extension custom-diffusion-webui but the girl is not my friend. I couldnt make it working.
On TI the best is to have a setup of good images without the background ("delete" it with paint), some people uses this in the captions "a photo of a man as [name]" and this 0.005:25, 0.001:250, 0.00001 with batch 4 and worked... but not perfect. It depends on the model used afterwards. Protogen doesnt make it perfect... I prefer 2.1 or 1.5... or maybe is the setup of the pictures? I dunno, I just wanted to share with you my little aknowledge, hope it helps. Any thoughts?
On TI the best is to have a setup of good images without the background ("delete" it with paint), some people uses this in the captions "a photo of a man as [name]" and this 0.005:25, 0.001:250, 0.00001 with batch 4 and worked... but not perfect. It depends on the model used afterwards. Protogen doesnt make it perfect... I prefer 2.1 or 1.5... or maybe is the setup of the pictures? I dunno, I just wanted to share with you my little aknowledge, hope it helps. Any thoughts?
Fthese above are from TI
