the full body image ?
the 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
Fsame setup in this video
FYouTubeSoftware Engineering Courses - SE Courses
In this video, I am explaining almost every aspect of Stable Diffusion Textual Inversion (TI) / Text Embeddings. I am demonstrating a live example of how to train a person face with all of the best settings including technical details.
If I have been of assistance to you and you would like to show your support for my work, please consider bec...
If I have been of assistance to you and you would like to show your support for my work, please consider bec...
Cyou said in the video to use 0.005, to leave it as it is, with batch 1
Cwhat about Filewords for images?
Fin TI i didnt use any
Fyou define what you dont want to teach
Fsince i have only my face in training
Fi dont need
FTI sd 2.1 512
Csome people uses this:
"a photo of a man as [name]"
"a photo of a man as [name]"
Cmaybe the solution is to use SD 2.1 512 instead of protogen
Cseriously, I have been thinking about it today
Fyou really should watch entire video
Fthat is my most informative video
