Software Engineering Courses (SECourses)•3y ago

oof.... need some advice on TI embeds. i went through the tutorial on SECourses... even using the s

oof.... 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....

alenknightOP•1/30/23, 9:27 PM

but when i try with my own source images .... horrible. i'm not sure what i'm doing wrong

alenknightOP•1/30/23, 9:27 PM

i'm trying to train an embed on images for my wife

alenknightOP•1/30/23, 9:27 PM

but the results are coming out very poor

HHaashim I don’t think that’s a good base model for training. Maybe try Dreamlike Photo...

IVG•1/30/23, 9:28 PM

and what kind of settings would you recommend for training?

FFurkan Gözükara SECourses Click to see attachment

Haashim•1/30/23, 9:29 PM

Interesting. My 2.1 TI training was subpar

alenknightOP•1/30/23, 9:29 PM

but the results come out like this.

alenknightOP•1/30/23, 9:30 PM

Original message was deleted

Colored Tonalities•1/30/23, 9:30 PM

remove bottom right

alenknightOP•1/30/23, 9:30 PM

the full body image ?

alenknightOP•1/30/23, 9:30 PM

the others ok?

Sandstormleader•1/30/23, 9:30 PM

have you tried fixing it with negative prompts as well?

Aalenknight the full body image ?

Colored Tonalities•1/30/23, 9:30 PM

yes precisely

alenknightOP•1/30/23, 9:30 PM

i tried using the embed for "bad-prompt"

alenknightOP•1/30/23, 9:31 PM

the images come out more "appealing"... but still totally different person

Colored Tonalities•1/30/23, 9:31 PM

these could also be optimized by being replaced by something full frame

alenknightOP•1/30/23, 9:31 PM

what's driving me nuts is when i do same thing with the sample images from SECourses... spot on perfect

alenknightOP•1/30/23, 9:31 PM

you mean full face? like up close?

Colored Tonalities•1/30/23, 9:32 PM

alenknightOP•1/30/23, 9:32 PM

ah ok so all of them up close

Aalenknight what's driving me nuts is when i do same thing with the sample images from SECou...

Furkan Gözükara SECourses•1/30/23, 9:33 PM

that is good

Furkan Gözükara SECourses•1/30/23, 9:33 PM

that means your training dataset is not good enough

alenknightOP•1/30/23, 9:35 PM

how's this?

alenknightOP•1/30/23, 9:35 PM

alenknightOP•1/30/23, 9:35 PM

removed the bottom right from before and added more close ups

Aalenknight Click to see attachment

Furkan Gözükara SECourses•1/30/23, 9:37 PM

i think should learn very good

alenknightOP•1/30/23, 9:37 PM

will try. i'm trying to keep it to 10 like your video

alenknightOP•1/30/23, 9:38 PM

and samples

alenknightOP•1/30/23, 9:38 PM

to match as close as possible and then i'll play with it tweaking as i go

IIVG what xformer version do you use in your setup?

Furkan Gözükara SECourses•1/30/23, 9:38 PM

i am using this one i explained in tutorial : xformers: 0.0.14.dev

IVG•1/30/23, 9:39 PM

so 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?

alenknightOP•1/30/23, 9:41 PM

any issue with using xformers 0.0.16? i'm on whatever build comes through auto1111 when you do git pull

IVG•1/30/23, 9:42 PM

can't say anything about local installs of a1111 but on runpod xformers 0.16 do not work...

IVG•1/30/23, 9:43 PM

also on the DB github wiki they say only 14 are supposed to work

IVG•1/30/23, 9:43 PM

https://github.com/d8ahazard/sd_dreambooth_extension/wiki/Extremely-Experimental-Libs

GitHub

Extremely Experimental Libs

Contribute to d8ahazard/sd_dreambooth_extension development by creating an account on GitHub.

alenknightOP•1/30/23, 9:44 PM

oh for dreambooth

Aalenknight any issue with using xformers 0.0.16? i'm on whatever build comes through auto11...

Furkan Gözükara SECourses•1/30/23, 9:52 PM

ye that is why. dreambooth is not compatiable with above 14 i think.not sure for 15 though

Furkan Gözükara SECourses•1/30/23, 9:53 PM

Furkan Gözükara SECourses•1/30/23, 9:53 PM

now i will try different prompts to beautify

Furkan Gözükara SECourses•1/30/23, 9:57 PM

Furkan Gözükara SECourses•1/30/23, 9:58 PM

coma difference

Furkan Gözükara SECourses•1/30/23, 10:01 PM

Furkan Gözükara SECourses•1/30/23, 10:01 PM

with comma

Furkan Gözükara SECourses•1/30/23, 10:01 PM

Aalenknight oof.... need some advice on TI embeds. i went through the tutorial on SECourses...

CoqueTornadoMusic•1/30/23, 10:02 PM

well, 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?

CCoqueTornadoMusic well, I am on TI aswell... today I have tried the newest extension custom-diffus...

Furkan Gözükara SECourses•1/30/23, 10:03 PM

these above are from TI

Furkan Gözükara SECourses•1/30/23, 10:03 PM

same setup in this video

Furkan Gözükara SECourses•1/30/23, 10:03 PM

https://youtu.be/dNOpWt-epdQ

YouTubeSoftware Engineering Courses - SE Courses

How To Do Stable Diffusion Textual Inversion (TI) / Text Embeddings...

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...

CoqueTornadoMusic•1/30/23, 10:05 PM

you said in the video to use 0.005, to leave it as it is, with batch 1

CoqueTornadoMusic•1/30/23, 10:05 PM

what about Filewords for images?

CCoqueTornadoMusic what about Filewords for images?

Furkan Gözükara SECourses•1/30/23, 10:06 PM

in TI i didnt use any

oof.... need some advice on TI embeds. i went through the tutorial on SECourses... even using the s

Similar Threads