Software Engineering Courses (SECourses)•3y ago

with dreambooth collab I managed to make a model looking 100% like the trained images, on my own mac

with dreambooth collab I managed to make a model looking 100% like the trained images, on my own machine on local, never (yet)

CoqueTornadoMusic•1/30/23, 10:17 PM

so she is a little upset, hahah, dissapointed...

CoqueTornadoMusic•1/30/23, 10:17 PM

because of the batch thing right?

CCoqueTornadoMusic so she is a little upset, hahah, dissapointed...

aidiffuser•1/30/23, 10:17 PM

been there

CoqueTornadoMusic•1/30/23, 10:17 PM

hahaha

AArtyom with dreambooth collab I managed to make a model looking 100% like the trained i...

Furkan Gözükara SECourses•1/30/23, 10:18 PM

ye machine scripts still inferior

CoqueTornadoMusic•1/30/23, 10:18 PM

I can only use batch 4 in TI, I dont know about dreambooth really too much, done 3 in december

ArtyomOP•1/30/23, 10:18 PM

if using the actual dreambooth version on Auto1111, it's so bad that my trained model (a 6 year old boy from a videogame), using the prompt, becomes a 40 year old black man

CCoqueTornadoMusic I can only use batch 4 in TI, I dont know about dreambooth really too much, done...

aidiffuser•1/30/23, 10:19 PM

1 or 2 yielded me the best results. With 4 I never got more than a 70%

aidiffuser•1/30/23, 10:19 PM

Im talking about DB

ArtyomOP•1/30/23, 10:19 PM

it's going to get better for sure, but in the meanwhile I'll stop investing time in drembooth on local

CoqueTornadoMusic•1/30/23, 10:19 PM

this was the most accuracy picture of Emily, maybe 80%

06166-1682460919-a_closeup_picture_of_a_emily1x-3000_1.5_elf_woman_with_neon_dress_the_wind_blows_her_epic_hair_1.8_fantasy_art_he_loo.jpg

aidiffuser•1/30/23, 10:20 PM

CoqueTornadoMusic•1/30/23, 10:20 PM

that person appeared when used the newest amazing extension

AArtyom if using the actual dreambooth version on Auto1111, it's so bad that my trained ...

Furkan Gözükara SECourses•1/30/23, 10:20 PM

it is not that bad atm

CoqueTornadoMusic•1/30/23, 10:20 PM

https://github.com/guaneec/custom-diffusion-webui

GitHub

GitHub - guaneec/custom-diffusion-webui: An unofficial implementati...

An unofficial implementation of Custom Diffusion for Automatic1111's WebUI. - GitHub - guaneec/custom-diffusion-webui: An unofficial implementation of Custom Diffusion for Automatic1111&...

CoqueTornadoMusic•1/30/23, 10:20 PM

if you try it with persons please let me know if you made it working properly...

CoqueTornadoMusic•1/30/23, 10:21 PM

custom-diffusion-webui

CoqueTornadoMusic•1/30/23, 10:21 PM

that is the name

aidiffuser•1/30/23, 10:21 PM

this is a 99% achieved in DB

CoqueTornadoMusic•1/30/23, 10:22 PM

which model did you trained?

CoqueTornadoMusic•1/30/23, 10:22 PM

did you merge it with any other model then?

aidiffuser•1/30/23, 10:22 PM

sec

CoqueTornadoMusic•1/30/23, 10:22 PM

aidiffuser•1/30/23, 10:23 PM

seekartmega

aidiffuser•1/30/23, 10:23 PM

batch 1 gradient 1 lr 1.3e-6

CoqueTornadoMusic•1/30/23, 10:24 PM

I am going to try a couple of TI before dropping the idea... TI on 2.1 maybe is the solution... finding a good model in 2.1 then... I dunno

aidiffuser•1/30/23, 10:24 PM

200 steps

aidiffuser•1/30/23, 10:24 PM

I didnt do a merge, i trained over seekart

CoqueTornadoMusic•1/30/23, 10:24 PM

cool

aidiffuser•1/30/23, 10:24 PM

no prior preserv

CoqueTornadoMusic•1/30/23, 10:25 PM

what do you mean with prior preserv?

aidiffuser•1/30/23, 10:25 PM

no classification images

CoqueTornadoMusic•1/30/23, 10:26 PM

captions?

aidiffuser•1/30/23, 10:27 PM

nope

aidiffuser•1/30/23, 10:27 PM

classic

CoqueTornadoMusic•1/30/23, 10:28 PM

cool, thanks! will try today

aidiffuser•1/30/23, 10:28 PM

let me know if Fátima is happy with the results

aidiffuser•1/30/23, 10:28 PM

hehe

CoqueTornadoMusic•1/30/23, 10:28 PM

anyway, the thing about custom-diffusion-webui is driving me crazy... they say that it works flawessly... blablabla... and then...

CoqueTornadoMusic•1/30/23, 10:29 PM

hahah will tell you... well, Fátima is fine, it was Emily... she is dissapointed... I have trained her 5 times at the moment

aidiffuser•1/30/23, 10:29 PM

there are faces harder than others

aidiffuser•1/30/23, 10:29 PM

i think

CoqueTornadoMusic•1/30/23, 10:30 PM

what is her name? your friend looks fantastic!

CoqueTornadoMusic•1/30/23, 10:30 PM

hahaha

aidiffuser•1/30/23, 10:30 PM

haha ill keep that to myself

CoqueTornadoMusic•1/30/23, 10:30 PM

maybe it is because of the 3264876324873 parameters we can introduce

CoqueTornadoMusic•1/30/23, 10:31 PM

S.D 2.1 with a good improved model maybe is the solution

Furkan Gözükara SECourses•1/30/23, 10:31 PM

prompt emphasis test on textual inversion sd 2.1

Furkan Gözükara SECourses•1/30/23, 10:31 PM

xyz_grid-0007-1283833964-photo20of20testsd21_asdf20very20high20detail20photorealistic20cinematic20lighting20awesome20quality20anime20style203d20render20rpg20game20charact.png

CoqueTornadoMusic•1/30/23, 10:32 PM

what does that 1.x number mean?

with dreambooth collab I managed to make a model looking 100% like the trained images, on my own mac

Similar Threads