#Shirakami Fubuki | Nagishiro Mito TI

1 messages · Page 1 of 1 (latest)

polar raven
#

Shirakami Fubuki/Nagishiro Mito's style.

V1.1: scroll down to download, trained with cafe's TI repo colab. 3800-5600 steps. #1048266381264683098 message

V1: Trained on Waifu Diffusion 1.3, mixed with Anything V3, 15k steps, generated with Anything V3.
First upload. I have no idea how AI works so if you have questions I don't know if I can really answer them.

#
1girl, fox girl, white hair,(fox ears:1.3), (animal ear fluff:1.1), (Shirakami Fubuki:1.4), portrait, (full body:1.2), green eyes, intricate eyes, (looking at viewer:1.2), standing, (facing viewer:1.2), detatched sleeves, (ass visible through thighs:1.3), (small ass:1.3), standing, small breasts, (midriff:0.9), skinny, tall girl

low ponytail held with a black hair ribbon and sidelocks and a side braid, strapless hooded top with a front slit and a (blue neckerchief:1.2), white detached wide sleeves, (black shorts:1.2), short shorts, a single thighhigh, thigh strap and black boots,

(morning:1.1), (beach:1.1), (palm trees:1.1), (tree shadows:1.1), (ocean:1.1), clouds
Negative prompt: nsfw, gun, weapon, holding weapon, sword, extra hands, extra arms, extra limbs, extra fox ears, bad anatomy, fat, thicc, loli, panties, panty peek
Steps: 100, Sampler: Euler a, CFG scale: 7, Seed: 2964469581, Size: 640x960, Model hash: 2700c435```
timber tartan
#

Oh, how did you manage to train on Anything?

lime flume
#

actually

#

yeah

timber tartan
#

For me it always became a nightmare

polar raven
#

Like this process, but with embedding instead, 0.05 https://youtu.be/1mEggRgRgfg

HYPERNETWORK is a new way to train Stable Diffusion with your images and the best part is: it's free! If you can run it of course, since you need at least 8GB of Vram to be able to run it. This neat technology allows you to insert any character or style you want and have Stable Diffusion generate brand-new images. So in this video, I will show y...

▶ Play video
#

I initially trained Fubuki/Mito on hypernetworks, then I decided to try embeds.

oak pewter
#

thank you so much for this!

polar raven
#

Added extra 0 when it starts to look good

timber tartan
#

Mh I tried this but didn't worked on Anything for me

patent compass
#

oh u didnt use the ti repo

#

or did u

polar raven
patent compass
#

0.05 learning rate?

#

anything doesn't usually work, even in the webui

#

that's weird, u sure its anything lol

timber tartan
#

Maybe its a miracle

polar raven
#

or 0.005 initially, whatever is the default

#

I also put 20 vectors per token, I think

#

Is it working, or is it generating incoherent garbage? Can I see?

nova patrol
#

anything doesn't support training though?

polar raven
#

It's either Anything V3 or Waifu Diffusion, I lost track of things, sorry.

#

or mixed both up by accident while training

timber tartan
#

Ah, should be WD then

polar raven
#

I couldn't generate coherently or high quality in WD, so I switched to Anything V3.

patent compass
#

trained in WD, used the embed in ANY

#

got it

#

here i thought someone figured out how to train in any

#

sadge

polar raven
#

Cause I am really new.

nova patrol
#

😭

timber tartan
#

You can tho use the trained TIs with anything

polar raven
#

and have no idea what I am talking about.

timber tartan
oak pewter
#

Mikeneko TI when

timber tartan
#

Sadge, I hoped someone solved the mistery

polar raven
#

The lower quality ones have too much variation in style.

oak pewter
#

nuuu

polar raven
oak pewter
#

maybe in a year or so, artists will have made more for us

#

sadge

polar raven
#

Nice to see Hololive fans here.

oak pewter
#

working on sakamata TI right now, can't make a single preview because of hands so far haha

polar raven
oak pewter
#

yeah, in my experience the collab gives better results somehow

#

and it's faster (when you have access to a gpu)

polar raven
#

2080 ti

timber tartan
#

Try the colab, or running the repo locally, it's really good

polar raven
#

I am right now

#

So you're telling me I can do 6x the work in 1x the time

timber tartan
#

It's free real estate (if you have enough vram)

oak pewter
#

yeah that's what intelligent people have said

#

it's apparently 6x faster

#

you can go even faster if you have more vram

#

with batch size

polar raven
#

Any help on the yaml file editing?

#

what do you recommend?

timber tartan
#

Editing the 27. And 29. Line

#

Also if you want, you can change at the bottom the preview image generation frequency
Idk exactly what lines but close to each other, modify the 100 to 200 or bigger

#

Two entries

polar raven
#

training now

#

I leave for work in 20 minutes, I guess I will train it more when I come back.

timber tartan
#

Okay

polar raven
#

Where do I see the training output images?

#

this doesn't work

oak pewter
#

go to your drive, and follow that hierarchy

#

they are in the logs

polar raven
#

Generated with Anything V3, with Nagishiro Mito V1.1, upscaled