Aine_(apocalypse_witch)

[wary kayak]

Dataset only has four images hence why I guess it would be hard,

5* images

[cunning roost]

https://media.discordapp.net/attachments/1038249717001359402/1093237375297142784/kasendance.gif

[wary kayak]

@brave tapir if you could please.

💗

[brave tapir]

Why me

[wary kayak]

because you're good.

[brave tapir]

https://cdn.discordapp.com/emojis/1058482280152174632.png?v=1&size=48&quality=lossless

[wary kayak]

pls

@quartz bramble if you ever have time

[cunning roost]

if only ik how to do char loras

[quartz bramble]

shouldn't you wait for more images to be made/released?

[wary kayak]

I don't think there will be

[quartz bramble]

how so?

[wary kayak]

it's a manga

but there's very little art of it

at all

[wary kayak]

It just needs captions

[quartz bramble]

yeah because on that other sample it was getting overfitted as hell 😄

[quartz bramble]

retraining with the new images

[wary kayak]

Yeah, I'm running one myself aswell

[quartz bramble]

then I'll do something else

here is the data I was using

[wary kayak]

were you using 128 dim ?

[quartz bramble]

yeah, too lazy to change

😄

[wary kayak]

have you ever used Dadaptation ?

[quartz bramble]

no, Adam works just fine

[wary kayak]

I stopped using Adam a while ago since Dadaption works better for flexibility.

[quartz bramble]

the optimizer plays a very small role if your dataset is good. Also I never needed to use free lr.

[wary kayak]

but you always use the default lr anyways

[quartz bramble]

kek

style overfitted

seems decent. It even doesn't do gloves if I don't prompt for them

might be pose overfitted

not the best

it works tho

[wary kayak]

I guess dataset really does matter alot in these situations

[quartz bramble]

Yeah to get it accurate you need to train enough, but that will overfit on pose and style. If you want to avoid it you need to undertrain, losing accuracy

Uhm

Well a possible solution is to generate accurate images with an overtrained lora using multiple models

Then use that to train properly

Also does she have "hair horns" or animal ears?

Or solid horns?

Because my realistic gens think those horns are made of hair

[wary kayak]

my gens seem to be having the same problems

[quartz bramble]

@wary kayak retrained after recursively feeding generated images

no need to weight pose prompts more

this one makes meat horns tho 😄

[cunning roost]

how did you guys train it to do so good? i tried, becuase i wanna learn to do characters

[quartz bramble]

this is actually pretty bad to the point I'm ashamed of publishing it

😄

[quartz bramble]

yeee new pose

uploading this thing

https://civitai.com/models/89570?modelVersionId=95351

[wary kayak]

VERY brave of you

[cunning roost]

dude its better than mine

[quartz bramble]

I'm just surprised AR has this level of accuracy with anime lora

NED is dead

I'll probably have to test DS more

[wary kayak]

@quartz bramble I fixed the flexibility issues

It's perfect now.

(almost, still can't do photorealism)

But it can do anything other than that

[vague drum]

Whats AR

[quartz bramble]

absolute reality

[quartz bramble]

did you increase the dataset with generations?

[wary kayak]

no, I merged my LoRA with yours

and it just somehow works

[quartz bramble]

oh cool. More ca layers lol

[brave tapir]

My toes hurts

[wary kayak]

Btw, this thread is finished. It's no longer needed. Thank you to everyone involved.