🍥｜anime

[barren steppe]

pls ignore hands

[alpine epoch]

Fire. I don't like the cold

[barren steppe]

aqua?🤣

[alpine epoch]

Aqua LoRA, yeah.

Needs quite some handholding, but ok.

Mostly hectic, weeping, mouth open, yelling, delirious, blank stare, leaning forward, looking at viewer, close-up Plus this LoRA.

Looks like there's an Animagine version as well, while I used the Pony version on the Animagine model

[bitter latch]

that just creeps me out

[alpine epoch]

Yes, that was rather rating:q

[bitter latch]

rating? huh? lost me

which might be a good thing

[native halo]

whats rating Q

[bitter latch]

questionable?

[native halo]

*quality

[bitter latch]

suppose to have blue hair. hmm

progress

there changed models, thats better

jams to 90's music

hha inception

[alpine epoch]

Looking crisp

[bitter latch]

throws more odd effect into the mix

well looks like the background is being sucked in but not her

😄

and there she goes 😄

[alpine epoch]

Event Horizon is strong on this one.

Can't recommend

[bitter latch]

thats better

[bitter latch]

I had the more detail lora set to 2. oops

getsa better idea

well that didnt' go as planned

but then again the most interesting effects don't

shrugs

changes to restart and hoo boy

makes cyberpunk style and damnnnn

thats better

yep i'm just being random

good night,

[gloomy lodge]

what model/prompt is this

[nova remnant]

Dalle3

[gloomy lodge]

ah sucks

is that the paid for one?

[nova remnant]

It's free but it's limited and censored

[gloomy lodge]

ew

thanks cat

[nova remnant]

[nova remnant]

I'm bored

[supple raptor]

have you considered being not bored instead?

[unkempt cape]

so after way too long than id like to admit

i have fixed issues and reduced mid-ness by 42%

man 💀

[nova remnant]

Woman

[near elbow]

neko!

[bitter latch]

firefox?

hmm sgmuniform versions are good

sampler

[bitter latch]

something this way random arrives

-goesback to bed-

[whole coral]

fire

H欧文

How to use

[keen hollow]

whoops, dupe there at the end.

mornin'

[steady grail]

morning

[versed jay]

I don't understand why when generating drawings, they don't come out looking like an anime, I put the same prompt as civitai with the character's lora, but it never generates an anime style model, it looks more like a 3d or modeling style... No It looks like a Japanese drawing, and it happens to me with all the parrots and models, how can I solve it? https://civitai.com/models/160086/android-18-dragon-ball-z-super

If you notice, this image does not have an anime style, and if you enter the link you will see that the examples do

[keen hollow]

put "source_anime" in your positive prompt

oh wait.. this isn't a pony model

not sure why you have the score tags in there

oh.. your example image is

you linked one model, but the image you generated is confetti...

similar prompt to my posts earlier across a few other models...

discord cropping.. seriously...

[steady grail]

the colors depend by how the model is mixed

you can try and fix it with loras

[nova remnant]

[versed jay]

but if what I use is parrots, I don't understand what you mean

[steady grail]

parrots is a gui?

[versed jay]

parrots = lora

[steady grail]

oh, just load another lora on top of parrots

[versed jay]

How could I do it? Is it good to mix both parrots? But how do I know which lora I can mix?

[native halo]

idk what is a parrot but do u like em fried or boiled

[versed jay]

Lora is a parrot, Google Translate accidentally translates it for me

[steady grail]

there are lora that add anime stle to gens

[versed jay]

Can you tell me a good one please?

[steady grail]

this one maybe?
https://civitai.com/models/4982/anime-screencap-style-lora

https://civitai.com/models/99619/anime-realism-control-lora

or this one

these are for 1.5 models

[versed jay]

The problem is that it is for SD 1.5, and the lora of the character that I pass to you is for XL

[steady grail]

oh, xl

xl? or pony?

https://civitai.com/search/models?baseModel=SDXL 1.0&modelType=LORA&modelType=LoCon&sortBy=models_v8&query=anime

look there

there are many examples or lora style

for xl and pony

[versed jay]

this #🍥｜anime message

[steady grail]

that's a lora, the base model?

[versed jay]

lora

[steady grail]

remember that you need to use the trigger token to use that lora well

sysdeep_android18

a lora is another model on top of the base model

[versed jay]

Yes, and even using the same prompt as in the images on the web, it generates something similar to anime

[versed jay]

Of course, and that's a lora, right?

[steady grail]

yeah, i understod that you are loading this lora model and following another existing prompt, but i'm asking what's the base model that you are loading the lora on

yep

[versed jay]

animaginexlv21 or autism

31

not 21

[steady grail]

https://civitai.com/models/260513/04betterthanwordsv30-06animaginexlv3v30

that one?

the style of this model looks pretty 3d, that explains why the 3d effects on your gens

[versed jay]

That's not very anime style, if you look at the texture it seems like the same thing happens to me, it seems as if it were not drawn, as if it were 3D

[steady grail]

yeah, that's the effect of the base model

[versed jay]

and which one should I use?

[steady grail]

it's hard to fight back the style of a model

autism is pony based, so the lora may not work properly

[versed jay]

And what model do I use?

[steady grail]

look for another xl model closer to your preference

this for example https://civitai.com/models/9409/or-anything-xl

or the model used to gen your reference image

if is present in the data

[versed jay]

And isn't there another version of Animagine that doesn't have that effect?

[steady grail]

the model you were talking about is this https://civitai.com/models/260267/animagine-xl-v31?modelVersionId=293564

i presume this is the reference image you where using https://civitai.com/images/6042860

[versed jay]

I don't remember now, but I think so

[steady grail]

that model can gen 3d images too

so, the problem may be in the pompt

[versed jay]

Is there any dragon ball style?

[versed jay]

which model?

[steady grail]

#🍥｜anime message

it's pretty flexible as i can see

[versed jay]

And there isn't one that doesn't generate 3D? Is it just anime style? and any dragon ball too?

[steady grail]

it's not so easy

try to add high saturation

to the prompt

or medium quality

[versed jay]

thanks

[serene spoke]

Hi, should i set clip skip to 1 for generate anime like images?

[keen hollow]

there's often really no 'should' - you can just mess with it. I'd say clip skip 2 is most common.

[steady grail]

depend from the model

some models require it

some other don't

just read the description of the model, or follow the metadata of another user using that model

[serene spoke]

thanks, I need to learn what sampling method e schedule type. Where i can find an help?
Why sometimes appears checkpoint in hires.fix meanwhile other times don't?

[serene spoke]

How does it?

[steady grail]

on civitai, where do you download the models

clicking on an images, most of the images have still metadata

[serene spoke]

ok, thanks can you explain me even this: thanks, what sampling method e schedule type. Where i can find an help?
Why sometimes appears checkpoint in hires.fix meanwhile other times don't?

[steady grail]

same answer of before

look on the description of the model, or metadata of other user

for the hires fix i don't understand what do you mean

do you mean, when using highres fix on the giu, sometimes appears the checkpoin option to select?

or do you mean whenn looking on metadata of other users?

i have experience only with comfy and auto1111, or his branch forge

i don't know about other gui

[serene spoke]

This is before

This is after

[steady grail]

on auto1111?

[serene spoke]

yes

[steady grail]

looks like an update, now you can upscale using a different model, sampler and scheduler

[oak spoke]

smash or pass?

[nova remnant]

Bald

Ugly

[oak spoke]

they'd probaly think the same about you

[steady grail]

smash?...

[versed jay]

Something strange has happened to me in the stable, suddenly I get this error if I use autism RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

However, if I use animagine it doesn't give me an error, does anyone know why?

[steady grail]

autism is pony, animagine is xl. very ofter, stuff for xl is incompatible, or partially incompatible with pony. maybe you used a lora, or another model for xl, with pony, or the other way around.

[oak spoke]

lmao how did that loli got there

[steady grail]

this reminds me the character of a game i don't remember

[oak spoke]

Agony?

[steady grail]

i remember in this game, she used her bony wings as spears

she was a real pain

[oak spoke]

or Scorn?

they're both hellish and creepy

I should play scorn one of these day

[steady grail]

nah, never played them

[oak spoke]

idk what game you referring to then

[steady grail]

must be something very old

maybe ps1

or something around 2000

maybe an enemie from parasite eve

or maybe not

[versed jay]

no, it happened to me suddenly

[steady grail]

just once?

[oak spoke]

shit it's cooking

[steady grail]

maybe was one of the old snes games

crono trigger or ff4

oh well, who cares

[oak spoke]

who knows lol

[steady grail]

yeah, i played so many games. happens to forget something

[nova remnant]

maybe this time im finally free from broken internet

[steady grail]

changed provider?

[nova remnant]

same one but I talked to them at least it seems stable

but discord is taking literally forever to load images

like I clicked to open this image when I first sent the message

[steady grail]

i feel you

[alpine epoch]

I'd spare some of my overkill connection if I could 😦

[nova remnant]

I can load up stuff like normal its literally only on discord

I'll resinstall it maybe it does something

[alpine epoch]

I just use Discord in a Private Browsing tab. No bs, less ads/tracking, always fresh.

[steady grail]

some servers refuse to load on web browser

[wispy canopy]

i suppose holding maps is hard in pony-land

[steady grail]

ahhhhh... you and your dreams

again...

ah, eface, your style mixer gave you problems?

i fixed it a little because locked like mixing different styles gave the model problems

[wispy canopy]

in that specific workflow, i kinda disabled it

i think

it's still just there, not connected

[steady grail]

yeah, i noticed. i activate id but i noticed the results were weird

i think loading different styles at 1 overload the model an give weird results

so i added a way to automatically lower the weight of every style token

[wispy canopy]

it worked fine for me, but i kinda use weird setup anyway

[steady grail]

it's just don't understand what that column of image text overlay does. i tested many things, but doesn't look doing anything

[wispy canopy]

both are the same model, same seed, same dimensions (minus upscale). I think i kinda even made more improvements to output quality.
(old workflow -> new workflow)

[wispy canopy]

ah, it's not connected

[nova remnant]

I think it took 15 minutes for me to install and open discord on my PC

I still can't see the servers

It's so broken

[steady grail]

yeah, i know, i attempted to connect and compete in various way, but i don't find any use for it lol

[wispy canopy]

it's to put in the artist name of the first artist block on each corner, alternating colors

[steady grail]

ohh

so, it should add a text

[wispy canopy]

when you have gazilion of these, it helps. A lot

but you can't add them at the same time, so that's why it's so convoluted

[steady grail]

that's cute

made with your workflow

i even placed better all the nodes

everything is more understandable

[wispy canopy]

oh for sure, i'm only debugging to improve upon it. new workflow is even crazier xD

[steady grail]

uahahahah

[wispy canopy]

i've found a VERY good use for kohya's deepshrink

[steady grail]

oh, really?

[wispy canopy]

yup, in the upscale step -> before, when you would denoise at sigma 1.4, you'd get artifacts

that's the lowest sigma you can go to completely resolve upscale latents

[nova remnant]

[steady grail]

oh, it works

[wispy canopy]

i think that's 0.56 in a1111? dunno what it would be

[steady grail]

usually i upscale max at 0.39

[wispy canopy]

with latents it can fuck up, so i suppose you're using an upscaler

[steady grail]

after this, i get weird mutations

this at x2

0.19 at x4

yep

[wispy canopy]

but yeah. new workflow -> pretty good detail, even if i say so myself 😄

[steady grail]

too havy on my small vram

yeah

[wispy canopy]

i mean, it's a friggin pony model, they're not supposed to do that

[steady grail]

it's good

your workflow is insane XD

you should play chess

[wispy canopy]

so many debug shit left in there lol

[steady grail]

you need such multy tasking to follow such a mess

[wispy canopy]

my desktop is the same -> i never look at it anyway

[steady grail]

ehehe

well... that's a mess to me too

they are always a mess, unless are for work

[nova remnant]

is it pony or autism something

after 1 minute loading that image

[wispy canopy]

same model i've been using all the time

but i've been able to squeeze out a lot more detail from the same model

@steady grail am cleaning up my workflow atm ^^ remove all debug stuff

[nova remnant]

autism I assume

[wispy canopy]

my own merge -> red revision

[nova remnant]

[wispy canopy]

i posted an image using the default pony turbo, on another server you also frequent xD

i just can't post it here...

[wispy canopy]

aight, i've been doing enough research on this dark tech, imma go to bed for now ^^

[keen hollow]

evenin'

[oak spoke]

"oh ho ho you dare to approach me?"

🤣

[green temple]

Looks like hanami from jjk

[nova remnant]

hey peoples

[brittle rampart]

hi hi, Cat!

what a news?

[nova remnant]

I'm.. tired

Big news cx

@brittle rampart how are youuu

Here's some floppy

[brittle rampart]

tired after a hard day or night?

[nova remnant]

It's 2am now, I was just on my PC

I'm lazyyy

[brittle rampart]

@nova remnant btw. I saw you trying to grab my mouse? bad Cat!

[nova remnant]

Is this animagine

[brittle rampart]

XL model?

[nova remnant]

[nova remnant]

Animagine is xl yes

[slate narwhal]

Neta Art XL, the model that's supposed to replace Animagine

[brittle rampart]

I feel like vintage!

all here use SDXL

[nova remnant]

Interesting did you use animagine before it?

[nova remnant]

You're not

[brittle rampart]

not!

[nova remnant]

Meant to reply to this one but still

1.5 looks super good still

It's just not as good with hands and some stuff I think it looks better

[slate narwhal]

Yes, but didn't use it for very long, went straight to Pony

[brittle rampart]

btw. Pony for version 1.5 is also good?

i have a Pony model for SD 1.5

[nova remnant]

I heard it's weird

Trying it shouldn't be too much of a problem

There's probably a tutorial on how to use it on the model page

[slate narwhal]

I miss some stuff from 1.5 that I can't replicate with SDXL, Pony or whatever. It was easier to create our own mixes with very specific stuff. Good times.

[brittle rampart]

mixes - models and LoRa?

[nova remnant]

You can merge models and loras

Mix is model + model

I think

[slate narwhal]

No, only models. SD 1.5 still has better tools for mixes like for example runtime block merge.

[brittle rampart]

@nova remnant @slate narwhal no. no.. for me it’s generally a “dark forest”. I don't understand the creation process at all..

[nova remnant]

[nova remnant]

I also don't know much of it, I leave it for the smart guys

[slate narwhal]

Yeah, and SDXL training/merging is harder, so people just go for Loras. Problem is that you have to use lots of Loras. I prefer to prompt directly as much as I can.

[brittle rampart]

Yes. I love long prompts too

[nova remnant]

Am I the only one that uses this for nsfw

[brittle rampart]

"That" - what?!

[nova remnant]

You can't post nsfw here, just a heads up

Or else !

[nova remnant]

SD

[brittle rampart]

really?! 😁

[nova remnant]

Make all the grabby gens

[brittle rampart]

it's horrible!

[nova remnant]

Grabby

[brittle rampart]

no no...

[nova remnant]

Yes

[brittle rampart]

just a quickie

[nova remnant]

Thighs 👍

[brittle rampart]

very strange. I made a post of 3 pictures on Civitai and only two were published

[nova remnant]

send me your account

I want to see your posts cx

The image probably had something in the prompt, it should give you a warning saying what it was, if it wasn't accepted

I'm pretty sure

[brittle rampart]

https://civitai.com/user/LSLisa

[brittle rampart]

the system didn't like my dog

[nova remnant]

I'm banning civitai!!

[brittle rampart]

m?

[nova remnant]

Just kidding :p

How could they do that to you

[brittle rampart]

it a crazzy

[nova remnant]

Sleepy ?

I am sleepyyy

[brittle rampart]

nope

[nova remnant]

Confused then?

cx

[brittle rampart]

yes..

[brittle rampart]

How can you not sleep whole nights?

[nova remnant]

I think I sleep late because I can't when it's hot

@brittle rampart your favorite style

[brittle rampart]

and how hot is it?

[nova remnant]

:p

[brittle rampart]

yep! i like this is styles

[nova remnant]

Not much today, it's common to be hot sometimes 40°c or more then it gets really bad

[brittle rampart]

+40!? in night!?

[nova remnant]

When days get to 40

The nights go down to like 35 or something

Some days were impossible this year

@brittle rampart more bigger ears!

[brittle rampart]

do you really live in Africa?! i have only +7 in night, and now +11

[nova remnant]

No x3

I'm from Brazil

[brittle rampart]

i am from Ukraine

[brittle rampart]

1 and 2 style cute

[nova remnant]

Ball of fluff

[nova remnant]

Miku comes out of space ship landed in the desert

[brittle rampart]

Miku broke space ship

againe

[nova remnant]

@brittle rampart I found the magical book about floppy ears

[brittle rampart]

send me Fidex!

[nova remnant]

Coot

[bitter latch]

combines the two

[brittle rampart]

this better

[wispy canopy]

same prompt, on default pony. who needs finetunes anyway? 🤷‍♂️

on zavy pony -> ok, finetunes do help image quality along a bit ^^

[bitter latch]

what is this pony thing people are speaking of? MLP?

[wispy canopy]

pony is a lewd model

basicaly best anime/furry lewd model there is atm

[brittle rampart]

SDXL model

[wispy canopy]

yeah, we dont' talk about the 1.5 model 😄

[bitter latch]

ahh

[brittle rampart]

why not?😂

[wispy canopy]

there's actually people using it?

[brittle rampart]

i am not use SDXL model. so... 🤣

[bitter latch]

screen going blurry think my brains finally tired enough to sleep... thud

[steady grail]

they trained the 1.5 model with the same data of the xl

but being the 1.5 smaller, the model is overtrained

hello guys

for whati read, it was just an experiment

[brittle rampart]

oh those “they”. took it and trained something?

[steady grail]

yeah, it's on civitai. but i proved it and i find it unmanageable

i get burned or rough images very often

the results are unstable

i personally don't like it at all

[brittle rampart]

you mean PonyXl model?

[gloomy scroll]

pnginfo?

I'll give smug in exchange

[steady grail]

pony 1.5

districated eface workflow

i'll study and work on it.
probably i'll set it for 1.5, or for a pony low step

[brittle rampart]

ah this is...

[steady grail]

i can see very interesting settings, but i can improve some aspects of it

[brittle rampart]

I tried to do something on this model (Pony 1.5) a couple of times and still didn’t understand how to work with it correctly.

[steady grail]

me too

i just surrendered

[normal nova]

Summoning ritual complete

[brittle rampart]

Now smooth out all the wrinkles.🤔

[normal nova]

I kinda like the sketchy look personally

But I understand anime is usually smooth 🤔 I'll try to make a smooth variant

[brittle rampart]

no no it good!.

don't pay attention..

[gloomy scroll]

What's that model?

[normal nova]

AutismMix, the pony version

No lora, I added some style tags tho

[gloomy scroll]

What's with the EMS-294683-EMS name?

[normal nova]

I use tensor.art to generate, so probably a quirk of their backend

Makes all the models show up as EMS something in meta

[gloomy scroll]

[normal nova]

One man spam pog

Will check back lata

[normal nova]

smüth

I made some smooth ones myself but didn't post them

[gloomy scroll]

Is there difference between smooth and non smooth prompts?

[normal nova]

Mostly switching between autismmix and pony I'd say

But I also got some style tags and loras to go one way or the other on the same model

[barren steppe]

fixed seed sometimes give interesting results

[normal nova]

Sdxl isnt much better at cocktails than 1.5 lol

[normal nova]

What’s the tag for these eyes ? Last time I checked it was something like « +_+ » but I think I ended up with the conclusion that it didn’t work

Eldritch roombas pog

[brittle rampart]

she ate a blue octopus!?

[gloomy scroll]

I used slit pupils

@normal nova @brittle rampart Those blue things are obliviously roombas

[junior quest]

https://tenor.com/view/kupo-gossip-anime-give-me-money-gif-9766049

prompt

[vital raptor]

getting cross shaped pupils is hit or miss with autismmix, seems to be pretty random, lots of others work though

[ivory pebble]

How do you prompt for that ":3" mouth?

I've had some success using "smug" but it's kinda random 😭

[vital raptor]

i don't, but 'cat mouth', :3, or 'uwu face' work on some models

with autismmix it's harder to get it if you don't have cat ears in the prompt, but I got it to work with cat mouth

one of the artist tags I use makes that mouth if the image has cat ears and smile/happy/smug/etc

[ivory pebble]

I think putting "cat" in the prompt might give better results along with "smug" but sometimes it just generates random cats in the background instead

I'll try adding cat mouth and see how that goes

[vital raptor]

using 'pointy ears' instead of cat ears i get the first image (left), adding 'cat mouth' gives the second image (same seed)

it does also change the image on her shirt to cats though

oddly enough, you can try putting 'cat' in the negatives with cat ears/cat mouth in the prompt and it will try to remove the random background cats

but it doesn't always work, just something i use sometimes if it's struggling

[vital raptor]

same seed/prompt with (:3 face:1.4) instead of 'cat mouth'

[ivory pebble]

Yeah it likes adding cats in the background or on the character for some reason

For some reason I feel like negative prompts don't really work all that well with Pony, tried it a couple of times and it either changes the style I'm using or it just doesn't work at all so I've stopped using negative prompts

[steady grail]

yeah, i noticed sometimes even exacerbate something i'm trying to eliminate

for example, adding crooked eyes, made the eyes more crooked lol

[vital raptor]

it's also artist dependent with pony based models, this is the same model seed and prompt just changing the artist tags, I had to weight 'cat mouth' to 1.4 just to get it to work, but then it adds ears

[ivory pebble]

Well it does say on their Civitai page that the model doesn't require negative prompts to work, and considering how the training with PonyV6 was messed up I'm guessing that messed up the negative prompts as well

[vital raptor]

sdxl as a whole doesn't work well with negatives, pony models are just more obvious

[ivory pebble]

I keep forgetting weights are a thing lol, I'll definitely have to try that

[vital raptor]

but, 1-4 very specific keywords can help push an image in the right direction sometimes

with this artist combo, (cat smile:1.3) starts to work, 1.4 weight adds the cat ears again

it's just a lot of trying different things until one sticks 😄

and then i change back to the original artist tags i was using and i don't even have to prompt for it 😄

one of the reasons i use this particular artist combo though

the expressions are so good

[native halo]

https://tenor.com/view/free-money-meme-fast-gif-gif-16935040457953192433

[supple raptor]

@wooden coral it's ban time

[terse ore]

Is autism as versatile as pony, just not as random and more controllable? Does it skew towards anime more than comic?

[vital raptor]

from my personal usage, they behave about the same for everything besides style
if not using artist tags, pony will randomly change styles almost every single generation and it's rarely consistent (depending on the prompt of course)
autismmix_confetti on the other hand will pretty much stick to a single default style

If you specify an artist, autism tends to stick to it pretty good, if you have 2 or more artists, there's slight deviation between them. For most combinations I've tried with 2 artists I get about 4 subtle, but distinct (to me), actual styles styles and it just rotates through them. Generally the differences are pretty minor like as if it's changing the contrast or saturation just a little between images

on the other hand, I haven't used pony even remotely as much as autismmix because autismmix stuck to my test prompts better and delivered consistently better results for what I wanted

[jade violet]

Has anyone tried out Neta? https://civitai.com/models/410737/neta-art-xl

[vital raptor]

is there a reason to?

it doesn't seem very good

[jade violet]

It’s supposed to be the new SOTA for anime in SDXL, just dropped today so trynna get a consensus. Some are saying it’s considerably better than animagine

But also I’m not sure if it’s just astroturfing, the Civitai reviews seems sus

[vital raptor]

to me it doesn't look very flexible and the (lengthy) notes on it pretty much say it's bad except for one specific prompt structure/style, so unless it happens to do the one single thing you want from it, it's probably just another model in the pile of garbage on civit

I mean, I have my own maybe ok model on there, it was merged from other models for a specific use, and if that's what you want, then it's great, but if you want something more flexible then use a good model

but the only way to know if it works for you is to try it and see, everyone has different tastes

[jade violet]

Yea great point, guess I’ll try it for my own use case 🤷‍♂️ appreciate your advice!

[oak spoke]

accidental hunchback

[twilit portal]

Lucoa (dragonmaid) in Indian Saree ✨

|| @nova remnant meow~~||

[nova remnant]

Google AI, not SD

[nova remnant]

[terse ore]

I copy pasted this from a comment which has some positive feedback, anyone try it? :

If you want pony (and are running a local install) but HATE anime and want photo realism:

You've heard great things about Pony XL/Pony diffusion, but you're not into anime. No problem.

Here's how to get the benefits of Pony XL, without the drawbacks of art-style.

You'll need two checkpoints...

"Real Pony" checkpoint (there are three, you want the one with the most upvotes on civit to start): this will be the primary checkpoint.

Second, you'll need the photo style SDXL checkpoint of your preference. If you want NSFW capable, make sure you use one capable of that. Use your current preference.

On to generation...

With "real pony" as the primary, you'll set up your other checkpoint as "refiner." Refiner should start at the last 30 or 40 percent for a 50 step generation. (I know, 50 is a lot! But it's to give realpony a chance to build the bones of the photo before the refiner has it's go to cut out the remaining anime-ness.)

Give it a go. 50 steps. Refiner start at .6

This combo can be used with or without standard pony prompting to differing utility.

If you want, try these positive and negative prompt preambles.

Positive: (score_9, score_8_up, score_7_up, score_6_up)

Negative: (score_3_up, score_4_up, score_5_up), sketch, drawing, cartoon, greyscale, monochrome, anime, 3d

[ivory pebble]

I've seen some people use this same method with Pony (or rather AutismMix) and Animagine, mainly because they want to get rid of the default thick lines AutismMix generates, however I haven't tested it out myself since it didn't look that good honestly

[steady grail]

burn the witch 🔥

[terse ore]

I see. Its supposedly great at realistic imagery while using pony for the coherence. But I haven't actually seen what the outputs look like

[steady grail]

this suggestion should work to the other part too

to make 3d or realistic images more anime

[nova remnant]

It looks okay I think

I didn't see much but it looked like a realistic models

The guy was saying that using only the real pony model makes weird faces

Which normal pony also does when trying for realistic stuff lmao

[terse ore]

I'm sceptical tbh, the vast majority of realistic models still have that artistic tinged skin for example, so how would something that starts off with pony look more realistic than a dedicated model? Might try it tonight but prepared and actually expecting disappointment

[nova remnant]

Because pony can do realistic stuff

But kinda bad

[ivory pebble]

Where did you find that comment? Was it on Civitai? I wanna see if that person posted some examples. From what I've seen that "Real Pony" still doesn't really make realistic looking people, all of the examples have that plastic look to them. If that method actually does work then maybe it's worth using to test on Pony with different anime checkpoints

[terse ore]

https://www.reddit.com/r/StableDiffusion/comments/1c9lenu/comment/l0mxygk/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button this is the comment. A few people say it's good but no imagery posted

[ivory pebble]

Yeah I can't find anything either sadly

[primal furnace]

would merging pony w/ a realistic checkpoint work?

[steady grail]

merging another model on pony is not the best idea

would be better to train it with more realistic stuff

[primal furnace]

Why? It'd get confused?

[steady grail]

pony is not compatible any more with other xl models

i don't know what would you get out of it

[primal furnace]

oh ok

[steady grail]

to merge 2 xl models you need a big a** machine

and for last, the training on pony is already very solid.
merging it with a worse model, would make it more dumb

so, better trying to train it to make it more realistic

[primal furnace]

ahh, interesting

[steady grail]

usually the other pony base model are models merged with lora, or model with more training

[primal furnace]

oh neat, I didn't know you could merge a lora. (I just started SD last week)

[steady grail]

yep, but honestly, i prefer the original

[primal furnace]

training in general takes a ton of computing power, doesn't it? Unless it's LoRA, textual inversion, hypernetwork, or dreambooth?

[steady grail]

it's more responsive

yeah

[primal furnace]

yeah my current workflow for more realism would probably just be img2img with a different checkpoint or something, haven't tried it yet

[steady grail]

someone sait that to train an xl model you need at least 10tb of suff

[primal furnace]

stuff? Like training images?

[steady grail]

i suppose so

[primal furnace]

10 tb no problem, peanuts 😆

[steady grail]

imagine tagging all that images manually, one by one lol

[primal furnace]

oof

[steady grail]

that's what astra did on pony

a champion

[primal furnace]

no way, that's amazing

[steady grail]

or at least that's what i heard

anyway, that model is a very big surprise if you love anime stuff

[primal furnace]

if that's true then they need some serious lauding

even if they didn't do it 1 by 1 lol

I am not so into anime, but I don't mind it! That's why I was considering options for realism mixed with it

I used to watch Ranma lol

[steady grail]

it started as a passion project, someone trusted him, sponsored him and now behind that project there is a very passionate community

[primal furnace]

nice

[steady grail]

i grown with Ranma 1/2 too

too bad, don't have a real ending

[primal furnace]

sweet! I still have the mangas somewhere, should dig em out

[steady grail]

yo nubby

[supple raptor]

this is a bit of a misunderstanding of what I was talking about. what I was actually trying to say is that in order to create a highly flexible full finetune, then you're looking at datasets on the scale of terabytes of images.

[steady grail]

oh, but i heard such number elsewhere too, it's not the first time i heard about 8 10 tb of stuff

looks like is a common conception

[supple raptor]

8tb is probably a pretty common number that gets thrown around since that's the size of the danbooru2023 dataset which is a popular dataset

[steady grail]

follow afterward by the statement that is such an insane and mind numbing work that it's better to train a lora with 100 150 images lol

[steady grail]

oh!

[supple raptor]

fwiw, large scale training is very possible on consumer grade hardware with LoRA variants. Kohaku-XL models are some great examples of this.

[steady grail]

training a base model, merging it with many lora?

[supple raptor]

kxl-delta and epsilon at the very least were trained as LoKR on 3090s then those were merged with the model for distribution.

[primal furnace]

cool, I'll check out Kohaku-XL for more details about that. Thanks!

[supple raptor]

here's the model page for kxl-epsilon https://huggingface.co/KBlueLeaf/Kohaku-XL-Epsilon
which includes a link to KBlueleaf's discord server (same person who trained the Kohaku models and who is responsible for https://github.com/KohakuBlueleaf/Lycoris

[primal furnace]

awesome! So these are models trained with LoRAs? Is that what a LoRA variant is?

[supple raptor]

Yeah, specifically they were trained as LoKr which is within the LyCORIS github repo

[steady grail]

the last lora merge of pony i saw, lost that gummy style and wasn't able to gen some pones any more 😦

oh, that's interesting

[primal furnace]

cool! New to LoKr, too

no ponies? Tragedy.

[steady grail]

the ponies where still there, but their pose where more rigid and harder to change

if only it was compatible with openpose 😦

[primal furnace]

oh I still have to check that out, too

there's sooo much stuff

is that like a control net?

[steady grail]

yep

[primal furnace]

cool

[steady grail]

i don't like it too much, if the poses are not compatible with the model, the character takes weird forms

[primal furnace]

lol I can imagine

[junior quest]

do we have a model that can create neon glows like this?

[versed jay]

https://civitai.com/models/288584/autismmix-sdxl I have tried to make the texture look like an anime drawing but there is no way, this is my model, nor does mixing loras with anime style work either, however with the animagine mixing an anime style lora it does work for me, what can I do to make it do anime appearance with autism?

@steady grail

[steady grail]

no idea, i'm terrible in merging XD

i know the theory... but, i'm just terrible

[versed jay]

Isn't there another model of autism that is more anime?

[steady grail]

confetti is more anime

[junior quest]

https://civitai.com/models/317902/t-ponynai3
https://civitai.com/models/282341/4th-tail-hentai-model
https://civitai.com/models/280321?modelVersionId=315675

what bout these?

[versed jay]

well it doesn't work for me, not even using an anime lora and doing the prompt like on the lora website, it doesn't even do it anime style

[steady grail]

yeah, it's hard

mayhbe if you ploay with weighted block merging

but it's a little complex

[versed jay]

what does that mean?

[steady grail]

is a way to merge only some blocks in percentage to another model

[versed jay]

and how is it done?

[steady grail]

imagine to move the anime style to another model, without touching his flexibility

is a little complex

there is a mergin extension to do it

i could only do on 1.5 models because i don't have enough memory

but it's pretty fast

[versed jay]

what extension is it?

[steady grail]

https://github.com/bbc-mc/sdweb-merge-block-weighted-gui/tree/master

i just understood that merging is not my stuff, but that process looks funny

i love understending how such stuff works, more than using it

[versed jay]

wow, nothing is understood

[steady grail]

eheheh, i just suggest to try and see for yourself XD

if you have enough memory to load 2 xl models, every merge don't require too much time

[versed jay]

But how do I know what I have to put so that it only takes the anime part of one, and the rest of the other?

[steady grail]

by try and error

that's what i did

[versed jay]

Can it be done with a Lora? I mean, get the anime style of a Lora, and then have autism use the anime style?

[steady grail]

i left all the blocks to 0, except 1, set to 100

i don't know

i don't think so, but you can try to put a lora in the checkpoint try starting the merge

maybe it works and does something

that's the funny part

[versed jay]

What do you mean by checkpoint?

[steady grail]

another name for safetensors

the base models

[versed jay]

In any case, if only 2 models can be merged, the problem is that there are no more anime models for autism, and if there were, I would only have to put the other model, or can models from different versions be combined?

[steady grail]

no

only same version

autism is based on pony

you can merge it with any pony model

[ivory pebble]

Not really but you can always try using "Pony Diffusion for Anime" and "Js2Prony", both of those are nicer for anime compared to Autism

[versed jay]

thanks will try

[ivory pebble]

Also make sure to use the V1 version of Pony Diffusion for Anime, heard some mixed results for v2 so i'm not sure if it's that good, but you can try it out and see how it goes I guess

[versed jay]

https://civitai.com/models/315596/pd-for-anime this?

[near elbow]

mixing ponies, you cruel folks...

[versed jay]

@ivory pebble

[ivory pebble]

Yup

[normal nova]

Here you go
positive : score_9, score_8_up, score_6_up, otic, ssal,
1girl, gothic, horns, purple skin, wings, scales, facial scales, glowing eyes, extra eyes, frills, bedroom, curtains, skirt, smug, leg band, lace, white hair, bangs, pigtails, on floor, wooden floor, wariza, on knees, arched back, narrowed eyes, (pentagram)
negative : score_6_up, score_5_up, score_4_up, (mgfx, zvu), qcz, greyscale, 3D, tail, sitting, squatting
ponyXL
concept art lora at 0.3 weight

[nova remnant]

[normal nova]

Oops I replied lol. Get pinged for no reason

[normal nova]

I was trying to only get one character but it gave me 2, I kinda like the contrast tho

[nova remnant]

I'm immune to pings

Have it all muted

[normal nova]

I almost sent you a super loud pipe sound but then decided it would reflect badly on me

Man I can't seem to prompt head down. Kind of like this but looking at viewer :

If anyone magically knows a way to make that in pony models I'll be happy

[green temple]

🤣 damn

[green temple]

wooooooow

legendary

[normal nova]

They kinda look like you'd have to fight them together in a boss fight or something

The left one would attack with claws and the right one with a big ass sword

[green temple]

yep

[normal nova]

And for the last phase they would fuse o_o I'm gonna try to make the final form

@green temple best I got

[green temple]

awesome !

[oak spoke]

"hey"

[normal nova]

She's like "don't look at him don't look at him don't look at him" 😆

[oak spoke]

🤣

[normal nova]

Wait I know exactly what this reminds me of actually

https://tenor.com/view/death-note-mentalwards-s4nri0-gif-18770878

Actually it's not exactly that similar, just watched death note recently so it's fresh in my memory

[oak spoke]

it reminded me of that one anime where girl can see ghost

and she must make sure the ghost doesn't know she can see them

🤣

https://tenor.com/view/mirukochan-gif-23329760

[normal nova]

Oh yeah that's way more similar lmao

What anime is that

[oak spoke]

Mieruko-chan

[normal nova]

Alternative final form

[terse ore]

Any recommendations that's good for doing img2img on backgrounds) scenery/buildings?

[terse ore]

3 sets of images. Please let me know which one is the better one from each set. Testing for a manga (although I will probably have to do more testing)

[supple raptor]

you need to add at least basic complete ground and backdrop before any of them are really worth trying in img2img. img2img is just using the input img in place of random noise so it's primarily providing color locations and general shapes. the grey backgrounds won't work well.

[vital raptor]

wtf, I just noticed that 'random' seeds are not random..

[nova remnant]

Come here grabby

[vital raptor]

I made the above image, then generated 30-40 images.
Then I changed the prompt a little and went back to the seed from the above image to compare, then a1111 "randomly" generated the exact same 30 seeds in exact order

[nova remnant]

[nova remnant]

Illuminati confirmed

That's it

[vital raptor]

hmm, must be something to do with the unprompted extension because the prompts are also exactly the same "random" prompts

actually though, it uses it's own seeds, so maybe not

it's just generating the same prompt because the seeds are the same

[nova remnant]

[vital raptor]

trimmed up the selection a bit, but it's pretty obvious the seeds are repeating

i'll have to test a bit to see what's causing it i guess

[nova remnant]

Hello I recommend a VAE

[supple raptor]

oh what's that? i'm new to sd.

[vital raptor]

ok, confirmed it's the unprompted extension, it's doing something to make seeds deterministic for some reason

[dire bridge]

Bruh

[brittle rampart]

An interesting monochrome sketch with an emphasis on one color. I would like to find LoRa that implements this for 1.5 or key tokens for the prompt

[nova remnant]

@brittle rampart funya

[brittle rampart]

nyaaa

[nova remnant]

When you say that

[brittle rampart]

😁

[supple raptor]

sketch, graphite \(medium\), (monochrome:0.9) with a single color prompted somewhere in the prompt will often do it if the model isn't too overfitted on full color images. Sometimes adding spot color can help.

[brittle rampart]

must try it . thanks!

[nova remnant]

Big eyes very

[supple raptor]

This LoRA is also one of my favorites for 1.5 models. I use it around 0.6 strength. Gives a very nice sketch look, but doesn't really help with the spot coloring. https://civitai.com/models/110244?modelVersionId=118849

[brittle rampart]

yes i have this is. it really good LoRa

[nova remnant]

Smol

[brittle rampart]

Well, we can say that this color accent effect was almost a success

thank @supple raptor add this is to my "tips and tricks" notes

[brittle rampart]

not to smol! just... just a perspective distortion effect 💖

[nova remnant]

I'm eepy

[brittle rampart]

gn!

what a long tail!

[nova remnant]

I'll go in like 20 minutes

[nova remnant]

To grab all the mouse girls you are genning!

[nova remnant]

Pikachu x3

[brittle rampart]

looks almost like kigurumi

[nova remnant]

I dunno who that is

Tail bed

grabs you into the dream world

[brittle rampart]

she calls you to sleep

[bitter latch]

slept to much today. i'm tired but brain will not shu tthe fk up

falls into the void

[nova remnant]

Grabby

👍

Flowy flowers

[brittle rampart]

first look postcard for holiday March 8

[brittle rampart]

@normal nova last better!

[normal nova]

Which one, the one I removed ?

Or you mean the 2nd one

[brittle rampart]

i know. before last was not good. now last better

[brittle rampart]

except to remove the vampire fangs. it will look cuter without teeth at all

[normal nova]

Yeah, I removed them now. Trying to make more

Overall it's not easy to get good cheese

[acoustic locust]

What model?

[normal nova]

AutismMix Pony version

[acoustic locust]

ok

Why are the hands so good?

[normal nova]

Several factors I guess ? Prompt, hi-res and some luck

It's not like I get hands this good every gen at all

[acoustic locust]

I know

[normal nova]

moar mice since chat ded

[jade violet]

Wow what checkpoint is this?

[normal nova]

Also AutismMix Pony version

[normal nova]

Prompt was also really simple :

score_9, score_8_up, score_6_up, uum, mjm, bomm,
wink, mouse girl, holding cheese, notice lines
Negative
score_6_up, score_5_up, score_4_up, mgfx, zvu

[brittle rampart]

uum, mjm, bomm, - this is magic spel!

[normal nova]

It actually is 😆

rabbit version

[jade violet]

Thank you!!! Do you prefer it over animagine?

[normal nova]

These days yeah, but I havent used animagine all that much. I like pony based models because I'm quite experienced with them

Also pony has the aforementioned magic spells and those are fun

Nothing wrong with animagine tho

[jade violet]

Gotchaa, do you use the confetti version much? Trying to gather what the difference between that vs pony version is

[normal nova]

Ive settled on the pony version, but it's hard to tell which is strictly better. I use the pony version because instead of being a mix of pony with another model, it's just pony with loras merged in, which means it works "more like pony"

oh yeah I didnt even realize

[brittle rampart]

😂

[jade violet]

Yea that make sense, appreciate the insight!

[normal nova]

I quite like this style you use !

[brittle rampart]

ah this is just oldschool 1.5 🤗

[normal nova]

Yeah I saw you mention that. Especially nice since it's 1.5 !

[brittle rampart]

I just can't remember magic spells 😉

[normal nova]

Omg no, she let go of her carrot basket !

[brittle rampart]

Well, she stole one carrot! the rest of the basket is just evidence!

I'm embarrassed to ask - is she from Pokemon?

[normal nova]

Yes, as long as you have one carrot, you can be happy. Thank god

[normal nova]

Nah, I didn't prompt for a specific character, just a rabbit

One of the magic spells is probably based on some character though, it keeps adding these hair ball things

[brittle rampart]

I'm just talking about this strange hairpin

[normal nova]

When my prompt is just "bomm" I get this

It's definitely a character from somewhere, I just don't know which

[brittle rampart]

now she look like VTuber

[normal nova]

Yeah, sort of like gawr gura

But not quite

[jade violet]

Is civitai down for yall?

[brittle rampart]

he work very very slowly

yep. not work

[bitter latch]

😄

hmm .. o. k then

sighs

[brittle rampart]

I want to go to that flower bed that is behind her!!!!

with a fork and sugar!

[bitter latch]

umm... wait a min!

gotta say these SGMuniform samplers are good

even at low steps and cfg

sets 1 and 1 presses generate and waits for horror

yep .. ew

10 and 3 work 😄

[brittle rampart]

strawberry crown!?

[bitter latch]

shrugs

[brittle rampart]

and why not!? you are an artist - you see it that way!👍

[bitter latch]

well I didn't prompt it so.. shrugs

[brittle rampart]

best!

damn! now I want strawberries!

[bitter latch]

changes it up , grinz evilly

😄

[brittle rampart]

strawberries!!!!

space strawberries!!!

[brittle rampart]

Princess Leia's childhood?

[bitter latch]

turns it the other way

ok now it's went nutso haha

[brittle rampart]

Guardians Of The Galaxy Awesome Mix Vol. 1

first track

[brittle rampart]

wow! Has the concept changed?

[bitter latch]

cackles

[bitter latch]

sneeezes, hmm someone's talking about me

the fruits alive

[brittle rampart]

now cute!

now kawaii!

[bitter latch]

ok

1 more and i'm laying back down, my sleep is not good today. blah

[brittle rampart]

gn!

[brittle rampart]

new queen!

[normal nova]

She will make all strawberries 20% sweeter

[brittle rampart]

but she looks quite menacing!

[normal nova]

It's because she just tasted strawberries and they weren't sweet enough

[brittle rampart]

and to blame....? artist!🤣

[bitter latch]

she sleeps

[brittle rampart]

they a sleep

[brittle rampart]

girl - garden bed!

[brittle rampart]

Better try a pose in which she will hold a large strawberry in her hands in front of her. slightly tilted back under its weight. and peeking out from behind the fruit. pretty smiling.

[normal nova]

Seems hard to prompt

Maybe doable if I img2img a sketch

[normal nova]

@brittle rampart It's a start !

[brittle rampart]

👍

[normal nova]

Had to use a blue dress otherwise it was too hard to make a sketch

[brittle rampart]

can show sketch?

[normal nova]

lol

I used img2img on this

[brittle rampart]

interesting.

how?

never try this is way

[normal nova]

Another :

[normal nova]

Well first I draw in a random drawing app, and then I img2img. I don't know what you use to generate but surely it allows for img2img ?

[brittle rampart]

hmmm .. some like that?

[normal nova]

Yes !

So cute

Although its not always easy to make SD "understand" what you drew

And since I only draw in simple colors, it doesn't work so well for 3D or 2.5D styles

[brittle rampart]

very interesting method....

[normal nova]

I'm glad you like it ! Just need to spend a little effort drawing something and as I said it tends to simplify the image a bit

What denoising did you use for those ?

[brittle rampart]

first resolution how a big?

[brittle rampart]

0.75

[normal nova]

wow that's low, I'm impressed you got so much detail

[normal nova]

What do you mean ?

[normal nova]

They're 1152x768

But it's an XL model so maybe more appropriate

[brittle rampart]

forgot.. you all use SDXL.... but

how you upscale in this is method?

[normal nova]

https://www.reddit.com/r/StableDiffusion/comments/xkjjf9/upscale_to_huge_sizes_and_add_detail_with_sd/ I guess this explains how

But I should mention that I don't use A1111

[brittle rampart]

ah. now i am understad... just sent result in Img-to-img and Upscale

[brittle rampart]

variation on a theme

[nova remnant]

steals your strawberry

[nova remnant]

Model and prompt?

[normal nova]

Model : MS PAINT
Prompt : #🍥｜anime message

[nova remnant]

Cx

[nova remnant]

I'm mspaint too

My profile picture I mean

[nova remnant]

Took a bit too long to make this safe version to post here