🍥｜anime

[rough oxide]

nyoooo

[lyric zealot]

[rough oxide]

the only "smaller" thing u can substitute here

is

the CLIP

but i'm using full clip

also

even if you meet the vram req you need 64gb ram

Lol

cuz the full CLIP uses 40gb RAM

well

edge is probly using some of my ram of course

but over 32 for sure.

[lyric zealot]

Jesus

[rough oxide]

even the fails r fun to look at

Lol

[lyric zealot]

Is it the something clipt5

[summer night]

is there going to be a way to optimize it without gimping it, like LLM quants?

[rough oxide]

https://comfyanonymous.github.io/ComfyUI_examples/flux/

[lyric zealot]

Try your signature waifu

[rough oxide]

t5xxl_fp16.safetensors AND clip_l

[lyric zealot]

I can’t even run the t5 clip for sd3

[bitter latch]

haha

[rough oxide]

migu.

[stable crane]

I really dig the flux eyes

[stable crane]

floral print is 👌

[slate narwhal]

First is Pony + Loras etc. Others are Flux Pro. I really like it.

[lyric zealot]

[lyric zealot]

daymn

the tie is so good, almost like you have to segment it out, and then inpaint it

[lyric zealot]

the second one is really good

there's still room for improvement, but damn

open source is back

but then you need giga amount of vram

[slate narwhal]

Yeah it's really good for a base model.

[lyric zealot]

imagine the finetune

[stable crane]

quite good

[slate narwhal]

Yeah, it will be epic. This is better than the SD3 Ultra pipeline. Here is a near perfect generation using the scorpion prompts (really difficult for models to get right; all SDXL models fail this one). I've asked for an anime scorpion, though.

[lyric zealot]

i can see the midjourney feel in those two pics

[slate narwhal]

Yeah very similar to MJ's aesthetics

[rough oxide]

omae wa mou shindeiruuu

lol

[stable crane]

me on the left

[slate narwhal]

cyberpunk anime centaur girl tests.

[wispy canopy]

flux is amazing

[stable crane]

agreed

[stable crane]

you doing local rendering?

[rough oxide]

flux ish blowing my mind

lmao

[bitter latch]

hmmms

[green temple]

ohhh !

[rough oxide]

military kitties.

[bleak cradle]

art

[rough oxide]

missing one word

but

wtf

flux is op

[green temple]

is it ok to post this ?

[summer night]

i am making creatures

[bitter latch]

ok about my bed time

one last one

there we go

thuds

[wispy canopy]

yes.

[brittle rampart]

I decided everything! I'm selling a kidney! I can’t live without FLUX locally!

[rough oxide]

(the potion prompt was actually fulfilled 100% almost every single time.)

[zinc flax]

she just wants to talk trust

[nova remnant]

Pupol

[bitter latch]

goes back to bed. thud

ok so i'm not in bed yet.. 😛

[bitter latch]

show off

[bitter latch]

haha

[fiery shadow]

is that flux?

[zinc flax]

That's pony with concept lora

[fiery shadow]

okay thanks

[crystal atlas]

making lora instead of doing homework...Yay me.

[gloomy scroll]

Happy vampire noises

[stable crane]

nice armor

[bitter latch]

there we go

[desert island]

damn it... my friend just exposed me to silly tavern... smh... days will be lost

[rough oxide]

doesn't seem that gud.

[desert island]

it's just a skin for your llm's with tons of very usable and well thought out extensions for interacting with your llm's. Loaded up mistral-nemo, hooked it up to comfy to generate images, and later I may hook up tortoise to it for tts. character portraits for dialogue with built in sentiment analysis for emotional avatar switching... I think it's pretty slick for rp'ing... I think it's pretty cool. Just figured I'd mention it.

[stable crane]

I need a reco for a 10gb vram

been wanting to try it out for a long time

[rough oxide]

i just kinda lost interest in like the level of text LLMs that can be ran locally ;_;

imagegen is like the only thing that has been still usable on consumer hardware (24gb vram or less/64gb system ram or less)

and has been evolving like crazy

[stable crane]

this? https://imagegeen.vercel.app/image-gen

[rough oxide]

a reco?

nonono

when i say imagegen i mean image gen ai

like flux

[stable crane]

oh yes

[rough oxide]

what do u mean by reco

[stable crane]

LLM reco for 10gb vram

or is it even worth it

I'm OK with slow

[rough oxide]

ah

[stable crane]

but if the quality is shit

[rough oxide]

that's the problem. i have no idea. something like 13b barely runs on 24gb

unless u use like quantization stuff

and lose quality

i completeeely lose interest at that point

[stable crane]

I guess you can hook up sillytavern to APIs

privacy tho

[rough oxide]

if i could hook it to like claude/gemini/gpt4o api it might be usable but i feel like ai roguelite is already an all-in-one kinda thing

[stable crane]

it does advertise openAI API hookup

[desert island]

you can

[stable crane]

https://docs.sillytavern.app/usage/faq/

some of the OSS models look trash but some seems nice

like llama's

like some ----ing c-nt-bot?
what

[desert island]

<- nods

[stable crane]

the level of docs is nuts

quite a project

[desert island]

I haven't engaged with them yet but it seems like a very active community. I'm still learning all the things I can do with it and familiarizing myself with the basics while I setup comfy to specifically build out workflows for character creation.

[stable crane]

ok I'm going to finally take the plunge

aka stop being so lazy

[desert island]

what are you getting into?

[stable crane]

sillytavern

[desert island]

oh cool I set it up on my local ubuntu last night so ping me if you have questions

[stable crane]

does it keep a history log for you?

feels like it must

[desert island]

yes it does... if you're local it all stays local, it's just an interface with tons of bells and whistles.

[stable crane]

installing the oogabooga now

[desert island]

cool I went the ollama route but that's because I already had it installed but I was close to going the oogabooga route 😄

[lyric zealot]

[stable crane]

man how many copies of torch am I gonna have

[desert island]

I really wanted to go with LMstudio but I have an older CPU and my architecture isn't supported.

[rough oxide]

gta reference comes out well kekw.

1.3mp doesn't come out as visually clear as 1024x1024 tho

hmmm

[desert island]

still developing my template for silly tavern... it's getting closer... I think once I get this done I'll use florent to automagically mask out the face and change the expressions.

and now posting it ... I realized it still didn't make the background totally transparent... 🤦‍♂️

[wet elm]

dayum what artist tags did you use on the jotaro and edward image?? they look great

[rough oxide]

i know they're gimmicky as heck but these come out detected on the main ai image detector i use while these 2 for instance are only in single digit percents. too many artist tags can add lots of like visual noise, there's such a super fine balance between it... if a model can be aesthetic on its own that's a lot better imo

style's gud and #2 looks badass but like #1 and #3 still look like typical ai pics to me

[wet elm]

but what model did you use for these images?

[wet elm]

idk i would never think the edwark elric image is AI honestly

[rough oxide]

used flux

most of the stuff still comes out ai detected tho

my only major point is i like when a model is capable of having style without needing like

a paragraph of text/tags/etc

[ivory pebble]

Used like 6 different vibe transfer images, no artist tags

[wet elm]

yeah real, nai looks bad without artist tags

[rough oxide]

this is just pure prompting w/ no artist tags or anything of the sort

[wet elm]

oh ok, i wanted to try to replicate that style for some hard images lol

[supple raptor]

the background removal node in the WAS suite works best out of all of the ones I could find a couple months ago.

[ivory pebble]

I get what you're saying but those anime screencap images just look really boring to me, to me that's no different than the usual stuff you see from different local finetunes, it's just very bland imo

[rough oxide]

i guesss . i still like that kinda thing

i can do 'not' that kind of thing

but i prefer screencap style stuff xD

[ivory pebble]

Oh yeah I know you like the anime screencap stuff, but to me I like when an image is more aesthetic I guess

[rough oxide]

this one had a kind of aesthetic

from that batch

[wet elm]

me too, i really like those different styles, but i also like anime screncap images in some situations

[ivory pebble]

Not sure how to explain it but these type of images don't really do it for me, idk 😭

[rough oxide]

that's y i din' post that one

originally

xD

it's still ok as a fail tho

trying to do deeper style now but

50secs per gen is the major pain point of this model lol

[ivory pebble]

Yes I also add the anime screencap and official art tags to my prompt sometimes so that the characters look a bit more accurate to the anime

[wet elm]

yeah when using non-oc characters

[ivory pebble]

Ohh is that from that new flux model?

[rough oxide]

ye

[ivory pebble]

To be honest I just mainly generate popular characters so the anime screencap tag can be pretty nice

[rough oxide]

ngl. i have no idea how to prompt things that aren't anime screencap or photograph. i so totally specialized in anime screencap xD

i just know it's capable of doing it.

;_;

without any artist tags

[fluid torrent]

cute 2d anime style, colorful lines, cute girl, 18yo, anime outfit, adult, black ornaments, happy smile, cozy, warm, close, asian garden gate, at the wooden gate, outside, sunny, shed in the back with open door, dark doorway, (intense black area behind the open door, big black shadow, creepy, eerie, horror, anime screencap, empty area

flux-schnell, online

oh damn, I forgot to remove a ( from the weights

[ivory pebble]

It can definitely be a pain to figure out even with other models, every time I find something that I might like I'm always like "but what if there is more" and I just end up prompting for ages tryna find different styles 😭

[rough oxide]

yea

this one is back to

llm natural language prompting

[ivory pebble]

Go to AIBooru and search for has:metadata, you can find some good artist mixes there if you're using NAI

[fluid torrent]

cute 2d anime style, colorful lines, fine lady, 28yo, anime outfit, adult, black ornaments, happy smile, cozy, warm, close, asian garden gate, at the wooden gate, outside, sunny, shed in the back with open door, dark doorway, (intense black area behind the open door, big black shadow, creepy, eerie, horror, anime screencap, empty area

[rough oxide]

this is still in the territory of things i like tho

xD

hmm

[ivory pebble]

That is NOT a 28yo 😭

[fluid torrent]

its anime 28yo

[ivory pebble]

Well nvm, it might be considering this is anime

Yea 😭

[rough oxide]

now iz gettin doll-like ogod

save me

[fluid torrent]

blame flux, blame the dataset, maybe just a little bit my prompt

[rough oxide]

i almost bought a dollfie dream doll once

this reminds me of that

xD

[fluid torrent]

I mean I wish cute could fairly proportional apply to older ladies, and certainly cute images of older ladies exist, but they are hard to generate

[rough oxide]

i can get older chars just fiiine

[ivory pebble]

To be fair I think it does anime a lot better compared to other base models, if only someone could finetune if tho so we could do proper anime

[rough oxide]

i use the word woman for older chars

and it usually works

[rough oxide]

aside from specific concepts (characters, specific danbooru trained poses)

it is by far the best anime model i've seen so far

lol

[ivory pebble]

Try mature female, not sure if it works better but might as well give it a shot

[rough oxide]

oh no i totally used girl for this prompt

pokerface.png

also

flux does that text too

[rough oxide]

which means

i posted this 1 b4

but i can do soooo many fantasies

i couldnt' do before

i always wanted to do adventurer guild stuff xD

or just adding subtitles so u don't have to

in pretty much any font style you want

[fluid torrent]

serious 2d anime style, colorful lines, fine lady, 48yo, anime outfit, adult, black ornaments, happy smile, cozy, warm, close, asian garden gate, at the wooden gate, outside, sunny, shed in the back with open door, dark doorway, intense black area behind the open door, big black shadow, creepy, eerie, horror, anime screencap, older face, warm smile, wrinkles, grey hair

flux-schnell

[ivory pebble]

I feel like you could get better results without using so many commas

[ivory pebble]

One thing that's really nice about flux is the fact that it generates good hands almost every single time

[fluid torrent]

serious 2d anime style colorful lines fine lady 48yo anime outfit adult black ornaments happy smile cozy warm close asian garden gate at the wooden gate outside sunny shed in the back with open door dark doorway intense black area behind the open door big black shadow creepy eerie horror anime screencap older face warm smile wrinkles grey hair

[ivory pebble]

Also the model will probably do better if you don't use it like an average anime finetune, meaning that using danbooru style prompts probably won't give you good results

[fluid torrent]

ohhh, you ment that! I see

but i really love the almost command prompt style of small clusters of words, as a dreamy stream of descriptive words

it's like chaining linux commands using the | pipe symbol

or at least it feels that way typing, not the principle behind it

[ivory pebble]

Yeah cause like, prompting something like "A 48yo woman with a happy smile..." will be better than something like "48yo, anime outfit, adult..."😭

[fluid torrent]

I need a language model to convert my prompt style to something the new models would get

[rough oxide]

u wanna do

actual nlp

think old stable diffusion

[desert island]

48yo, anime outfit, adult < tokens

[ivory pebble]

I agree, I love prompting with danbooru tags but these models will probably do better with natural language prompts

[rough oxide]

And then what some ppl do after

[wet elm]

Tysm, do you by chance remember what artist/image you used for that style? Also how do you set up your vibe transfers for style? Whenever i try to do it it either has zero effect or completely dominates everything

[rough oxide]

is like Tags: <all tags> in a new paragraph

true hybrid prompting

u need a mix of nat lang & booru tags

basically for me nothing changes, i always used hybrid prompting

o ye these r more examples of why it's so gud for anime

it can follow instructions like

near exactlyyyyy

i tried to informaiton overload it on the potion prompt

yea no it didn't fall for it

all 9 details are exactly in the right place

every time

u can put stuff on monitors chars are looking at

etc

[fluid torrent]

im back to my custom sdxl model

[ivory pebble]

Set the information extracted to 0.1 and the reference strength to somewhere around 0.15-0.3, it really depends on how much you want your image to look like the vibe transfer one.

[wet elm]

Tyy, i just got nai like 4 days ago and ive been trying to get some cool styles

[rough oxide]

i can try to do this in like the 'new' prompting style on dev local and show u what comes out lol

[fluid torrent]

that would be nice to see, thank you mimizukari

[rough oxide]

first output

[ivory pebble]

Can't you just ask chatgpt or something like that to turn your current prompt into a natural language prompt?

[fluid torrent]

yes, I already did a few times

[ivory pebble]

Or is that not how that works?

Ah okay

[fluid torrent]

to turn it into a romantic story, based on simple 2 word prompts

[fluid torrent]

not tried it with flux yet

[rough oxide]

the only thing i tried copying was that exact pic i quoted xD

i'd need a full nat lang prompt for the other part

[ivory pebble]

Yeah AIBooru is where I got that style from. Not sure which aritst mix it was but it was one with the artist quasarcake

I'm sure you'll be able to find it quite easily, people love their style

[rough oxide]

o i added some other parts

ooo shadow grill

[fluid torrent]

ah damn, you beat me to the idea

[rough oxide]

it was ur idea.

[fluid torrent]

But wait, I might have another

[rough oxide]

when i said i added more i went to the prompt u posted

[fluid torrent]

the shadow ghost thing, that was new

[rough oxide]

and worded it in a way that'd produce sumthin

lol

well "eerie shadow" kinda vibes

[fluid torrent]

yeah, it was sure heading this way!

[ivory pebble]

But can the model do big booba goth girls that's the real question

[rough oxide]

yes but you can't use those words.

[ivory pebble]

Wtf

[rough oxide]

i mean

u can use goth girl

obviously

but not the first part

lol

[ivory pebble]

But big booba...

[rough oxide]

write around it. it's possible

i've gotten huge chars

just randumbly

lol

so there's probs a way to describe it consistently

[wet elm]

nice! do you not post your images here with metadata? Ive tried to put one of them in inspect because the style was so clean but nothing came up :(

[ivory pebble]

Probably because I edit the images myself, NAI messes up the eyes when the character is not close to the camera so I love to draw the eyes/face myself

[rough oxide]

that's a nice thing about this new modellll

[wet elm]

ooh thats so cool

[rough oxide]

flux can make good eyes even if char is further from camera

[ivory pebble]

And I usually delete the original images cause I hate having duplicates 😭

[wet elm]

lol

[ivory pebble]

That's that 16 channel VAE doing its work

Hopefully we'll have some finetunes soon with a good VAE

[wet elm]

i was referring to that jalter image you posted some time ago, also that goth marin kitagawa one

[rough oxide]

don't count on ittttt...

200gb vram to finetune

xD

well

[steady grail]

what's that flux?

[rough oxide]

2huai probly has a stake in it eventually, neggles said they might eventually do a waifu finetune

[rough oxide]

yea. furry communities r already crowdfunding

[ivory pebble]

Yeah I couldn't really tell you to be honest, it's been so long since I've made those ones

[steady grail]

a new furry model?

[fluid torrent]

jumping back to my sd15 model, Cute Lunar

[steady grail]

after pony, i'm starting to love furry communities

[fluid torrent]

oh no

you are infected

[ivory pebble]

Yeah realistically I think the only ones who could finetune it are wdv and nai, not sure if it will happen tho

[rough oxide]

i mean for now i'll just assume it's gonna be the best model for like 6 months or so

unless a dall-e 4 or something comes out

i love how it also has like

mtg/yugioh/hearthstone/

shadowverse card frames

i wonder what other card game frames it has

[ivory pebble]

I agree but a lot of people also want to make models/loras with it, and since that seems to be nearly impossible it kinda sucks honestly

[rough oxide]

probly pokemon?

also i clearly need to give this more height xD

wait no is the quality just like

gatherer quality

lmfao

ye

i used gatherer resolution

came out kinda clearer

lol

but also u can just get stuff like

[rough oxide]

so much potential

the ultimate xyz monster with level stars for some reason

[fluid torrent]

sd15 seems to be more dreamy-immersive

[quiet bramble]

Any XL models that can mimic this style?

[fluid torrent]

something like this? or to much lens glare?

[wispy canopy]

flux too stronk

[fluid torrent]

flux-schnell, online, doing some text to go with the anime title screen

[wispy canopy]

flux stronk

[stable crane]

@desert island silly tavern setup went well

it's kinda fun, can use them for coding tasks too

can ask a soulless LLM from big co, or an anime girl for coding help

[stable crane]

no regional prompting right, it just understands half-half?)

[wispy canopy]

it understands:
A highly detailed and intricate elemental golem stands tall in a dynamic pose, radiating immense power and presence. Its body is composed of swirling masses of fire, water, earth, and air elements, seamlessly blending into each other. The fire element glows with an intense orange-red hue, the water element flows and shimmers with cool blue tones, the earth element is rendered in solid brown and gray shades, and the air element whips around in swirling white and light blue patterns. Studio Anime's distinctive art style is evident in the sharp lines, bold colors, and dynamic composition of the image. The golem fills the entire frame, leaving no space for any background or surrounding environment, emphasizing its imposing presence.

[stable crane]

hyped for the user versions of this thing

I need a better GPU

[wispy canopy]

that's on my own gpu... running in fp8

about 14gb of vram needed

[rough oxide]

i literally think that if you can't get something (not too uber complex or specific) to appear then you're just prompting it wrong. flux is sooooooo strong i haven't found a serious failcase yet

[stable crane]

I don't think it knows spidergirls

[stable crane]

I don't have that much 😢

[rough oxide]

monster girls r difficult to pull off but if u keep adding relevant things to the prompt it might work

like for a slime girl adding like melting helps

to turn it full slime girl

[wispy canopy]

it's a generic model, not an anime/booru model, so it's just not good at that

[rough oxide]

i wouldn't underestimate ittt

i got this far

if u refine a prompt enuff, something will happen

just need to gib her spider pupils and maybe sum facepaint

xD

[desert island]

glad to hear it 😄 sorry I stepped away for a moment. Well 16G is just enough for me to run my model and comfyui so I don't think I'll be able to run tortoise as well unless I setup another machine in the house to serve up either the llm or image generation. I have a workflow that is running creating all the expressions for the bot character I'm going to use as my assistant. It's just taking soooo long generating all the faces for all the expressions right now.

[tawny helm]

Ah, Eface, you're still here making good images as always.

[wispy canopy]

never left, only started path of exile grinding hard 😄

[tawny helm]

I'm barely ever even here anymore. lol

But PoE is good though!

[wispy canopy]

grinded 20/40 challenges this league -> it started last friday

[desert island]

so fun... I got like 20+ different expressions now to put it all together in my silly tavern 😆 . Squee

[wispy canopy]

and other than that, i've been in civitai chat a lot

[stable crane]

lol that's not bad

[stable crane]

cool

[fluid torrent]

night chat!

[rough oxide]

1080p actually works great.

i think comfy updated something and now it works bettr.

[lyric zealot]

[desert island]

looks like Ino 😄

[gilded anchor]

It is. lol

[desert island]

oh well that make sense then 😁

[gilded anchor]

haha 😄

[steady grail]

this was pretty much an incident

[rough oxide]

sum moar 1080p examplz.

hokuri reimu is best waifu

[steady grail]

what's 1080p?

it should be a res. a model trained on 1080p images?

[rough oxide]

1080p is a res

1920x1080

it was trained on 2mp

[bitter latch]

haha

ok back to bed

-stare-

sleeps

[vital raptor]

heh, typical young girl size switch next to a me sized switch

[rough oxide]

too bad it didn't do I K E A

ended up I K E X

;_;

that other switch looks like idk a wii u or sumthin

[bitter latch]

she says "hmm?"

[bitter latch]

donuts

haha

[bitter latch]

the new art piece

goes back to bed for a bit

[bitter latch]

angry with explosion

hmm interesting

oooooh

same seed and everything except diff model

random seed

holy.. faints

uhh oh she's coming!

haha

oh umjmm

ok then

[desert island]

@bitter latch Looking good loved alot of your stuff today ❤️ .

[stable crane]

is this flux? really nice one lol

[steady grail]

what acat world

[bitter latch]

thx

[desert island]

is it a cat lady? or a cat-lady?

[bitter latch]

yes.. or Neko

[steady grail]

chibi is my favored style after all lol

[bitter latch]

haha extra chibi

oops caught again

ok no more donuts for her

[dawn glacier]

i dont know how to use this

i want to make my own pics

[steady grail]

charming

not what i expected

maybe i should look for a chibifier lora 🙂

many model i like are not trained very well in chibi characters

[green temple]

try cuteyukimix

[steady grail]

i wonder if i can mix it in my model

i'm always looking i more ways to make cuter XD

this is as far i can push chibi with this model

this Kurisu is so charming lol

damn, i totally forgot what she is missing >.<

kioma-sama will never forgive me XD

the model is not able to gen a lab coat...

or the lora is resisting the change

that's what i get when i try to ad a lab coat

no idea what's that

i know, kuris without a lab coat is simply wrong

shouldn't be approved under any ceiling or sky

a true sin

yeah, it's the lora

too much inflexible

[steady grail]

i like this prospective

[brittle rampart]

who offended you again"?

[green temple]

no one

[steady grail]

she is offended by the treacherous lack of lab coats in my model

i know, a true tragedy

how can she be a woman of science without a lab coat?

[brittle rampart]

It’s not true that she, like me, is offended by the presence of “Big Booooooba!

😁

[steady grail]

cute

[brittle rampart]

By the way! I've been wondering about this question lately. why are generations and models based on Danbooru tags? why not gelbooru tags? or Pixiv?

Gelbooru, in my humble opinion, seems more diverse

[bitter latch]

well if it's a furry based model it usually goes by e621 tags

[brittle rampart]

apparently because this resource is more aimed at this style

[bitter latch]

yep

[brittle rampart]

but we're not talking about Furry style

[bitter latch]

was just saying 😛

[brittle rampart]

no no. I'm just interested... not a specific style. because, for example, I like the furry resource furaffinity.net

[nova remnant]

Put "scientist outfit"

[nova remnant]

Furrynya

[lyric zealot]

https://tenor.com/view/makise-kurisu-makise-kurisu-spin-kurisu-kurisu-spin-gif-8134591790627908365

[nova remnant]

Fursus

[lyric zealot]

https://media.discordapp.net/attachments/1072697081443131476/1217133054636851382/evilrv1.gif

[desert island]

sooo good.

[lyric zealot]

I didn't make it, and it's almost definitely not ai

[keen hollow]

so what's this flux thing? I always go on vacation when new stuff releases... 😐

[lyric zealot]

3 models, 2 open source, 1 closed, the dev version (which is one of the opened ones) is widely regarded as great (from the community)

you need a lot of vram for it tho

[keen hollow]

I've got 12 gigs, maybe it'll do it. Have you tried it?

[lyric zealot]

12gb vram might be enough if you are willing to wait, and offset it to your system ram

i've tried it on their site

their pro version (which is closed source) works good, great with text

[keen hollow]

oh nice, didn't know they had a pro plan

comfy support is meh for me... looks like swarm supports it though

[lyric zealot]

[lyric zealot]

good luck

[green temple]

puffy cheeks

not i wanted but i like this style

[wooden sail]

Flux can do style like that after all

[lyric zealot]

The tattoo is sick

[wooden sail]

And unintentional 🤷‍♂️

[lyric zealot]

Prompt for dragon tattoo like some yakuza

[wooden sail]

2000s style rendering

[green temple]

3rd

[steady pumice]

it really do be that way

I couldn't get it to look right but liked it so

[green temple]

lol the 4th

[rough oxide]

the crystal ball revealing the futureee xD

[lyric zealot]

would gladly let ai take my job

[rough oxide]

same

ubi when

[keen hollow]

whew, this flux thing is neat, but MAN is it slow....

schnell is usable...

[lyric zealot]

schnell is kinda like turbo, only needs a few steps (1 to 4)

[keen hollow]

yeah, I've been messing with it

[keen hollow]

flux doesn't know Tifa, so my use is limited. 😄

[rough oxide]

flux bettr for original concepts. it can do practically anything original

was testing a coloring book thing too and it worked well

can do like everything i couldn't do in prev models xD

that was in pro but i am able to get it to sorta work in dev now

just need to prompt it moar xD

[zinc flax]

Cleaan

[bitter latch]

hmmm.... was that | | close

sigh

haha

[rough oxide]

iz gud, i was going for like coloring in the lines/color book style thing

[bitter latch]

gotta say I'm loving this cutelunar model

she's like "what?"

goes back bed

sleeps

goes back to bed... thud

there.. thud

[near elbow]

having fun with flux (sorry SD3)

[unkempt cape]

another nai3 wannabe

its not any better than the stuff now BUT

but it has a free unlimited gen test site with good speed so 10/10 peak sdxl experience

[zinc flax]

What is it this time

[nova remnant]

free stuff ? I want

[summer night]

i am going to wait for forge to make generating with flux possible on 8 GB of VRAM...

[bitter latch]

@gloomy scroll HA! wanna donut?

[bitter latch]

she found the donuts

ok this model keeps suprising me! (cutelunar v10

hehe

it actually did hte peace sign. wow

ice cream!

evi lcackle

[bitter latch]

"bye gotta go!"

goes back to bed.. .thud

[nova remnant]

smol

[bitter latch]

merge of furryverse and cutelunar I would say was a success

[stable crane]

very cool, you got a lot of high frequency detail out of these

[nova remnant]

I will say it...

[green temple]

what

[nova remnant]

Big

[rough oxide]

og->nai inpainting

realized fingers were messed up too

xD

[nova remnant]

Now fix legs

[rough oxide]

lol

i thought the legs were silly

[nova remnant]

They are noodles

[rough oxide]

xD

[nova remnant]

https://tenor.com/view/subnautica-octodad-gif-18948332

[rough oxide]

nevrawrrr. i like how silly the legs are

but the half closed half open mouth was kinda bothering me

[slender barn]

thank you!

[green temple]

is thjis cute lunar ؟

[fluid torrent]

lol no! this is sd3 + copax lora 😄

https://civitai.com/models/517485/copax-cute-style-stable-diffusion-3

This is Cute Lunar, almost the same prompt

[green temple]

copax sd3 is cuter

[summer night]

@native halo i published it
https://civitai.com/models/540591

[terse ore]

I'm seeing a lot of posts about a thing called flux on Reddit but idk what it is exactly

[steady grail]

a new model with a very big library that require an high end pc to run it

the model is something like 24gb big

you need a machine able to run a server, just to gen boobs XD

they seriously have gone too far lol

but i suppose that image genning is just a byproduct of his real use

[nova remnant]

New model kinda like dalle

[nova remnant]

You can run on 8gb

And there is a smaller version

[terse ore]

What's so great about it?

[steady grail]

oh

[nova remnant]

It's like dalle

Looks and responds to prompt well

[steady grail]

yup

[nova remnant]

Its a bit dumber than sd3 for a few things I tried

Like it doesn't do styles all that well, some concepts just don't exist

Idk it's a really good model

[terse ore]

Will it have controlnet and stuff? If it's such a big model it can probably do lots of different styles

[nova remnant]

But really weird that it doesn't know some things sometimes

[terse ore]

I prefer dedicated specific models tbh flux sounds like a jack of all trades but master of none

[steady grail]

well, 1.5 too are biased toward a concept

[lyric zealot]

i managed to run it on my smol 8gb vram

[steady grail]

without loras, many things wouldn't be possible

[green temple]

can it do smol ?

[nova remnant]

What I liked most about flux is it's coherency

On things like text

[lyric zealot]

the dev version took about 3 minutes per generation, and the "turbo" one took like 1 minute

[nova remnant]

yes

On sd3, same prompt

Idk what's wrong with flux lmao

I wanted her showing a cat drawing

[rough oxide]

u need more space

give it more space and u can do it

1024x1024 is like one subject

or like two human subjects

lol

don't forget it's trained up to like 4k or something? 1024x1024 is just like a single slice of a whole pic. (not entirely sure on that, but 2mp for sure)

last 1. but they all came out.

[nova remnant]

but just a person showing a drawing shouldnt need anything

[rough oxide]

just tried 1024 pro gen and this came out

iunno just not enuff space

lol

gib moar horizontal spaceeee

;_;

[nova remnant]

why is there a cat

[rough oxide]

wat

wait

[nova remnant]

x3

[rough oxide]

i thought u were saying

u wanted to show a cat a cat drawing

...that is a lot simpler. xD

one shot

and in dev too

what prompt r u usin

[nova remnant]

was just like anime girl, girl showing a cat drawing

the hell am I supposed to prompt lmao

[rough oxide]

iunno i just did "anime girl happily showing off her drawing of a cat, illustrated in an anime style"

keep the action in the same line as the character

maybe

that's how i usually prompt i think

[rough oxide]

i keep brute forcing until things work then change my habits on writing xD

flux is a bit diff than what came before

[native halo]

hmmm tasty new pony lora

[wooden sail]

Housewife Reze ❤️‍🔥

[oak spoke]

mushrooms

[steady grail]

very cute results

i expecially like the bg

[oak spoke]

now make them hold nanaya's torch

👀

[bitter latch]

who what?

[oak spoke]

this

now I really want to make lord of frenzied flame

🤣

[steady grail]

😱

[oak spoke]

anyway have some fish

[bitter latch]

hmm that didnt' work but ..

[vital raptor]

this model doesn't know what a spine looks like 😄

[oak spoke]

I mean it's pretty close

[bitter latch]

there thats better

ok I need a new theme... think i've run this one dry

There we go

slows down

https://tenor.com/bcMSu.gif

[steady grail]

very sweet

reminds me so much of orange road

or Maison Ikkoku

anime of my childhood

ok, a little more than childhood lol

[bitter latch]

clip skip 2, clipskip 1

[steady grail]

flux?

[bitter latch]

no

[steady grail]

pony?

[bitter latch]

merge of cutelunor and NOMD 😄

[steady grail]

i like the details a lot

[bitter latch]

but I sorta did borrow part of hte prompt from a pony prompt

turns clipskip upto... 12

waits for nitemare

huh....?

[steady grail]

mh

that's a good idea

[bitter latch]

wow

[steady grail]

i wonder what happens if i play pony skipping more clips

how much lose will become the prompt

what what i lose in response, i'll gain in speed

baybe i can find a good balance

[bitter latch]

turns cfg scale downt o 3 from 6, and sampling steps from 30 to 150.. generate

[steady grail]

trying to compensate the mess with more steps

but it's not pointless?

[bitter latch]

shrugs

hmm... ok...

changes to 2 and 80

[steady grail]

the increase of stepps added too many meaningles details

[bitter latch]

nods

prob doesn't help i'm using lora:more_details:0.75

[steady grail]

eheheh

[bitter latch]

ok not totally madness

[steady grail]

cute

[bitter latch]

goes random seed, changes to boy, 2, 8

gah .. no no

blah

hmm 1, 50

and removed the lora

ok whats with this alll of a sudeen. wtf?

removs the pony prompt stuff

ah ha

hmm suppose to be boys

[ivory pebble]

I love this! Is this khyle's style? Looks kinda familiar

[bitter latch]

hmm trainstation bar?

CS 12 does some odd things

turns down to 6 and goes back to cutelunar

hmm

[bitter latch]

but not anime enough.. hmmms

[zinc flax]

Is this modare?

[ivory pebble]

Yup , I specifically used one of his Asuna images for vibe transfer just so I could get a different angle

[zinc flax]

Nice, love their muted colors tbh

[ivory pebble]

Yeah that's another reason why I used the image

[steady grail]

holy carrot. she has gone too far with her obsession with carrots

establishing an entire religion about carrots.... well, for rabbits my have sense lol

[stable crane]

😄

sounds like the words of a nonbeliever

some flux->pony pipeline pics

[stable crane]

starting point

[brittle rampart]

yes that's right one of the artists Khyle

[brittle rampart]

Head to the stars - Captain!

[finite herald]

can we please veer away from imagery such as this

[bitter latch]

ok

[frank peak]

Haven't been here for 8 months, see no progress at all, still broken hands?

[zinc flax]

Hands are still the most ok stuff

[brittle rampart]

hands like hands. located in the right place 😁

[frank peak]

This one looks fine. But I scrolled up a bit, and 9 out of 10 has bad hands

[frank peak]

You know I meant fingers

[zinc flax]

Tbf it depends on the prompts lol

[fluid torrent]

from my new sd3-sd15 (cute lunar) workflow

[brittle rampart]

oh, well... what a picky guy...😁

[fluid torrent]

sd3 fingers

same image, refined with Cute Lunar (sd15), fixes hands! 😄

but the next comparison SD3 seems to win. not cuz of brighter colors, mostly facial expression. clear winner.

(btw, using slight bit of copax lora ...)

a bit of flux-schnell in between

[ivory pebble]

Depends on the model honestly, most of the stuff you see with the broken hands/anatomy is from older models

[fluid torrent]

yes, still broken hands. Though also a bit more fixed hands

[zinc flax]

Or just ppl who wouldn't bother fixing the hands

[fluid torrent]

with flux I guess you need only 2 or 3 trials, depending on complexity of hands

[brittle rampart]

not at a generation competition. it's for fun.

[ivory pebble]

I don't really have issues with hands or anatomy anymore, but that's probably because I just use NAI and Pony models so idk about the others

[zinc flax]

Ye, I could understand

[zinc flax]

nai still clears

[ivory pebble]

That's true, but it's different for everyone. For example, I hate seeing an image that looks good aesthetically but has messed up hands

[brittle rampart]

I don't think anyone likes this

[fluid torrent]

flux with hands

anime style, female hands holding a pen, golden rings, shiny, realistic lighting, lens glare, horror style, white-blue top, sleeves, clothed, trick

[ivory pebble]

You might not but I do

[fluid torrent]

the online replicate version gave some nsfw warnings in between, idk why

[ivory pebble]

New AI base models and censorship, name a more iconic duo

[fluid torrent]

its just the website I think, I still believe in the model!

[zinc flax]

I've seen some nsfw gens from the open source version

It's quite meh unfortunately lol

[fluid torrent]

Im still waiting for mini-flux, if it ever comes