🍥｜anime

[gloomy scroll]

Mofu is ready to launch hug barrage

[merry tendon]

can someone do the vatican mascot?

[modest fog]

Grabby

[wet elm]

is there any extension to get random prompts from like danbooru while keeping the character? like a random background + pose + artist generator or something like that

[gloomy scroll]

You can look into wildcards

[keen hollow]

Nice colors. I like the more muted matte looks.

[sly girder]

Thanks, I have been having fun playing around with noobai

[terse ore]

Is there a consensus yet as to which is the best illustrious XL derived model?

[vital raptor]

probably noobai

[vestal ferry]

https://tenor.com/view/트릭컬-리바이브-트릭컬-코미-코미-backflip-computer-gif-13479146599765963639

[vestal ferry]

noobai

[keen hollow]

that's a lot of shine

[wet elm]

hmm ok ty

[bitter latch]

A wolf going clubbing

[vestal ferry]

can i try this one?

https://tenor.com/view/memeoshi-wakaru-sorena-soyana-seyana-gif-11880477

[rough oxide]

wdym try it?

the metadata all still on file

"1girl, nepgear, nepgyaa, neptune (series), <o> <o>, nervous sweating

and then "worst quality, bad anatomy, bad hands, lowres, text, artist name, signature, jpeg artifacts, simple background, white background, abstract" for negative

hmmm originally i was trying to get the actual nepgya look but

white pupils didn't really help it much xD

https://tenor.com/view/nepgya-neptunia-hyperdimension-gif-23230922

[vestal ferry]

oki i will try

https://tenor.com/view/frieren-fern-beyond-journeys-end-macdonald-burgers-gif-10253262265477048425

[rough oxide]

i still have no idea what i was trying with my most recent gen lol

[rough oxide]

cute teto

[vestal ferry]

https://cdn.discordapp.com/attachments/1266209175772594188/1302467186367004762/ok-nodding.gif

[vestal ferry]

[vestal ferry]

look

[rough oxide]

adult nepgyaaaa?!

xD

[vestal ferry]

do you want to see something scary

im not sleeping

[rough oxide]

matsuri will save uz.

[bitter latch]

unexpected but funny

[bitter latch]

tired...

sleeps

sleeps

goes back to bed.. thud

[bitter latch]

nite

[modest fog]

blue

[modest fog]

Brown

[green temple]

best color for smol

[vital raptor]

i prefer purple

[green temple]

purple is cute also

[modest fog]

[modest fog]

@green temple

[green temple]

big blue

[green temple]

evil

[modest fog]

she is cute

[sly girder]

the big blue cartel

[vital raptor]

she didnt have anything blue, so I lent her my hoodie 😄

ironically I used to have a hoodie that color, but my niece stole it.. it was so big on her she kept tripping on it 😄

[sly girder]

loool kids can't resist grabbing anything on sight

[vital raptor]

I ended up giving her a bunch of my old shirts for sleeping in, I'm 6'4 so my shirts were like ankle length dresses for her 😄

[modest fog]

6'4

you're 2 people on top of each other

[vital raptor]

[modest fog]

Grabby

[modest fog]

@green temple pat her

[rough oxide]

@jagged field if ur still playing around with mochi, comfy updated a lot of stuff with it so u might wanna git pull + check out new video example, might be updated workflow https://comfyanonymous.github.io/ComfyUI_examples/mochi/

[modest fog]

licks

[dire bridge]

oh

[sly girder]

I am really starting to like noobai

takes a while to know which artist tags work best though

(also I may have attempted too many times doing gens without artist tags..)

[modest fog]

its a cool model, im not a big fan of having to use artist tags for stuff

[sly girder]

yeah it does kill a little the freedom for prompts

[modest fog]

I tried a few without artists and things were...

[sly girder]

Ironically one gen I did without artists did look cool

but it didn't resemble a human so..

[vital raptor]

I spent some time yesterday figuring out some artist tags for noobai that I like, but it would still take a bit of prompt tuning to get it as good as my current gens

pretty much got to the point where it's not worth the effort, it knows some concepts missing from pony but it also loses a lot (especially nsfw concepts)

if you want to gen specific anime characters it's a lot better at that

[sly girder]

true, depends on what you want to trade off for

[modest fog]

The only tag we need is top down bottom up

👍

[vestal ferry]

i like to make it more simple

like

artist tags,

then space

(masterpiece), (best quality), (ultra-detailed), beautiful color,

space again

then normal prompts

[sly girder]

yuh, I'm slowly getting the idea behind on how to prompt it better

[vestal ferry]

and easy example of mine is like this

artist: okpriko,

4k, (high detailed skin), (masterpiece), (best quality), (ultra-detailed), beautiful color, amazing quality, very aesthetic, illustration, absurdres, newest,

dynamic pose, perspective,solo, dark-skinned female, dark skin, 1girl, scheherazade (fate), fate (series), long hair, breasts, black hair, thighhighs,

[sly girder]

niiice

oh it's pretty much how I been prompting, well sorta because I keep experimenting (no matter how bad the idea is lol)

oh right

if anyone needs it this is the doc with the 1000 artist styles that noobai uses (examples are nsfw) https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=9v7zie

[green temple]

oh ty

[green temple]

shit the examples 🌚

[sly girder]

Yeeee I forgot to warn it

mb

[vestal ferry]

https://tenor.com/view/doro-dorothy-nikke-goddess-of-victory-nikke-gif-12656566622012438896

[green temple]

🥲

[sly girder]

[vestal ferry]

https://tenor.com/view/doro-dorothy-nikke-goddess-of-victory-nikke-creature-gif-17030571333828640261

[supple raptor]

noob has a lot more than 1000 artists. that would only reach down to artists with ~650 posts on danbooru alone and it can do artists with ~100, sometimes less. then there's artists from e621 as well.

[sly girder]

damn

well if I find a more completed doc I'll share it here

[supple raptor]

danbooru artists alone is ~18k artists that it should do reasonably well

[sly girder]

https://cdn.discordapp.com/attachments/942176642674098179/1205710253921009685/artist-noted-artist-note.gif?ex=65d95c2d&is=65c6e72d&hm=0fc297941f055734bedfd2d407ba65708a4a36ecfe9d4cd5b17bbac33aba868b&

[supple raptor]

i have no idea what the e621 artist selection looks like, but i would assume it's somewhat similar, though probably less accurate than the dan artists since illustrious didn't have any of the e621 training and illustrious already did most of what noob can from danbooru, just not quite as cleanly

[sly girder]

that makes more sense actually

I did find weird for it to be 1000 exact

[sly girder]

Yeah my bad I found a way better list here

https://mega.nz/folder/Vd8yTCTL#FEpX0epT6cAcMZQI-QuBAA

the examples are safe for work this time

[green temple]

nice

[vestal ferry]

https://tenor.com/view/doro-dorothy-nikke-goddess-of-victory-nikke-santa-gif-78923470987598656

do you know why its divided in two? like 例图(无furry负面) and 例图(有furry负面)

https://tenor.com/view/doro-dorothy-nikke-goddess-of-victory-nikke-creature-gif-17030571333828640261

[sly girder]

I'm wondering the same thing, not really sure why that doc divided it

[sly girder]

but honestly you might as well just use this one instead

if you are looking for artists specifically, they got examples there too

[past junco]

Somebody miss me?

[jagged field]

Thank you so much the issue is that those other fp8 workflows are very slow on my H100

It takes less vram yes but turns out to be slower i can generate up to 121 frames in 270 seconds but i get this error that's out of vram but the video generates and everything is good

I can run 2x H100 together but I don't know how to connect them together with 1 Moochi is it like i'm gonna load the VAE and text encoders to a GPU and the other will have the model ? or how's the setup gonna look like

[modest fog]

steals it

[modest fog]

No smol examples

[modest fog]

Bad

[modest fog]

You were gone for so long

[modest fog]

Smol

[supple raptor]

https://tenor.com/view/now-kiss-touhou-marisa-reimu-mwa-mwa-gif-17590771652541408635

[modest fog]

Bad

not my sfw channel

[vestal ferry]

https://tenor.com/view/izumo-saki-d4dj-d4dj-petit-mix-anime-surprised-gif-20531241

[past junco]

My laptop charger broke, was fritzing out and glowing red with smoke

[modest fog]

Reported

[zinc flax]

need more

[modest fog]

Go do that in a nsfw channel

It's kinda weird seriously

[zinc flax]

should be fine as long as it's not tongue kissing

[green temple]

+1 😱

[modest fog]

Help me

[zinc flax]

LMAO

Basically

[jagged field]

Hello Guys

[jagged field]

I need help running Moochi with multiple gpu's on ComfyUI please

[modest fog]

[jagged field]

Can someone come play with me a little bit ? I'm trying to play with some models on comfy

[modest fog]

Thank you for smol

cleaned my eyes

[jagged field]

cute :3

[modest fog]

I stole her ice cream, lick

[green temple]

lol

[modest fog]

Shes a fish

I'm hungry fish

[modest fog]

galaxy smol

so.. big ?

[sudden hemlock]

Hey everyone! Lemme ask something: what (locally-runnable or free) model would you guys recommend for Anime, but that can do more composed scenes and a variety of characters (men, adult-looking women) decently instead of just being amazing at cute girls with simple backgrounds and nothing else?

[desert island]

give us a sample of the kind of prompt you want and I'll do some stuff to see if that's what you are looking for. Also give us your system specs so we know what kind of parameters we are working with. Should we prioritize quality or speed?

[bitter latch]

😄

slows down

[sudden hemlock]

Okay so going into some detail:

I use generative AI mostly for TTRPG characters as I run a lot of games. This time I'm doing one that's very anime-inspired (especially on Solo Leveling) so I'd like something that looks a bit more anime.

I tried PonyXL and some variants, but it was really bad at anything that's not a cute anime girl, and the art style always blends some western in. It also straight up didn't understand most fantasy concepts (even the ones that have booru tags), and was quite bad at backgrounds.

The closest I got to in terms of the compositions I want was actually Bing's version of Dall-e 3. I could tell it to make a character casting a spell, or shooting lightning, or holding a spear, and it would understand. However, it sucks at anatomy, faces and textures, it's heavily censored (even for non-porn), and doesn't have tools for hi-res, inpainting, etc.

[sudden hemlock]

For example I made this with Bing

solo leveling anime style, solo leveling, manhwa, drawing, a beautiful mature woman with dark skin, brown skin, tall, fighter, burly, romanian, green eyes, short hair, uneven straight bair, black hair, necklace, earrings, pants, black t-shirt, green jacket, fur-trimmed coat, smug smile, confident, crossed arms, sunglasses on head, explosion magic, explosions, fire, explosion background, city, night, beautiful detailed eyes, full body

Great composition but the face and textures are kinda weird and low-res

[sudden hemlock]

I have an RTX 3060 with 12GB VRAM, and definitely prioritize quality over speed

[desert island]

let me tweak some things and I'll see what I can do.

[sudden hemlock]

Thanks a lot for your time!

[bitter latch]

model is paragonv10_v10vae

😄

back to regular anime

[near elbow]

just for fun, generated in 0.6 seconds with flux schnell (living up to its name). not as detailed but not half bad

same prompt with flux pro, more than 10x slower, noticeably better but 10x so? hmmm

[bitter latch]

better

[desert island]

Just so you know I'm still working on this I'm just trying out a bunch of different things to try to find something that I think will resonate with you and produces consistent results.

[past junco]

Glad to see that people are still chugging away

I'm still using my old laptop specs (new charger) but I'm getting far better results with Forge and a larger model

Still doing some funky merger experiments

But at a higher scale than before

Gonna save up for a dedicated system

[green temple]

[sudden hemlock]

No hurries, and thanks for all the help! I'd test all this stuff myself but I'm not even 100% sure where to look, and it would probably take ages with my PC :x

[past junco]

Oh btw

Never thought about this before

But since I do have a very crazy setup with lots of custom fiddling and experimental gen

I wouldn't mind doing some prompts for you guys

[modest fog]

playstate

I was cat bot

[past junco]

Cat bot?

[modest fog]

Yes

[past junco]

Oh lmao

The user yeah

Hey man

Sorry I haven't been active

Laptop died for like a month

[modest fog]

Feels like you have been gone for more than that

Were you doing other things than talking here

[past junco]

I do a lot of things, always busy. Work, game development, moving houses and now living by myself

It's been a crazy time for the last few months ngl

I do miss my homies here ngl

Nice to see a friendly face

[leaden stirrup]

yo how cani create a img help me

someone plz help me here

[past junco]

Honestly depends

[leaden stirrup]

depends

[past junco]

What system are you using?

[leaden stirrup]

what do you mean what system

[past junco]

The specs determines what you can generate, and what you can't

[modest fog]

Sounds good

[past junco]

Windows?

[modest fog]

Better then me just a all day long

[past junco]

What kind of GPU?

[past junco]

NGL been kinda down lately

[leaden stirrup]

Ryzen 9 8000 series Ge Force rtx 4070

[past junco]

I'm gonna go on a limb and say that if you can't generate art, you're doing something wrong

[sudden hemlock]

One piece of info that might be relevant is that because it's for TTRPG illustrations and I'm gonna have a single image for each character, character consistency between different images is not a big priority. I'm more focused on getting a single really good one for each

[past junco]

Btw I use SD for a lot of my TTRPG stuff as well

Pleasure to meet you

Atm I'm generating images for cards for use in Tabletop Simulator

[sudden hemlock]

Same! I think it's one of the coolest personal uses for AI art, my players really love it

[modest fog]

Look up on Google how to install forge stable diffusion

There should be some easy tutorials

[leaden stirrup]

okey

[past junco]

I highly recommend Forge, it's what I'm using

Made my life easier

[modest fog]

Forge, swarm and SD next are pretty good

And easy to use

[past junco]

I still to this day have not used Comfy

[stable crane]

comfy is pretty nice

I can't go back now

but I do want to try that inpainting one

[modest fog]

Comfy only really useful if you want to experiment with weird things

Not much difference on text to image

[stable crane]

https://github.com/invoke-ai/InvokeAI/releases/tag/v5.0.0 looks cool as hell

[past junco]

I believe that with my field of expertise Comfy makes the perfect sense for me to use but I just never took that leap

If it ain't broke, why fix it?

[past junco]

Sounds like me

[stable crane]

for that guy who's brand new, he should not use comfy

[past junco]

Yeah no

[sudden hemlock]

I'm a software dev and I still find comfy kinda confusing ngl 😵‍💫

[stable crane]

it's messy

[past junco]

I love writing shaders and using nodegraph and lots of stuff like that

[modest fog]

What's so different about it

[past junco]

And still never touched it

[stable crane]

it's designed for iterative inpaints

check our their youtube vids

[modest fog]

Ugh

[stable crane]

I honestly wish I could just kinda write could using various building blocks

(not in python)

[modest fog]

I really hate clicking on anything AI videos

Incognito we go

I'm too dumb for that

I just click generate

[stable crane]

yeah it seems more like an art workflow

kinda like a basic photoshop where the only tool is layers and generate

I'm so lazy

[modest fog]

I'm lazy and useless

[desert island]

Just so you know this is what I'm currently trying to get work.
I'm trying to merge this
https://huggingface.co/sd-dreambooth-library/solo-levelling-art-style
with this
https://civitai.com/models/904111/sd-35-large-modern-anime

[modest fog]

for a second I thought this was on the wrong server

[stable crane]

really impressed with these couch textures

[stable crane]

lol

[modest fog]

I... Am too
Grabs

[stable crane]

looks really soft

[modest fog]

Big

[stable crane]

and comfy

[modest fog]

Lmao the dog on the painting

[stable crane]

haha yeah

[modest fog]

https://tenor.com/view/awkward-introuble-ohno-shocked-whoops-gif-5687730048162886939

[sudden hemlock]

Oooh, that's nice! I barely know how merges work (technically), but that seems like a lotta work so thanks again!

In case it doesn't work perfectly, I'd say getting anime style good compositions with magic and stuff and being able to actually do not-teenage-girl characters well is definitely more important to me than looking exactly like Solo Leveling

[desert island]

it's really not but it's just taking some time to download the models. waiting for downloads

[desert island]

if you're doing this for ttrpg you'll probably also want to create lora's for the mc's in your game.

[sudden hemlock]

That would be cool, how hard is it to do that?

I know what a Lora is and have used some but no idea how they're made

[desert island]

I've only created flux based loras thus far but I'm sure there is tons of documentation out there for sd based models. And it's done downloading now I'm off to go try this out and see what it produces.

[sudden hemlock]

[past junco]

@modest fog How have you been lately anyway?

[past junco]

What server? Lmao! Actually, forget I asked

[modest fog]

Let's just imagine if lewds was a server owner

https://tenor.com/view/sus-cat-cat-stare-stare-bruh-cat-really-n-gif-25597234

[modest fog]

Very

[past junco]

Sorry to hear

[past junco]

Reminds me actually, been getting some really interesting realistic gens. Not sure if it belongs in this channel exactly but it's something worthy to note

About to arrive home one sec

[bitter latch]

ok... the hair went wonky

[past junco]

This gen freaked me out a little bit, wasn't what I was expecting

Was expecting something a bit more... anime

Considering the model

Forgot to share these ones earlier

[desert island]

still working on it but here's the first attempt.... I gotta long way to go but I think the style is kind of there... sorry I haven't ever worked with the straight sd models before so I don't really know how to prompt them.

[sudden hemlock]

Damn I really like the style and the details on the clothes and flames

And isn't SD 3.5 super new? I think very few people know how to prompt it xD

[desert island]

Yeah I scratched the 3.5 and went back to a 1.5 model

[bitter latch]

hmmm just put silver hair but it wanted all of it silver.... ok then

hahah

[desert island]

okay the hands are trash but let me put this together in a workflow for you. DM me so I don't clutter up the chat.

[past junco]

I'm divided on whether this is cute or creepy

[bitter latch]

the answer is... yes

[past junco]

Lol

[past junco]

Strange enough, I can't replicate that realistic look from the earlier gen... I'm just getting anime now

[past junco]

Ruined a great gen...

[bitter latch]

nah.. they are an ailen

[past junco]

Parasyte the Maxim?

[past junco]

Btw check this cool thing

[past junco]

I like the look of this

[modest fog]

tasty

like carrots

[jagged field]

Hello

[jagged field]

Oh thank you <3

😻

[brittle rampart]

I’m even scared to imagine why she needs such a big carrot! Jack I would be afraid of her! 🤣

[modest fog]

https://tenor.com/view/bunny-bnuuy-steals-your-bunny-steals-your-bnuuy-gif-26209043

[brittle rampart]

what tag to describe a pose in 3/4 turn?

not profile and not from the side

[vital raptor]

i've had some luck just combining terms with some models/poses, like from side behind if you wanted slightly turned away from viewer, or from side, facing viewer for front facing, it always takes playing with the weights a little

[modest fog]

that is pretty RNG haha

Put both perspectives and hope for the middle ground

[vital raptor]

it happens a lot, sometimes there just isnt a token for what you want

[past junco]

Tried fixing it

[jagged field]

Hello guys

[green temple]

hi

[jagged field]

Do you use Comfyui ?

[green temple]

no

i use forge

[jagged field]

can you help me with image generation ?

what model do you use ?

[green temple]

i use illustriousXL_v01

[jagged field]

ty

[modest fog]

Grabby

[vestal ferry]

HI

https://tenor.com/view/nikke-doro-dorothy-pipe-bomb-gif-4643514375044855612

[modest fog]

[jagged field]

wut

[modest fog]

don't stare at me

[jagged field]

k sry

[modest fog]

Very big

[vestal ferry]

https://tenor.com/view/maid-gif-21987731

[vestal ferry]

neko loli?

[green temple]

no she is smoli as @modest fog said

[vestal ferry]

mmmm

https://tenor.com/view/nikke-doro-dorothy-pipe-bomb-gif-4643514375044855612

[modest fog]

Points to smol

[vestal ferry]

https://tenor.com/view/anime-smash-lesbian-punch-wall-gif-4790446

[jagged field]

can someone help me with comfyui maybe some ideas ?

[vestal ferry]

i use forge ai

[jagged field]

all using forge no one using comfy :'(

[vital raptor]

technically I'm using comfy, but only as a backend because nodes are dumb 😄

[jagged field]

Okay i can install forge

[vital raptor]

if you're going to install something new, i'd suggest swarmui

[vital raptor]

it uses comfy as a backend so you can still use nodes if you need/want but has a better frontend for easier general use

what do you need help with?

[jagged field]

can you join diffusers longue ?

nvm I'm going to sleep

[vestal ferry]

i was using a1111

then forge

i didnt need to delete a1111

https://tenor.com/view/nikke-doro-dorothy-pipe-bomb-gif-4643514375044855612

[supple raptor]

when reimu steals your broom but you have places to be:

[desert island]

if you have questions about comfy you can ask me, it's what I use. @vital raptor nodes aren't dumb 😜 they allow you to test and create weird workflows and allow you to interact with your models in ways that can fit whatever use case you have. You might want to create your own applications that can do whatever you want and exactly how you want them to work. Example: I don't think I could have gotten mochi to work on my 4080 if I was using swarm or forge. Someone needs to create the workflows that allows for the frontends to do these kind of things right?

[vital raptor]

thats a big part of why i use swarmui, it has a simple UI for generating images, but if i want to experiment with things then i can use the nodes

[desert island]

so you aren't saying don't learn comfy, just don't use it exclusively? Use swarm when you get workflows you like rather then sit on the comfyui 'backend' to gen your images and stuff... I think that's what you're saying correct me if I'm misinterpretting since I don't know any better.

[vital raptor]

i'm not saying either way really, if you want to use comfy then go for it. I don't think comfy is good for beginners or for non-technical people, but they can use whatever they want

[desert island]

I guess let me clarify since I'm unfamiliar with swarm or forge and haven't taken the time to install them or test the use interfaces, is that how you use swarm? I'm curious cause I want to understand how other people work in this space.

[zinc flax]

It's just a111 but even simpler

Gets the job done in the end lol

[vital raptor]

swarm was made to be a user friendly front-end (the web interface) and it can use multiple things for the actual image processing, comfy, a1111/forge, novelAI, other web services apis, etc

by default it uses comfy in the background to generate the images, but without having to actually change nodes around, it has usable menus and buttons

but, it also allows you to use the normal comfy interface and play with nodes if you want to

[desert island]

thanks that helps me a bit... I think we're saying the same thing.

[supple raptor]

Swarm builds a comfy workflow automatically for you with the vast majority of features that most people use. If it can't do something you want natively then it also has a comfyui tab that is literally just comfyui with full functionality including custom nodes.

[vital raptor]

forge is a fork of automatic 1111, both are much more user friendly, but have their own issues, generally speaking comfy (and therefor swarm) has better performance and compatability

[desert island]

Thanks @supple raptor and @vital raptor this was very helpful. 🫂

[vital raptor]

the biggest upside to comfy is that when a new model type comes out, like SD3.5 or Flux, it's as simple as making a node or 2 that supports the new code, and it's available same day, swarm generally benefits from this as well, a1111/forge sometimes take weeks to get updated for new things

[supple raptor]

Swarm's built-in image, model, LoRA, etc browsers are also excellent. Far better than the equivalent experiences on Forge/A1111

[vital raptor]

i like the inpainting better in swarm as well, the learning curve is a bit higher, but it ends up being better overall (though could be better)

[supple raptor]

Forge/A1111 LoRA browsers fall apart if you have more than like a dozen loras

[vital raptor]

and, on a personal level, I like swarm more because the developer actually listens to feedback and allows people to suggest ideas/features. I've had issues with comfy's dev and a1111 might as well be talking to a wall

but swarm has not only implemented things I've suggested but I've done some of the code myself because the dev takes the time to actually look at it

[supple raptor]

Yeah, I don't think there's a dev in the gen-ai space that's more responsive or open to feedback than mcmonkey.

[vital raptor]

and without the elitist arrogance that some have...

[supple raptor]

I've seen him patch niche problems that 1 person reported and add features that someone requested within literal minutes of getting the report/request on multiple occasions

Obviously not every response is that quick

[vital raptor]

speaking of that, looks like he added Mochi text2vid support yesterday

[rough oxide]

it's been able to run it fo ra while now

but def more stable now

was avail back oct 26th as basic support

[rough oxide]

oh swarm

my bad

comfy has had support for like 2 weeks already is what i meant then

[vital raptor]

yeah, and you could use the comfy nodes in swarm too, but native support now

[rough oxide]

i sleepi now. i should sleep before i misread anymore

[jagged field]

Maybe if you have time we can talk

sometimes i want to play a little bit with ComfyUi but i don't have ideas

[desert island]

sure DM me whenever

[jagged field]

can you join any vc ?

[desert island]

vc?

[jagged field]

Voice channel

[past junco]

I've not heard of swarm before

[desert island]

Here's the whole channel dedicated to it if you're interested 😁 https://discord.com/channels/1002292111942635562/1134546054872842290

[past junco]

Thanks for the info!

[modest fog]

I want a dog

[modest fog]

would not buy this smol

[green temple]

others are smoler

[green temple]

🥲

[modest fog]

Grabby

@green temple

[modest fog]

get away from my car

[rough oxide]

i can't catch this pal (species) with this spherical object, wat's going on?

[green temple]

trying🥲

[rough oxide]

whai illusv0.1?

have u tried the much further finetune (32 epochs more) noobai v1? https://civitai.com/models/833294/noobai-xl-nai-xl

[modest fog]

what are you trying

[rough oxide]

whatever ur trying to pull off should work better in noobai v1 at any rate but u might have to lower the like 1.4 weight

cuz it's a lot more sensitive since it has way moar knowledge

[green temple]

i was trying to edit the images a little bit

delete things

[green temple]

i`ll try it soon 👍

[rough oxide]

removed entire bottom half of prompt

tho

lol

also

i think the tag is "Alternate" not "Alternative"

[green temple]

😂

[rough oxide]

u dun wanna do that if u get into noob ai v1

lol

it's way more trained

even more than novelai is xD

[green temple]

niice

[rough oxide]

but also wut r u trying to do

[glacial coral]

A fren is asking what’s a good SDXL or Pony model for anime that isn’t relatively difficult to prompt with what’s should the answer be?

[rough oxide]

heytea outfit?

[rough oxide]

noobai v1

literally the purpose of model

[green temple]

i was trying to delete somethings from the image

like the butterflies and the hair

[glacial coral]

Understood, u got a comfyUI workflow for it or know of one he can use?

[rough oxide]

o

[rough oxide]

uh

but

if u use this

rename the model to NoobAI or expand the checkpoint

and load it manually

i have like everything collapsed

[glacial coral]

👍

[modest fog]

how dare you

@green temple giant prompts must be giant for the smallest smol

[rough oxide]

isn't necessary in noob. lol

minimal negative prompts r better in noob too

[modest fog]

no

[rough oxide]

but anyways aside from noobai's purpose being smoller prompts & easier to use it's literally waaaay more trained than illus on 32 epochs for 12.7m images which i think was more than illus training too

..i heard it cost like over $60k

weebs shur advance ai fast.

[rough oxide]

also it should be noted you quite literally, for the most part, don't need LORAs for this model, it can do anything with 80+ tags on danbooru (as long as if it's at 80 the posts are quality)

and can do up to... alisa mikhailovna kujou

as for timeline

like mid july

just works

[glacial coral]

i need to learn how to prompt

[rough oxide]

wat is "95th masterpiece"

[glacial coral]

i fucked it

[rough oxide]

well, the same will be with illus, but you'll usually wanna use characters/artists

[glacial coral]

i knew something was off

[rough oxide]

uhh

the char sheet is over 6100

[green temple]

is it the "early access version " ?

[rough oxide]

but good for ctrl+f

Nooo!

v1

ur using the right one

[glacial coral]

i cant read for shi also thank you for that

[rough oxide]

if u grab a char from the list and like 1-2 artists from artist list

sumthin'll happen

probably.

if ur making oc

u can do that

just use artist

with oc

like mixing hews, club3, i like what that produces often even for ocs

[green temple]

oh i thought i should download another version

[rough oxide]

it's this eps 1.0 pred version

[green temple]

oh ok good

[rough oxide]

brb

[green temple]

thx 👍

[modest fog]

@green temple I'm that thing in the shirt

[rough oxide]

i r back

[rough oxide]

do you have any specific like questions about prompting?

[modest fog]

Ugly

[glacial coral]

im not the best at it in gen like what do i use to start with and what do i use to end it

[rough oxide]

1girl/1boy/1other, <character name if defined character>, <copyright if using a character>, <details/artists> is how i do it

[glacial coral]

so if i wanted to do a original character would i just prompt what i want the character to wear and other stuff?

[rough oxide]

ye; like the (sort of) custom rukia i made here[since it's a specific form not catalog'd well by danbooru yet], just minus name/series, and it'd be like that

ideally you might still wanna use an artist

model really shines with at least 1 artist named

[glacial coral]

i see

[rough oxide]

like this is just desc

but also a couple artist

this is without any artist at all

major quality change

still usable but more flat/not as expressive

[vestal ferry]

https://tenor.com/view/dorothy-nikke-doro-gif-7509545823082567005

[glacial coral]

this is pretty cool

[vestal ferry]

female sensei of blue archive

https://tenor.com/view/momoi-blue-archive-cute-finny-sensei-gif-4689091049969952688

[modest fog]

https://tenor.com/view/cat-cat-lick-fast-lick-cute-cat-cats-gif-27624519

[vestal ferry]

https://tenor.com/view/nuh-uh-punch-gif-14008503396845178733

[rough oxide]

commence licking... jk

[vestal ferry]

https://tenor.com/view/sensei-blue-archive-anime-gif-14298477291726793689

[green temple]

i like the colors more

[tranquil sentinel]

Hdllo

[modest fog]

Bloo

[modest fog]

grabs

[keen hollow]

I dunno how ot use this noobai thing

results are poopy 😛

[green temple]

how dare you !!! ... jk

[keen hollow]

still kicking the tires 😛

[green temple]

[keen hollow]

ugh ... I can't get decent quality...

[green temple]

add any artist tag and it will look good

[zinc flax]

They kinda fried the colors on v1 huh

[keen hollow]

artist tag, good idea

I toned down CFG to get the colors down, and increasing resolution on first pass also helps a lot

euler a seems to be yielding best results

after battle scene... dialing it in

more battle scene (post battle warning so spoilered)

[vestal ferry]

https://tenor.com/view/momoi-blue-archive-cute-finny-sensei-gif-4689091049969952688

[keen hollow]

trying to do some ultrawide fighting stance ones now...

[modest fog]

ugly artists

[vestal ferry]

need the numbers

https://tenor.com/view/momoi-blue-archive-cute-finny-sensei-gif-4689091049969952688

[keen hollow]

whatcha mean?

[vestal ferry]

of the resolution

[keen hollow]

3840x1600

[modest fog]

Reading that my first thought was finger count

How many fingers? Need the numbers...

[brittle rampart]

@green temple is the left option the result? and the one on the right was the original image before the changes in Img2Img?

[green temple]

not img2img but a site

[green temple]

yup

[brittle rampart]

some online service?

[green temple]

yup

[brittle rampart]

link?

[green temple]

ok

[keen hollow]

time to sleep... g'night

[green temple]

shinyuu no neko-tachi

[brittle rampart]

huh?

[green temple]

best friends ( cats)

[brittle rampart]

although I love dogs.

[green temple]

big lemon

[modest fog]

Noo

[green temple]

oh right , u hate friren

[modest fog]

Ball head

@green temple where smol to make day better

[brittle rampart]

Friren a smol!

[green temple]

no , but fern is

[modest fog]

She's ugly

[modest fog]

Those monsters in the sky will eat her @green temple

[green temple]

[brittle rampart]

@green temple The cat is now one step away from going to HELL!

[modest fog]

I'm innocent

[brittle rampart]

Why is Friren terrible?

[modest fog]

She's not a cute elf

Ugly head and eyebrows

And worst of all people sometimes call my gens like her

https://tenor.com/view/lazy-cat-fat-gif-18347421

[brittle rampart]

Which Elv character do you like?

[green temple]

agree , fern is better

[modest fog]

I don't watch anime

I like 1girl!

[desert island]

Deedlit 😜

[green temple]

friern is better

[desert island]

I haven't watched it yet... it's on my to-do list though because of all the love ❤️ it's gotten.

[modest fog]

@green temple

[green temple]

10/10

[modest fog]

giant water lizard

[green temple]

i like aqua/cyan more

[modest fog]

It's all bloo

[green temple]

nooo

[modest fog]

I am eating bread

and tea

just a plain piece of bread

[green temple]

pog

[modest fog]

Im like in medieval times

butter or anything just makes you fat im too fat cat already

[green temple]

milktea is better

[modest fog]

I never drank pure milk

It is so bad

[green temple]

[modest fog]

@green temple I tried asking for a small aquarium

[zinc flax]

https://tenor.com/view/fern-pout-frieren-beyond-journey's-end-anime-gif-15807147379354972881

[modest fog]

is this fiern

or fern

Fien?

I forgot

[green temple]

👍

[modest fog]

not an elf

[green temple]

🤣🤣🤣

[brittle rampart]

The story is pretty good. and it was nicely shot.

[brittle rampart]

You're deceiving!

[modest fog]

what

[brittle rampart]

you drank pure milk. even the purest.

[modest fog]

[brittle rampart]

Yes. Yes...

[modest fog]

[green temple]

🌚

[brittle rampart]

🤣

I wonder if the cat is already 10?🤣

I'll have to clarify...

otherwise he can't

[modest fog]

[brittle rampart]

are you already 10?

[modest fog]

I was born and was already drinking tea

[keen hollow]

ok i gotta go do today 😛

and eat something

maybe a sandwich

[green temple]

smol tea

[brittle rampart]

did the tea have a fluffy, white, foamy “head”?

tag for the expression ">_<"?

[vital raptor]

pretty sure that is the tag

[brittle rampart]

I'm not sure it will work.

[green temple]

it worked for me

[vital raptor]

might work better with closed eyes and no other tags for eye color or anything

[green temple]

yes

[vital raptor]

that said, it doesn't work for me with the artist tags i use since they're not as cartoony

[green temple]

chibi artist tags

[brittle rampart]

(closed eyes, >_<:1.2)

🥴

[near elbow]

literally "a random manga artwork" as a prompt 🙂

another

and another... nice variation 🙂

[supple raptor]

@finite herald happy friday!

[finite herald]

thanks nubby

[supple raptor]

Coke

hope ur doing well

[finite herald]

mood

[near elbow]

"a series of manga panels, story of a kitsune marrying a mecha" - because we can!

the sad truth is in the black square

[modest fog]

eats her teddy

[modest fog]

@green temple

Bite

[vestal ferry]

https://tenor.com/view/frieren-fern-beyond-journeys-end-macdonald-burgers-gif-10253262265477048425

[modest fog]

Plastic food

Also ugly fake elf

[modest fog]

[vestal ferry]

https://tenor.com/view/frieren-anime-elf-cute-kawaii-gif-1391919008270779198

[gusty palm]

havent genned in a while

upscaling changed this a lot im not sure which i like better

[keen hollow]

oh wow it's @gusty palm

been a while

[gusty palm]

hey

hurricane killed my computer and i had to fix it and lost a bunch of work over it

[bitter latch]

welcome back

[gusty palm]

thanks

currently using ux-forge but the img2img thing is weird and i think imna go back to normal a111

[rough oxide]

u up to date with latest anniemay model?

[gusty palm]

no was dat

[gusty palm]

kakigori still on top

[clever shore]

I have an OC that I want to replicate, probably as like a LORA? It's a robot with pretty basic features that need to be replicated exactly every time, but it has different facial expressions. I will probably use it with some sort of anime/comic style model. How would you guys go about making something like that? I do have a dataset of ~10 images. I feel like this channel is most relevant to the style i want so i wanted to ask here

[gusty palm]

cool

[supple raptor]

lora is the right choice here. 10 images is a small dataset for a character lora which will make it more challenging, but not impossible. you won't want to use flip augment either since that will result in the handle on the door in it's chest being on random sides and will make the numbers in the side less likely to train correctly. for captions, e621 tags would probably offer the best accuracy. i know the furry crowd has some automatic captioning models that utilize e621 tags, but i don't know what they are off the top of my head. as for models, i'd go with NoobAI since it's the best anime style base model by a pretty wide margin and also happens to already be trained on both danbooru and e621 datasets so it won't have to learn those tags from scratch.

[clever shore]

thank you for such a detailed response 😭 Honestly I'm not too picky about wanting the numbers on the sides, so thats okay even if it doesn't work. do you recommend anything i should add to the dataset to make it less challenging? i can quite literally just draw some more. ive done some loras before for like anime characters but they've always had very varying poses and different outfits and this guy's just pretty much the same thing everytime

now that i think of it i'll also have to familiarize myself with how to make loras because i bet the process has changed since I last did stuff like a year ago. I also used colabs but i have a dedicated gpu now so i assume it'll be easier

[supple raptor]

varying framing and angle of the character and distance from the viewpoint as well as some images with full scenes with other objects would be helpful. it wouldn't hurt to just give it a try with what you have first though as long as you don't have to pay per training run. if the first attempt doesn't come out quite good enough you can consider generating images with it and keeping the very best ones, maybe even touching them up if needed and adding them to the dataset to train a second lora.

idk what the current best notebook is for training on colab, so that will require some research. something based on kohya_ss would probably be the best choice

[clever shore]

oh i meant i used to use notebooks but i can run everything locally now

[clever shore]

okay thank you

[supple raptor]

oh cool that makes multiple training runs easier since you don't really have to worry about the time it takes or number of runs

here's the most recommended trainer for local training https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/main

[vital raptor]

I ordered two M40's for my server.. they're a bit slow compared to the 4090, but 24gb vram each should be good enough for LLMs or slow training 😄

Should be here tuesday, will be fun to play with

[near elbow]

have fun!

[vital raptor]

I'm hoping they'll be at least fast enough to take over the LLM from my 3080TI so I can use that for something else

[jagged field]

Hello

[near elbow]

flux fusion with "an anime character design sheet, section 1 is the character name and the key visual, section 2 is design notes, section 3 shows the same character from different angles"

[jagged field]

Wassup guys

[near elbow]

no loras were harmed in the making 🙂

[jagged field]

I've tested Flux Fusion V2 it's very good

[near elbow]

"gundam" added

[jagged field]

Anyone wanna play with me with ComfyUI a little bit ?

[near elbow]

"an annoyed kitsune female, poking around in a bowl of spaghetty, bowl label "Comfy", speech bubble "not really""

like magic 😉

[vital raptor]

lol, comfy noodles

[jagged field]

This is flux fusion ?

[near elbow]

right

[jagged field]

oh nice

[near elbow]

8 steps, 6.6 seconds

can't complain 🙂

first try, no cherrypicking

another output

the "annoyed" carries over well 🙂

[vital raptor]

I kind of wish there was a pony type finetune for flux

[near elbow]

and regarding GUIs...

pony is broken on so many levels, starting with the magical quality tokens... but yeah, haven't seen a so "specialized" flux yet. pony does it all

assuming "all" is specific anime characters, tentacles, shemales and more 🙂

[vital raptor]

the hardest part with pony models is figuring out the tags you actually need or dont need, I only use score_9, score_8_up, score_7_up, source_anime, plus a couple artists interpolated, the rest of the prompt is image specific

I don't do specific characters, but the nsfw capabilities are next level when you want them. I mostly gen sfw though

[near elbow]

yeah, pony is for k1nky stuff (stupid server censoring bot 🙂 ) and special poses and super specific charaters... anything else manga/anime is fine with flux nowadays

"manga drawing, a fight scene, headline "flux or pony?""

[vital raptor]

i would use flux more for my sfw stuff, but i couldn't get a style I liked out of it since it seems to basically just do 'generic anime style'

with pony I can mix and match multiple specific artist styles and tune it to the point of having my own style

[near elbow]

but it's amazing how well flux does combos 🙂
that's the better language model etc, no chance with pony i think
"manga drawing, 4 panels, the story of a pony converting to flux, panel 1 shows the original pony, panel 2 and 3 show the transformation, and panel 4 is the flux pony spewing electricity and power"

ok, 4 vs 5 panels, but hey... really nice

[vital raptor]

yeah, the context and comprehension of flux is pretty impressive for sure

[supple raptor]

i only see panels 1-4 and panel "s" seems fine to me

[supple raptor]

yeah, sadly it pretty much does instagram photo with heavy filters, generic anime style, and generic "illustration" style. even things you'd think should work on any base model have nearly no effect like oil painting of..... i've even tried adding descriptions of what oil paintings look like with minimal effect. i was very disappointed in that aspect of the released model. some pre-release versions could at least do stuff like that.

[vital raptor]

I've debated playing with some custom workflows using flux and pony/noobai to try and get the best from both but I've been lazy

[gusty palm]

i feel like i dont really like forge

[supple raptor]

SwarmUI is better

Coke

[vital raptor]

[gusty palm]

looking at tht github now

[supple raptor]

just make sure you grab the up to date SwarmUI, not the old StableSwarm repo

[vital raptor]

hangry noms

[supple raptor]

‼️

a wild yue has appeared