🍥｜anime

[bitter birch]

sure you can show us your results

[twilit portal]

[dense raptor]

Hey guys, quick question — do you also get a lot of bad outputs like extra fingers or weird hands in anime images?

I’ve been working on a small tool that automatically detects these issues and filters out bad images so you don’t have to check manually.

If anyone wants, I can test it on your images and show results 🙂

[bitter birch]

I love it

[bitter birch]

hacker, don't click the image (image already deleted)

[smoky leaf]

Does clicking the image actually do something?

[bitter birch]

What I know is if you click the image, they get somekind of discord ID number that is long af, and it could be used to log your account in bypassing the 2FA, or something, I forgot

I totally forgot how it works, but just don't click it

Just find it on YouTube

[pale mural]

Its same as mr beast scammers

Watched one honeypot server go from about 16k phished user accounts, botted, to 30k± in 48 hrs. Industrious lil buggers 😂

[bold ether]

Did it got nuked already?

[bitter birch]

You mean deleted, yes

Thank you for informing me

[tawny wasp]

Nisha The Snake
https://civitai.red/models/2569730/nisha-the-snake?modelVersionId=2887410

[wooden sail]

Pixel-perfect background generator

[near elbow]

nice - what are you using for that?

[wooden sail]

Flux.2 with some tweaks

[near elbow]

just for fun your pool plus "change the setting to summer noon, keeping the details and composition unchanged" with qwen image edit

"snowy winter evening"

"retro neon alien invasion" 🙂

not sure how pixel perfect that is, but looks rather convincing

including the snowy tree reflection and all

even got the ice in the pool right

[wooden sail]

It will work (just don't make changes abrupt in the game) - simple transition animation will hide mismatch 👍

[near elbow]

spam... though i don't get the business model behind that... prepay the shipping?

[wooden sail]

It's man-in-the-middle type I guess. Someone post real listing then scammer copies it, sets higher prize and then just collect the difference. Or may replace the item with cheaper copy

[near elbow]

you join the server and the first thing you do is post a giveaway in a channel that has nothing to do with any of this? come on

i've seen the same story with a used macbook on another server by others

no one in the internet is so kind as just to give away things. there is another motivation behind it

sorry if i'm highly sceptical, this is the internet and usually the free lunch isn't

[junior zodiac]

Hi... I'm new here! I'm a developer recently introduced to game dev and I'm looking to gain more experience. I'd love to help anyone who needs assistance with development or creating assets for their games. You can check out my first project, made in 15 days, on my profile.

[shy karma]

I’m Akeem, a full-stack developer with a strong interest in AI and generative models.

[near elbow]

typical anime measures... sorry, blame the model 😉

[tawny wasp]

Nisha The Snake (Illustrious) OtakuStorm_Ai
https://civitai.red/models/2579891/nisha-the-snake-illustrious-otakustormai?modelVersionId=2898372

[wispy canopy]

if you made that on anima3 -> you did it on purpose, and you know it.

[near elbow]

literally hot

[dusky linden]

Flaming even

[near elbow]

mixed messages here

[dusky linden]

Oh shit dope glitch effect, anima model?

[wispy canopy]

yeah, anima

[dusky linden]

Maybe i should switch from illustrious nowadays

[wispy canopy]

you can't do highres with it yet tho

[near elbow]

upscayl.org?

just for fun

[near elbow]

yes YOU

[dusky linden]

If i miss a spambot you guys see earlier then me feel free to ping me

[bitter birch]

Sunset beauty

[snow sapphire]

ok i'm not really in the anime scene but it seems like there is a lot of interest with this kinda stuff so will see what i can cough up

[wispy canopy]

as cool as it looks, that's a scheduler/sampler mismatch if i've ever seen one

[dusty sigil]

Whats that

[wispy canopy]

basically, different models want different settings. and that kind of artifacting usually means the selected sampler got a schedule (both a specific kind of setting) which it didn't like -> you get very heavy artifacting

[dusty sigil]

The thing is that image was made using gpt image 2.0

[wispy canopy]

ah. yeah, i saw smth on that on reddit. that gpt image sometimes does that

[dusty sigil]

@wispy canopy how about these mixed between nano banana pro and 2 and gpt image 2

[wispy canopy]

those don't have that artifacting

image generation came a long way, making those so easy xD

[dusty sigil]

I thought it went for painty or pixelated style didnt knew these were scheduler artifacts

[wispy canopy]

it could be. but if you look at the guys' face on the left -> something went really wonky

[dusky linden]

With gpt image you can't really control it like how you could locally so who knows what gpt decided on in the background

Thankfully we got open source so i can mess up my own images dang it

[dusty sigil]

Yeah it definitely doesnt seem right

[wispy canopy]

this. i'm pretty sure they got some really impressive tech working behind the scenes. but something's off

[dusky linden]

Though i mean what they do have is impressive though, i just wish i could control it more

I kinda ignore closed source lol

[wispy canopy]

i mean, i make the lewd gens -> open weights / local gen all the way

[dusty sigil]

Oh this was the original image

[wispy canopy]

yeah, that's VERY heavy artifacting

have some local gen xD

[dusky linden]

I mostly use it for drawing references so i dont have to slog through the damnation called Pinterest

[cobalt canopy]

Generated myself with Stable Diffusion Illustrious. Character art for TTRPG game I'm playing in.

[near elbow]

horns maximizer

[bitter birch]

Flat color prompt?

[near elbow]

not explicitly - anima does a crazy variety of output for the same prompt if you stay vague enough, it's really fun 🙂

[cobalt canopy]

Why is she always holding a chair? Cool images by the way.

[cobalt canopy]

Why XD

[dusty sigil]

||That chair is the male lead he was converted into a chair by a "cat" and shes chasing the cat to get him back into a human|| (extremely simplified and theres so much more to it didnt spoil much first 10 mins only)

I know it sounds stupid but trust me the story isnt as stupid as i made it sound its super fun and the anime itself is beautiful

[near elbow]

well, for anime stories this is one of the more logical ones 🤪

[bitter birch]

Oh so it's just randomized?

[near elbow]

same prompt, different seeds (she's supposed to say "anima loves variety" 🙂 ) so yes, it varies wildly, and i love that

of course the more closely you describe the character and scene the more reproducible it gets

[bitter birch]

Oooh, I think it's the model you used

[near elbow]

https://huggingface.co/circlestone-labs/Anima

[golden crest]

https://tenor.com/biY19.gif

https://tenor.com/bfox2.gif

[bitter birch]

Back to Retro Nostalgic

AI can only make flat colors when without background

But currently I only use the SDXL model for now, especially since I run it with an RTX 3050 Laptop

[near elbow]

hm, not really overpowered, right

more anima

[bitter birch]

Can this model run on Forge UI?

Might try it

[dusky linden]

Anima can run with 8gb vram but i think forge neo can maybe

[bitter birch]

Forge Neo is somehow 10% slower than Original Forge on my laptop

And I don't care about VRAM, I'm just gonna put anything on my laptop until it OOM

[near elbow]

forge neo does https://github.com/Haoming02/sd-webui-forge-classic/blob/neo/README.md

[bitter birch]

Hmm, will try

I havent learn more about other features.. What does text encoders do and do we really need it?

[bitter birch]

Lol

[dusky linden]

Hmm strange, im not a forge neo user so i recommend that you pop into #🤝｜tech-support

[near elbow]

are you on the latest version? forge neo?

you need all the models it says, e.g. text encoders are for prompt understanding

[bitter birch]

nope, I'm using the basic Forge

But, is Forge Neo as fast as the basic one? or is it Faster?

I wanna try ComfyUI too, but is it also as fast as ForgeUI?

[near elbow]

well the old forge does not support anima 😉 - neo should be very close

comfy is a different infrastructure, hard to compare directly

but should not make a huge difference in any direction

[bitter birch]

OK got it

well, I'm not really using the AI oftentimes, and my hardware is also limited too, so I'm gonna stick with the basic Forge due to it's fast generation than the A1111

[near elbow]

hang loose

[near elbow]

may 4th 🙂

[near elbow]

armed and dangerous

[wispy canopy]

is that kinda level of spice ok nowadays?

[near elbow]

no idea - there's comparable artwork in this channel. if you think it's moving the red lines too much, i can take it down, no offense taken?

[wispy canopy]

nah, i'm fine with it ^^

just wondering, i had timeouts for less at times (years ago) XD

[near elbow]

yeah, i still remember the beta when i played around with "succubus" as main part of the prompt 😇

nowadays there's not so much going on here anyway, unfortunately

[wispy canopy]

i think a lot of people got fed up with a lot of the stuff happening

[near elbow]

and SD is not really top tier anymore either, unfortunately

[wispy canopy]

stable diffusion just got us all in

yeah, now it's so many different models

[near elbow]

yeah, 2022... when i got the 4090 rig just for AI images. still works nicely 🙂

[wispy canopy]

i'm so glad i upgraded from my 3080 to 3090 when i did

regret i didn't buy a 5090 when it was at €2k tho

[near elbow]

but it feels like the models are converging towards perfection, LLM based prompt understanding, and censorship of the most trivial things... all very similar

[wispy canopy]

the cheapest it ever was

[near elbow]

before the AI hardware madness rush, yeah

[wispy canopy]

and now it's easy 🥲

[near elbow]

right, mandatory image - chainmail plus coat 🙂

[wispy canopy]

but i really love you can feed anime even raw bs as input with the right formatting, and you still get cool stuff xD

[near elbow]

i remember contests like "an anvil" because the AI couldn't draw one back then, or the first "what it can do text?"... and nowadays, well... like cheating

[near elbow]

or AI image editing, supercool, like having your personal photoshop assistant

[wispy canopy]

that's not anima, i suppose?

[wispy canopy]

but for real, as far as open models go, anima is seriously crazy good

[near elbow]

seedream 4.5 via replicate API (that has a very convenient "disable safety check" flag 😇 )

[dusky linden]

Personally if noone raises a fuss i wont either, and yes theses been more risque

Posted before without issue before i became onboard

[near elbow]

yeah, more relaxed now definitely - thanks for the confirmation

[dusky linden]

I'm thinking of reviving the dailies, since this is the most active channel im checking here if its something you guys are interested in

[wispy canopy]

it was so close to perfect text xD

[wispy canopy]

xD

well, anima is still running on t5 (internally), that's a big difference

[near elbow]

and so much better than older models anyway (many folks are still clinging to illustrious or pony with SDXL underneath, that's antique now 🙂 - ok, i get it, tons of mixes and LoRAs out there)

[wispy canopy]

agreed. "just let it go man"

gooners be like

[dusky linden]

I cling to Illustrious because i use it for style references i cant get right in anima yet

[wispy canopy]

that's gotta be very rare then, anima in my expereince responds better to artist tags than illustrious does, expecially overtuned finetunes

[near elbow]

i mean, anima was trained on illustrious and pony output AND has the explicit artist tags - you probably know this? https://thetacursed.github.io/Anima-Style-Explorer/

yikes, they added a paywall for p3...

[dusky linden]

Hmm maybe i should give it another go

[near elbow]

just try it - personally i love the crazy variety for the same prompt. if all you want is character consistency, you probably need to describe them in great detail to ensure that (or train a LoRA of course)

[wispy canopy]

actually -> if you do follow the prompting guide they have on the HF page, it listens decently well, and will generally just ignore character tags if it doesn't know it

[dusky linden]

Hmm i used to llm prompt build but i think Gemma should be good for it

[wispy canopy]

be sure to use the guideline on the hf site, prompt output and consistentcy increases by a lot if you put things in the right spot

[quasi shadow]

Someone who I can't seem to find made this image a good while ago

Yours reminded me of it

Even though it's just the scythe

[wispy canopy]

oh yeah. i know who made that <- that was me xD

i recognize that image ^^

[wispy canopy]

goddamit, which model did i make that on? ah, found it.

flux crea didn't listen to all instructions, but i liked the output nonetheless

prompt still delivers

[near elbow]

"color shift silhouette black background female grim reaper scythe" 🙂

[wispy canopy]

now that's cool shit

re-used an old sd3.5 prompt on anima. anima being smart as it is... is actually trying to strike at the feels?!

[near elbow]

well, anima has qwen 3 6b as CLIP so it does understand a lot more than SDXL based models

but it gets most creative with short prompts - e.g. "grim reaper silhouette surreal" and that's it

same prompt, different seed

and another - all over the place, i love it

and a bit more elaborate

[humble hatch]

help on how to use anima? despite following tutorial i can't get any decent image, it's all a mess

wai-illustirous worked right away without much effort, but not wai-anima for exemple

[bold ether]

hard to help when the only info we have is "it doesn't work"

what kind of output ? what settings ? hardware ? etc

[slender barn]

Getting back into it its been a while

[humble hatch]

well i followed the tutorial here : https://civitai.red/models/2458426/anima-official
Same CFG, same resolutions, same sampler, and so on

using forge neo

tried both normal anima from the link and wai-anima

[near elbow]

maybe go for the original? but the latest forge neo (have you updated it?) should work just fine

https://huggingface.co/circlestone-labs/Anima

totally different style

[bitter birch]

Oh wow

[bitter birch]

Daaaamn, Good shit

[wispy canopy]

actual path of exile build

[main orbit]

Warning: Blood.

[near elbow]

rasterized look

[near elbow]

and b&w

[wispy canopy]

just some anima things

[near elbow]

wholesome kitsch

[twilit portal]

my dazzling queen ✨

[near elbow]

chibi sticker

[wispy canopy]

modernized elf

[near elbow]

ernie turbo (no idea why it uses duplicate bubbles, but the rest is neat 🙂 )

[wispy canopy]

Double minigun, double bubbles, but I like the charm of repeating it to herself

[near elbow]

got it, "on the left a thought bubble text ..." gets rid of it

[wooden sail]

So, how is base Anima model to you?

I see great finetuning potential (for concepts mostly). Also it's kind of sampler dependent and can be easily trained

[near elbow]

how much of a difference to p2 or p3?

[wooden sail]

Better details

[near elbow]

same speed and size? sounds like a no-brainer to upgrade, right

[near elbow]

got it. still can't spell reliably 🙂

[near elbow]

ah, even managed to get a nice cup without any people in the image

[humble hatch]

so i updated forge neo to try anima, but i get this error : The size of tensor a (2016) must match the size of tensor b (512) at non-singleton dimension 1

number change depending on the resolution i put

[near elbow]

yeah, google says "Most likely its the problem with the resolution, it should be divisible by 16."

try 1024x1024 to be on the safe side. if that works, play around a bit

and a forge courtsy of anima

[near elbow]

chaos

[stiff temple]

chaos

[humble hatch]

finally managed to make anima work... Now how do i prevent specific characters from blending together? I tried using BREAK but it doesn't really work

[stiff temple]

interest ing

[near elbow]

just describe them in their physical location, no BREAK etc needed - e.g. "red skin female demon horns kimono next to a blue skin male angel halo wings, a green cat at their feet". anima is quite good with telling what attributes belong to which character

[humble hatch]

it's the only way? I usually like the character to be randomly put in different place with general prompts of what's happening, sometime it works but a lot of time it just blends them

but thanks i'll guess i'll try

[dusky linden]

different models have different prompting rules

[humble hatch]

btw on what database is anima trained? danbooru? just to know which character and artist (and what name they use exactly) i can do natively without the need of a lora

[near elbow]

booru AND normal text (i'm very thankful for that). and a lot of artists and characters. check out their HF page https://huggingface.co/circlestone-labs/Anima

CRT look with ernie turbo for a change

[near elbow]

nice but not exactly anime?

[slender barn]

it is?

[slender barn]

how is this anime? may want to delete it by your own standards

[bold ether]

Let s chill out.

[steady grail]

that's more manga style

[bold ether]

It s anime enough in my book. If we see people posting too much of such or such style in the stable model channels or here. We ll create new Dream Communities for it. I m all for it.

[steady grail]

oh well, anime manga, manha, manhwa, visual novel, novels. same stuff after all

even if novels, only have a very small content of images

[bold ether]

don t say it too loud, I can hear purists sharpening their knifes from here. /s

[steady grail]

i started with anime, and delved in all that stuff in the same way

[bitter birch]

Takes more than 2 years to learn AI image generation, to create the art of elegant beauty

[near elbow]

that's one single image, i really like what anima does as a model

another juxtaposition. slightly different style

[near elbow]

... just saying

[wispy canopy]

is that yoshitaka amano style?

[wispy canopy]

anima punches waaaaay above it's weight in parameters. it's less parameters than sdxl but in terms of capabilities it's not even close anymore

[wispy canopy]

my poor 3090 tho... it's been cooking since sd 1

[near elbow]

anima system requirements are not so bad, especially compared to ernie or other more recent models - still on my 4090 here and loving it (never managed to get a 5090 or rtx pro 6000 when they were still semi-reasonably priced)

[wispy canopy]

a while back 5090's were available for €2k. i'm now really regretting not getting one....

[near elbow]

and i love the wild variety of styles with anima, and every result is so nicely random and new

i mean, 10 seconds for such an image at 1024x1024-ish is totally ok

[wispy canopy]

i'm pushing 50s for a 1280^2 -> which is kinda expected, since i can't run fp8

[near elbow]

still reasonable-ish

more styles 🙂

[wispy canopy]

for 50 steps sure

not like hidream that takes 5 minutes for 1024^2

[near elbow]

why 50 though? 30 work well enough for me

[wispy canopy]

i'm testing various schedulers atm, and there's a stark difference between 30-50

30 isn't bad. at all

but with the right sampler/scheduler -> 50 steps just fckn obliterate the same prompt/seed as if it iwere 30 steps

usually -> ancestral sampling

if you're using deterministic/ODE sampling, 30 is generally good

[near elbow]

sounds like i should play around a bit - quite happy with the current results though

[wispy canopy]

ancestral/SDE sampling -> 50

[wispy canopy]

you really should. the scheduler is SO extremely potent in terms of image composition, it's nearly almost a required factor in image generation

blurred one out because it was nsfw

[near elbow]

hm... same seed and all, 30 steps and 50, and 70. interesting how it switches poses mid-generation

not a quantum leap though

[wispy canopy]

A) you're using any form of ancestral sampling
B) which scheduler?

[near elbow]

#🍥｜anime message

[wispy canopy]

simple is one of the worst schedulers, imho

[near elbow]

still gets nice results, i never saw a reason to change - what would you suggest to try instead?

[wispy canopy]

https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler

i'm picking this one out, even in blind tests

but this schedule is optimized for flow models -> what is anima -> a flow model

[near elbow]

too lazy to install yet another node and change my workflow though - what existing samplers would be worth a try?

[wispy canopy]

generally, it has the best between the tradeoffs, better detail, or better composition, and this one pretty muchs sits perfectly in the middle

@near elbow i tried sending 9 images and got a timeout, so i sent you them by dm, lemme know if you want details some of those exotic schedulers

[wispy canopy]

anima -> which technically makes it anime. (anima can do some pretty decent realism tho)

tried making cat out of water. got jesus cat instead.

[near elbow]

got them, thanks a lot, will take a closer look later

[near elbow]

"a cat made out of water liquid translucent ripples" works kind-of-ish

[wispy canopy]

i just replaced fire with water in my prompt, but i really thought it was funny how it completely changed the resulting image

[near elbow]

ah nice one

[wispy canopy]

fyi: (lowres, blurry, chromatic_aberration, bad anatomy, bad proportions, bad perspective, extra finger, missing finger, worst quality, text, artist name, logo, jpeg artifacts, spiral eyes, censored, border, bad hands, abstract, aged down:-1), @dino \(dinoartforame\), cat made of water, a blob of water resembling the shape of a feline is walking (i'm using negpip, so my negatives also go in my positive)

[near elbow]

"masterpiece, best quality, score_7,
(((safe))),
recent,
cat made of water liquid translucent slime"
(negs: worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, censored, censorship, child, loli, underage)

and a lot of NSFW instances where anima decides that a catgirl is much better 🙂

[wispy canopy]

. yeah. your prompt is catgirl heaven. i tried it. got 5/5 7/7 catgirls

[dusky linden]

animal cat made out of water

or was the booru tag cat/animal/ i forgot

[near elbow]

anima is also natural text (fortunately), not just booru - otherwise i would be totally lost

[near elbow]

two for the price of one

[bold ether]

so that's mi and mini mi ?

https://tenor.com/view/midget-dr-evil-austin-powers-dancing-dance-gif-5137925

[slender barn]

yes I am using him with a few other stuff

[humble hatch]

for anima what shift value should i use?

[near elbow]

hm, i'm using their default workflow (with modifications for file saving) in comfy... isn't there recommended values on their HF page?

links to other discord servers by freshly joined users are ALWAYS suspicious 😉

[cobalt canopy]

Trying out a free trial of GPT. Really happy with the art it can make so far.

[cobalt canopy]

Those Ai collar bones, wowsers >_<

[wispy canopy]

#justanimathings

it's a 3 year old prompt at this point, and it never gets old ❤️

@slender barn of course, can't do this one without invoking a yoshitaka amano xD

[humble hatch]

i have an hard time doing NSFW stuff with anima, unlike illustrious, like it's difficult to make it show nipples or penis, even if the image is super explicit, it just hides those features, even if it's in the prompt, with NSFW added and so on... Any helpful prompt i should add?