#🍥|anime
1 messages · Page 252 of 1
I know but still, it's funny how these models purposely limit their full potential
agree.
Hmm maybe, but astralite is using auraflow for the next Pony model, the truth is I don't think anyone has/wants to spend a huge amount of money to make a proper finetune on flux, but we shall see
Now that would be a game changer
if its not the original, some other pony might walk along. Or maybe some other fantasy creature.
who knows.
Yup, pretty much.
Tbh, if hands are your primary concern, you can always use a model fine tuned for hands. But then everything else will be garbage. In most pictures hands are such a tiny part of the whole composition and are such complicated multi part things it’s a wonder they come out halfway ok at all
it bothers me ....
azunyan
Whats a good model for hands? I just wanna use something to inpaint hands
cute lunar seems to improve on sd3 hands 😄
what are the sources you want to inpaint from? real photos? anime images? AI gens?\
Just 2d anime hands, although it doesn't need to be 2d because I can just color over it if need be
then I guess any anime model of your liking would do, just mind that you only need to prompt the inpaint image (not the complete picture!)
yawns
The best option for inpainting is to use the same model to regen a masked area, it’ll blend in best that way. Just takes time and sometimes a lot of genning
Maybe also use controlnet depth with hand models. I’ve used that with good success in the past
i usually just crudelly fix them and than regen the part wth low denoise
this usually fuses the fix with the image
But all else being equal, a SDXL model will inpaint hands better than 1.5 right?
Generally, but like everything else it’s model specific
i used some 1.5 model very god at genning hands
till they didn't handle any object lol
Maybe I should use it to inpaint hands, what model?
I have a custom trained 1.5 model that does really convincing realistic images with hands and everything, but I wouldn’t use it to inpaint anything else
I just inpaint with whatever model I made the image with, it works fine
the image will not fuse if you use inpaint with a different model
anyway, for anime, should be aniverse
i don't delve too much in realism
i don't like it at all
this one, but it's a little stiff
i dont really do realism stuff anymore, it started getting a little too uncanny
test
omfg finally i can talk
Anyways heres a pic that suspiciously look like it came from another model
wow that got dark
ooh spooky
that came out better than I thought. wow
now thats a full moon !

that's where this came from, the cute, fluffy, blue monster in the background, I just can't figure it out!?!? there is not a word in the prompt.... HOW and IN WHAT WAY!?!?!
I won the seed gacha apparently
IIne Asoury is the keeper of the library!
the most anticipated movie! 😁
looks like made in abyss lol
This is a character not included in history. A! No! it is the narrator of the story itself! 😁
So is this a very handsome boy or just a beautiful girl?
Well, that doesn't happen! this completely changes my knowledge about parallel worlds!

I suggest putting it up for voting!
It's definitely a girl!
Come here little mouse

Bonk!?

even smol laughs at me 
come to terms with it!
Smol
what is she gonna do 
Take your soul and sell it

Banana guns
@supple raptor When did you get so smol
2
What problem
"we are made of stardust"
same prompt clip skip 2
steps 10, cs 3
"Mr stark.. I don't feel so good"
got ya
Donuts in SPAAAACE
she got some donuts
haha
gasp
have some desert. i'm gonna go take an old man nap
hmm lady bug woman?
"what was.. that?"
.
COBRA!
snake trooper?
Cobra Snake trooper
Miku Cobra Queen
armored up
Cobra Clan Warrior
shakes the channel
ooh I like that
getting some interesting results . I like it , but time for some bed. night THUD
sleeps
and model is cyberrealistic25D_v10
Still awake... but what did she see?
falls asleep
her thoughts at this time - “Stop treating me like a little girl!” I’m already big!” 😁
anime
Yes. this is an anime channel
we are the anime
how you guys doing that ?
we write prompts and generate them. we generate. we generate.
so it is based on your real photo ?
of course not. 😁
he can.
how and what the tool name ?
There are quite a lot of tools.
thank you i will search
Good luck
looks like steampunk
foxyball
hehe yep
Share this cute style!
no style it was just "young petite toddler child, fox ears, bushy hair, orange color theme, play, simple background" no style just random words
Yes, I'm talking about the next picture. or share the model
tea time!
boys are under represented
not the prompt but ok..
haha well thats a family photo that didn't go as planned
o.O
Through the looking glass?
nah just to many styles making random
flux 
not flux
sansei !
Don’t you call... the Cat will come!
🤣 ❤️
this seems too much to me. Aren't models trained on a certain value of this parameter?
then again everything goes odd at 12 😄
shrugs
haha
there, ok laying back down. feeling blah
what did I do now
more miku!
english please
we post images we create elsewhere. no bots here that i'm aware of
Oh I can't believe it, I'm trying to create images 😭😭
So I'm going to look for another server, bye
bye then.. don't let the door knob get stuck on the way out
ug
cute and funny
it's not me. It's all orange juice!
new model under review, unpublished. a bit more mature, so I've called it Sexy Lunar for the moment. it's a bit nsfw out of the box, almost requiring some stuff in the negative prompt to tame it down.
Sexy Lunar seems to have a nice range from semi-real to 2d anime. Here's some realism stuff.
Sexy Lunar can stil do a bit of cute anime style
she stole my shirt! 
so cute
interesting style
gives me idea's for new styles.. ponders
combines style with some lora's
have 2
I am liking this style
haha
goes back to bed.. thud
anyhow.. back to bed fo reals
flux is kinda nice, but the style is a bit generic so far, will have to play with it more..
I could not get styles to work on that model lmao
It's either photograph, that anime style you showed or something weird like 3d figures

yeah, it's very generic, which is probably fine for some people
but, it's also a base model so i wasnt expecting much
also, the faster schnell version of the model is pretty bad for quality compared to the dev model, so you have to put up with the speed reduction
for me it's a difference of ~3s for schnell, ~13s for dev, 40s for 2x hires on dev

compared to sdxl/pony models running ~3s, 9-10s for 2x hires
ugh, loading the models takes forever, stupid 24gb files

I, of course, didn't download them
Didn't realize that would be so much to load lmao
schnell version takes about 17s to do hires 2x, so about 20s total
it's also doing 4 steps because it's a "turbo" model, the dev version is 20 steps
i think they're actually the same speed if you use the same steps
yeah
Some people were saying about exchanging weight of stuff in-between the models
Supposedly higher quality on the fast
In the #🆕|sd3 channel
🤷♂️ given it's apparent training dataset, it's not really making anything i want to use, so I'll just wait for a better version
and for me the speed doesnt matter enough to mess with it, a 12-15s gen for the dev model is fine
seems to generate blurry images a lot of the time, not sure why
It doesn't seem to be trained on any real anime artist iirc
Judging by how generic most of the anime gens look like
yeah, they basically used a dataset that was as non-artist specific as possible
mainstream generic TV anime #2347
ok what character is that?
just a cat girl
orders breakfast
waits
hmmm ?
interesting
haha ---> cup noodle chibi 1girl
haha
lol why is it all anime girls?
i mean i didn't see a generation of a boy
it's just that there are more anime girls that are popular and prob most here are guys.. so yeah...
yeah lol i'm aware of that
simple background,wavy background and just a lucky seed
hmm ok, thx
interesting 😄
combine enough styles and well it goes HS mode:
Bros and sisses. A bro on civitai yapped true this statement, apparently all you need in negative prompt is inequalities:2 and this will fix nearly all issues
not that I could see
Literally anything in the prompt has an affect though
true
Some random keyword that magically fixes most things though? Not actually possible
with the word inequalities in neg
same seed prompt and settings , without
shrugs
well, prompting cute, prety much fix most of the problems
lucky gen
nice angle 
Flux has controlnet now
model is too cute to be cute 
it`s chromasparkle v2
i can't find it on civitai 😦
1.5
cuteù
tomorrow i try to merge with aniverse and some other stuff
oh, that model is good in genning portraits
i wonder why i can gen good chibi only in this format
is there any way to make ip adapter work with pony?
i don't think there is a pony model of ip adapter
it should, does it not?

for what i did read, results are unpredictable
let me do one with vs without same seed and everything
ferals are so much trained in it, that it messed up any connection with other models
i dont think so, unless i have installed something wrong and/or have bad configs, it seems to not affect the end result at all
let me do a gen with vs without
but my controlnet is on d drive so it's gonna take awhile to load the model

maybe its because im using autismmix instead of pony, but i figure if it works with one it should work with the other
autism is not xl?
based on pony
i dont think so, is it? The civitai page says its an pony merge
kinda technically sdxl in a way but so far it's kind of a new thing afaik
look at the cute nun trying to summon something with handsigns lol
with vs without
im using tponynai, a merge based on pony as well
hmm, what model/config do you use? also, nice style on first pic
i mean ip adapter model and config
first is on ip-ipadapter
where did you get this preprocessor? i dont have that one but the model is same
might be a built-in stuff for my current version of forge
ps: do not update forge since it's in the mist of major overhaul now
oh i changed to a1111 dev some time ago
then you have to check your controlnet i guess
should be somewhere in there
i think the preprocessor is in the extension
so either update that or something
should work like this right?
what kind of preprocessor do you have under ip-adapter?
try the
clip h
since you said you liked the style
here is the pic for my ip-adapter
ps: pic is not ai generated
yeah it looks really cool, i will try it rn

is clip_h best for copying styles?
been so long since I used ipadapter in comfy I forget how
i think the big g might be better
but i think someone tested it, and it's not that big of a difference
not worth the extra generation time/vram usage
not sure which one is it (or where i read that)
oh yeah, now its definetly different, it just does not really look like the style, might be my prompt's problem tho
well, depending on your strength and stuff
but at least it works 
you can have more than 1 ip-adapter as well
ip adater burns the image very easily
it's kinda like NAI vibe transfer
yeah tysm!
you can't have it at high strength
im really new to ip adapter so i have 0 knowledge of what im doing lol
but mixing different references at the same time can make very cute unpredictable results
just test different weight and stuff, start weight low, like 0.1 to 0.2, then if it's still too much, tweak the starting control step
yeah thats why i wanted to try it out, since vibe transfer can be so cool at times
0.5 for starting control step means start ip-adapter at half the steps
yup yup
i remember i could gen without burning till 60%
with 2 references, 15-30% each
even less
but it's funny
the only donwside is that it's very slow
well, really depends on the picture as well
so yeah
it's a lot of trial and error
as slow as depth
tyyy, it does not look anything like the style but ig ill try some different configs
looks like that one pokemon lol
good luck good luck
true 
it's unpredictable, you are not the first one lamenting about it
btw what is this command to check metadata?
you meant like this? #🍥|anime message

naga u
so dark lol
i tried it ( Again) 
2 mins difference 
nu
jesus, that's crazy long
just the model is very cute
my gens are normally 3 minutes 🤣
I'm using autismmix but not as cute, still cute
this took 5

mine, at this res, 15 min
yeah
jesus, that's kinda like how i run flux dev version 
6 min at 1024
well, you can just click on that bot and add it to your user
Mine are like 4 mins
i coded that bot, so you can just use it anywhere (if you add it to your profile)
it's already a miracle i can gen at 2048
then right click a pic
damn i never knew that was a discord feature lol
it's a new thing called user-installed-app
tysmmm
ohhh nice

Actually so useful
Hey there! I'm happy to explain how Stable Diffusion works. It's a fascinating piece of technology!
Stable Diffusion is a type of AI model called a text-to-image diffusion model. This means it can create images from text descriptions. Here's a simplified breakdown:
1. Noise and Images:
- The model starts with random noise, which is essentially a bunch of meaningless pixel data.
- It then learns to associate this noise with real images.
2. Learning from Data:
- Stable Diffusion is trained on a massive dataset of images and their corresponding text descriptions. This dataset could include millions of images from the internet, books, and other sources.
- As it learns, it gradually understands the relationship between visual features (like shapes, colors, textures) and the words used to describe them.
3. Generating Images:
- When you provide a text prompt (e.g., "a portrait of a cat wearing a hat"), the model uses its learned knowledge to create a corresponding image.
- It starts with a noisy image and gradually removes the noise, guided by your text description.
- This process of removing noise and refining the image is called "diffusion."
4. It's Like Painting with Noise:
Think of it like painting. You start with a messy canvas (the noisy image). Then, you use a brush (the diffusion process) to add details and structure, following your instructions (the text prompt). The final result is a coherent and meaningful image.
Why "Stable?"
The "Stable" part refers to the model's ability to produce images that are consistent and predictable. It's less likely to generate strange or unrealistic results compared to some other AI image generation models.
Key Points to Remember:
- Text Prompts are Key: The quality of the image depends heavily on the clarity and detail of the text prompt.
- Creative Freedom: Stable Diffusion can be used to generate a wide range of images, from realistic portraits to abstract art.
- Still Developing: The technology is still evolving, with ongoing improvements in image quality, realism, and control.
I hope this explanation is helpful! Let me know if you have any other questions about Stable Diffusion or AI art generation.
ooh nice so we can ask things too
it's using gemini llm if you know what that is
so yeah
just a small little fun bot
🌚
@nova remnant
flux doesn't seem to like "from above" or "high angle".. not sure how to make it do that
'overhead view' seems to work most of the time
kinda cartoony
nice lighting
gets out the shrink ray
charged.. pressed button.. fire
ok more cute
goes back to bed.. thud
leaves one last one
dozes off
what a cool house he has! remove the stones. put things in order. and you can live almost forever!
where is the white dragon!?
I'm kind of hinting at something like this
ideal. only a head with a mouth and wings!
it turned out interesting...unexpectedly
cute
or maybe there's a dragon on the back of her head
looks like a korean or chinese ilustration
So uhhh is this competition? https://huggingface.co/Laxhar/Freeway_Animation_HunYuan_Demo
seems to be missing text encoders and i can't tell from their model card if they just didn't train them and expect you to use the ones from original Hunyuan. also seems to have some pickle imports that the original model doesn't share which makes me suspicious of it. i'm interested, but i'll wait for some clarification from the team and/or a complete .safetensors one.
Yeah ill wait as well, but also interested to see what this is about for myself
This model with unreleased loras
https://civitai.com/models/611079/nai3likediffusion
all 6 of them lol

charming
comfy made .safetensors for us. https://huggingface.co/comfyanonymous/Freeway_Animation_Hunyuan_Demo_ComfyUI_Converted/tree/main
Might give it a go next week
what is this?
Hope this came out alright for what it's worth. Used the Asethetic (anime) lora for Flux.
so it looks like putting "anime screencap" in front of any prompt gives a very reliable anime look.
(without any lora)
Yeah, I agree. I actually got some decent images just from putting something like "anime still", something to give it the go-ahead to treat it more like a screencap in whatever ways you wanna word it.
same but with anime still. so screencap seems to give a more flat look, anime still has more shading.
neat stuff
modern vs. retro i suppose.
also with anime still... or maybe it's just random and both work in the same way (but seemingly more than just "anime" by itself.
yeah just ran a lot of seeds, anime still has a more modern anime look across all of the images.
anime finetune of Hunyuan

I have no idea what this means 😝🤣 I guess I'm over here playing with flux like a chump
you know, we should add more lines that visually display the movement of the tail in this case
yes we should !
flux is the new hotness to be fair. hunyuan is an open source diffusion transformer model made by tencent which is more comparable in size and requirements to SDXL. https://huggingface.co/Tencent-Hunyuan/HunyuanDiT
ya do know this is stable diffusion server not Flux server right....?
do you see the sd in flux? that stand for stable difusion
exactly my dear friend, exactly 
shrugs

well, people do it all the time, me included
we talk about all sorts of stuff about image generation
shrugs.. i'm just in a meh mood
really trying to get booted arn't ya?
i would say the left one is a little too nsfw for this server
actually even the right
I would say both are but shrugs i'm not a mod
yup, even the right 
think positive, we can get cuties thanks to our analytic mind
i scavenged for 2 hours and cant find name 😢
well, first, it might be like some chinese traditional dress?
half boob shirt?
i wonder how people tagged it on booru
how can i get the bot 
Hey @green temple 👋 Hope you're doing well!
there
you can click the bot and add app
just know that's the bot is partly coded by me, so don't share "private" stuff to it
tho, i don't keep any data of that as well
but you will have to trust my words for it

great work
just for any reason you wanna remove bot, go to settings > authorized apps
how can i use it after adding it 🌚
for viewing metadata, you can just right click the image, and view parameter
oh ok
if you wanna talk to the bot, then ./ask for "one time" question
you can also add it to server to talk to it, with history
now that i think about it, i didn't implement the history properly, and it's probably shared across all servers
that might be bad

HUGEEEE
🙂
how to use
novelai is best way
was the pic nsfw or wut

It's gone
A mod went and like
Squished too hard it disappeared
if even you are aware you've had worse pics in here, I think that says something lol
Stocking
O.o wtf?
why did you offend her?
this me rn , not her 
haha
eh!? did the cat hurt you?
Stitch is not the same anymore
yes 
anime stitch
did he take responsibility?
and the cat hurt you?
no, she just got told no more cookies before dinner 😄
yeah! and who told her this?
me
This reminds me of how they put a treat on a dog's face and don't let him take it
here it is... the same thing
I want to be able to get star-shaped chocolate pieces when baking cookies!
yeah, i dont know why this model makes all the chocolate chips stars
this is not fair! My chocolate loses any shape when baking!
you can get chocolate that has a higher melting temperature so that doesn't happen, but it's usually really expensive and it tastes bad 😄
and sometimes glow on its own?
the second one is just super!
several types and general appearance of the character! bomb!
what tokens are responsible for this output?


