#🍥|anime
1 messages · Page 221 of 1
What’s has everyone found to be the best pixel upscaler?
nubby
Depends heavily on the task. If you mean for use in hires fix, then you don't really want the best upscaler. you want one that changes color, saturation, contrast, and brightness as little as possible while also leaving the image very slightly noisy because that noise helps the 2nd diffusion pass by giving it something to sort of bite onto and add more details. This one is my favorite for hires fix on anime style images. https://openmodeldb.info/models/4x-NMKD-YandereNeo-XL

People won't go near her to cut it
Silly stuff when she has such friendly looking companions
Have you had time to check out the spice and wolf remake?
Yeah, I haven't watched yesterday's episode yet, but I was up to date until then. It feels a lot like rewatching the original, and I consider that a good thing.
Yeah, I like it so far. The first arc is really picking up now. I hope they can adapt more than the last one.
Yeah, the place that the original ended was my only real complaint about it. I wanted more so I hope we get that this time around.
I bet someone has already made a lora or whatever new type of those there is for the show.
yes
more specifically (hair over eyes:1.9)

@green temple you dare skull emoji me

do you want the whole prompt or something

congrats on the baby
scary
"picking up slime girls in dungeons"
dark strawberry girl
😄
ok strawberries
well umm something went wrong
there thats better
time for work laters
Less crazy room
More thick latex

@nova remnant more 
I still can't get that eye shadow look

You always lurking
Do you gen anything cx
nothing that i could show here LMAO 
I know the feeling
When you get bad hands and stuff like that all the time
Better not post here
I'm gonna send you a friend request so we can fix them together 
exactly. 
How do you get a decent img2img output when generating buildings? Mine has artifacting and a lot of the lines that are meant to be straight, aren't
If they are on a small resolution that is very hard
For that you probably want controlnet keeping the image "straight" or increase the upscaling
I do 1536x1536 then 2x upscale but perhaps that isn't enough when there are details in the image
Maybe different type of controlnet. I usually use canny or anime lineart
if you're generating at 1536x1536 then that could be part of the problem, XL models are trained at 1024x1024 and 1.5 models are 512x512, possibly 768x768
but also, no matter what you do, you're not going to get 100% perfect lines, some stuff will need to be fixed with inpainting or something before the final upscale
the general idea is to generate an image that looks good, then upscale without changing too much
Yeah... 1024 is also too low res so I don't think that would work either. I use blender screenshot and using a toon filter sometimes looks better than the img2img. To have background and character and generate will need img2img and that will make it faster instead of picking and choosing. But would need manual work either way I guess.
1024 is fine, for the initial generation, upscaling comes later
generate at a resolution the model can actually do, then upscale it. generating at a resolution too large for the model isn't helping at all
Oh. I always thought it was better to input the max size
for example, i generate at 1024x1024 then, if i like it, i x2 upscale, if i want it for a wallpaper, i upscale further
in fact, more often with autismmix i generate 768x1024 (or 1024x768) because the compositions seems to be better
I want to make colored panels, composition with blender screenshots and 3d anime characters. Any recommendations for upscale workflow? It takes a while to upscale iteratively, I always just used esrgan anime 4x at the end of the processing
I'll have to try making the screenshots 1024x1024 and see how it turns out
the way i do it is:
Generate -> img2img (low denoise) x2 -> SD Upscale script x2
if i still want mor eresolution i use the SD upscaler again
this is an example of something generated at 1024x1024 originally
4K x 6K final upscale
lots of detail
this one is 6Kx8K, generated at 768x1024
Yes it's quite detailed. For me I think just minimizing the artifacts and reducing the need of more manual adjustments is the main thing. Perhaps upscaling will make those problems disappear
Like this for example, one is through a toon filter and the other is the filtered one+img2img:
you can get proper background on img2img with those imgs but to do that u need several controlnets + high denoising and XL controlnet sucks for these kind of stuff
I don't mind using 1.5. but problem with 1.5 is that in empty areas of plain color, it will try to fill it and it turns into weird swirls and stuff
try a lower denoise value with upscaling
High denoise with SDXL makes it change too many things. But it still isn't too bad at .5 and canny imo. Relatively
Ok I'll try
the higher it is, the more it changes from the original
Yes
it wont change if u use controlnet
i usually use 0.2-0.3 range (0.25 mostly) for my images, it changes just enough to clean up the upscale issues, but leaves it allone otherwise
I'll have to try that, yeah. That's probably better actually. SDXL denoise more than 1.5 when the value is the same I think
I use controlnet but it turns a solar panel into a mattress for example
if it's just for upscaling, controlnet isn't needed at all
well u cant fill the gray areas with stuff without a very high denoising
and u cant use high denoising without controlnet
unless i missed something, filling in the grey area isn't the goal
just a proper upscale
The gray area doesn't really matter but for example if the character is wearing a plain colored T-shirt and it's a close up shot then weird stuff happens there
I'll grab a few upscale workflows from the internet and test em out
Other issue is that 1.5 controlnet works well enough that it remains even facial expressions, features, and eye directions. SDXL... doesn't
because 1.5 has more controlnets like depth anything/marygold
I use only anime lineart controlnet as the main and it alone is enough lol. I think perhaps it's easier to make a more simple styled manga. More basic flat colors and 1.5. because it's simple colors, it's quicker and easier to edit
For SDXL, canny is good for the larger details but it fails more in the fine details
i say play with the diff controlnets,the strenght of it and combine multiple ones til u get the result u want
I have done testing on that for both, SDXL can't do it but I am happy with the 1.5 controlnet. It's just color that's the issue. But I did recently make a workflow that is pretty good fairly consistently. Its just simple output but it is good and fast. Maybe I'll start with that
Will at least minimize manual adjustments after generation
The img2img backgrounds of SDXL look better though:/ maybe generating different parts with different things and then overlay the layers and cut out the parts that aren't needed
to get the same color u need to use another controlnet
What controlnet controls color? I was looking for ways for control control
i think shuffle/recolor/reference work for color
I'll check those out. I think I tried one or something like it but must have been a whole back
you can apply it just fine, but the results aren't likely to be very good.
here must be water cat!
Nay! grab ONLY strawberry!

😂
Wow!!! it so so cute!!!
yep. a bit .
The super secret and cute funya's style
why?
hmm.. this is bad or good? not sure about anime style this is
Good
👍
why he evil?
yeah! grumpy! it means you're angry!
It's because he will dominate the world but he's there writing nyaas first Cx
So he got grumpy
red-haired dominant!? funny 😂
What model?
Pony
thanks
Now it's a game of "hide and seek"
i can't gen any more 😦
Why?
i need to change ventination of the vram
is not sufficent any more
even genning normal images, put my vram at 64 ° c
with time, it will literally cook it
Damn that's crazy
in a month or 2, i will change it with something better
what gpu do you have?
something old gtx 1650 4g vram
i can gen with xl a 1024x1024 in 15 min or little less
💀
and 1.5 in 15 20 sec

WHAT!!???
It's like 30 seconds for me
inevitably xl leaks on the cpu
20 sec on 1660 4Gb?
maybe he doesn't use many steps
yep, if i have 4gvram, how can it load a 6g model? Xd
20 sec ain't bad though
i learned many tricks
We can tell
then where is my video card deceiving me?
Limitation breeds creativity
playing with comfy or forge helps a lot
or in this case "Shit graphics cards breed workarounds"
1660 6Gb. in my case 1.5 gen still 2,5 min with HRfix
, oh well, i idn't consider upscaling
I have a hiresfix and a facefix and it gens in about 30 seconds
of couse on 1660 6gb?
the upsscaling requires 2 or 3 more minutes
Lmao no, I have a 3070
It'd probably be excruciating on a 1660
usually with xl, i play a lot with turbo
now it true
on me, i'm saying
with XL model it was 15 min - yes
but i don't maintain the same steps of before highres fix
me too
yes
Why is the genspeed jump so big?
mine only goes from 30 to 40 seconds between 1.5 and xl
is it vram limitations?
yep
Figured
when there is not enough vram, the model loads on cpu
I did that with an LLM and bluescreened
SD3 is going to be fun
but I'll still collect on 3070! there's still a little bit left
So the biggest model will be 8gb?
i don't remember
8 gb of vram
yes! 12Gb
This isn't a 3090 lmao
oh ok
neither
yep.
If you don't mind it being used you can get them for less than 300 on ebay
useg gc are not the best
but now I'm just patiently waiting for the generations to finish!
they are the fastest hardware to burn
Have you guys tried tiled vae to reduce vram usage?
for me 848*1280 with hires uses vram about half out of 12GB with tiled vae
when you use xl, tiled vae is not an option, as the minimal res supported is 768 and the results are still meh
with 1.5, the story is very different
i use a tile of 688, max before leaking and degrading the image
I use in xl I don't see any differences, even the gen time is about the same
that't one of the tricks for example
with 12gb you don't need a tile, so, it just gen in a single tile
upscale in img2img, if I don't use tiled vae I get CUDA out of memory
this error hasn't happened for a long time
well, if you upscale 848*1280 x2 or x4, it's obvious
unless of course the resolution is sky-high
oh, didn't notice the hires
x1.5 only
i honestly don't know the limit of 12gb
maybe splits the tile in 2?
so, instead of genning in 10 steps, you gen in 20
not much of a difference on 12gvram
yeah, there is some leak in cpu
but it's minimal
wow, disillusion is my death >.<
This video is for non-profit purposes, we do not claim any rights to this audio/video.
Disclaimer: "Notwithstanding the provisions of section 106, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecord or by any other means specified by that section, for purposes such as criticism, comment, news reporti...
i so damn love this series
and his songs
i should learn this song on piano, but it's so hard >.<
maybe making some sort of jazz, i would get a good results 😱
that reminds me ai can gen songs now as well
trap is not music!
https://soundcloud.com/mikebedder/mr-blue-sky
today my fav..
A cover of E.L.O's Mr Blue Sky.
Worked on this over the weekend of the 31st august 2012 to 2nd of september 2012
I recorded all the parts myself, im happy with the outcome and hope all you E.L.O fan
the guys of this song are so dumb XD
but as the bear Baloo said - “oh, what a rhythm!”
tried to upscale without tiled vae
yeah, but the intonation oh god, that intonation, i can't take it seriously XD

damn, this numbers lol
I want as much memory as indicated at the end of the message!
good idea, I thought you do by photoshop
i do that too, but i correct the smuggles left behind with an img2img
perfectly correcting high res images is hard
gotta go
see ya
did she wake up and go out?
hmm what? sorry brain not grocking
""she fell asleep"" first img and second img - "did she wake up?"
Well, she just switches to Chibi mode when she sleeps🥴
umm ok
hmm doesn't look right... almost
cackles
that looks odd
switches model and vae and tries again
thats better
iron WOman?
nah more of an explorer
random extra leg, but i can edit it out
ohhh it might be my 'legs' tag from when i was getting tails.
strawberries again 😉
it's hard to get consistent hair colors with gradients 😭
again!
ok that interesting
she went into the night!
remind me of this , but yours better 🙃
The author of the model loves strawberries very much!
meh
cool
hinata always cool 
oh.. hinata
now i think better
🤣❤️
based on this is?
yes
nice
yep!
😍
I wish I used a platform that allows for simple post processing of images
I always want more saturation
For instance :
Okay, maybe I overdid it on this one but you get the point
Anyway, heaven and hell version 
Don't count 
Lazy mouse 
based on paint sketch?
No, but it's not very reliable, carrying can have a lot of fails
just a pretty difficult pose
wow !!!!
I was trying to do piggyback initially but apparently it doesn't work so I got that instead
more
img2img of piggyback instead?
rena from higurashi
Grabby

@finite herald have a small sample of my collection of miku's i've been making while testing artist response on my model merge ^^
gf 
Want some of the whackier outcomes too?
Impasto/Pseudo impasto test;
i wish sophie got the anime instead of ryza ;_;
Whats good inpainting for anime
had to screenshot it -> didn't save those images seperately lol
I think from the number of Miku

Listen to Asobi ni Iku yo -SPACE MAN by Gabryel Cote #np on #SoundCloud

I listened to it
I give it a hoshi/10
Understood nothing of the lyrics
👍
Now bath time

😻
we need more 🍓

why two bells in the tray?🤣
To fill your bell-y

@cerulean bane
Go dm me tofu

now you have my attention.
no bell! I want🍓
well, I'll give you some disguised by bells


realistic anime?
look nice, what model you using?
bottom right 2= animagine xl 3.1, rest = naidv3
shinx is so kyuuuuute
thanks 🥰
no offense but wtf is realistic anime????
idk man
.. is that a trick question?
You either have realistic or you have anime
It's also called Live action adaption of anime style
Oops thought it was about style
usually involves CGI of some kind
but in this case it's all cgi
tho frankly when I do it it's not total realistic
You know I'll get kinda mad if AI people try to blend realistic and anime styles
Because the result is very mid and then they post it on reddit resulting in more ai haters
you mind the haters?
shrugs. doesn't really care, I just make stuff, like it, leave it, wateva
The average people don't care if art is ai
thats a valid point but you mind people hating on ai?
Why they start banning it is because they look mmmmmmmmid
I mind if an average Joe hates ai
because that means for sure that it sucked you get what I mean?
why? don't waste your brain power. doesn't matter
artists hate ai yeah ofc nobodu cares
my point is to understand why people hate it, and to also not mind people that hate it, its their choice to not like it because its sometimes unethical,
The thing if they hate and actively shit on an image, big chances are I'll probably hate it too
the reason is not only because an image is bad.
again.. doesn't matter, move on
Right do whatever you want just don't post it on fandoms or something
don't feed the universe more bad energy and thoughts , not worth it
Keep it to yourself and the community
I just wanna ask, why are you so mad about this topic
Because when I browse art online I get really mad when a piece looks OBVIOUSLY ai generated
okay, anyways how do you do those weird stickmans
.. and why should it matter.. don't answer that.. cause it doesn't
and make it into a person
It's rubbish it's bad it hurts my eyes
somethin like this
sounds like a you problem
Not just a me problem, a good portion of the world has this problem
that's a tool called control net that helps with posing
sighs... and yeah i'm gonna quit feeding you more of my energy you are not worth it, blocked and bye bye bye , hope ya have a better day
thanks
ahh quieter already
Look sure you wanna post sure go ahead, you are just burning people's eyes if your art looks bad (most of the stuff seen here is nice post that)
reinstalls FO4 in prep for update
u can always post your lewd stuff on pixiv (except if u are american/british) they are banned from posting 18 content by their gov
lol
not even touching that one
gets idea's for art and grinz evilly
ai creating ai art using ai art
Wtf are you typing
Please get help
From trusted professionals that can fix your mental condition
it's not generating number on the arm
How bad is this
ok FO4 installed so now i can make real art instead of cheating with bing create
this time on the leg, I used arm tattoo🤣
nice
cackles , sets extra detail lora to 3
huh wonder what happens if I set it to neg
tries it
well.. ok then
oh no
depending on the model you may need to add arm tattoo
actually I did, but it appeared on the thigh
need some luck to get it the right location
no, it's Kohaku
👍
decides to create before bed.. ponders
checks for idea's
hmms
tells furry side of brain to be quiet
default settings
Miku - the captain A?
miss marval auto bot america
or something
brain winding down
hehe
strawberry!!!!
now thats art
colorful
What I kind looking girl, hope nothing bad happens
That looks very cool
thx
you have died of dysentery
looks like from that abyss anime.. can't think of name
backup has arrived!
as always, something cute, ugly, kawaii, fluffy... with ears and tails... in a school uniform, a maid outfit, lolita bows... or in full metal armor.. 🤣
ok thats trippy
And with... Grabby thighs for me
👍
😉
I brought your opponent

@brittle rampart he's a big and strong warrior
and when they gather in a detachment, they are just wild animals!
if they play good music, then with pleasure
smol?
smol!
As smol as funya
🫣 💯
what happed?
It's just crashing I dunno
Just have to wait a few minutes for it to come back
It's not on my PC 
Smol ears this time
yep!
dangerous rabbit
still would! try hugging and died!
smol needs big ears
This one is safe
true!
I understood! This is a strawberry conspiracy against me!
"From above", "From below" crap! how to remember this?! 🥴
I do those all the time
Cute fox temple
We need to visit
Or cat
I mix them up

Coot?!
#🏞|general-with-images a sinple illustration with plain colour of a rose garden, not too dense,

We can't do it that fast!
#🏞|general-with-images In Pixar's 3D style, seven monkeys are hanging upside down from a tree at dusk, one monkey after another. The last monkey reaches out to scoop up the moon reflected in the lake with its hand.
@serene cosmos @grim kestrel We can't do it that fast!
yep. as it was ordered. I really didn’t get 7 monkeys
It looks like reflection on the water and at the same time from below looking at the clouds
I'm going crazy
yes. it look like a reflection
They are going to steal all strawberries

I used controlnet
fast PS?
huh?
I mean this was put together from three pictures in Photoshop?
nope
ai generated
added , comic frames overlay
tho I do have a Manga creation program
token in prompt?
yep
hmmm.. must try it...
epic
Mecha
huh... all I did was put lens flare in neg. same prompt and seed. wtf?!
goddess fall down on earth!
the queen has arrived
"I'll adopt you"
yea went random strawbery
strawberry Queen , lora:more_details:.8 more detail (Fine glitch art:.5), Cel Shaded, Comic Book , red-pink strawberry overlay, sci-fi strawberry, apocalyptical (red-pink fog overlay) , theme: Fruit scene
there took off style and addedd strawbery red after apocalyptical
takes out comic book
there
no just regular
model also regular?
😂
attack of the strawberries
replaces strawberry with lemon
oh and art by HR Giger, vivid gothic pastel colors
horror again
🥻
what model? pls 🤞
what model pls?
naidv3

Some c-word 
I have yet to try just throwing tags at sd3

Sd3 bikini armor 
I promoted full armor 
It doesn't seem to understand tags well, cowboy shot is a very popular one and doesn't work
Im gonna cut dat leash
are you using the api or running it locally?
Only the api is available atm
some people got the model
probably the guys who are making it ? 

there was a page to sign up to get it early for developers and stuff iirc
I mean they probably got the api too
odd style experiment
No joking
Looks like when I was screwing around with "low quality" on the positive prompt on 1.5

Olrite then 
Frfr
I know I can make ugly stuff sometimes. I think I already know how to make pretty stuff so most of the fun comes from experimenting with weird settings these days
1.5 🗿
thanks 😁 mind sharing prompt? 😬
Without the pony bullshit it's :
black skirt, thigh highs, (blue wings, red wings), neon, arched back, art nouveau, frame, bows, ribbons, bracelet, sitting, throne, skulls, fire, halo, heels
miharu, 1girl, solo, pilot suit, gloves, looking back, standing, from behind, looking at viewer, moon, starry sky, 1990s_/(style), <lora:miharu_V3:0.75>
SD3 result could be interesting, depending on whether it does art nouveau well or not ? Might wanna add some anime tags too
Or maybe it wont understand much of it anyway since it's not made for tag prompting 
Its not very good at tags
@normal nova
I tried just throwing tags at it with "In modern anime style"
I just always do that it makes more anime stuff cx
pretty decent
Have to admit I'm kind of let down by the anatomy in sd3
Will try making some anime stuff with it
It feels like they went all in on prompt adherence and ability to make text
I guess prompt adherence will prove useful, so that's nice, especially if quality can be added in by the community. Not entirely convinced text is worth the hassle but what do I know.
However as the models grow larger and more advanced a drop in community work is to be expected imo. It already cost 50k to train pony on xl, I imagine it'd be even more for a "pony 3"
I don't think it will be much worse to train on
even if its larger
the technology also getting better optimized and stuff
also astra, the pony creator has more support now to train the next model
so thats good
SD3 attempt
Like the anatomy is so discouraging. I know I should be grateful because it's open source and stuff
It just doesn't seem like the DALLE / MJ killer it was touted as
nyo, they just got access to a discord bot in a private server that runs it. nobody outside of Stability has been given the weights yet.
Woooooooooooow !!!
What model ?
Is crowdfunding a model going to be a thing? 🤔
It already is in a way. There are several groups and/or individuals that fund their training through patreon or some other donation platform.
😔

I like this style…. Could you share the prompt?🙏
Sure ! Prompt :
score_9, score_8_up, score_7_up, otic, yew, bomm,
1girl, futuristic, wires, tokyo, neon lights, shoulder pads, blue hair, pale skin, mechanical, robotic armor, mecha, plugsuit, bust portrait, solo focus, sunset, smoke
negative:
score_6_up, score_5_up, score_4_up,
mgfx, zvu, cigarettes
on ponyXL at scale 6
with Vintage 1990s lora at 0.3 and a retroscifi lora from a friend at 0.8
idk if this counts as anime 
Thank you, been looking for this kind art style 
catagod!
magicat
just a romantic catagirl 😁
although I requested an android cat
I'll give her something delicious
catpizza

not cataroni!
no cats were harmed in the making of this pizza
maybe because they were already in the form of salami?😱
Don't do it
he cute
sometimes I get the impression that after a restart, YV needs a long time to persuade to do something worthwhile
maybe
just maybe
we are going crazy


I just have the impression that he himself learns from his own generations. The more you generate in session the better the result SD produces.

@nova remnant last cool!
space rabbit
The lonely space rabbit

I like this view the best
What model?
completely agree
Fox not for sale


how long for an AI to transfer this video to an anime ? https://www.youtube.com/shorts/UpttYjF9kcw
#viral #nghile #nghile30_41_5 #vanphonggenz #dilamcogivui #gogovnz
why?
||Asian illustration, round face, (A male programmer in a checked shirt with a computer in his hand, )white background, hd mod, glowwave, simple, clays, 3d, blender, oc render, in by pop mart, blind box toy, c4d, Disney Style, Matte||
👍💯 😉
why cann't use "dream"anymore?
because there's no bots

Grabby 
How can i solve it
spend money
soga
or, if you have a good gpu, just run it locally like the rest of us
ok. thanks 
have a money - have a GPU! have a GPU - have a gen! Have a gen - be happy, dont wory!
😁
that is life!
just fiddling with some idea's I had while my brain woke me up with them. (mutters bastich adhd)
Prompt: Illustration vector fine detail, (shadow behind:0.2), Miku Bio-Roid, cyber theme, city canyon, street view, (wide angle perspective:1), hi-res, three tonal color shading, soft colors, black outline style, elegant, very detailed Cel-Shaded With elegant shaded detail's, soft shadows, illustration, Midjourney style, hypernet:mjv4Hypernetwork_v1:0.8octane render : Steps: 60, Sampler: DPM++ 2M SDE SGMUniform, CFG scale: 6
everything cost money now. SD,mj,runway too
that's good. Someone really likes this
I do
Here comes the grabby comment when I like something
Grabby 
Did you controlnet for the pose to stay the same cx
Not really. Everyone here hates me!
of course.
shut fk up girl u are great. everyone is great
Put inception in your prompt
Don't know maybe it goes something
