#🍥|anime
1 messages · Page 175 of 1
That is the point i tried making Itadori Yuji form JJK as a curse but it didn't turn out as i would have loved but thank you glad you like it
guys... who know what it that and how working with it?
i am see it all time after restart A1111 UI
Guys, tell me how to edit metadata in png using regular Windows tools
I really need to start using wider aspect ratios more often
why?
that's highly subjective, but different aspect ratios can change what else is in the image besides the main focus
The main thing is that in the end this main focus is not lost.
please check my model out https://civitai.com/models/170390
PLEASE AFTER DOWNLOADING UPLOAD ATLEAST ONE IMAGE plis T-T THE PICTURE I USED WAS MADE IN COMFY UI , IF YOU WANT TO GET THE PROMPT JUST DRAG AND DR...
oki-doki!
thanksssss T-T
@brittle rampart it would really help if you could dm your reactions but its fine if you can't tho T-T
what is the easiest way to train an anime model?
cute style
charming
she is a ghost?
It's super cute I love it
yeah 🤣
Diabetes inducing
looks like blessed vae
just f8-anime2 🙂
never used it
what's 1 novel view? a 2d image with 3d colors?
view from top,front,back and sides in a single image i believe
that could be really useful
too bad i already have my hands full to test even on it
food time
well enough, depending on the definition of well? 🙂
evenin'
I'm trying to fix hands via inpaint but each gen looks bad. Do I have to keep trying like a hundred times or something?
inpainting hands without fixing them first or at least adding a depth map is usually a hopeless process.
If they were fixed, then why inpaint them at all? And when you say depth map, I assume controlnet?
fix as well as you can then inpaint at low denoising to blend them into the style of the model better and make up for lacking art skills. also, yes by depth map I mean with controlnet.
Ah I see. I'm using the PS SD plug-in and it has a controlnet inpaint option but I'm not sure how to use it. I'll have to research I guess
Yeah, I'm not sure on that. I use the Krita plugin which sadly doesn't support controlnet.
If you ever need to know how to get the "free" PS version, you know who to call 😉
idk if it's a good idea to say that in a server like this one
I won't say it here
Not really new
Turn outfit to santa costume and the gems to ordainments pls for science
and... merging again?
Making the crystals into ornaments is going to take some manual edits and I'm away from my PC for the night so here's Flandre in a Santa hat for now
wow, i really wanted to do something like this
but prompting a character laying properly on a bed is hard
expecially embrancing something
Really?
i couldn't do it properly 😦
i couldn't get a proper bad
cute XD
the surface my character were laying, was always bare and plain
it was unnatural
maybe try some thing that (sleeping in bed in the fetal position:1.2),(hugs:1.25 an electric guitar),
let's try it out
oki-doki
last masterpiece,best quality,absurdres,detailed background,perfect correct anatomy,portrain:1.35,1girl,[solo:medium:1.25],a young preteen (red foxgirl:1.5 sleeping in the fetal position:1.25),([hugs the planet earth:medium:1.45] in space:1.3),serious,severe eyes,fluffy foxtail,space,stars,galaxy,nightie,(brilliant red:1.2 hair),long hair,wild hair,pastel colors,soft shadow,chromatic aberration,ambient oclusion,cimenatic light, and some LoRa`s
spring, sakura, and ..... Fuji
epic shot, anime, dramatic cherry blossom, official art, portrait, flat shading, wind, sideways glance, looking at viewer, , kind eyes, sensual smile
I would also add petals in the wind
yeah, it should have that ^^
POOF! instant improvement
what a shame, i can't post this 
What's happened?!
young girl on night dress with too prosperous breast
heh! this will be appreciated by many here
show it, don't be tormented!
you do it!
yep 
it's so charming, it's warming my soul 💖
ai can be so charming
coadaptor repth is incredible
pretty sure if i add even concat it will became even better
I'll be happy to see the results
me too 
i love playing with ai
but i have so many things to do
i wonder if i should just to use lcm. end result are way different, sometimes less cute, but way faster
taming it and with the use of lora. maybe i can gen better results
LCM? what is that?
a sampler
they made a lora that let you gen something with steps as low as 5-8
at max 16 steps you get very good images
LMS Karras?!
after 16 is rarely worth adding more
or LMS
is promising, when tamed
ah... Comfy....
if i'm not wrong, you can use it on auto1111 too
if you don't have the sample, maybe you need an extension
load the lora, drop cfg between 1-2 and steps at 16
at first, try max 8 steps
I think you can, since last release, else you'd have to install animatediff extension, they have the sampler built in
oh ok
it's just insane how many things you can do with comfy
his community is fantastic
so many deloper creating in tthe name of art
or porn 😮💨
i suppose so
'science'
oh, that's that too, right
Got pixart working on comfy too, that model is also very fun to work with
But it's just not as well trained or finetuned
oh... abstract art
absolutely too much
With comfy you can load it in cpu ram, which leaves only the inference model in vram
So it's not all bad
But you do need some normal ram to do it then
i don't have 20g of ram
it's even more otragious
i never expected to need that much ram, i'm not managing a server
@steady grail LCM - lost details! i am love a many details!
yes lcm has less details,u trade details for speed
lcm need to be tamed, that sampler is more focused on the prompt
then it's not my choice!
it works great if u like creating images in the flat color style
i'm generating with lcm, but for now without lora, because my workflow is still incomplete and very chaotic
oki-doki....
two fuji! fantastic!
the other is his lost brother raiji
I think the Japanese would have kicked us by now!
uahahaha
ok, my puns deserve this and more

i'm the nightmare of every social group
you are on a roll today!
eheheh
just to say i wanted to rest
i really don't like result from lcm
i just leave it for the pose gen
i suppose that my gens today reflect my desire to just rest lol
good night then
i'm generating, i'll stop when i'm dead ❤️🔥
how to generate a picture?
this kinda looks cool
using a prompt
damn, looks rich
the light effect looks pretty neat
rawwr
thanks
She's scary when she's angry!
well i know that, but in this discord plate, how can i draw a picture via prompts?
@brittle rampart
i am newbie
I've never asked this question here. I use local Automatic 1111
@brittle rampartok, I know that, thanks a lot!
by the way, the cartoon girl is really charming
thanks. she was born to be charming!
your name is rare, a little bizarre.
source?
m?
kpop girls look like anime cuties
yes
true!
I found an ingenious way to correct crooked hands and fingers when generating! Draw big floofy cat gloves - paws!
how?
lol
thats not gonna help with cute girls
really #🍥|anime message
since we spaming non anime girls
cute
nawwwww not cute
no forgiveness for you!
eww dude
obsessed over obama?
with trump hair too?
this is the end sign
the world needs more cute pics
guys.... Tell me, is it possible to change PNG Info using standard Windows tools?
hate this is!
not sure you can edit metadata like that
i dunno lol, i didnt think it was possible to do that
just edit the img in paint and then save it again and it will remove all the sd tags
but thats not editing
it will remove all the sd tags not good idea
thats the best part
only one word!
Чёрт!
render the same image with same seed and edit out the tag
it change the image
don't generate with that word to begin with, use young girl or something
only tools I know of off hand remove the metadata completely, I've never looked into editing it
this is was gothic lolita dress
so they block lolita keyword on civitai?
yep
so called western experts when they find out about lolita fashion
sounds odd but i think it could be due to depicting minors
but tbh most anime girls look minor too
specially the ones with 4 arms,tails or animal ears
depicting minors with certain poses could be the cause i guess
i wonder if it would catch it if you used gothic_lolita instead
they dont like girls with mouse ears,its offensive to mice
ExifTool gui is better but i dont remember how to install it
yea that one
pretty sure you just download the .exe files for it and exiftool and put them in the same folder somewhere
this is for A1111 UI extention?
ah
although, I'm not sure if it will do what you want, i'm poking around in it right now
Well, I’ve already achieved results with the program I showed you. so thank you very much for the moral support!
by the way. It’s quite strange why the A1111 doesn’t have such functionality.
lol clever.. 👍
for anyone else interested, exiftoolgui can do it, it's just a weird interface
The legendary ban hammer
omg mods
nah jk
someone should use the whole novel of lolita as a prompt
Damn. Looks like someone fell for it in under 2 minutes 
what does it do?
😔
It steals your discord and probably steam account too.
not with a single click on the steamcommunity domain it doesn't
certainly it requires further data entry or clicks
it asks you to insert you steam data
ah
facking a steam login
jerks
if the same data are used on the mail, they log in your mail and steal accounts associated with it
pretty scummy, yes
Yep. It's a common scam in many different circles. They often also ask for 2fa tokens, just like the real site would so even that doesn't protect your account.
much more fun to just generate images of anime characters eating cheeseburgers
ok returning afk, see ya 😴
there are a lot of factors in gen speed, the biggest is your GPU and drivers, SD vs SDXL, various optimization settings depending on how much vram you have, etc
first thing I would check is if your vram is maxing out and overflowing into system ram, if it is then that will cause a major slow down
as an example, an image that normally takes me ~15 seconds to generate will take almost 2 minutes if i go over my vram amount, it's a massive speed change
it would be even more severe on a slower system/gpu
the version of cuda, torch, and xformers could also be a factor, but unless you're running a fairly outdated and not updated version of a1111/comfy it should be updated enough to not really matter much
version: v1.7.0 • python: 3.10.10 • torch: 2.0.1+cu118 • xformers: 0.0.20
This is current for a1111
or rather, what I have for a clean install of 1.7.0
Trying to make her eat the burger results in horrifying results, so I'm settling with holding a burger instead 😄
those are the versions that got installed when i installed a1111, i didn't do anything special
I did a clean install for 1.7.0 and left my 1.6.0 folder alone
but, i think it's the same versions for everything else
latest gpu drivers should improve speeds, I haven't heard of any recent driver issues at least
yeah
that's what I use
NOTABLE LOOT: 25 DIV VENTOR'S GAMBLE
😄
we decided to use it ourselves
next map -> covenant, badge, divinarius, 2 raw div
2 maps of total suck, 3rd map -> 2k purple and 2k blue essence
this league is SOOOO rewarding
I've been surviving off of id'd rares from the ground
I found a 8d ring not sure if self for that price

Weren't you mfing
we weren't
right now i'm a duo -> a mana cursebot and hexblast miner
he removes my hexes all the time, but then again, i grant him tanky stats
fixed his build so he can run that ventor's, 10 quant + 55 rarity + 100% ele resist in total
the 3rd man of the party is missing tho 😦
What gpu are you using?
1070
seems about right
with a 4090 i get around 6it/s with sdxl, hires passes are 1-2it/s
oh, so the 1024x1024 is from hires fix on a 512x512 gen?
What commandline arguments do you have
try something like hires 1.5x instead of 2x
my guess is you're running out of vram and it's dumping it into your system ram and slowing it down
no
Do you have xformers and medvram
I'm pretty sure you can have it
I tried comfy but didn't work for me
It should be normalvram on comfy I believe
I tried on a Linux thing, and there I could start it with normal vram
Bath time
brb
based on a quick google, medvram is default in comfy
supposedly comfy will auto set those settings based on your GPU as long as you're using the default checkpoint loader node
did you try using a smaller hires scale?
i would try that first, would make it easier to confirm where the problem is
if 1.5x is still that slow, try 1.4x, 1.3x, etc keep going smaller until it's faster or hits 1.0x
evenin'
looks like a turkey burger 😛
AI-eye is a gift! Do not squander it, our new gods have blessed you!
What is this? The daughter of Hulk and Vision?
thanos girl 😄
more nightmare fuel
X23?
i prompted 'wolverine girl' but all the ones with proper claws I can't post here, and I didnt bother to fix these ones
Tried to give him a little anime flair.
You know what would be nightmare fuel? The child of Superman and Violator (from Spawn)
not the movie version
that one is just hilarious
they're working on finally rebooting Spawn as a movie in 2025
should be cool
how can people actually think Dwayne Johnson is natural and doesn't use steroids or at the very least TRT?
cause he has like a body not even possible for a man in his prime of testosterone levels and working out for 10 years with the best possible genetics
Hollywood has warped peoples image of what men and women look like in real life
oh there are those who believe that
they believe him when he says he's natural and doesn't take anything
when he suddenly at the age of 50 gains more muscle than he has in his whole life in just a year
like...I know they might not know a lot about steroids, but all I can think about is Dwayne "The Tren" Johnson
hes 100% natty,vegetarian,only drinks pure milk and eats lots of eggs,great genetics
he actually has said that? lol
he's not vegetarian though 😛 He says "chicken and rice bruv"
yeah he has said he's natural
and that he's never taken steroids in his entire life except once when he was 18 to try it out and then never taking it again
oh yes
Mike's another total natty
just like Chris Hemsworth
the only actual natural actor with muscle is Henry Cavill
he's worked out since he was like 18 or 20. He never had a sudden spike in muscle mass. Only time he looked bigger was when he put on fat and got bigger arms cause of it 😛
but looking at Chris Hemsworth sudden muscle spike for Thor Love and Thunder, it's like...dude, that doesn't happen
guy is 40
they probably told him it was flintstone vitamins
of course
but I can't blame the actors for taking steroids
it's like "Ok you're gonna make 20 million dollars, but we need you to put on some muscle. Can you do that?"
"No I don't want those 20 million bad enough to take testosterone"
probably destroys your liver or something
cause they have access to doctors that would actually prescribe it and give them health checkups
nah not with the stuff they have access to
they have access to protective medication as well as pharmacological level testosterone
so they can do it as safe as possible
and testosterone, even at supralevels, won't fuck up your liver unless you already have like borderline cirosis or some weak liver. Or receive some tainted testosterone, or do it without protective medication
and they get doctors to give them the test and stuff so, if anyone can do it safer it's them. That's not to say they're completely safe. There's always risks, but they probably think 20 million dollars makes it worth the risk. I would
supergirl hyped up on steroids
then I am 1 million percent sure that the result will not be missed 😂
snap there goes cali
Are the authors in a hurry to make a model quickly without worrying about quality? so to speak, on the crest of a wave of popularity of the format
i think the sdxl doesn't make it easy for authors to train models cause sdxl is resource heavy
and this too.
its been a few months since sdxl release and you will see there are lot more newer version of sd1.5 models coming out than sdxl models
What do you think influences this? complexity? resource intensity? unwillingness to lose popularity sd1.5 models?
money
its cause sdxl isn't as accessible as sd1.5, in terms of both training and usage
of course
Solas.. in my case I can’t even afford to think about spending several thousand to replace my computer.
1000-3000 euro thats crazzy! in my case
yeah we all feel that way
technology that are coming out are not in alignment with consumer buying power
rather the strong influence of "mining" over the past few years.
because of this, prices for video cards and hard drives have skyrocketed
but the demand is not the actual problem, its the engineered scarcity this economy creates
Perhaps you are right, I am very weak and far from economic issues.
difference between a girl in a jungle and a girl walking through a jungle.
yeah... walking through
if you add closeup with walking through a jungle it will create an interesting visual effect
example
closeup, a girl walking through a jungle this is ?interesting...
i think you'd have to mention the path she is walking on as well
like rocky road / path etc
im playing with nubby's model
a girl walking through a jungle at night. (closeup:1.2).. the path is filled with vines and rock. the visual effect like this one .....
I played with her yesterday as Miku
1girl. pond
btw, . or ,
wouldnt matter
really?! I always thought it mattered in the prompt....
nah not really, the ui takes the words and interprets them
the sentence structure can have effect with or w/o , or .
i must remember it!
Is this AI? HARD TO TELL
How did you do it💀 near damn perfect 
me?
No but i love your miku 
then use the reply feature
Yea
Sounds new 
I haven't been on Sd for a few months now
Whats the reply feature?
And how to use it
you literally used it earlier
Making it my lockscreen
I'm talking Discord btw

Ahhh my bad my bad
you sent this 20 seconds after I sent my image without reply so I got confused
Yee my Fault 
It does matter. A period is said to be better at keeping ideas separate from one another, particularly when using SDXL.
What about ,
,
Commas work just fine, but theoretically to a lesser extent--at least with SDXL.
yeah i was saying the comma's wouldn't matter much with simplest of prompts
Technically the best way to prompt and save tokens is space and dashes instead of commas and space
hmmm
Is SDXL the model you use or stable diffusion?
This isn't a matter of saving token space, but rather best prompting practices in helping CLIP to understand what exactly you're trying to do. Natural language processing, as it were.
I didn't say it was only to save tokens tho
exactly
idk if I should argue
When I say SDXL, I'm referring to the latest Stable Diffusion generation of models that are out there. The ones that are made to work at 1024x1024 base-level generations, and different ratios of a similar pixel size.
since sdxl is more relevant to context using contextual sentences with , and . would make a difference but prompts that are 3 - 5 words won't have much complexity
The best way to prompt is by hijacking in my brain
Well, anyway, I should just check it on the same prompt and SD1.5. Which is better, a comma or a period?
SDXL often works best with sentence-based natural language prompting anyway. Though anime models are a different story, depending on the creators' training and model merges.
With 1.5 it's going to matter a lot less. Either/or should work just fine.
Commas are more popular there.
Ah my bad ty for the info though I'm still teaching a break from SD
yep, i find the use of comma meaningful to structurize an idea
but sometimes for the sake of lazy fun i use . to separate a couple or 3 words lol
Does anybody have a high end PC?
Define high end, lol.
What comes to mind when i hear High end is expensive technology
guys, you have completely confused me!
I have what I'd consider to be a high-end PC within a non-AI-centered environment. But AI "high-end" is a scale.
In short, with 1.5 it doesn't matter so much.
Choose what you want.
oki-doki
to avoid confusion, i have the obvious answer for it, speak to SD in language pattern 🙂
but even then... this is what i used in that image 1girl, rocky terrain, sunset.
Ah yes. Midjourney prompting, lol.
you know lol i never used MJ except for a very few try on discord bot
Same. But that's just about the level of prompting you need for a "great" end resulting image.
ahh
i kinda learned prompting by watching people on sd
though it started with chatgpt lol
didnt have sd then
language pattern -and now, as for a girl, can this be explained?
I certainly didn't understand it, lol.
sure, just speak the mind :p
If I can translate, perhaps you simply mean "tell it what you want"?
yes lol
i stopped using lcm lora
lcm lora is neat for references but for the actual image quality you wouldn't wanna settle with lcm lora
If only more people would realize this.
lcm less details
Though it's usually good enough for a base-level gen if you're upscaling it after that without LCM.
yeah that too
interesting method
It's not practical in my case, as the time it saves really is only a couple seconds, and I'd rather have the little bit more cohesion one gets from genning with a 2M Karras sampler at 20 steps.
WHY not a 5 fingers!!!!! damn!!!!!
im back to using 30 steps 🙂
With which sampler?
mostly dpm++ 2m sde karras
It's said that exponential instead of karras works best with 2M SDE/3M SDE.
didnt know that, i should try it out
the difference would be strongly visible with higher res
the image tone i think has a visible difference from sde karras
finally!!! On the 62nd attempt to fix right hand, I persuaded SD!
62 attempt? that's determination
yep!
good job... i would give up if i had to try more than 5 times
did you inpaint?
yes
ahh... i hardly ever inpainted to fix hands
i delete a lot of the images lol
radically!
Anime Illustration Diffusion version 052 and the new 06 comparison. Long story short, slightly better prompt following, but a lot worse cohesion. Not worth the upgrade. https://civitai.com/models/124189
Model Introduction (英文部分) I Contents In this introduction, you'll learn about: Model information (see Section II ); Instructions for use (see Secti...
in childhood the Demon King was a sweet, kind girl... and then bad companies... forest robbers....
the struggles of money corrupts
It's all the bad influence of adult things!
not that love. that adult thing is the persuasion of money.
money is evil!
there is a saying 'money is the root of all evil'
but as for necessity money is not to be confused with abundance of intrinsic resources
and here! how correct is it to use constructions of this type [mystical dragon:medium:1.25] or constructions of this type (mystical dragon1.25)
I think this is for a more accurate follow. Well, a number for strength. in my understanding of this.
but I'm not completely sure
so, you can set low and high?
something like
i think soft, medium or hard
but not sure
why again not correct hand!
grrrrrr!
are you using any embedding for hands ?
for positive lyco:GoodHands-beta2:1 and for negative : negative_hand-neg,bad-hands-5
maybe conflicting embeddings and lora are conflicting with each other
of course
maybe
try using the embeddings in negative box only and see if that improves w/o lora
oki- doki...
or just impaint
again try 62 steps or more?! its sadistic!
62 steps are just too much, you will get relatively not differences
I mean the number of attempts to fix
ohhh
as in the previous illustration
you can try with manual editing
or adding correct hands on the image, lineart controlnet and another try
good hands are hard to elaborate. they are too small, too different from each other and not particurally trained
regarding controlNet. With the latest versions, when using it, my generation is slow as hell.
auto, right?
i moved on comfy, because my mid low end hardware was too slow of auto
ok, gotta go
hmmm.. for me 2-3 minutes on generation not a prombem now. but with controlNet now it 20 min!
last suggestion, try to play just with lora, embedding are very maginar with results
with 4gvram in 20min i gen a 2024x2024 image
yeahh, auto, is easier, but not very performing
more better
I WANT REALLY BIG DRAGON!!!!
who know what is the difference between these models?
8s is lot stronger
What does this mean?
the default settings need to be adjusted to lower settings for it, otherwise that might have distorted effect
try it out
the inpaint denoising strength might have to be lowerd from 0.4 to 0.2
oki
Crap! I want a dragon but.... I get are elf hugs!
interesting.... don't their ears bother them?
hmmm.... that's where I'm wrong?!!?a young preteen child elf (hugging of big dragon:1.35) in the (deep mystical fantasy forest:1.2)
where my big dragon!?1
The asset
even her evil characters came out quite cute
heeelp... neeed help! how to make a magic circle shine?!?
novelai just goes and makes the best image models and then doesn't wanna freaking share!!! 😭
m?
they added nai diffusion v3 and it's literally better than every anime style SDXL model available
everything will melt over time
as it was originally with novelai
Doubt anything will come close to nai v3 unfortunately
It's the best anime model for now
until they release v4 that is lol
let's wait...
i think the only reason community models surpassed nai v1 was because of reverse engineering of [REDACTED] and the ability to use extra networks
the period of earning money will pass and it will become available to the masses
nai v3 can do a crap load of characters and styles without needing extra networks
the most popular anime-styled SDXL models on civitai don't even come close
this is scary! You can’t use one button for creativity with the inscription “I want to do it beautifully”
because novelai train their models on h100 clusters,comunity models dont have the money to do that kind of training
stretching
but the three sleeves are confusing
AI loves them extra sleeves
AI love all extra!
from old?!
?
i see it before.. or quite similar
That's new but I have done something similar
oki-doki
I keep trying to summon a demon but I only get multiple hybrids
HOW your magic circle glowing!!?!?!? it`s a magic?
(anime style:1.4), a demon girl with (demon wings:1.1) wearing a dress sitting in the center of a (circle of glowing runes:1.4) on the floor, (eye level shot:1.2), (from side:1.1)
winged demon (sitting in (glowing shine magic circle:1.25) kneeling:1.45) magic! it`s really magic!
prompts are weird
why?
just that sometimes they do exactly what you expect, other times it doesn't
TRUE!
your prompt on my model doesn't work well at all, i can't even post that here..
this is only part of the prompt
and here he is all
,masterpiece,best quality,absurdres,detailed background,closeup shot,portrain:1.35,body focus,(looking up:1.45),perfect correct anatomy,solo :1.5,a young preteen child winged demon (sitting in (glowing shine magic circle:1.25) kneeling:1.45) in the (deep mystical fantasy dark dungeon:1.2),(long shiny red hair:1.35),floofy hair,curl hair,wind hair,(crying,angry, screaming:1.5),cry from the heart,tears,(face in agony:1.4),(light fantasy long demon clothes),(chains on the wrists:1.35),(demon huge:1.3 wings:1.45),demon horns,(dramatic pose),dramatic illustration,(medieval theme:1.1),night time,gold rays from sky,obscure light and shadow,soft shadow,chromatic aberration,ambient oclusion,cimenatic light,reflection,refraction,
of course no LoRa
crap! The model from @supple raptor is at its best here!
Moree... MOREE
I thought it was a bowl of dumplings.....
Boy I can't wait. 1.5 plugins on SDXL, Controlnet tiling, etc.
Adding Universal Compatibility of Plugins for Upgraded Diffusion Model
What does it do exactly?
cross-platform compatibility for the extension?
think its something to do with being able to use sd1.5 extension with sdxl with that adapter
something ppl been wanting
Ahh make sense 
More info here: https://showlab.github.io/X-Adapter/
She didn't paint the clouds correctly. 5/10 
artistic license
she's learning to paint 🙂
also her view would be beyond the canvas which we can't see 😛
She's literally painting the sun what are you talking about lol
but the sun in on the left of her canvas and the right of our view
yep kinda is
she is an artist, she sees it that way!
Eh, that's too young for this server. You really need to delete that.
this is a little too much
I haven't tested yet but they dropped a 2.5 "anime style" safetensors a few days ago.
@untold glacier @steady grail why?!
Just delete it please. I don't want to see you get banned.
the costume emphasize too much the forms
i only see this on civiai
Thanks.
always check huggingface first, usually updates way faster https://huggingface.co/bluepen5805/blue_pencil-XL/tree/main
Now, regarding the very super important discussion we were having, heh.
yes, super important
this already looking sweet
Oh there's a 2.5 available?
obviously, she is cheating
yeah not sure what the 'anime style' one is I haven't tested that yet

2.5 is wildly good btw
perspective is different, shes looking in the same direction as your top arrow
BluePencilXL v2.5 and merry christmas
which rotates the clouds and see things we cant
But then she wouldn't have the sun in the center of her painting. 
wow, cute colors
I'm going to have to try it. Test it. Put it through its paces.
or maybe when she started the painting, the sun was in that position
and honestly, you can't see the sun with naked eyes
So it... moved horizontally... across the sky? Lol
it's not centered, and it's not a 90 degree change, she's looking 45 degrees between the way she's facing (since thats canvas) and what we see
it does suprisingly well with hands
which version is it?
I use the base v2.5 not LCM or anime style
I'm starting to get flat-earther vibes from what I'm reading lol. I'm just joking with you, of course.
depending on location and time of year, it's a thing
if she is to the south, compared to the sun, it's possible
ahhh very nice, cause i was wondering what that version offered, can you post more?
also native 1920x1080 renders can actually come out pretty decent with 2.5
not completely horizontal obviously
fyi i render all sdxl images at 512x768 and hires by 2x
yep
I still say this is crazy.
that above is a native 1920x1080 no upscaling or fix
a crazy as the round earth
rainbow eyes one of my favorite modifiers
Where do you get v2.5? I only see v2 on Civitai.
lol why crazy?
huggingface is always more updated than civitai https://huggingface.co/bluepen5805/blue_pencil-XL/tree/main
Thanks, Civitai is just easier because they are all in 1 place
probably because 512x768 is a normal resolution for 1.5 models, most of the point of sdxl is 1024x1024 minimum
otherwise why sacrifice the speed
every version of bluepencilxl is in that repo
Because SDXL is meant to work at the higher scale. Aesthetic placement and flow of the image will be better at the higher res base gen.
but, there are quality reasons too
post more so we can get a feel of that version 🙂
everything I'm posting is 2.5 renders rn
we are risking to get our community flattened by flat earthers if we don't change argument lol
ahh you meant that in a surprising way, yeah cause i also was under the impression before that we needed to stick to base canvas 1024
I'll see how it handles IPAdapter, the only Anime type model I've had so far that doesn't do weird things with a human face in IPAdapter Face is Counterfeit, all the others make them too real.
Just a multiple of those. That's why there are prescribed resolutions for SDXL, as it's much more prone to issues outside of those.
so i wonder what the anime version has to give, im gonna download that
How many models have you tested, exactly?
Can't remember completely, about 5 or 6. I deleted them when they didn't work how I wanted.
11 models for 66GB makes me cry
right, but you know Emad came out saying on X that everyone could render at 1024x768 with sdxl turbo
SDXL turbo is based on 2.1, right?
It's not actually SDXL.
no, they have sdxl too
no it's based on SDXL
Then what am I thinking of?
turbo is a SDXL 1.0 model that is optimized to use less denoising steps
well its built on sdxl model
but only renders 512x512
SDXL-NotReallyXL-ButFaster
i used turbo with 1024x768 too
aka, 1.5 😄
Perhaps your prompting was the issue. I've found that if we don't ask specifically for anime, even with anime models, it would get that extra step to actual anime style reproduction. Try Hiwamarimix XL-V2. Very good at following instructions and amazingly sharp and cohesive.
Nah I did. Just some of the anime models have real stuff merged in with them and they aren't overly pushed towards anime. So when you IPAdapter a real person, it tends to pull out the realism stuff still in the model.
the power of native 1920x1080 renders tho
It's fine though Counterfeit 1 and 2.5 work really well for it
I'll have to try Counterfeit XL. I absolutely detest the 1.x versions of that though, as the methods to put it together are atrociously done, so I don't have high hopes for the SDXL version in recognizing proper prompting.
how about some (photorealistic) ?
Not sure how it is with prompting, I tend to use IPAdapters mostly now
all 14k+ local gens I've done are anime >.>
So you... just don't prompt?
I do, but only short prompts for what I want to start with and then IPAdapters to get the styles and stuff.
I think the IPAdapters are cool so I play about with them and I'm crap at prompting
can you try a wild fantasy beast out in the jungle and add photorealistic, im curious how well it does
I see. Well, SDXL relies more heavily on prompting than 1.x versions ever did, so that could be a part of any problems you face. To be fair, SDXL is much easier to prompt with using natural language, e.g. sentences that describe the scene of what you're wanting.
im getting the anime style
let me folder my NSFW from earlier and I'll try it out lol
lol
We need a proompter emote. 
So this is what these "Anime" models can do with a human IPAdapter image, they can make it look a bit too real, which looks weird.
It's a little better when you put "Anime" in the prompt
What's your negative look like?
I'll see what the "Anime" marked BluePencil looks like