#🍥|anime
1 messages · Page 257 of 1
I am bot
My first ever AI art... made almost a year ago ☺️
start what? you wanna generate images?
Neko is cute!👍 😍
I wish her tail was fluffier!
hmmm space station?
yep
this keyword extension and DTG ent , sure goes random, LOVE IT
same prompt and everything, I just hit generate.....
hmmm is that tokyo?
hmm no
age regsion, her expression says what happened? , haha
ooh interesting eyes
adds shiny heterochromia
no and no! just my older sister came for vacation from the magic academy😝
the older sister
oh .. takes out the lora
what`s the prompt for the style ?
(picturesque,aesthetic, ultra detailed background),(detailed skin texture,subsurface scattering:1.18),perfect correct anatomy child body,perfect correct anatomy hands,(age regression, age down, loli, child:1.05),nahida_ (genshin_impact),((artist: ciloranko,artist: tsubasa_tsubasa)),1girl,illustraion of girl,light silver hair,white shirt,bow,bowtie,genshin impact,gloves,gradient hair,long hair,long sleeves,multicolored hair, no shoes,pink bow,pink bowtie,pink eyes,green hair,shirt,((nahida_(genshin impact))),white gloves,white pantyhose,((green eyes,symbol-shaped pupils)),(obscure light and shadow ,soft shadow,chromatic aberration,ambient occlusion,continuous,lighting)
thx
I would cry from this ice cream
neat! Loli's revenge is very scary!
Of course not!
hmm ice cream muffin?
last muffin
no that looks confused.. kinda
wow !

grabby legs




dalle3 paid ?
there's a paid one?
idk i`m asking

@green temple
my gen now
Madoka?
Just random
Smol
"a sweet anime meme" - don't ask what she says (flux schnell)
another 🙂
"an aggressive anime meme"
you a son!
me too. me too.
Youa son...! X3
scary 🫣
little bandits!
and we know who the leader is
Don't touch grass
Cuute
I'll help her pick up the books
Guys, who knows, this message appears when you start generation
TypeError: '>' not supported between instances of 'int' and 'str'
something about those eyes
generation doesn't work
nice color palette and soft shadows
this error message basically just mean you are trying to compare between two different types of variable (integer and string). In your case, likely comparing something like steps or resolution to something. But that should not matter, since you should not fix that yourself.
what did you do to your webui that makes this error?
I sat and generated. and then I just restarted Forge. I didn’t install anything new from the extensions.
oh I got it! I accidentally hit this half-knock
and the value was set to 2
maybe that's why. Since gpu weight should probably not be 2MB 
yes true, but still such a user error is not excluded
true true

So cool , nice try 👌
trying to make seven of nine, comes out more like the matrix...?
how do you get that kind of effect?
dall3 is my cheat code 😄
I use various things like "splatter paint"
ran out of my full size imagen3 gens ;_; but yea splatterpaint helps with that
getting closer anyways
how do you use that
full size?
I saw it but don't remember where now 
select testers only i think
idk what imagen is
so its from google
ye
but
for the full size one
they lifted like
all the restrictions
i got so many good gens today
any interesting things to share?
these for nowh
i was testing lots of random stuff without direction
from a similar prompt as the last one
before it wouldn't gen anythiiiiing
it only denied like one request i had and it was for a pretty obvious reason i think lol

but it can even do classroom setting now
I love dall3 so much ... honestly more than flux I think.
meanwhile on local machines. thats not a problem
I got a new Pixel 9 phone which apparently comes with their advanced AI stuff for a year
borg!
cat if u wanna try it i could give u a 4month free trial of their advanced AI stuffs, but the ability to make hoomans is coming in idk, next couple weeks or so?
the problem is the local machine


sure
i'll dm u the invite
scam links incoming
beware of things like this, beware!!!
nuuuu
u can see the same kinda link if u hit invite friends
in geminis window ;_;
lol
this button kekw
but you might need advanced to even see that
my understanding is it gives a 4 month free trial
I think I am using the imagen 3 non full size or whatever ?
yeh
These look lower quality than what you were posting
they'r;e lower quality than the base 1024x1024 outputs
when u use those ones posting the 512x512 is better
with copy image
but yea i agree it feels lower quality so i'm not sure if it's like implemented within advanced yet
the curr iteration of model
anyways
the ones i posted wre from test kitchen
but it can still do
some stuff on the 512
lol
whats this rat called
shinx
and this is renamon
i think gem on the site is still using an older model of imagen. i have to wait for my 1024s to refill ;_;
@rough oxide I made him very annoyed 
i found out that's still imagen2. imagen3 you'd know if you have it because you only get a single image and it's way higher quality
i kinda suspected as much with some of the gens i was getting T_T so i still gotta wait to get imagen3 on my gemini sub, but can use it thru test kitchen still

But it says imagen 3
o
wait wut
okay
for some reason u and my friend have imagen 3 and i don't there xD
i can only use it in test kitchen
dang.
i guess it means it can't be that much longer b4 the rollout is done i guess
I remember using the 2 or whatever it was before
images looked way blurrier
the downloaded version is definitely better too
they feel like 1024x1024 with a upscale
the 512x512 is crappy 
to 2048x2048
the preview it gives
also
oh nvm
i was gonna say that means u can gen people
but that's advanced only
cuz people also encompasses things like anniemay
just trying examples they put out on the imagen 3 ad site thing

it does give me the warning only advanced thing can do people
cheating with pixel art 
@rough oxide help
what is the white hair elf
that is popular
called
ur def not gonna get it
frieren
it can do demon slayer tho
cuz i tried doing oshino shinobu
but
another shinobu came out
100% reproducability.
it can do like all the popular shounen stuff but it's not like specifically trained on the booru or anything
so only HYPER popular stuff
oh my god it did friren
100% her @rough oxide
lol
runs away
imagine if it quickly self-trains a lora based on the google image results of what u ask it to make
before making ur image
but nah that's waaaay too advanced for 2024.
yeeee
i'd say even longer but
it's only been 2 years
def 2028 shit
xD
if something of that scale isn't out by 2028 then the steam train stopped goin
what i wanna kno is if i'll get rate limited once i do finally get imagen3 on advanced
i wonder how many images it'll let me ask for
dall-e is like 200
the test kitchen is like 100 daily
knowing google probably a lot 
i am an absurdist power user
i sure hope so.
i do notice with imagen 3 u only get 1 gen at a time
instead of the grid of 4 like i was getting
with test kitchen it's 4 at a time
Sometimes when you click on the images
so i'd assume minimum 400 pictures.
it can automatically generate more
ohh
r u on google chrome btw
Yes
Then you can generate more fairly quickly
but its literally the same as asking "generate more" I think

this is probly because the pictures it genned had a visible human face in it
if ur not on advanced it'll not generate those and only show u the rest
cuute
Cute veggie
Why did you offend her?
i`m innocent 
then the cat did it?
yes... 
liar 

what`s that in her hand ?
Universe!
oh nice one 👌
she stole the dungeon core?!
how
I don't know...
I was getting 3.6s/it with flux back in August
latest update, 5.5s/it
Try swarm, forge is an experimental platform and will always change and break things
Or if you like the ui then a1111 or you can try reforge fork
At this point forge is just a1111 with experimental features and upgrades without most of the benefits it had before
yeah probably time for me to swap around
I've tried swarm on 3 separate occasions, and each time, ran into extremely frustrating issues. So that's a non-starter for me.
it does have its quirks, but it's been updated a lot in the last month or so
I prefer it mostly for the UI and I don't use a lot of extra stuff like controlnet or extensions, just basic inpainting and upscaling so it works great for me
did a pretty big rollback on it, we'll see what happens
no luck.. oh well.. perhaps an external factor
ah, well, actually the rollback and a reboot seemed to work. let's see if I inch back to latest commit now if it sticks
about 3.6s/it, flux nf4, about 70 seconds to generate at 1024x1024
woo, seems to be back to normal now after doing a checkout to revert and a hard reset back to current 🙂
How do you get those perfect generations I can't
I don't know if you're asking me, but if you are, the answer is simple: Flux dev, comfyui and a good Lora you train yourself
How do you train lora
ask a girl named Lora to train implying she is obese
i like the lights
the res is insane

There are several guides on the internet, but the easiest way is: Go to https://replicate.com/ostris/flux-dev-lora-trainer/train and watch this video and do the steps he teaches in the video https://www.youtube.com/watch?v=_rjto4ix3rA
Fine-tune FLUX.1-dev using ai-toolkit
Here's how to train your face into the new AI models. Here's the link for $10 free at Replicate: https://replicate.com/invites/572ae3c6-1385-41b3-afdc-1e99e11836b8 (NOT Sponsored)
Discover More From Me:
🛠️ Explore thousands of AI Tools: https://futuretools.io/
📰 Weekly Newsletter: https://www.futuretools.io/newsletter
🎙️ The Next Wave Podcast: ...
backdrops in imagen3 seem to have a lot more detail by default instead of just being blurred out/part of the scenery
anime

ugly 
mrw
a lot of models have pretty decent backgrounds, but you usually have to prompt for them
ye, you don't need to heavily prompt the bg in imagen3
from what i've noticed so far
pony models specifically default to pretty empty/blank backgrounds
its a themed background okay
its a silent hill background lora built-in

pony based model without background specified at all, versus adding detailed background, forest, mountains, lake, to the prompt
same seed also
added at night
i just added snow to the prompt
but yeah, there's more to the prompt
backdrop actually got worse when i described so much
this is my entire prompt mrtarou, solo, (child:1.2), petite, thin, pointy ears, short blue hair, braid, (tilted head:1.2), purple eyes, score_9, score_8_up, score_7_up, source_anime, happy, hoodie, shorts, detailed background, forest, mountains, lake, at night, (moon:0.9), snow,
imagen is probly one of the ones that needs boomer prompting xD
yeah, prompting is always so different for each model
I believe it goes through llm then it describes for you better what yuo want
kinda style clash
in first one
but i like these kinds of backdrops
i haven't seen pony produce that kind
joins in

the second one is like my ideal "mountaineous forest" background
with how it looks painterly in the back
similar to how anime does it xD
first one would be good too if it didn't clash with the style of the character lmao
char is trying 2 hard.
do imagen images have something inside of them that prevents them from being recognized as jpgs?
the key is, if your model can also do decent realistic images, it probably does better backgrounds
they upload on discord fine but
i can't upload any to steam
xD
and actually for jpg these are massive files
hmm
i wonder if they're actually pngs
why would you do that
i use showcases a bit lol
i have some friends who like ai stuff and i show them stuffs xD
for some reason the download button downloads the images as .jpg, but the real file format is .png
steam doesn't autoconvert

I inputed the original prompt and then used Model keyword extension with inkpunk, and DanTagGen with short setting, and it gave me prompt: at night,
mrtarou, solo, (child:1.2), petite, thin, pointy ears, short blue hair, braid, (tilted head:1.2), purple eyes, happy, hoodie, shorts, detailed background, forest, mountains, lake, (moon:0.9), snow, InkPunk768, winter clothes, very long hair, long hair, nature, single braid, mittens, hood, snowing, night, open mouth, winter,
score_9, score_8_up, score_7_up, source_anime
model: cutelunar_cyberrealistic25d merge Sd1.5 with VAE Animevae
😄
turns DTG random upto 1.5
-hides-
teh kurumi test lol
takes out the score stuff, rewordsa little bit and uses a style he created and saved
DOH!
dont stop posting cause I am...
throws marshmellows
it ain't SSI level but it's good.
wasn't expecting it'd get that close off just "anime harley quinn"
reworks a little bit, doesnt use any saved styles and turns off extensions...PROMPT: solo, (toddler, chibi:1.2), thin, small pointed ears, braided blue hair, (tilted head:1.2), heterochromia eyes, smirk, hoodie kemono, mountains, night, detailed winter lake forest background
ooh mispelled kimono
fixes
furisode>kimono
didn't check if imagen has concept of furisode yet tho
hmmm
sleeves rn't long enuff really
but dang i like the brushed on background of this one
sometimes it comes out like a furisode
but also festivals r normally associated with kimono so that might be giving it noise. i should try a furisode in a normal scene
not logn enuff sleeves for furisodeeee xD
well it knows a furisode is a kind of kimono at least
but sleeves don't come out right xD
wll Kemono is a japanese furry style yes
ahhh !
oh this kinda reminds me of something that's REALLY hard for image models to comprehend
oops to much clipskip -3
i forgot what model i did this in
but trying to make someone like smaller than a coin
and then also have the coin in the same image
lol
it took me so much prompting to do that
oh
it must've been dall-e 3
given the timeframe
before it became impossible to do anything in dall-e 3
no vae
whoa....
fixd
generic fantasy gold coin
lol
for prompt adherence we def entered next era with flux/imagen
cuz that was so impossibly hard to do on dall-e 3
never heard of dream
dream by wombo
i kno ideogram surpassed midjourney tho?
oh
that
do they use up to date models even?
i don't think u can get this result without a NLP model
but i could be wrong
I just used anime v3
this is near 100% reproducability tho
with a way easier prompt than i used to use
lol
the aura is missing sometimes tho
(most of the time)
Restart polyexponential stps 80, cfg 7
nono
I could do that easy
I dont do easy ..hehe
the ones with dancing girl are done on Forge on local machine 😄
shrugs, just messing with settings
oh.. yeah dancing in kimono at park night festival
hmmm...?
tries shorter prompt
sigh bored
never thought I would say Ive got to many days off... sighs
goes fishing in NMS
done fishing
i like that water what model?
flatlunarsushi
where ya got it?
you are a blessing thank you
welcome
using the clearview_v23 vae
works suprisingly well
ty
animevae
haha
cookies
hmm thats... diff.. and confusing
brain.. sleepy.. should.. mm sleep.. but....
B...But donut
no ...
once I get one I like ,then ill goto bed.. maybe
donut room?
ok thats it for me. 😄
I think...
hmmm?
ready for takeoff? Donut-VTOL?
magic donut ride!
morning. so whats the theme for today?
oh "pan" automatically showed up ?
Nope, that's edited
oh ok
This is original
👍
hehe I love thi Geeky rem B ext
rhats alot of bread
hmm
closer
hey @wooden coral
hmm almost
ah ha
😄
yep I think that does good..
full body portrait, random background, random clothing
I think this merge is a success

ponders
Looks familiar
yeah just cut out background and replaced it, with a ext 😄
What model? That dark one is badass!
Nai

Just using the vibe transfer tool on a dark image kinda gives that
Really?! Oh wow. Nai is so variable. So many different styles
You can try controlnet reference model
I like that she looks like a grown woman instead of a kid. I’ve been trying to find models that don’t give my characters the huge eye squishy face syndrome 😂
u can change the way characters look in a model with a style lora
Definitely, just finding that a lot of models have the same kinda anime look to them ( fat chubby cheeks with massive eyes ). There’s a few good ones out there though that I’m experimenting with.
dang.. image upload being weird.. oh well
My image was using a reference that has a girl squishing her cheeks with her hand
Kinda like the image tora just posted lmao
So that is very much what my style is trying to do, they look cute cx
If you want less of that, look into korean or Chinese style loras
That is pretty much the Japanese style of characters

Big eye 
@green temple
smol goblin fairy 
just avoid using "big cute eyes "
( do not do it ) 
👀
chubby cheeks and big eyes are cute! and cuteness will save the world!😍
n word mentioned
nove***
i sure do hope mystery model gets open sourced (99% its not happening)
kick nai outta the window fr

nai is like cool
now vibe transfer tho
(basically controlnet) is really cool lmao
The way I like to use it is very questionable

Sometimes i just put a goblin smiling and make creepy stuff lmao
what's it made of, SD3? does it work with MistoLine or Anyline, for exact render over a sketch? looking to turn a line art sketch into final render but with perfect precision of lines and prompt
https://civitai.com/models/733653/real-crystal-pencil-xl-yf
it's a mix of about max 40% pony and sdxl models. You might want img2img or even controlnet uncanny for your goal. If you use comfy, I can look for a workflow for you.
@rustic oriole I think it's working, check out the included comfyui workflow in the image. Example animated gif will follow in a bit.
thanks, i'm trying to decipher what am i seeing in this one , found some workflow that does sketch based guidance https://openart.ai/workflows/cgtips/comfyui-basic---convert-sketch-to-digital-image/rSmBJheN9S4ToS3Zfc6N i hope it'll run on my 6gb vram 🙈
Created by: CgTips: The Canvas Tab node enhances the creative workflow in comfyUI, offering a versatile space for uses to draw, sketch, and prototype ideas seamlessly within the interface. Click on below link for video tutorials: https://www.youtube.com/@CgTopTips/videos
which workflow did you use btw?
I've added the controlnet stuff to my workflow, and cleaned it a bit
controlnet is a little picky. if the model don't recognize the form, it feel the lines with random stuff
i attempted to force chibi stile on a model without that stile and generated a blob horror
my experiments generated so many horrors
so many horrors
Cuute
sometimes canny controlnet seems to give a cloak of invisibility style
@rustic oriole thanks btw! You made me try controlnet canny again since a very long time - since A1111 in fact. Since many months I'm now using Comfyui. It's great now to use it to upscale my 512px local Flux Schnell images, and apply the best stuff that Pony / SDXL models bring! 😄
ahah no idea what this all means but sounds cool and useful 😄 it helps with upscaling and also styling or what? how do you make it apply things from another model to existing image, kind of the styles get mixed together by the second pass in the other model ?
i noticed all workflows are very different, there are so many options in how to link those controlnets to other steps, my head spinning 😵💫
do you think possible to pick an image like this for example and then with openpose apply different steps to this character so it will run for a sprite sheet of frames? like an attack animation made of 10-12 frames let's say
same principles should apply. Just try to set up a workflow manually following an example, maybe twice. Then (wait not longer than about 10 to 60 minutes) try again from scratch without any example.
Once you done that a few times, you'll get a better grip on how the nodes interact. It's not all logic, you just have to get the experience.
Some nodes have just stupid names, or work different that expected
Watching examples and figuring out how it works, that does it for me, thats how I learn
img2img plus openpose is something that i'm looking for i guess 🤔 its funny, there's one way to make it into anyline/mistoline for control, from the source image, and another way from source image to openpose sheet directly into running steps, without using line art frames.
i think i'll need to make a sheet of frames from openpose sticks and apply it somehow
yea gotta learn a lot
one sec
download this image and drag and drop onto your comfyui canvas
Is this 2.5D anime ?
🤣 turned out quite ok, maybe ill wait a little and post the source image used ... it was from flux. but this one has the pony/sdxl layer over it.
The sauce ...
whats she thinking about
tomorrow`s exam

