#🍥|anime
1 messages · Page 150 of 1
there are more things happening than just the prompt, seed, sampler, scheduler, steps... everything affects
are you generating here on the bots?
Hello wyz 
hey, how are things around here?
everything is fine I think
hello pppppppp
Hello
I'm . Not ppppppppp
Fun with injections
Just testing the possibilities
👀 feel free to experiment but please don't post nsfw or if you go too far... blood, veins and such. medic is ok, a junkie not
anyone with a GPU of 'Radeon RX 7900 XTX 24 GB'?
I don't but what do you want to know >.>
generation time for a 512x512 --> 1024x1024 image using high res fix
mine is 2 minutes for using a GTX1060 6gb
theres this chart from months ago
Its a very fast card, it actually has AI Cores or whatever in it for AI stuff
not sus 
unlikes the others 6xxx and below
and i can get to 20it/sec using tensorRT
but it's a pain to use it
Yea sure
You mean tensor cores

I dunno whats called
Ah so you mean meme marketing term cores
Yea they're decent
she's coming to kick butt
yep insert thing that no gamers care about but makes it more expensive
Sus body proportions
hmm not sure why that first one looks like she's jumping at the viewer
shrugs
yeah just being random
cause I can't think of a prompt.. brain... not awake.. fully.. thud
assuming iteration=image, then I have 0.013 iterations per second for a 100 step 512x512 image
or 1 image every 76.54 seconds
@tame tree
Iteration is the amount of loops the data makes per step and the amount of time it takes to do 1 loop of data
10 iteration/s means the data goes 10 times in a loop per second to do 1 step
why does that sound complicated help
most models go to the next step at 100 iterations finished
oh, its in that command window thingy, nvm, its 1.33 iterations
at what settings
100 steps at 512x512, gpu is gtx 1060 6gb
I was trying to match the settings in the tom's hardware graph
50 steps is pretty much the highest you'd have to go .. anything more won't have any significant improvements
I do 20 or 30 99.99% of the time
32 steps + 15 steps upscale
hello kuromi 
1070 
Those “ai” or “tensor” or “cuda” cores are actually doing the heavy math stuff, they are specialized on multidimensional math, key for inference, training, etc. on games they do the DLSS and upscaling etc so they’re pretty useful for what we do
Tensors are kinda like huge arrays
For sdxl is 1024 x 1024 dimensions
1024x1024x1024 (Tensor also calculates Z dimensions.. which is great for perspective and depth in image generations)
forgot how to prevent these issues 😦
so random when it does the many heads/limbs stuff O_O
Honestly
You just have to accept it's a universe where many heads and limbs exist
Then, you might even start seeing it as cute
what about a universe where people use ppppppp as a nickname
Peasants get the veggies market and the oligarchy gets a comically large castle
Not fair bros
but theres at least pretty girls 
You want to know how she got these scars
And these broken fingers
low resolution got her the scars
normal things in the stable world
it's a trap
Nice
What a beautiful woman in there.
Is that Makima? 😮
Cirno had a bit of a rough day. She burned down the cabin trying to cook, tried to run away so Reimu wouldn't find out and fell in the mud and then had the realization that she is now homeless.
LCM inference time: 15.95922589302063 seconds
wished i could use an anime model for this 😦
16 second generation is pretty fast on my gtx 1070
later I bought her an ice cream though so all in all turned out to be a good day for Cirno.
cute
lol that last pic, Cirno fiending for another hit
images with 2gb vram
we're gonna let this one pass, but no external links and no nsfw, part of our https://discord.com/channels/1002292111942635562/1040299965647437905
going to delete
ah yes I'll dm it
is real
A, B or C?
B
her shoulders are showing 
what's your inpaint prompt
oops
wrong @
@nova remnant
I explained here #🍥|anime message
anyone tried this? https://huggingface.co/segmind/SSD-1B
model?
not anime 
But what if you can gen Steven Seagal
i only need anime steven seagal hugging miku 
The Segmind Stable Diffusion Model (SSD-1B) is a distilled 50% smaller version of the Stable Diffusion XL (SDXL)
just like every other model is 1.5
obviously it's not literally SDXL
I just found funny
it's in diffusers archive format, so it might just work in any of the WebUIs that support diffusers since that includes the configs in the model rather than it being separate like it is in compvis like we normally use.
Short-stacked women in there.
This is a sign of either using a model with the VAE pruned and not loading a separate VAE or not doing enough steps.
correct me if im wrong, but im sure the only LCM model now available to the public is the dreamshaper

it's not really worth using it atm, unless you have a super slow gpu, or are just testing it
i have one of those seals irl
looks like makise O.O
a bit yes!
fox girl go bonk @lyric zealot 
anyone know what these side hip things are called
the clothing holes
seems prominent on wide hips
@stable crane booru tag guesser/idk what it's called usually puts them as side slits, hip windows, side window
side slit?
thigh cutout https://danbooru.donmai.us/wiki_pages/thigh_cutout
not sure if AI would do good with that
A Miku Adams movie
oooh spooky
o.O
It works well with dresses though
Good deal
hupppy aluwen
hmm 4am I need to get to bed
she's suppose to be just sleeping not hte never ending sleep.....
anyhow, night. thud
(snores)
tummy

@finite herald @gritty spruce @wooden coral @dull jackal ^ #🍥|anime message


the old free nitro link! people still click on those? 😜
thanks for the ping, tried to spam more channels
this does something right
.>
not sure if anybody click those, but just incase someone accidentally click it so...
yeah usually reporting works fine but sometimes breaks, bot hiccups

trying out the color gradient stuff, it's great for stuff like hair and clothing
yeah with kohakuXL and yandere upscaler
Trying out a random model
None of the images came out sfw 
Was doing 9 at a time lmao
time for some darksushi 
What were you prompting lol
hahaha tell the truth, you prompted that 😂 you can always neg prompt nsfw
@zinc flax just 1girl to see what comes out lmao
Oof, that's always a gacha lol
hmm how can I do clothes that look like they are imbued with magic
"1girl, cape, (clothes imbued with magic:1.2), black clothes" surely
surely
I'd like to see anyone beat this amount of fingers
its like 10 in one hand lmao
what would that even do exactly 
I'll try the other one
LCM to I2I
she's doing a spell, fast fingers

better i2i
@silk narwhal one of the 9 images came with some glowing markings 
big magician 
perhaps with this ? https://civitai.com/models/51686/glowingrunesai
Twas the night with a ping
Oh, love these!
blergh .. internet is being slow
I can't download loras 
why not?
Not using my pc atm
aah
bewm?
full image wasnt good but AI got me a bird cat the in the corner 
If you're asking about that girl above I have no clue, just random 
With robotic fingers, you can do a lot
You know... Perhaps I found the reason why the models keep funking up finger count... REMOVE NARUTO FROM THE DATABASE


hello p
blerghhhh
We have a channel where all reports come in for us to see, from there we can take further action without you needing us to ping actually. Pings just make us act quicker probably. So to answer your question: Yes reporting always helps 
hello boto of all botos 
say AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
No I'm just making a random noise
What sd is that?
kohakuXL
Ahh ty

that's why you fall to the big b00bz 😂
experience or a LoRA 😛
Experience in breaking dall-e leftist corporate dystopian censorship
Just kidding
what else would you rather, they master it on doing marketing ads?
Theme: Miku goes to the halloween party.
what would you dress her with?
go for it? lemme see
hahaha at least you didn't say melons hahaha
but she's not like that
Now I wanna try pumpkin dress
go for it
tweowwwww
isn't that cool!
and it doesn't censor that!! interesting
speaking of asuka pixel style
I could say that she doesn't censor worse, but I can't post these in this respectful chat

😄
dunno if this is cursed or not
asuka screaming with giant mouth,giant teeth,anime pixel art style
adetailer for hands is not good
in halloween spirit, it's fine
better to fix hand maanually and than to use controlnet to add details
i just do this every time
some images in pixelart are really good
penguin
dalle3 
Same prompt with bing's DALL-E 3 integration vs BreakDomainXL0.5g (SDXL-based). I'd say considering that the SDXL model is running locally, not bad. Still not great though. 
I wonder how much further we can have the context make sense to the AI, when its running so few parameters and from so little context comparatively to the giants that are DALL-E and Midjourney. In my mind, that's the biggest difference right now between SD and the closed-source server-based alternatives.
my new haristyle
cause i put big bald shiny head in prompt
prob cause its a girl and theres almost no images of a bald girl in the dataset
skill issue
you can put all monster teeth and eggs in mouth cave and no bald? come on! lessgo
yea not fully bald now im out of credits 😔
ill try it when i get home 👩🦲
1060 6gb 
its ok i can create 1080p images in 4mins 👩🦲
i already have a microsoft account i had 200 credits on release and now they all gone 😔
next year 
do you use control net or anything ?
That was pure txt2img with hires fix. then passed through tiled diffusion upscaling with canny controlnet to upscale it even more.
Here's a couple more without upscaling that you can pull prompt and parameters from
awesome
an innocent and family friendly painting of a conservative trad girl.
4th impact incoming
let us pray
This is amazing.
who's sick of bad shooting stars being everywhere in their night gens?
idk what happened here tbh
wow still use same model ?
not sure what the same model you are referring to, but here's the pic with metadata 
basically just embedded data with parameters on how to re-generate the same pic
so like learning thing
parameters? 😵💫

Sadly this only really works on the model it was trained for. With other models it will change the image a lot by the time the strength is high enough for it to actually work.
not sure if you are joking or just confused with my choice of words 
well, one easy way to look at metadata from a pic is to use this https://supagruen.github.io/StableDiffusion-CheatSheet/
that's the easy way without having to open webui
or just open the image with a text editor. It's all in there in plain text near the beginning.
confused 😵💫
yup, or that, or using promptdog 
well, "parameters" are just settings. Like hires fix denoising strength, cfg scale, clip skip, and etc
is this self trained?
the sceney looks very kimi no nawa
sorry but today we doing succubus instead 🤠
So, the lora is actually a LECO. It's trained without any images. In this case it was trained to remove the shooting stars from this model because they were everywhere on all of my nighttime gens. https://huggingface.co/AshtakaOOf/Ahnn/tree/main
More info on LECO here: https://github.com/p1atdev/LECO
That example is the LECO that was trained with Ahnn used on Kakigori_v3 to test it's portability.
The good news is that retraining it for other models is super easy, just have to change the model path in the config file and retrain. The bad news is that it will take around 40 minutes on my system for each model I want to train one for.
Futaba Sakura, is that you?
so basically, you use it like a LORA, but training method is different?
perks of it is that there isn't a need to gather images?
just random gen 
I'm on the case, large- I mean succubus incoming
miku succubus 
yep, no need to gather images, just list the concept you want to erase, related concepts that it appears in, and a concept to take the difference of. The whole prompt file for it is 9 lines long with most of that not needing to be adjusted.
so if the concept isn't present in the model, then this training method doesn't work? 
interesting method tho 
wait how do we upload images in discord with metadata, i thought discord strips the metadata
The idea behind it is to remove something that is appearing in your gens when you don't want it. An example would be umbrellas being automatically added whenever you prompt for rain in some models. That was happening with shooting stars in this model whenever I prompted for night sky.
do we upload as files instead of image?
The model that you linked is very pretty btw
Throw the PNG file in discord, not copy
thats how i usually do it, gotta be careful next time
ahhhh, i mean make sense. If you want to add concept to model, it's not possible to do it without providing images
Discord hasn't stripped metadata for months. It only did it for a couple of weeks. You just have to drag and drop the file into discord instead of copy/paste from the browser.
i thought discord removes it like facebook
you can use it to reinforce a concept that is present, but somewhat weak though.
I might try to train one to remove human ears when animal ears are prompted in Kakigori.

so this is basically going to end game of finetuning model
I'd say more of just a tool in the toolbelt to try to fix some problems that occurred durring finetuning/merging.
@native halo is this the right place for succubus

she needs bigger horns like this
ignore the backwards hand 
I had a dream I was resurrecting someone using AI and songs
i love when the dragon girls comes out
You know those stores that have cold air following from the ground up on the entrance
That's it
might be hot air 
not that type of store

no fucking idea 
She looks like a monk
@native halo ever heard of the thighs church of hell ? 
i am now a believer of that religion
forgive me goddess for i have sinned you may punish me now
smol and angy
It's like Lotta and your oni girl merged 
that was the idea behind it 

Where's the other part
It's missing

You should ping LottaLewds in one of them. I'm pretty sure he would like it.

I usually don't like to ping people unnecessarily, but here goes
#🍥|anime message I gen your OC partly merged with my oni girl
Nubby said you might like it @stable crane
trueeeeee
and a lot of highly good generation, those kinds that you like 

help me make this better
(masterpiece:999)
please master
Don't actually do that
(steven seagal : 4.5)
but maybe try using hires fix as well if your gpu can support that
touhou you say? I don't know any of those prompts. 
i try make this but i just get random thing
Have you ever tried 2 touhou characters in one
Mixing them
so are you trying to make the sticker style pic, or to get a better pic of the first pic you posted?
I mixed remilia scarlet with flandre scarlet. Let me find it.
i once made a drawing about 2 touhou chars mixing each other 👍
I have no idea why, but the way this read, kinda sus
🤠

Each other... Food?
salad
Oh okay
More thighs 
i just want the same picture result not a random
i mean same hair style same outfit
this regional prompt
ahhh, okay, then there two general ways to go about doing this. One, to just describe everything with prompt, if it's kinda generic, then that might work. Two, to train a lora of said character, that might be a lot more effort, but if done right, can consistently get good result even with different models
yes, plus the left_right lora
hmmm train lora maybe is good thing
it's definitely better to train a lora for character that you wanna recreate over and over
but i just have 5 image can i make lora with that ?
might be too little, but not sure since i haven't train any lora yet 
@supple raptor 
I should try again, I could do better.
didn't ask for me but I've trained very obscure characters on dataset before
yes, you can? but it'll take more work / iterations. train little by little and use produced images as more data. just make sure the hands are okay or it will learn bad hands
if you meant giving a pic, make AI give you prompt that might gives the same generation result, then yes. Go to your img2img tab, drop the pic, and press interrogate clip or deepbooru
It's possible, but will be more difficult and very prone to overfitting. You can potentially train one that's just ok, generate a bunch of images and pick some of the best to augment the dataset and train it again to try to improve it.
Or also run the WD_14 tagger on Kohya lora trainer to generate and scan for tags. if it's an original, drawn image
I would suggest if there are even things that look like the character use some maybe
so same hair style
similar kind of outfit
just so it doesn't overfit
Lora training is it's own beast. lol
I'd recommend just googling first or looking #🔧|finetune
@supple raptor ive made these couple of months ago
LLES
idk what model should i use for sd to generate tohou stuff
Meanwhile.. I'm making thighs 
Kakigori v3
Does Shrek and touhou very well

nubby how into touhou are you
I was into the games but there's so much media idk even...where....what to do
I just like the characters. I've never played the games.
same here
Ah
love the characters, arts and songs
Actually true
So many trials and errors 🫠
took me a damn month to figure out how to work it based on trial and error yeah
well a few weeks but. longer than I thought
i dont even have a decent pc to train lora
done,(steven seagal : 4.5) u owe me $500 💸
my fav miku
civitai so buggy, i cant even find character loras
Forget Steven Seagal, we are all sinners 
It's been almost a week and I'm still making her lora
Having issues with her bangs ig
Hmmm
I'm noob
@zinc flax I offer you magical thick thighs , they will increase your luck by 10% on loras
Very nice
im using counterfeit for the first time now, is it normal for it to take a long time to generate an image? ive been using blazing drive all this time and the generation time was really fast
idk if different model takes different times to generate an image
not really, unless you are comparing SD1.5 model with like SDXL based model
if not, if the steps are the same and hires fix upscale is the same, then it should be about the same time
it might be because you are swapping model?
unloading one model and loading another, if it's in hard drive, then it's gonna be slow
its in ssd, i swapped from blazing drive to counterfeit and generated few images, i think its getting faster?
also sd 1.5
then there's almost no reason for big time differences
model
test
latex,bodysuit,skin tight:4.6
ok putting midjourney style and using a hyper network for it too.. makes it odd...
but very artsey
First attempt at the extra ear remover LECO was a complete failure. Will try again when I'm not in bed
Fried gens

going home
ooo nice 🙂

Wrong Vae?
@dusk seal hello stupid p

Hmm
I dunno
Images like that are usually weird resolution or vae
Dall-e 3
extra ears remover for kakigori_v3 
if you do negative
does it add more
It will most likely make the human ears bigger, and gradually shift them towards elf ear shape
DALL-E 3 would be amazing if they did not sensor it so heavily. Or even if they instead of telling you sorry I can't do that, just gave you something close that it could do.
it's still quite amazing
it'd be nice if it just retried generating until it gave you something it thought was OK
or they could stop being little bitches and just relax it 🙂
hope your insurance is up
there goes the neighborhood
mecha!
adding "chibi" for fun 🙂
It sucks for us that it's so restricted, but with the amount of attention that it's getting it's probably a good thing that they're restricting it as much as they are. We don't need to give governments any more of a reason to crack down on AI as a whole.
oh, governments are EMBRACING it a lot to create nice troll campaigns, social media disinformation, overall FUD... just the wrong ones 😉
hello everynyan
"a government agent creating fear uncertainty and doubt, instruction poster saying "FUD""
I just calibrated my monitor yesterday and holy shit this is so much better
I recommend that you all calibrate your screens if you haven't already
also @supple raptor how you dealing with discord no that they changed their internals. I heard a lot of betterdiscord/vencord plugins broke
the "calibrate your monitors" text is not perfect but recognizable 🙂
this is literally me
"callrate"... well...
they were broken for a couple of hours, but it's all back to normal now. The only one that was really a problem was image zoom was causing multi images to show at the very top of the screen with the top ~15% of the image cut off.
but she is cute
some of my plugins are still broken
it is up to each plugin creator to fix theirs
so yknow
cu later 🙂
do a manual update, I had some more updates this morning that seems to have fixed the last few that were not working right out of the ones that I use.
oh ok I will
ok nice
it didn't fix all of them
but now custom css is back to normal
very nice

BIG Ants are BIG!!!!! 😮
Okay what mutated the Ants to be so huge?
Is there this kind of thing for cat ears?
They aren't very portable, so you'd want to train one for the model you want to use it on, but they're really simple to train and no images required. When I get home I can share the configs. I have (hopefully) better versions of that one training for V2 and V3 of my model right now.

took me like a couple of minutes
I didn't manually calibrate it
my screen is not that expensive that it is worth manually calibrating
Hello dog
Feet
what prompts should i use to generate these types of image? minimalistic yet detailed, flat colors, idk how to explain the details, its like a sketch?
how about the 3rd pic? the character looks detailed but the background looks dystopian, i try giving prompts for specific backgrounds but always end up with something generic like this
even said walking through deep forest and stuff but doesnt work
try to change model,some models cant do backgrounds
1st one is blazing drive, second one is counterfeit
try with breakdomain,cutemix or aurora
when i wanna test if a model can or cant do a certain prompt i just do a x/y/z plot with all models i have because some are good at certain styles while others fail miserably so only way to know is to test them all
darksushi is also good with backgrounds
(Steven Seagal background:3.8)
Try without quality prompts
Remove high quality, best quality, low quality and worst quality

Hello torqx

yeah not sure what happened there
And she's changing
well then
is it lewd if it's made of bone and tattooed?
wait if I have to ask it prob is.. sorry
look into my eyes...
ur stuff is always so cute...
how did u make that one
also curious about the bug girl with green wings
thank you ❤️
it's all about the lora selection
pick cute ones, pick realistic ones, mix em
After more testing, I've determined that Kakigori_v3 just doesn't really have a problem with creating extra ears on animal ear images so I can't confirm that one worked very well. Here are the configs for my shooting star LECO that I trained on Ahnn though, which did work very well. Credit to aha in the Touhou AI and AnimeAI servers for providing the configs and here is the repo for LECO training: https://github.com/p1atdev/LECO
@lament lion is here too 
Yup! Hello! 

Thats really good :O
Ahnn is really nice https://huggingface.co/AshtakaOOf/Ahnn
Thats the model? 
yesh
Its pretty cute 
@supple raptor
is 3x hires really bad or do I just have bad settings for it
yesh 3x is too much. 2x is pretty much the limit to maintain coherency. more than that needs too much denoising to make up for the upscaling artifacts and that counteracts all the benefits of hires fix. Best to do 2x in hires fix then 1.5x in img2img upscaling with either Tiled Diffusion or SD Upscale if you want 3x upscaling.
I do 2x hires fix then 2x tiled diffusion when I want beeg img.
pwease
Its just a cute model :p
cuteyukimixadorableechodimension

Its the name of the model
I don't know where it is from 
I know some cute yuki are on civitai
O, tis one 
@magic wave I believe its some obscure chinese model
I have a friend who was hunting for the original
Fair :o
Theres clearly a style here tho 
model's really adorable
Searching for my phone for 10 minutes
I left in the kitchen window

Impossible to not see it, I'm just blind
I mean that is weird place to put it tho xD
I am gonna buy a domain today, anyone got any suggestions?
am gonna use it for an email as well
man
happens to me every time
until my mom comes






