#🍥|anime
1 messages · Page 258 of 1
all these are the results of the work of the model, lora, style?
Guys, tell me, has anyone encountered such a problem in Forge? Sometimes when adding several tokens, the result looks like blurry spots or completely black. If you shorten the tooltip or remove tokens, these problems go away...
sometimes such an erroneous result appears even if you rearrange the tokens
boom?
totem magic
if this is with SDXL it could be the math problem that was in a1111, should be an option to change it but I don't remember what it was called, just changing around the tokens usually fixes it
It is true that this occurs on SDXL models
and even on fairly common models - animaPencilXL, animagine...
there's a setting called "Emphasis mode" in forge, it should be set to "No norm"
it's a calculation bug that's been around for awhile, but something about animagine-v3 made it happen a lot, so anything based on that model will cause it
By default this parameter is set to "Original". Thanks for the advice.
yeah, for some reason both a1111 and forge have it broken by default
and on top of not just fixing the math error, they made it an option and then called it 'emphasis' so no one could find it 🙄
Apparently everything is for the convenience of the end user 🤣
pretty sure a lot of the developers just have ego problems and make things harder for no reason just so they can make people feel stupid
by the way. I can’t understand in the new version of Forge the parameter in the SDXL interface - “Clip skip SDXL” - what does it affect and how much?
and here’s another interesting question about prompt tokens. which option would be more correct mouse \(animal\) or animal \(mouse\),
legendary !
oh this is why ! thx 🥹
mouse \(animal\)
niice
黄河流域
Lin?
extra ears is the tag. Good luck though. Most models are very over fitted on human ears and even with negative prompting you can't consistently get rid of them.
Does fake ears work too ?
that will make them wear fake ears like a hairband with them
if it works
Oh ok
Big 
It might. fake ears isn't a danbooru tag so it would be more likely for CLIP to associate it with the tag fake animal ears (on anime models at least) which refers to animal ears on headbands, hoods, or hats. In turn, that would possibly fight against the animal ears instead of the human ears if used in the negative prompt, which sometimes happens with extra ears as well.
Oh , so it depends
yesh, ultimately most of the problem is simply that most human heads that the AI sees during training have human ears which are typically not tagged so it learns that shape should be on the side of their head but not really what that shape is. the number of instances that it sees animal ears is relatively small, but consistently tagged, so it learns what the animal ears are, but not that they shouldn't be accompanied by human ears. i wonder if a negative lora could be trained on an extra ears dataset with non-extra ears regularization images to help the problem.
sleepy smol or sad smol ?
yup
charming
niice 😂
hey does anyone else have an error that makes your image just random noise/breaks it completely? im using autismmix
animalist
and here we have the maid of honor dress
and the flower girl, 😄
ok enough of that. hmmms ponder next
oop she mad
not perfect, had to inpaint a shirt.. cause umm.. yeah
ok last one. to bed. nite THUD
and one to 'grow on' 😄
thanks for the tips.
I'm guessing you're using A1111 WebUI or Forge? If so, turn off emphasis normalization in settings. It's a math error in A1111's prompt weighting code that causes some XL models (especially pony based models) to output noise.
when they said "breaks it completely" i was thinking of black image
cuz if that was the case it was the half vae i think
if that does happen to be the problem then using this SDXL vae is the better fix since it retains the quality of the original vae but doesn't overflow when run at fp16 (meaning no more NaN in VAE) https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
alternatively, there's the blessed versions of that fixed vae that I made mostly as a meme but later found that i actually prefer one of them for base pony https://huggingface.co/nubby/blessed-sdxl-vae-fp16-fix
normally i'd recommend just using the one from madebyollin which looks effectively identical to the original SDXL VAE, but if you find you want to play with some different contrast and brightness adjustments a couple of the ones in my repo can be useful. Pony was too washed out for my taste with the normal VAE though so one of the ones with increased contrast and decreased brightness worked pretty well.
for right now i'm just messing around in imagen 3, awaiting the next big anime model xD
if it happens to be XL based i can check back into those
imagen 3 - model or AI?
it's an ai model
thanks
aside from text it's SOTA
but sometimes it does text well
lol
has pretty much anything popular in it too
but like
has to be
really popular
and better at changing styles than flux
It’s strange, but for me it’s not yet possible to even just try it.
This tool isn’t available in your country yet
on gemini maybe?
gemini rolled out to more people than imagefx i think
imagefx is mainly for testing
gemini advance can only generate images.
titanic
didn't expect this char to work on imagen xD i still dunno where the popularity threshold is at for being usable

🧀
I keep trying to make it a yellow dress but hmm it likes white I guess
ya gotta pay the cheese tax
white is pure
goes good with yellow flowers and purple eyes
hmm looking at that. umm better not 😄
aww
🙈
hmm xmas in Jamaica?
grabby
this solved it! tysm
what did you mean by this? i've been using gemini advance for almost all of the gens i've been doing recently. it even has less restirctions than imageFX
I mean that without a “subscription” the image generation functionality in the Gemini service is not available
it should be available but just not for faces, but ye, pretty big block if you aren't able to get it
Yes. There may be restrictions for some countries.
Chino Kafu!?
yup
I like these visuals, also how have you been?
snek
mystery model delayed ):
finally the merging comfy model looks like is working >.<
merging with comfy is.... a mess...
many vaules are automatically set by default on auto1111
as the order of their activation
had to study what everything does, step by step XD
i wonder if i should try to make a workflow to automatically test all the block weights and save an image to have a comparison with the best result..... why i ask myself such question?
why comfy, why you push such ideas in my head?
would be easy to do, you can just copy/paste the workflow without the model loaders multiple times, then connect the model loaders to each of the merge nodes and adjust the parameters on each of those. I think comfy wants to keep all models used in the workflow in memory though so it might require a very large amount of ram/vram.
for clarity, each version of the merged models would be a separate model in memory.
i was thinking about converting every block with a widget that increase by one each time
but i need to find the right nodes
writing it by myself is out of question lol
screw the refuses!
which model is it uwu??
a giant merge 😄
it's like umm paragon, some anime, some realistic 2.5D one's . shrugs
and alot of weighting of prompt parts
haha
well time for work again. sigh
prof plutonium sure looks different than I remember
ohhh !
looks like that's the only crayon shin-chan i can get
i realized it was limiting me hard i dun think it liks making shin-chan cuz child
lol
still some bothersome limitations but overall tons more freedom than i'd expect in imagen
for sum reason i was in the middle of generating sum1 eating pizza.
i got this far
Make it pineapple pizza

for image models u have to
be very specific about that tho
but this model CAN do pineapple pizza.
unlike others i've tried
xD
just have to be specific about pineapple chunks
and i already failed cuz i forgot i needed to use the word chunks
still the problem of not actually having a removed slice
maybe i have to be specific about that
yeh that works
my favorite pizza combo is pineapple & garlic
anything else is optional
xD
grab
the hair ?
if pizza came in boxes like this i'd buy pizza more often
Cutes are overflowing can't hold with my hand
ohh nice 
niiice
that's super clean, what model did u use?
AutismMix_confetti. But this result is after 8x scale down and up (to add pixel effect).
Original image looks like this:
Oh
that's an interesting workflow
base models are getting a bit better at pixellation at least
looks like imagen's breaking point might be when u get the char closer to the frame
xD
but that's also cuz i'm using an existing char maybe
sorta pulled it off by heavily editing prompt xD
new version "Don’t Starve"?!
shrugs
if I were a producer of goods, I'd make an anime themed one for sure
pizza pizza
you turn a corner to find
oop i've seen this movie.. no no
iz random!
Got her!
oh .. wtf... ooh boy
oh she mad mad
haha
see ya later (goes to bed)
haha
goes to bed.. thud
"mom can we get tycoon bujinsword at home?"
tycoon bujinsword at home:
Plagiarist! 😁
She easily finishes the marketing war 🤣
it turned out very cute!
thanks but it was very random but it worked out
The content filtering logic on CivitAi surprises me
and this was deleted
human stupidity knows no bounds
Huge
Is this a diver?
not mine
floaties
Anyone try Cashmoneyleakaiautismpony? It's apparently a successful merge of pony and SDXL models which can use the loras of both pony and SDXL

yo havnet been here in a while
I'm trying to inpaint with controlnet. The composition is correct according to the controlnet. But there is discoloration and splotches and it doesn't look good. It goes away if I don't use controlnet
Bonk
lucky gen 
Is that an Angel?
yeah
why did they break her fingers? 🥴
model not yet ddpo
horror 😱



Huge
I can help people find video's on how to make $50 for free and I want none of the profit you make. it's all on youtube. 😄
but I do take donations of art 😄
no u get 0%
why does she eat grass? 🥴
That's why she has such sad eyes!
Did the hieroglyph on the hoody carry an interesting meaning?

wow
with regional prompter ?
No. pure prompt separated by BREAK
oh nice
feel like sd3.5L isnt better than flux, at least for anime
didn't even knowo there was a 3.5
there isnt one yet
only until lykon approves and it gets safety tuned
grabby legs 
the merging workflow is a failure 😦
dealing with the weighted blocks without causing weird mutations is the model is hard
but thats whta makes it art
yeh one of my favorite gens was from a terrible merge idid
it output the one good pic so i thought it was a good model
only to find out thatw as like the only good pic it did xD
but that one pic... was art.
this was that pic xD
turned out to be an omega super fail merge that never produced anything good outside of that tho. i shuld learn how to merge better.
WAHH boot to the head!

boo
huh....
hmmm interesting
And here we have a normal Jack o lantern.. totally.. normal.. nothing strange
Apy Halowwen
is random as random does
ponders a diff theme
haha
donald trumpin anime style,sunny v2
boy
can any one tell me that how to crate images within this server
you don't
nice
how much she has grown 😉
alot !
i'm looking for finetuning and just reinstalled all the stuff
just yesterday did anothere test where i genned bg and character separatelly and mixed the stuff
the results was pretty cool
with the right noise, the model maintain a close resemblance to the character plus gen an interaction with the bg
i'm pushing more in this direction, to have more control on the gens
tbh, i think a lot of people just moved to other servers.

🥲
guysm what version do you think is better for genning? 3.11? or 3.10?
how far you have come.. I'm still just enjoying SDXL... 🥴
damn, i didn't specify i was talking about python
lol
i was lost in thinking about kohya
what's wrong with that?
I still can't understand how it works
i'm experimenting with diverse versions of dependencies while studying fine-tuning
now i was curious if python 11 was better than python 10
for me it's "Dark Forest". but are there really such big differences in the results obtained?
dark forest?
in other words, I don’t understand this at all.
eheheh
for now i don't know, i have some improvement, but i still don't know if it's because the different version of dependencies or because comfyui got updated
something like loading the model and the workflow for the fist time is way slower, but then, every successive gen is faster
I see the same thing using Forge
well, forge is an experiment in implementing some of the stuff in comfy in auto
a definitive success
cute
thx ❤️
but damn it! why Forge works much faster than A1111. If this is an experiment!
with a bit lower quality that it doesnt even matter
I wouldn't say that. I like the results in Forge more than in A1111.
me too , i said that because i have seen some a1111 vs forge generation videos but it is not that much difference
could be 0.20%
"hand tp own ear " lol ( i wanted that )
Tp?
?
hand tp
I think it worked like hand own ear
to
oh maybe
no , now it`s "to own ears " and its the same
yup
It turned out to be a funny background.
big boob Elf
hihi first time to drop a pic
Catch me if you can
Huge

I walk through puddles - I don't need anyone! 😁
batch promp?!
Yes and no, its a list of stuff that randomly picks a line
Because comfy metadata
Weird. but apparently this is formed in Comfy by some kind of node.
When I used Comfy, it seems to me that the metadata was not like that.
Doesn't Comfy save workflow to metadata by default?
right. but all this is in the Inputs:Text section
That's why I asked about batch Promt
Apparently this is still some kind of Node.
i really love the speed of 1.5
I am trying to use this lora https://civitai.com/models/562141/pony-takeshi-obata-or-death-note-artstyle
But i don't know why is not working
Damn can you like drop me new servers
Touhou Project AI is where I mostly hang out. There's also the Aibooru server which is largely more geared towards things that wouldn't be allowed in this server.
I’d never block Yue
no, it isn't nai
If you remove the style and description of the character, you will be left with blush,;3,((red kigurumi:1.19)),hood down,animal costume,shoes,gloves,fangs,from above,((from side:1.27)) ,smile,standing,(proud, doyogao:1.1),head up,(head back, chin up:1.1),closed mouth,closed eyes,tail wag,((headpat:1.28)),leaning forward,grin,looking pleasured ,arched back,spoken heart,sparkle,motion lines
a female butcher
The gate is opening
that can't be good
uhh oh she's awake
hmm ok.. interesting
someone could make a manga out of some these...
basic armor
Battle armor
Coffee or tea or me?
hmm more streamlinesd
Get a beer on weekend!
wow!
updates website
hmm battle engineer?
whoa , not messing with him
armor ready for deployment
those lower portions look like bat'leths
oh umm swallows my... !
changes seed by 1 and turns DTG upscale prompt to very long prompt
WHOA !
o.. k
hmm super gundam mecha?
hmm mech dragon?
ok that looks more gundam
well thats not a gundaom.. wtf
guess thats what I get for putting mecha gundam varient
now this is more like it!
hmm looks like it might transform into a jet
ok 1 more and I goto bed
thuds
ok NOW I goto bed.... 😄
desert town
capture a mini creature inside a jar
Dm me Aibooru i wanna join
Someone picked up a new style where have you been? did u join a new server if u did send link huuhu
more like i've been taking breaks from ai art and such

I always see AI art in Pinterest like mostly focused on anime and specific games like ZZZ i wonder if they have a discord
It gets tiring i know omg
I just play mlbb
fwiw, touhou project ai's image gen channel is not really very focused on touhou. sometimes we go through phases of touhou meme's but gens of all anime and some realistic genres are appreciated there.
yup yup, but will still pop back to get updated about it 
Very chill place i like it there

how you guys been

I wrote anime art in the prompt, anime enough to post here
👍
"a totally cliche anime image, cluttered, sweet, stereotypical, japanese text"
"a stereotypical 1980s manga cover"
"a stereotypical 1980s manga inner page, several panels, action flow"
🤣
speaking of...
anyone knows why?
what model do you have loaded at the moment?
Pony diffusion
make sure it's v6, the previous versions were sd1.5 which wouldn't be compatible with that lora and would cause sdxl loras to be hidden. if it is v6 then the download might not have completed successfully and the lora file may be corrupted.

How does one do hi res fix on pony? When I do it, it elongates their bodies into a wraithlike horror
it shouldn't 
are your noise like super high?
Yeah I guess that's the reason 😅
yup, for hi res fix, you usually don't want your noise to be too high
There is lineart breaking up on on the high res fix, do you know why that could happen?
what does that mean?
Don't use latent
i'm super bad at reading comfy metadata, but if you are using latent upscaling like Sleepti said, then you need to change that
like for me, im using this most of the time
Anyone having luck using imgtoimg for recreating poses and such?
Use controlnet for that
Im using forge which i think comes with controlnet?
Then use it
Does it have "open pose"
That might be what you want
I forgot if there is a new better tool for poses
controlnet with openpose is the way to control pose
If i started with an image that i created can i use that in openpose?
yes, if the image in question has easy identifiable character pose
Youre the best, thank you so much!
you need to get the correct controlnet model for the same type of model you are using, example if you are using SD1.5, then you need to use SD1.5 controlnet
i know this is a dumb question but is the model number usually in the description in the drop down?
I can't remember if you need to get your own model or not for forge
You do, i have a bunch
oh wait, you are asking about a totally different thing. You meant the model type right? Like SD1.5, SDXL or others

Am I blocked?
