#🍥|anime
1 messages · Page 98 of 1
See ya
realistic ai gives me true creeps, they are grapically good, but, my god
cause it's to good?
the true photorealistic ones are indistinguishable from a photo
which got a guy in canada put in jail
don't know, honestly, some are really good, but they have something that just don0t make me feel so well
maybe uncanny valley
ok reinstalled speccy
people need to be virtuous and responsable, with power
ai can damage people, if used in a stupid way
ok enough art for now, gonna play that game I got on humble choice this month(I forgot to cancel might as well play some)
I'd rather use AI for anime stuff
maybe that's the reason photorealism gives me creeps
temtem
thats why I like the merge I created that is more anime cg fantasy
sorta real but not to far . sometimes
sometimes it does but shrugs
2.5d is also fine
@bitter latch have you tested the other games from the choice bundle? Like Roadwarden
kind of interested on it but no idea if it is good
no havn't yet
what's written there?
Trillion stone spine
i was almost sure it's some random words 
here you go cutie ||metadata|| 
Wait that text was actually part of it
sigh I hate that bug in some games where my direction key gets stuck and I have to hit it again to stop
yup, it's not inserted, that's why i was like, it's probably some random words or gibberish
oh I know I just like to translate things like that 😄
ya know I really should add some of the art I do to my site
How can I create something like this in A1111?
such a cool wallpaper damn

Why sad 
Oh oooof das fair, you need a upgrade :o
is SDXL going to improve anime things?
joins the train art .. event?
I'll make a train, let's do this.
Ooooof 
I just took wd1.4 tagger output and gave it as example for train

I merged it again with "BeenYou"- the base model is already lost.
ok lets try that again
how you did such a hight resolution??
No
Well, you should upload it 
a more sorta modern piece
Wth
people were doin' trains!
I just wanted to do something weird and different lol
That you did-
everyone always posts the same ol' portraits all day 😄
FINE .. here's a regular ol' train....
@shy totem what model are you using?
what a cute train, but, here is the passenger's vagon? lol
I merged it with a ton of other models to the point I don't even know anymore.
Name some
I don't remember the base, but I'm sure I used the pastel one, then AbyssOrangeMix and AnythingV5 then the latest would be BeenYou and ToonYou.
These are the settings I use:
hmm no train but interesting
wait.. two suns?
hmm almost looks like each side is a diff place/time?
Is this without hires fix
yes
Then of course it looks like this
ok let me do it with hi-res fix then
I dub this the horny train. All aboard?
Ah I should mention I use Token Merging and set the CLIP Skip to 2. Not sure if those make any difference but Token Merging speeds it up quite a bit.
can sdxl 1.0 run on automatic yet?
downloading it now 😮
gonna be a bit before all extensions and tools catch up
tried the animesh_prunedv21 model
sounds like comfyui can run it
hha
should be an emoji
ponders borrowing it for his server
😄
dang this gets realistic in an anime sorta way
yay my train
now lets try something different
I knew you all would love it
got me food on the way Hibachi Chicken, Fried Rice, Hibachi Shrimp, 2 Cheese Steak Roll
and she goes back to catch one.. or is she throwing it?
normal highres fix
it's freaking hot outside
move to nl, it's pretty damn good out here actually...
thats better
Today's Tifa
eats some chicken with fried rice while watching "Bubblegum crisis"
can't forget the Phiilly steak rolls. love these things
She stuck?
Everywhere I look I see that F
Kinda like a LoRA, but worse in every way. Even before LoRAs were a thing, they never caught on.
hypernets were god before lora
but it was strange they didn't catch on
I still use some occasionally
hypernets makes all the results the same, they are very boring
ahh
figured they would be better cuz of the cool sounding name
in the past, models where very unstable and generated too many mutations and they solved it like this.
any contaminated many models like a plague <.<
untrue, depends on the prompt, they would often just make everything a certain style
yeah, but the faces where pretty much the same, no matter the prompt.
often, even the prospective and their posture was hard to change
maybe because most artists draw the same face
in my experience it just felt like an early-day LORA
i feel them way less flexible than lora
lora?
oh
i love ai lol
Don't ask why she's leaning like that lmao
(Hope you don't mind me borrowing your girl for a few gens Vil hehe)
Someone else is challenging magic circle prompt
I tried lmao
Also again with the leaning
does circular platform count as a magic circle
is evident they enlisted her for her innate battle talent lol
i just had to re-seat my sdram sticks for some reason, taskmon was showing only 8gb of systemram, all fixed now
happened a few months ago too i think a piece wiggled free or something
@supple raptor you da best btw. I have been making magic all day with your model. I left a review for you on Civit. Thank you!
bone crunching sound
Thank you! I'm glad you're enjoying it!
When you high dive but miss the water 
lol
It’s been so awesome! I didn’t wanna share some of my favorite ones just in case people steal it or something lol. But they will be on my IG for sure. Lol
i finally trying kakigori
It's a very cute model, I think you'll like it
loving what im seeing
Whoaa! That's close. What Model and Lora did you use?
(why)
Hey Loved this one. Which model did you use?
It’s so good! I’m loving it. I’ll send you what I did with it
think it was lottamix
Yeah model name is in the png metadata
Good to know
got em thank you!
👀
I would watch CYYBRBICLER
bicler
cute!
this is very neat!
gonna gpu with this model all day tomorrow
gasps
umm.. phrasing. phrasing
blinks
add some ganz0 suits and you will be remembered as a hero ... 🙏 
ganz0? hmm not familiar
well that went to scary
💘
I know a Watson when I see one.
... hmm
I realized what WebUI is missing .. This is a short text description for Lores, embeddings, etc., In which at least keywords can be entered... 😭
There is the extension Model Organizer, that can somewhat do that but I tried it and didn't like how everything was in its own tab, separate from the interface I use to select and view my network extensions.
So, I just keep a very long, detailed .txt file up to date. 😅
Extension: https://github.com/alexandersokol/sd-model-organizer.git
where were you before, my savior!?
i am must try it now!
Oh, looks like they've updated it since I used it last. It may be a lot more usable now.
is there going to be any good Full Fine-tune for Anime on SDXL?
SDXL already generates pretty good Anime tho not that perfect but much better than anime's generated by previous base models.
So I bet it would make a great base model for anime generation. Considering that it got 3.5Billion Parameters meaning more room for superior quality then what current top anime models generate like anythingv3.
I'd imagine given time, yes.
well thats interesting
it was almost perfect, only if finger wasn't the part of her lips.
I just picked one at random most of the time i dont fiddel with the images so people know what they are getting with the prompts
but viewing without nitpicking, it's gorgeous
wow, is that made using SDXL?
turns off style pile
nope
then which one?
my ElegantRPGanimeMagick Mix/merge
based on 1.5?
It's built in. Just have to put a .txt file next to the thing you want it for with the same filename and enter the text you want to show up inside the txt file.
of course the first one I stopped on to take a screenshot would have that txt. 
So you have to keep up with multiple text files, all of which are separate from SD. That sounds like a major pain. Better to have an all in one list, or the extension that keeps things organized and is searchable in-SD.

Or LoRA trainers could just make these files for the user so it isn't such a pain in the ass.
Well, the Civitai helper does something similar.
one last thing about trainers from fantastic stuff!
It's a lot harder to read though 
Alternatively, if the lora trainer isn't an asshole that removes the metadata from their LoRAs, you can just look in here at the top few tags
that's not civ helper. that's built in. See my screenshot above
Then what are all the .info files doing alongside my LoRAs?
Being non-helpful? 
hehe she looks.. confused.. surprrised?
ok it's stuck on that armor
This is pretty painful for me.... where? what? and why?
It's the most indented section. The first few are what will reasonably work to activate the LoRA.
Aight. I figured out what was happening. There's a separate refresh button with Civitai Helper for the network extensions that I missed. Once I refreshed it, the network extension cards are showing the proper interface for accessing the .info files. It's still limited, but among other things, it has the option to replace your prompt with the prompt of the card, or to add the trigger words into the prompt--if they were given in Civitai.
But where are those keywords then?
If you use the Civitai helper extension, it will add them into the info file (if you downloaded the model from civitai) and you can easily access them as I stated above. The second icon from the right.
Otherwise, it's guesswork.
I think I tried this extension but did not understand how I can get the Civitai key. if I don't confuse anything
It should scan them automatically and find the models from the hash. It's super simple.
When you have it installed, go to the Civitai Helper tab, then ensure under the "Scan Models for Civitai" the model types you want to scan for are selected (usually I have them all checked.) Then you just click scan and it'll find the info, download it, and associate it with all of the network extensions and models that it can find.
ok. but i am not have this is extention
Your UI might look slightly different, as I'm using a different extension.
Yeah, that's the extension. 3rd one down.
Your tab. (Again it will look slightly different on your system, but the tab name will be the same.)
oki-doki! installed!
welp i'm going to bed
G'night!
g`night
You definitely have a vibe going. You're making me want to watch this anime, lol.
Scaning.......
Be sure to refresh your network extensions with the refresh button and the other refresh button once its done, then you'll get all the previews and .info options.
Civitai Helper!
yes, on a serious note, its a great extension
I've been stuck in ComfyUI for a week now tho
will this button be available after the scanning process?
I would imagine it would be there already, not that you'd want to refresh yet. But it could be that if you've just installed the extension SD might need to restart before the interface updates.
Alongside your network extension buttons, so on the main txt2img or i2i tabs.
ok before I really goto bed.... how does the rock pee?
He Dwaynes his johnson
runs
Scaning process is done. now i an click Save Setting?
You can, but it likely won't change anything. Now that the scanning is done, you should be good to go.
I think both extensions are more than enough for me to organize data. In my opinion they both complement each other. thanks a lot.
Awesome. Glad to help. 
very easy add new content from CivitAi... Nice.....
Does this feature work for you?
You have to hit the refresh button to the right of the refresh button in order for it to show up.
its not showing up for me, ive refreshed many times lol
I was already using it, but it broke today
me too
I haven't upgraded SD in a couple of weeks. Have y'all?
yeah, maybe its that
this is rather not an update button, but a button for adding extended information on the card
we got a really cool new feature too its called edit metadata
yeah this
It has been broken since this feature arrived
Yep, it refreshes the UI to see the extended info.
That's probably it then, yeah.
- click Refresh button
2 click botton with blue array near
imma go to sleep gn yall
Goodnight!
Yep.
yees!
Sadly, it looks like the recent Automatic1111 update interferes with the process.
When's the last time you updated SD btw?
Is there a way to install the old version?
I've read that you shouldn't upgrade to 1.5....
So you literally just got the latest then.
Sadly, my UI doesn't tell me my specific version, heh.
just the other day I put everything from scratch. and now use v.1.4.1.
Was talking to Onurcenko, since they're on 1.5.1
Which just released an hour ago
I don't know, it's doing automatic updates
I am rolling back to version 1.4.1.
@brittle rampart@untold glacier It was a big issue for me, it's resolved, thank you. 
Reverting worked? Good to know. 
Goodnight, peeps.
very cute LoRa! thank!
I used only Kakigori v2 model
btw. CivitAi removed one my pic for Kakigori v2 model page.. sadly
auuuuf
fff!!!!
didn't have all the lora's needed so I just used fluffyrock model 😄
So quick question... I downloaded a checkpoint that has a yaml file... I've never had a checkpoint that needed one before. Is there a special folder I need to put it in?
too hard to fix it

I haven't had issues with the latest Auto version - are people having problems?
Yeah it's been a bit buggy for me
Thing I've noticed most is my prompts don't stay in the txt2img boxes. Like I closed my laptop and they vanished lol
And they don't copy over like they're supposed to from png info either
I think you put it beside
I promise I use more than one prompt
oh.. in my pics LoRa`s just cosmetics.. Details, light....but model is Kakigori v2
Same girl. Same pose. Different background. I'm so creative
Gonna keep an eye on this SDXL stuff. Feels sort of bloated with the extra refining steps.
Tried out Comfy too, interesting but will definitely take some learning.
Enjoy this where's Waldo game. ||I gave up||
I don't get it
Adetailer detected 4 hands on that image
ahh
calvin from calvin and hobbs
dump of some SDXL gens. can't wait for the finetunes to come in
hmm calvin got a change as he grew up?
Whaaaaat fr? This looks too good for a non-anime model dayum
did you edit the text on the pool one?
Nope, that genned with the text i prompted it with (oops i replied to the wrong message, this was for the "did you edit the text on the pool one?" xD)
ok NOW it's calvin and hobbs (sorta)
thats crazy
I need to dowload SDXL
has higher spec requirement but very worth it to use imo
also, what prompt did you use for the pool one?
do you think 6gb vram eill be enough?
no; 8 minimum for sdxl if reqs are still the same
it's a much bigger model
and refiner goes to like 10-11 GB VRAM
although refiner maybe isn't necessary for anime gens
partly because sdxl is a 1024x1024 model too, i don't think 6GB can do that very well on any of the models
yes its enough in comfy up to 1440p
is it?
what are the vram reqs & stuff nowadays i thought 8 was minimum still
although comfy has done a ton of optimizations lol
you can do this on 6gb vram
you're using a bad vae for that image
or this workflow too
if you're not using it yet, this is the recommended vae to use with sdxl https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors (released by them as well)
that duck has a bad vae applied to it
aka one included with 1.0 i think
can't load the vae, i need the model with the vae baked in
oof... you'll have a bad time with anime gens
the bad 1.0 vae really affects illustrative stuff
which webui are you using SDXL on?
comfyuii
Made with sdxl
Then you definitely don't need a baked vae. Just use the VAE loader node to load the vae separately. and make all vae connections go to that node.
is comfyui better for 6gb?
yes
so coooool
downloads and installs sdxl
hmm looks simple enough. now what kind of results will I get or will my vid card go F U
we'll see
here goes
crosses fingers
it's moving... slowly
it's only at 30%. oh boy
well not going to be using this much until I get a better system
first gen
umm
huh
hmm getting error on refiner. odd
oh well
well tried mergin sd_xl with some preve models. yeah win10 blue screen. ok not doing that again
doh!
creates a new merg
merge
grinz
hmm how shall I test it
Dude, I appreciate your enthusiasm but we really don't need the play-by-play. You're just spamming random thoughts into the chat. Can we tone it down a bit? 😅
That's not a reason to spam the chat...
it's not spam if it's got a point to it and not selling or convincing anyone of anything. it's called chat for a reason. but ok
I'm not going to argue etymology with ya, bud. If there's someone just talking aimlessly into the chat, I know I often find myself looking for other places to be or talk.
shrugs. ok
you do you I'll do me, and all is well in the world
k... k.. moving on. next
I mean.. I am called randomtora for a reason
😄
anyone else awake ?
there we go, did a first pass then refined in img2img
with the new merge model. works well I think
just realized the reason random fish kept showing up in my renders is because of (fish eye lens) token

upsized it in img to img with some other settings. better?
or should i just try something else?
i like those starry looking skies
my first generation with SDXL
if you want better quality use this vae: https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors
should i be using vae? Heard it's bad

your output has lots of artifacts
chromatic aberration, etc
if you notice, even they went back on the vae, since they reverted the update to the vae weights xD
oh, so like, yesterday if someone gave the link it would've been bad, it should've been an earlier commit link, but today they updated it and reverted the changes to vae and now the main link is the correct link
so back to my point, even the sdxl devs realized the 1.0 vae was bad xD
Auto 1111 or comfy?
using auto1111
mmmm
Try naming the vae the same thing as your model
not very exp with using sdxl in auto1111
if it's in vae folder with same exact name as the base model
it might work?
<base model name>.vae.safetensors
or however it works
hold on im checking something
doing a plot to see if there's actually any difference
like there's literally no difference, no idea if i am applying this vae correctly or not

I see it, but it's slight.
Lemme point it out
1 is none, 2 is with
you don't ALWAYS have to zoom in
for this onei coul dsee it all over
without zooming in
like
it was present everywhere
there for example it was in the eyes
Has a much slighter effect than 1.x VAE comparison.
You just zoomed in though, lol.
what am i looking at tbh
i can see it a lot easier lol
like what am i suppoed to see
the rainbow effect & line artifacts


See #🍥|anime message
i guess im literally blind af
Tell me you don't see that
i have a trained eye for AI stuff lol
this i do see
i can see it at a glance
but to show a comparison i did zoom in, yes lol
i see it strongly on edges
Definitely a lighter difference than 1.x with VAE vs no VAE
no zoom without vs with
this is a HUGE difference imo
and this is no zoom
this is 100% scale
where is this even in the pic? lmao i cant find shit 
i can see that at a glance
No WONDER you're having problems, lol
pfff @untold glacier doubtin me. but that is non zoom regular scale
and i can see the chromatic aberration+lines
xD
im literally blind i guess
it's very apparent on edges
It's not a huge difference. It's just a difference, heh.
well, quality is quality lol
Look back at the 1.x no vae vs vae, and there's a big difference all around.
anyways, just a suggestion, pointed it out as i saw it xD
cuz i mainly saw it on the hands and face on that pic
can you like point out where is this in my pic? I wanna find it on my source pic and zoom in
tl;dr if you use the vae you just downloaded, it won't have any of that artifacting anymore, it produces the same exact image, no compositional change
learnt today that im blind af
but you see it in the hand example i posted?
that was 100% scale no zoom
Lol?
are you getting link to the wrong pic im talking about?
Ohhhh. Now I'm the one who is blind, lol.
or like am i seriously blind that this pic is at the fingers there
Thought you were talking about the other one

YOU GOT IT YYYYYYYEEEEEEEEAHHHHHH
i do see it now if i really zoom in
but can't tell any difference when not zoomed in 
guess im fucked in the eyes
well, just gonna use the vae
It's more apparent in some pictures than others, & i know what to look for

when you posted that pic, i looked for the most defined edges
like those in your characters hands
and i saw it along the edges of her hands
Ok. So, how does one use it if the VAE is already going through the refiner though?
You just connect load vae node to everything ... ah wait a1111
......Idk how automatic1111 does refiner
is controlnet only for poses or is there other uses
Tons of uses.
u dun use refiner vae at any stage of it
you use the sdxl_vae for all of it
anywhere there's a vae decode part
But... you need the refiner for proper upscaling? Right?
the vae is the same
base & refiner use same vae
ergo you use the downloaded sdxl_vae
what are the chances ill mess something up if i just try some random settings to test it out
if you look in files for base & refiner you'll see same hash for vae

The base and refiner don't use a diff vae, lol.
see here in this instance
the base + refiner use same vae
which means
sdxl_vae here
replaces both of them
If you look at the image I uploaded #🍥|anime message The refiner has a VAE output
vae for both base+refiner
should be unused 100%
that vae does not exist, it's terrible
xD
it's the same vae for both base+refiner
so you use just the downloaded sdxl_vae

here's my workflow if u wanna take a look at how i do it
But this is my first time using Comfy, so I mean
it's a full custom workflow tho, doesn't copy anything else
Thanks. I'll load it up
Base + Refiner have a baked in vae, this baked in vae is the vae that was so bad the SDXL devs reverted the upload of it
and instead made the main vae the 0.9 one which is now the one available at the main link
You use that 0.9 vae for every vae decode node.
you never use the vae from base or refiner, both are unused.
So... I believe that's the one I downloaded then? It was a separate VAE.
yea if you got it frommm
Further up in chat

this is the only vae you use
Now I just have to understand what I'm looking at. What is this?
text_l modifier
most workflows don't really separate it, yea? but
it's actually
REALLY IMPORTANT to separate it
for high quality results
ummm
it's actually kind of hard to explain how to use it properly too
ʙᴜᴛ ᴡʜᴀᴛ ᴅᴏᴇs ɪᴛ ᴅᴏ
uhhhh it's an additional positive prompt
using a different CLIP model
green = clip vit-g (biggest model, most comprehension)
yellow = clip vit-l
ummmmm
lemme load one of my good image outputs
here's an example of how i used it
but
uhhh
it's not really good to just
paste green into yellow
like a lot of things do
You can also just leave yellow blank sometimes
it's not required
and you would get a higher quality output
there's too much talking, so here's a pic
than filling it in with a copy paste of both
these pictures
would all have the full prompt loaded if you put them into comfyui
also IDK if top one had a nsfw prompt
i can't remember
xD
that would show examples
of how i use the yellow box
Lol. Thanks
I love the hamster pool, haha
lol
sometimes i use it just to try and emphasize some points more
xD
sometimes i just put
Alrighty, thanks so much. We should probably let the anime come back to the channel. I appreciate all the help!
True.
i call it "style text" because i sometimes put the style there, such as like, anime illustration, on paper, drawing, etc
Essentially, it'd be the text on an upscale rather than the original t2i. If I'm understanding potential application uses correctly.
perhaps
understand that any other workflow you used til now
likely didn't separate it
and the green box
was likely both green+yellow
but through much testing i found that's 100000% not how you do this xD
anime-style stuff comes out SO much better if you differentiate the two too
Considering I haven't used comfy until today, I'd say that I haven't dealt with anything like this before, heh.
u were saying refiner was using a different vae or something?
her lips 
idk how the workflow for 1111 works at allll
i didn't say anything about refiner, because that's like the 2nd step, and im still learning the first step
and for auto1111, you need to swap the model before using it, and it takes a long time to do that
Style: Anime Positive: anime artwork {prompt}. anime style, key visual, vibrant, studio anime, highly detailed Negative: photo, deformed, black and white, realism, disfigured, low contrast
Ah okay
already got that in my style
nope

and yea they work pretty good
yup yup and added it to a style stuff
you got to modify them? 
owo
xD
got to go get my food
What... what am I looking at
it's lunch time
ewwww the reddit post format
is not very clean
this is so much better
well, that comment below is for auto1111 style
Ah i see
Ah... so it's just additions to the prompt. I getcha.
yea what i do is take like
so if you are using auto1111, then it's great
the words i wanna use from each one

i usually use the full negative
but then cherrypick the words i wanna use from the style
I'd love to, but I have to wait for the SDXL 1.0 compatibility.
and then add my own style text too
Anyway, go get ya food
it's already working, no?
Wait what? With the refiner?
don't go back ;_;, comfy is so much less resource intensive too xD
havent try that tho
crazily faster.
okay, time to get food fr
ye
You don't understand the hundreds of hours I've put into making the best auto workflow and gen results.
it'll never be as optimiiiiiized ;_;
It'll be more useful though, and that's the most important part.
okay anyways pure sdxl/ui discussion might be getting off topic xD
I'm thinking that it's likely that most of the channels where spirited talk happens is having a discussion in the vein of SDXL, heh.
but yea if ur trying sdxl anime gens this + a little of your own modification to fit your own style
works pretty damn well
Nice
period after prompt cuz it's better padding iirc
than a comma
Been debating if i should try and add like KHFB/Auroranegative to this but
errr
Why do you say this?
i remember at least a1111 had a setting for it
so if comfy's dynamic prompting is similar
a period is more powerful than a comma
in making it keep focus with what came before the period
before continuing to the next stuff
for anime i still think we need a finetuned model... Kind of. I've gotten some amazing results that needed no alteration but they're pretty far apart 
can do chibis amazingly tho LOL
for, i guess... daily life prompts?
stuff that isn't just a subject and an artsy style/background/etc
takes a lotta work to make those work
can sort of be alleviated by some of the old tricks like saying detailed face to draw attention to the face tho i guess
Once Auto is up and running with SDXL 1.0, Adetailer will fix that issue quickly.
Auto inpainting faces has been a Godsend.
Sweet






back with food