#š„ļ½anime
1 messages Ā· Page 182 of 1

mittens, the easy way to fix hands š
i don't know if i should use the face adetailer for hands
i noticed that the noise have the form of a face
I've been using it with my current anime merge, it cleans up the faces nicely. I don't use it with some of my other models though, especially realistic ones
i noticed that it pretty much works, but i'm pretty that with a node for hands, i would get even a better result
I use adetailer for hands with depth controlnet and depth hand refiner with 30 Steps restart and 0,45 inpaint denoising
Gives really good results for hands so far
there is not adetailer specified for hands, right?
There is
on comfy?
that's what i'm doing, but i just noticed that that node have something in the noise thar resemble a face
i supposed they did so to improve the result
Thats what im doing right now, but a1111 has the new depth hand thingy inbuilt into the extension
Which works a lot better from my very small experience
there should be a setting to change detection strength, might need to increase it it it's detecting random things as faces
you are using squared lines, curse youuuuuuuuuuuuuu!
all the workload must rigorously be a spaghetti mess!
in the a1111 extension it's called 'detection model confidence threshold', defaults to 0.3 and detects all kinds of random stuff as faces sometimes
I need it clean for my OCD
I even enable grid lock in comfy so everything is in line
i found my nemesis O.O

you monster....
š¤£
i'm not much different from a chaos agent lol
that's how a workflow should be
a very big efficent mess
I have chills lol
because we're not allowed to have good hands
it was using the subject input latent image as reference instead of the processed input :<
downside when using anything anywhere š
ai really hate good hands
maybe we should do a hand model, tagged with hand poses XD
more a tool that has a left and right hand 3D model and it poses it to the estimated pose the BBoX detection sees and than generates from that data.. literally generate from observation than mathematically calculate where each pixel could be
some suggestions for and fixing prompts?
just focused for hands
use the most recent models SD1.5 i would say . as majority of bad hand data is pretty much diminished with the user trained data and merges
and all the usual negative prompts š
Very nice, what's the model and prompt?
there should be metadata embedded
should
or im just wrong af
1girl, nakano yotsuba, orange hair, blue eyes, green hair ribbon, bathrobe, onsen, fog, smoke, sitting,
<lora:nuanyang_20240104200822:1.0> nuanyang <lora:Yotsuba:1>
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, negative_hand-neg, text,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 22, Seed: 2876177536, Size: 512x768, Model hash: 51e4b0d5e1, Model: SoushikiV1.0, VAE hash: 0468eb8aab, VAE: kl-f8-anime2.ckpt, Denoising strength: 0.51, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.1.0, CFG Rescale phi: 0, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Lora hashes: "nuanyang_20240104200822: 8a9374cfc553, Yotsuba: bb3f8a3fbf3e", Dynamic thresholding enabled: True, Mimic scale: 7, Separate Feature Channels: True, Scaling Startpoint: MEAN, Variability Measure: AD, Interpolate Phi: 1, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4, TI hashes: "bad-artist: 2d356134903e, bad-image-v2-39000: 5b9281d7c6ad, negative_hand-neg: 73b524a2da12", Version: v1.7.0```
most of it is in my default settings so 
Gonna try shoshiki model
Damn right looks amazing
I'm still working on my negative prompts
Idk what to add there yknow
well, i should probably remove some stuff from mine as well, but too lazy to experiment
if the model is good, you usually don't need to have a whole lot in your negative
#š„ļ½anime message credit nubby
That's true
Model I use is centeurfeit combined with a dnd one
But I get alot of limbs popping out and multiple arms out of one
Any negatives I can do that?
Make that don't happen
I've heard counterfeit is hard to control in those terms
I'll try that but even when i add bad anatomy and easy negative embeddings it still happens
i guess so. For me, counterfeit model doesnt really follow prompts well
I've sorta got it controlled
Lemme get examples
@lyric zealot it's hard to control but makes very very good art
Available at https://tensor.art/models/597922149832853677 ! Good day, everyone! Thank you for your interest in my model, this model is based on my ...
It looks good but counterfeit gives more grit if that makes sense
BluePencilXL v3.0.1 dropped 2 days ago on huggingface. Looks like they fixed the 3.0 borked VAE
wow, pika looks insane
Guys doesn't what civitai did break the SD license?
I think it was called creativeML or something
Damn this one so good
Was just about to ask
Itās really mostly the work of the Lora
Pretty new Lora so 
Nuanyang?
Yup, that
Is it published?
ę¬ęØ”åéēØäŗēęę“ęåēå¤é³čęÆä»„åä¼ č¾¾ęø©ęēęč§ļ¼ęØčēę768*1024ēéæę宽å¾ļ¼å®å ·ęäøéēę³åę§ļ¼åē¬ä½æēØåÆä»„åØęé1äøļ¼äøå ¶ä»loraäøčµ·ä½æēØčÆ·åä½ęé. This model is suitable for generating a softer sunset backgroun...
This?
Does anyone know?
I'll have to try it
Still need to figure out prompts
I barely understand
(insert)
Like wtf that mean

Spinning anime girl I see I see
I see people's prompts have (bad anatomy:1)
Means (hello:1.1)
I see I see
Like if you really want something to happen example āsittingā
Like if you include āsittingā in prompt but subject isnāt sitting
You can try raise the weight
To like 1.2
Are you using auto1111?
Yes I am
So I'd have to do 1.1?
Okay, go highlight a certain word that you have written in your prompt
Then press
Ctrl + up arrow key
Or down
What would that do?
Did you try that?
Aren't at home rn so I can't....
wtf is this wizzardy? have you cursed my pc? where comes all that stuff? lol
It works right?
Like it will go from
So it would Increase it like 1 to 1.1?
like a charm
(Maroon hair:1.1)
1girl to (1girl:1.1)
it's crazy stuff
Yup
And
Another thing
Have you ever wanted to move certain prompt forward or backward?
Just highly it
Then alt left or alt right
I didn't think the position mattered???
And thatās where you are wrong
It matters
So last in is what is least focused
Well, yup
But auto1111 got a feature
Of normalising all weight
So like
If you have one at weight 1
Another at weight 2
Then I think it will be both at weight 1.5?
Or something along those lines
Itās a little confusing
But you donāt really have to worry that much
Nope
You shouldnāt
Ahhh I think I get the example
If you have all your prompts :1.5
Then itās gonna cancel out
And basically be 1
Since all the ratio is the same
So basically, donāt increase all the weight of everything
something like splits all the weight so to not overburn the image with too much weight
So I should have (bad anatomy:1.1), nsfw??
so if you have 100 100 100, they became 33 33 33
As an example
Yup kinda
And if I wanted no nsfw at all I'd do
Well, first I would usually try without weight
(Bad anatomy:1.1) (nsfw:1.5)
Then increase weight later if need be
1.5 might be a little high, but depends on model
yeah, nsfw as first token and then add weight to tune down the image
usually is enough to get good images
Wait so I'm confused let's say i put example (clothes:0.5) in a positive prompt what happens then does it make the have less clothes????
even if some models are an ecception
Nope, just less focused that prompt
yep
So let's say I do (nsfw:0.5) would it make that prompt less so it be more sfwm
?*
Cause I be congused
Confused*
the weight increase the relevance of the token
If you are putting nsfw in positive prompt
with a negative value, the token get less priority, confronted with others
Then itās always gonna be adding and not subtracting
Just how much of it
So 0.5
Means, give nsfw half the attention as compared to the other prompt
Ahh I see now
So should I add stuff together like (1girl, white hair, white eyes:1.1)?
That ties together
Itās kinda like different children
How much attention given to your children is different

That's a sad mood
Iām joking btw
I'm a horrible father
Well, if you want, then sure, but that should not be your own prompt
Cat dad
š
they are just generated first, in the same order you wrote them, or something like this
first a girl, then add hair and than eyes
but than you add a (fire:3.0) and this change the priority of the gen lol
than you add (fire:3.0) 10 more times and the gen totally forget you even mentioned the girl
Exactly
that's pretty much how weight works lol
Also what are some good basic negatives to use
that's how i go from a cute girl on a fire background, just a firewall....
sometimes increasing weight just doesn't work if the model don't have that concept...
most importantly imo is that tokens effect each other in a cascading manner. So if you have 1girl, long hair, red hair, blue dress and the dress keeps coming out red because of the "red hair", try changing it to 1girl, blue dress, long hair, red hair
and that's where lora and embedding comes
they add concepts missing in the model
oh, to prevent the leaking of the color
This is all I use on 99% of SD1.5 models. (Worst Quality, Low Quality:1.2), border, 2girls, nsfw, skimpy, grayscale, (watermark:1.2)
usually i don't care about that stuff, i love when chaos happens, but i suppose leaking color can be fixed
all of this in positive
There are a lot of problems that can be fixed simply by changing the order of your tags aside from just color. If something is being ignored, try moving it earlier in the prompt before increasing emphasis on it. If something seems like it's having too much effect try moving it later.
Then everything gets even more complicated once you go over 75 tokens, because then you're effectively dealing with the same thing in multiple prompts at the same time.
It'd be nice if WebUI would give some sort of visual indicator for where prompts are being split.
worked a little on hand adetailing, they still...
Imma yoink those
what's worse is that comfy don't even have a counter
Yeah that would be nice to see what ties to what and what influences what
Are you using this method? https://github.com/Mikubill/sd-webui-controlnet/discussions/2400
They have it too https://github.com/Fannovel16/comfyui_controlnet_aux/
Holy, that actually looks good
I'll give it a try
oh, the graphormer, that's a big no, that control net misses most of the hooks
decided to try controlnet yesterday... after 500 years
i experimented all yesterday and i got good results only on realistic not too deformed big hands
I'm currently having a blast
the control net fails to detect and fix small anime hands
I was trying it out yesterday a little in A1111 and didn't have much luck with actually fixing them. Detection was no problem though.
and for some reason i even got a weird error, so the workflow failed to load it on ram
it was a pain of a test
This was my first attempt 
looks like attempted to inpaint the hands with the prompt of the image
That's exactly what I did. According to the guide, you shouldn't need a hand specific prompt for the adetailer pass because of the controlnet. Clearly, that didn't work so well at the recommended 0.7 denoising.
so the process is, you get a mask from the hook, still from it and controlnet you get new hand, and than the model gives them color
the denoise was too much?
The preprocessor is creating a corrected depth map that approximately matches the original hand position. Since the idea is for it to be able to fix things like too few/too many fingers, it needs high denoising to work. Personally I haven't gotten it to work well yet though.
or maybe the controlnet didn't kick in for some reason
or his weight was too low
the new hand where correct?
before they get generated on the image
Unfortunately I can't figure out how to get it to show me the preprocessed image when using Controlnet inside Adetailer so I can't tell for sure whether it's working correctly or not.
Yeah, we need Bing-su to add it to the Adetailer extension.
I guess I'll gen an image without it and try using the controlnet for manually inpainting so that I can see what's happening better.
Ok, I see what's happening. The problem is because I'm using the inpaint only masked option. It's creating the depth map for the full image, but not cropping that when applying it to the inpaint.
Does it work now?
Original, Inpaint Only Masked, Inpaint Whole Picture, Depth Map
that's so cool
@lyric zealot ur a wizard for the alt and ctrl thing
bless you, you have made my life easier
wtf... I just switched back to Only Masked to try to grab a screenshot of the preview where I could see that the depth map for the full image was being used and it generated a cropped depth map this time.

hey nubby, do you have any tips for gens?
Way too many to list. If there's something in particular you're struggling with, I'm happy to try to help.
My hair colour is leaking into clothes
and background aswell from the looks of it
What's the prompt?
1girl, short hair, dark orange hair, orange eyes, business suit,lora:add_detail:1
very simple but is leaking in
maybe i should put short hair after the colour prompts
im not sure about it
try this: 1girl, black business suit, short hair, dark orange hair, orange eyes <lora:add_detail:1>
looks hella good still some colour bleed but way better
entire background is orange tho...
We're doing that here because either the hair or eye color is leaking into the suit. Tokens effect the ones that come after them the strongest, so by moving the clothes to before the hair and eye color in this case, and also providing a color for the clothing, it should help to reduce that chances of that happening.
I see so should I do eyes colour, hair colour, short hair?
so it might cut off the bleed
wasn't there an extension to fix this?
no, separating them more in the prompt will be helpful. I would recommend adding any description of the background between business suit and short hair.
The cutoff extension, sort of works, but also has it's own set of problems.
i see, thanks ill try that
oh yeah
was trying to figure out the name
trying to test the hands fix thing... proceeds to get perfect hands 5 times in a row
the fuck is this shit
i hope it's on auto lol
if not, i have messed my workflow pretty badly lol
wdym?
Big 
Smol 
you use auto1111? or comfyui?

auto
so, i have to blame the new nodes... no! I MUST BLAME THE NEW NODES! š¤£
the new nodes doesn't work lol
It's struggling really hard to detect the hands for me
š
in other hands, i'm setting the adetailer
i found out the adetailer is very cimpatible with lcm lol
my depth map is this 99% of the times
remove (dark room:1.3) from the prompt
no
so, we have to work in the dark i suppose š
try to generate a photorealistic image
i noticed this model works better with photorealism
nice
https://github.com/savbell/whisper-writer
this program is pretty cool if you want to try it
Just speech to text on your whole system.
You can run it using only local Whisper models, because the Whisper models are open source.
Well, it depends on how you define open source, but yeah.
Mine has rarely failed to detect the hands, it just hasn't been very good at fixing them. Feels easier and faster most of the time to just manually fix them and pass through inpainting at low denoising if necessary.
whats the sampler you use most?
DPM++ 2M Karras
It's worth trying them all out on some of your average prompts and models to see what works for you though. Some prompts and models just do better with others. If DPM++ 2M Karras at 20 steps fails me then I usually try DPM++ 2S a or DPM++ SDE at ~15 steps.
how do i do prompts to make them look older?
(older prompt:1.3)
speaking of hands
how the hells did you do that
text2img
pos(good hands:3.6) neg (bad hands:3.8)
looks real
I'm using a lora that I trained on some comic style art. Mostly trained on images that I made using img2img until I got a decent drawing. It usually adds a paper texture to images that I like. I need to cut down the dataset and re-train it to be better I think.
very cute lora
Some are nice, and some are ruined because of the hands. And sometimes it gives hrer extra ears. And sdxl doesn't know how a person hold a sword.
What?
The index finger on the left hand doesn't make sense in that image. And the top of the microphone is at the wrong angle comared to the hand. It's only ok because it's better than a lot of other sdxl generations.
it was a joke
well, it wasn't wrong
Oooohh
XD
there is a story behind this image XD
i was trying to gen Iori Yagami in a fighting pose
and that's was one of the first images i ever generated
it was nothing i was looking for, and all the pose where ridiculous like this
sick of it, i attempted to gen it with a mic in the hand
failing again
so i got bored and edited a mic in his hand with the word enka, under the influense of pure delirium XD
I knid of wonder how many images sdxl and it's finetunes have been trained on where the hands were just barely cut off. I get so many like this one.
I don't try to fix hands anymore. I'll generate something that almost has a good hand, then try to inpaint it and every attempt is wors than the original.
I don't think I've photoshopped a good hand in first though.
That moght work.
when the model don't know how to gen something, it simply cover it
usually i fix hands, by hands lol
simply redraw them
Square ratio will more likely result in less hands
True. You get a lot less full bodies with 1:1 aspect
expecially when you don't prompt full body, or something regarding the feets or boots
@steady grail
what is your picture
that one lol
that model failed me too many times XD
We need to up the bodysuit ratio

i would love to know if there is a way to stop comfy from loading and unloading the upscaling model š¦
tiling upscaling like this is hell
charming
@stable crane 
and a DBZ scouter
right now i only have reverse bunny suit pics š
ah, rabbits with human legs and hands I see
I'm feeling blessed
Why is that 
That bunny one?
this one

maybe I am just really digging the proportions
and I found a rather alright way to upscale
but I can use this style on any subject
name a girl-type and I'll try it
do green eyes elf
ultra muscular hatsune miku flexin biceps
Sir, bing blocked me with that prompt, it is extremely unsafe for works I'll ask you to try again

mfw creating my twitter acc
cool
Miku if she was a Baki character
Okay I give up on any Hatsune Miku, bing just blocks all prompt, maybe it hates muscles 
I toned down the proportions
X3
tyyyy

ono! did I miss a free steam gift card?
You missed the free-est of all gift cards


What is æŔāé˳ lora?
ę¬ęØ”åéēØäŗēęę“ęåēå¤é³čęÆä»„åä¼ č¾¾ęø©ęēęč§ļ¼ęØčēę768*1024ēéæę宽å¾ļ¼å®å ·ęäøéēę³åę§ļ¼åē¬ä½æēØåÆä»„åØęé1äøļ¼äøå ¶ä»loraäøčµ·ä½æēØčÆ·åä½ęé. This model is suitable for generating a softer sunset backgroun...
Have a nice day everyone! Guys, I have a question. I once read an article that there is a certain pattern in the stability of SD operation depending on the browser, is this true?
What do you mean?
For example, does it work faster and more stable in Chrome than in FireFox?
well, SD doesn't run in a browser, so it doesn't matter
depending on what UI you're using, it might make a small difference
More stable in chrome
I've never had anything be unstable in chrome, but I haven't tried anything else. If you're low on vram then disabling hardware acceleration in the browser might help performance, but I'm pretty sure they can all do that
And don't use adblock
Other browsers might have some bugs, nothing major I don't think
I use adblock, doesn't seem to cause me any issues
but, I'm not hurting for performance or resources, so that's probably the actual problem
Many people have had issues with adblock, me included lmao
hmmm
Just like sometimes something disappears in your UI or is delaying.. idk those things happened to me
I gen with youtube playing and like 12+ tabs open, hardware acceleration is on, etc etc
true.
but normally i'm not maxing out my vram for normal gens
I use Firefox. and sometimes there is indeed a delay in the response time of the interface. or, for example, a gradual, non-sharp increase in generation time after what period. and I have to restart the server
this is flower pollen mega close up
Funny as you said that, I got this image randomly
She's using the pollen magic

then she can be a elf!
eepy
i would eat that cake tbh
oh theres cake
š°
borgar
@dull jackal
me thinks someone got hacked
They probably typed in their discord acc there
and of course got hacked in the process lol
haha
Yummy
oops I typed mike elf instead of miku
even moar burge
Giant burgee
bites ear š

Experiments gone wrong
Burger Queen miku
this is actually another level scam. The URL looks legit, but if you hover over it, you know it's fake. Take note not to press this link
if you wanna press it, do it in a enclosed linux machine
Surely it won't suddenly steal my discord acc
I feel like I'd need to connect it to steam and then discord somehow
depends if you have 2fa
Taking one for the boys
if you do, use it on some browser that doens't have any saved password
and private browsing
Huh, what is this super secret link changer
dont hammer me tho
zombie
hammer smol sus he kinda sus
[www.google.com](<https://www.youtube.com/watch?v=dQw4w9WgXcQ>)
@mods

Wait, I know that yt link.
Xcq
The famous
That video won't give me up

LMAO
brother
he literally sang for countdown 2024
for some country
literally rickroll into 2024
A legend
This is pretty dang scary
It's the violence path for her now
well, just really always look at what's the proper link first before clicking 
I don't click anything 
I never click links from strangers as enticing as it may look
Shortener, yeah
Spam elf girls no clicks
zombie
Miku
Sushi
you're boys all so fast... my computer can't keep up with you
I'm just using old images
Wizzard!
I have few of these already
in one folder?!
hmm
Okay so I don't know why put all of those 90 folders had duplicate images for each month
my output have only 6500 files
I don't even want to know how many I would have if I saved each gen automatically
polka from Miku
Hullo, hullo! How is everyone is lovely morning?
I need to go buy groceries but for some reason don't feel like doing that today
and here itās only -14, but a blanket and tea raises it to +24
š¤
srry this is the only char on my dreams š¤
hmm who she?
a kunoichi
2/10
No horns to grab or ears to bite

but she does bite 
This is why I'm under a fuzzy blanket
Well I guess I'll just go get salad and sushi box from near by kiosk
odd style
why?
hmm kinda looks like a photo negitive of a black and white picture
creepy
if a creepy impression remains then the goal has been achieved!
creepy?
this guy thinks it's only 10k
bit bigger and bit cute
anyone know a good model for weapons?
also any god ones for a eldritch type of water spirit?
scene of the destruction of the great and terrible Cthulhu?
barak obama
If it's too specific then you'd probably need a lora
eldritch type of water spirit
cute loli with blue hair and eyes?
more intangiable spirit
Is this so that there is all sorts of transparency and so on?
@supple raptor how does adetailer work?
My simple understanding is that it uses a model to recognize particular features of a generation (like a face), and then re-renders that section at full res.
AI trained on recognizing faces/hands/people looks at image > masks those it found and passes it to the sampler that will only regenerate what is within that masked area at full resolution and blends it on the existing image> hopefully gets an output that is better than the input image
Model?
custom š
Based on?
i got a couple questions about merging?
what do you use to mrge?
merge*
lets say u like the look of one model but another has some pop that u like how do u take the pop
if u know both models recipe and know about both checkpoints weights u can choose what to merge if u dont u just gotta randomize and test if it didnt come out how u wanted then try again with another random weight merge > repeat til u get style that u want
i see i see, thank you?
cute
yo mate, what does vae actually do?
if i'm not mistaking, it decode from data, to image
like affect image wise?
so it dosent change the appearance at all?
it only affects the colors and some artifacts it cant change composition of the img
VAE is like extra shading and color correction
and merging pretty much this...
like someone said , find somet that compliment each other or maybe add to.. shrugs
they don't always work the way you want. good at some things bad at others
doh
she's melting
oh nm haha Miku Hastune
take 2.. eww
yeah thats a fail. ick
tho I'll try other samplers to check
yeah it's burnt
I just sorta try things and see what works. i'm random that way
huh this other new one I made , came out more realish
I have no idea who this guy is, not hulk hogan as spider man
tho I guess it's good for 3D realistic
shrugs
merges in a mix of furry anime to see if he can make it better, but does weighted sums instead
huh.. it's going lewd on me.. haha
yeah it went realistic with no styles
what happens if the same string is in both the prompt and the negative prompt?
~touches~
i love water now 
Huge
How long does the merger procedure take? and what does time depend on?
min to 3 or so
and settings I guess.
I'd read so much in that library
i dont understand what this book says, would u help me out?
Grabby book 
@native halo bing also likes books 
guys do you think the singularity will happen? when?
when you jpegš¤®
looks like Yukinoshita Yukino
Yukinoshita Yukino neko
she could help me with this massive book
the tertiary model, primary and secondary what do those mean?
I downloaded the thumbnail kek
primary is the first, secondary is the second and tertiary is the third
depending on what merging method you use, it uses them differently
so primary is what i want most from
look at the top
huh
since they chose the difference thingy, you can see an explanation of it on the top
i still dont understand it
i see so which should i use?
there is no should
what findings?
im merging counter feit pastelmix and lympus
see if it makes beautiful art
so what is the config
yknow what we ball
the interpolation method doesn't matter
ofc it matters lmao
also checkpoint merger saves the 6gb file then u have to load it and test it,model merge is better because it merges model to ram and u can test it right away without saving it to disk
i see isee
its already saved it
i hope i get all the good parts i like
counter has good details, pastel has good colouring, olympus has good shading
if u doing a blind merge u gonna have to create a lot of checkpoints and test them all to see which one u like the most
for example add diference mode its checkpoint A + (checkpoint B - checkpoint C)
i see
so ill have to try different gens to get what i like?
it always does a long neck
long neck? idk whats that
neck is long
she is a giraffe girl!

That convo reminded me of Mike Tyson's interview.
"I broke my back"
"What do you mean by that?"
"My back is broken"
Thpinal
What model is that? Looks so clean
NAIv3
wow i love the lighting of this pic and how it has that all green nature flair i would even set it as my phone background if people wouldn't call me out as a weeb then
thank you!
merging some models
good luck with that!
i don't even know which model i want to use
i have not been touching sd for a while and now it's hard to keep up with whats š„
counterfeit and pastelmix
i tried olympus aswell but didnt give me the deatil i wanted
you mean the pastelmix from meina?
yeag i liked that one pretty much
never heard of counterfeit and olympus so far
Did that newer ip-adapter stuff come to Auto1111 yet?
can't remember who was messing with that...
I think it was @gentle musk š
counterfeit is hard to handle
mixed with pastel it makes beautiful art
No idea, I don't use A1111
They keep updating stuff on it, they added a FaceID one for SDXL last week. A1111 doesn't update quick enough to keep up.
Ip adapter is out on a1111 yea
when you don't have to deal with the gui, things are faster
are you using a character lora, or just prompting for a specific look?
Just prompt for the character also that image should have pnginfo
nice, seemed like a lora because of how consistent the look is
eepy
Fellas is it wise to share a model you merged?
That's what most of CivitAi models are
hey vil
how do you make your own model
I don't. I gave up long time ago
That's probably easiest way
how do i make one?
the latest face one is?
oh nice, have you tried it?
Yeah, it seems similar to the SD1.5 one. Not perfect but pretty close. Highly dependant on the input image.
yea ive seen it around havent used the face as much i use ip adapater plus only
https://huggingface.co/h94/IP-Adapter/tree/main/models theres a face version on there too
Yeah that's the older face models
They have FaceID now https://huggingface.co/h94/IP-Adapter-FaceID/tree/main
Which uses Insightface
not sure about that one in a1111 would have to just use it and see
It likely won't work because its not going to try and run insightface, which is required for it to work.
Well it might "work" as in it won't error, but it won't work properly
BluePencilXL v3.0.1 appears to have fixed the baked-in VAE problems of 3.0 https://huggingface.co/bluepen5805/blue_pencil-XL/tree/main
can you guys gimme propmts to see how far this model can go?
1girl solo breasts looking_at_viewer blush black_hair hair_ornament jewelry medium_breasts collarbone upper_body yellow_eyes grey_hair multicolored_hair earrings parted_lips lying teeth japanese_clothes medium_hair hair_flower on_back armpits black_kimono
bnet
bet*
has a few words that models love to fail at š, without using lora's or other optimizers
love this plugin š
cant that one get you banned i heard?
ill do it in a second
Bad news... am going to bed all
GN
Haha guys I tried ip adapter in comfy and it didnāt work. But on webui it is working beautiful. Managed to transfer style, depth and pose at the same time for one character. Hehe will post an update in a minute
Here is the image, not bad, couonterfeit of course adds many details to the image along with the depth and pose,
Finally I can exploit those 8gb ddr6 that this card supposedly has
black and white, kinda different
I like when people do different stuff š
Attempts in SD to be riding a shark...
Versus Dalle
Started down the path of reinstalling comfy and setting all this up, but just gave up after like the 9th thing I had to manually pull down to get it all setup. Just don't have that kinda time. š
Can ayone help me out to set up the new depth-hand controlnet model for adetailer, is not showing up at all, I have a symlink running with the controlnet models on another drive but the adetailer one is not showing up ;c
New one 3600x1872 ^^
anyone got the auto prompter link
Not even sure what that is. š
(it's useless)

It's called auto complete
I believe
Just find in the extensions
@native halo is here 
im still lookin for that extension that writes the whole prompt for u then clicks generate and creates a masterpiece without me havin to get up chair 






haven't done one myself