#🍥|anime
1 messages · Page 147 of 1
gotta catch them all
same seed and model, just "cinematic film still a cute anime dog boy ((holding up a sign saying "i am lazy"))), cherry blossoms, impish grin . shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy" instead
ok back to anime
It works with no extra work in A1111 V1.6. Not as vram efficient as Comfy though.
borrows
tries some more xl models
hehe
that figures. wtf This video contains content from UMG, who has blocked it from display on this website or application
Watch on YouTube
the way her twintails flip outwards reminds me of Midoriko
xl just takes to long on my machine
xl takes too long even on my 3080 12gb
what's for dinner?
how long does XL take on your 3080?
whole tomato soup?
I think it was around 1it/s at 1024x1024 I don't remember for sure since it's been quite a while since I've even bothered to try it.
same prompt and model, different negatives.
the amount of change in neg 8 would rule it out for me. this is my neg vs no neg
what is neg8?
hang on
(monochrome) (bad hands) (disfigured) (grain) (Deformed) (poorly drawn) (mutilated) (lowres) (deformed) (dark) (lowpoly) (CG) (3d) (blurry) (duplicate) (watermark) (label) (signature) (frames) (text)
fun fact, you could remove all of those parenthesis and it would be the same since they're all being normalized to 1.0 anyways.
I'm surprised it's changing things as much as it is. You might want to check to make sure you don't have any embeddings that use one of those tags as a filename.
it's clear
"monochrome" does a lot, but something else there is a big one
it's also a very minimal prompt in some of those examples
here's one with a more descriptive prompt
the change from brown to red hair in this one is what really surprises me
prompt: masterpiece, best quality, BREAK, woman in an evening gown, gala, city lights, blurry background, ((modest))
neg1: ((watermark)), ((signature)), ((logo)), bad anatomy, disfigured, deformed, malformed, mutant, gross, disgusting, out of frame, poorly drawn, extra limbs, extra fingers, missing limbs, blurry, out of focus, body out of frame, cross-eyed, ugly, poorly drawn face, mutation, mutated, ugly face, extra legs, extra arms, blurred, strange colours, boring, sketch, lacklustre, extra_limb, poorly drawn hands, messy drawing, text, malformed hands, malformed face, crop, extra limb, two heads, deformed hands, mutated hands, bad proportions, cloned face, gross proportions, malformed limbs, fused fingers, too many fingers, wip, fanart, crude, rough, amateur, speedpainting, draft, misshapen, defective, weird, broken, disassembled, mangled, unfinished, disproportioned, disproportionate, unappealing, amateurish, basic, beginner, training, grotesque
neg 5: Frame, framed, ((text)), grainy, low-res, titles, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, ((extra limbs)), cloned face, disfigured, gross proportions, (malformed limbs), missing arms, missing legs, extra arms, (extra legs), fused fingers, too many fingers, long neck, ((ugly)), fat, obese, chubby, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra limb), (poorly drawn hands), messy drawing
just for reference
it like pads out the full character length of the first prompt set
someone like nubby can explain better than me, for sure 😄
Just did a test with my 10gb 3080. Not too bad. 15 seconds for a standard size gen, 15 for an 1.5x upscale, and after that a superupscale 2x. The total was 220 seconds, or three and a half minutes for all of that. Not too bad with ComfyUI. Auto1111 would've taken forever. Likely at least half an hour.
Assuming this is with AITemplate? Yeah that's perfectly fine.
? What's that?
Essentially, you're taking the tokens and stopping them where it says break and then continuing with new tokens after it. This means that CLIP's attention is split between before and after BREAK, where the weight on prompting is even between the two.
If you say "1girl, sitting on a bench, park BREAK cinematic, fall, colorful" Then the attention weight will be split between those two sections.
SD with Auto's build already does this after you hit 75 tokens, but if you want to make sure your gen's prompting is properly weighted the way you want it, it's best to use BREAK so you have that control.
So, CLIP is only able to process blocks of 75 tokens at a time. This would be our maximum prompt length, but someone thought of batching tokens to process 75 at a time then concatenate the results from each block. BREAK is a keyword in both A1111 and ComfyUI that fills the rest of the current block with padding tokens so that the next token will be the start of the next block.
c++ implementation of pytorch basically
I'm not certain on the code particulars of how ComfyUI does its thing.
maybe Applebees tonight
Stand up. If you're s/itting down, then it'll take longer.
you probably exceeded your vram and have things being offloaded to ram, or maybe even swap file.
do you have hardware acceleration turned off in discord and your browser? those can eat up a lot. Nvidia Broadcast is another somewhat common one. Sometimes dwm.exe will randomly decide that it needs 2+GB of vram and refuse to give it up.
- go to the details tab in task manager
- right click on one of the headers
- click "select columns"
- check the box next to both "Dedicated GPU memory" and "Shared GPU memory" and click ok
- now sort by the new "Dedicated GPU memory" column and you'll be able to see what is using vram and how much.
it's not super accurate, but it's usually close enough to find the problem app and either kill the app, change it's settings, or determine that the PC just needs restarted because Windows sucks.
mystery solved
Yeah, because of AI I usually use Edge now. I used to use firefox for all things--I still consider it my main browser--but most of the time I have edge open without hardware acceleration so I can gen without problems.
Wait
Check this first
Task manager > details tab > Dedicated/Shared GPU memory.
What's taking it up?
Well, if you've already checked it and saw that python was the only thing taking it up, then yeah. I'd restart.
Protip: Gradio hates pretty much every browser that isn't Chrome. Even chromium based browsers are known to sometimes have problems with it.
do you inpaint much? that's where I've seen the most problems with non-chrome, chromium based browsers
I can't remember the last time I inpainted anything outside of adetailer's automated process, heh.
Yeah, then edge will probably work fine for you. If you run into any weird problems, I suggest trying Chrome as a first troubleshooting step though.
- use more efficient optimization (xformers if available)
- lower resolution
- upgrade GPU
if anything that should lower your gpu usage overhead
which sampler are you using @nova remnant and is it the same one that you usually use?
karras is the scheduler. what's the full name of the sampler?
dpmpp_sde and dpmpp_2sa both do effectively 2 steps for every 1 step. it's completely normal and expected for them to run at almost exactly half the speed of other samplers. I believe Restart is the same way.
dpmpp_2m with Karras scheduler in my opinion. Should never need more than 20 steps and is as fast as any of the other fast samplers.
if you're using a hires fix workflow then I'd suggest 20 steps on the low res pass and 15 on the hires pass.
Nurse to give you luck abe
won't hurt anything, but you could go lower without giving up noticeable quality, especially on the hires pass. It doesn't need as much since the image is already mostly formed.
SDXL on my system is a no go lol ... 35 minutes to load any model
rip
do you not load everything for floppy disks?
Fooocus would work .. but now gives random errors and won't use python 3.10.11 for some reason ... and comfyUI just boggles my mind
I didn't get to use floppy disks that much
count yourself lucky. they are cursed.
step 1: cry
step 2: stop caring about good hands
step3: cry some more
step 4: prompt for specify hand positions that are well trained, such as waving

step 5: cry even more and learn to draw hands so you can fix them
feet gens are the real challenge mode
I don't see much feet from SD

We need more
I mean
low resolution hands and feet
a long time ago, I spent a few weeks trying to make an embedding for better feet. I succeeded in making them far more cursed than they already were.
Cat need 16k x 16k res feet
👍
hmmm..testing realistic models and using anime model as the refiner at 70%
A big problem with hands is that there are so many different shapes that they can make, and even if it's just a little off it can look completely wrong.
looks kind of video-gamey (not in a bad way) you could probably switch even earlier if you want more anime style.
lol..always fun when it starts at 2 It/s and drops to 1min/it
isn't it december already..need my new gpu for my bday/xmass
I have 6. We got 2 more kittens relatively recently to bump us up to that. I'm thinking of getting a third Litter Robot 4.
More fire and cat 
Pretty much.
6 brain cells running around in your house
Nah...they all share the same brain cell.
needs more dark 
Lmao
switch at 32%
that looks pretty cool even more of the video game look
@native halo dark kitte

too much cat posting in here
dreamscaper to aniverse 🙂
Those are models?
another dark kitten 
need a lamp to see whats in that blob 
I colorized one of the Flan frames the other day.
I am catman
idk what car is this too dark to see 
Need to go inside and check with her
I see
Thighs 
no oh my gucking god
HELP THEY TROLLED THE ENTIRE SERVER
novel ai mods

skill issue
prompt got hit by automod, you can grab it from the PNG info on the image though.
- open the image in browser
- on comfy, just drag the image into your workspace and it will load a full workflow for the image, on WebUI go to the PNG info tab and drag the image there.
my new one that I'm getting ready to upload 
not sure what happened
got a @ everyone ping saying new model
message deleted
but it doesn't seem to be a troll or incorrect
so. keep an eye out
yesh
@native halo hello there
only dark images allowed 🤠

her right hand is simpsons but her left hand is normal
looks like the book is between her right index and middle fingers as if she's getting ready to turn the page

new japanese subgenre ""manga"" where they have hand tentacles instead.
Hmm. How is this working for you?
I pretty much haven't used anything else since I switched to it
I've been getting a lot of normal hands for my generations even for totally wack ones.
Not sure why you're getting cursed hands
Well that's an endorsement if ever I've heard one.
tentacle hands?
I have no secret really, some models are better than others and my parameters typically work decently on the models that I use.
when a model switch fails 😄
oh, i think I know the one you're talking about and yeah, that's smol's
fennec ears are the superior fox girl ears
where to actually find that hires fix?
4x_fatal_Anime_500000K_G
I can do them at that resolution but I can't hires to close to 1000
google is no help
Wow, I haven't been to the new site since they moved.
Search for fatal in there might be that one I dunno 
meh
Call me crazy, but I'm just not impressed? I'll hold my opinion until I've tried it though.
Yep we'll see
of course until... something is leek-d
not really anything to do with it
higher res free for $25/month O_O
How could that ever happen
surely
surely
Need an accidental leak again
dl'd
Time to gen shrek with it and post
this is my brocolli. it will end the world.
kinda wished there was a way to switch refiners multiple times ...just to see what it would do
technically you can with comfy
oof 😦
pregante shrek?
playing Bad Day by Daniel Powter in the bg
seems overall much more coherent and just..trained on higher quality. but like I said until anything is leaked there's nothing we can really do with it
meh
lol
Whats that
v1 vs V2 nai
oh
Is it SD 1.5 ?
yep
trained on 1024 though too
thought it was SDXL base?
what about it? 😮
probably because this is the prompt:
[anime flat color:0.2], [digital art|a storybook illustration] by ([Jan Theuninck:0.3]:1.3), [anyone:0.1], [[rasquache|romanticism]:0.1], BREAK You could tell me what to say, you could tell me where to go. But I doubt that I would care and my heart would never know If I make another move, there'll be no more turning back Because everything would change and it all would fade to black
improved it even further, so it goes even whackier 😄
DAYM
actually pretty epic
the actual prompt: {[anime flat color:0.2]|}, [[__medium*__|__medium*__] by ([__artists__:0.3]:1.3), [{1-3$$__nouns__|__nouns__|__nouns__|__nouns__}:0.1], [[__movements__|__movements__]:0.1], {abstract|surrealism|realistic|absurdism|} | {2$$ BREAK $$ __songmaster__}]
is this ur sdxl model right
yeah
She looks so wounded here, literally.
i could leave it on overnight, and interesting things would come rolling out eventually
You could tell me what to say, you could tell me where to go. But I doubt that I would care and my heart would never know If I make another move, there'll be no more turning back Because everything would change and it all would fade to black
actually that's pretty good
lyricsprompting is pretty good for setting themes
what actually works really well is the alternating prompts
[ prompt A(the thing you actually want) | promp B (fitting song lyrics) ]
link your model again
According to Breakdomain, this is the "Songmaster"
my hidden sauce, the actual wildcard
ah okay
Partly optimized settings, partly luck of the draw.
I can't even explain I just love it so much it's like RuneScape
Are you... using a VAE?
looks like an highres fix with a bit too little denoising in 2nd phase
also no vae
Low CFG?
a1111 right
check settings and make sure the vae is overridden or set
@nova remnant what upscaler are you using, and what denoise?
Kakigori is a little more rock and roll with "Songmaster". Heh.
pmgogmgomg
might want to use ultrasharp or something similar, might be better
it 'feels' blurry
where it shouldn't be
I recommend one of these, or anime6B.
Usually I don't do Ultrasharp for anime though, it tends to introduce patterns into the finer blended elements.
All the good ones are going to be pretty much 4x and up. If you have it set to only upscale by 1.x to 2, then it'll downscale as is appropriate.
@supple raptor Model Reviewed 
That's outta date now Eface. They moved sites.
ah
It shouldn't make too much of a difference, as the upscalers are usually pretty fast compared to regular genning. It'll still gen at your preference.
I recommend fatal, which I linked. Or Hugepaint.
well, gotta go to bed ^^ cya guys later
Later Eface! G'night.
(well... 1 more gen xD)
gnight
that's what we all say and then it's 2 am
aight, now i'm going to bed for real
skill issue
If you're not using Automatic1111 how the heck should I know? Lol.
What're you using?
comfyui is much
harder
well, more complicated and you have to know more what the process is
why did you choose it first?
With that you're upscaling with the latent aren't you? Not the upscale model.
so u are using an esrgan upscale model but also using a latent upscale? thats bad
img2img and use a smaller generation resolution
there is but sorry not comfyui support here
time for bed .. g' night all
G'night!
she got sleep dusted
ppl share upscale workflow nodes in #✨|sdxl chat thats where u will find all the good ones and also where u will find most ppl that use comfy
You need to upscale it with your upscaling model, then downscale it to 2x or less, then VAE Encode it to output the latent to your sampler at a 0.5 or so denoise so you can have the end resulting image.
I just checked your review. You monster! 
I was not expecting that
LMAO
this is the correct way to do it. That's what A1111 does behind the scenes when you choose a 4x upscaler but only want it upscaled 2x.
there's a different one. hold on, i'll load comfy to get the name. You don't want to use latent for the downscaling.
review posted 😉
I believe that's the node you want.
You'll still need to vae encode it back so that it can go into a sampler again.
change the method to area
this will simply resize rather than doing latent manipulation
Here's my HiresFix workflow for Comfy if anyone wants it. Cells can be enabled/disabled with CTRL+B, the LoRA loaders are all disabled by default.
That's a different node. You're using the "Upscale Latent By" node. You're looking for the "Upscale Image By" node.
As I said before, you're going to need to VAE encode it again to run it latent through the sampler.
Help I flattened my fingers on the barbell.
Nice. Loving the look. Very unique.
#🍥|anime message
A little over an hour ago*
nice
but but I left 5 stars 
i have fallen and cant get up 
Civitai. If you click the link I linked and follow Nubby's link in the linked link, it'll link you there.
If you take and download the image, popping it in your PNG info tab, you'll get the details.
fair enough i do need to learn these things
Speaking of...
I'm not going to baby step walk ya through everything, bud. The info is there.
@untold glacier did you like the cuteshrek

Never been a Shrek fan, myself.
Yeah, it was Aurora, with hires upscale being BreakdomainAnime.
Aurora is the best booru-tag model out there imo.
It'll get you closer to what you want prompt-wise than anything.
Lemme link my review I did of it
5 star review:A fantastically versatile model, with the ability to generate images of all kinds in high quality. I love it. Luna di...
You'll see what one is able to create.
civit hates people with poor wifi
AI is sadly not a low tech-friendly business. 
i have a okay pc but its just the wifi
civit sometimes just breaks and nothing loads
btw wut the difference, this matters
That could also just be Civitai, lol. They have so many problems with their site--they even capitalized on it by having a contest for people to submit 404-error type "this is fine" images for them to display whenever it's down.
Get the smaller one.
Oh yeah civit bugs out for me like... every time I try to use it lmao
my curiosity compells me to ask the obvious question, but why not the bigger file?
if you're still looking for it, here's my hf link https://huggingface.co/nubby/Kakigori/tree/main
I believe it's the pruned version, which, if you're not training anything, is what you want. Lemme check the file I have to see if it's different.
"Gee, thanks Nubby for linking me your model. I already found it but I appreciate the effort."
must be nice having wifi
i have to wait a whole sleep
share some with the whole class abe
The 2gb version is what I have. That's the one you want.
The model, that is.
i got lots of downloads for tonight then
bro do u know what starfield is i got no disk space left :((
Weird hair but I kinda dig it
pets him gently
Merged the baby hairs in with his eyebrows lmao
pets gently
btw y u petting him?

my hand hurts
my latex suit is too tight and it hurts 
Wanna trade lmao
Not sure, running a 6-7ish year old gaming laptop
I can gen, just... slowly. Very slowly lmao
needs more ( dark : 4.5 )
depressed cuz these awesome pics were spoiled cuz of their limbs
least oversaturated ai image
what even is going on
adjusting the colors in post process is wayy easier than adjusting broken limbs
my hand is not broken i swear 
@native halo Are you sure they are not broken
u right i need proper medical treatment 

Pro tip: "thighs and face close up" won't end up well
(face on thighs:3.5)
@native halo time for your medicine 
(ultra,mega,epic,tight latex bodysuit:4.5)
Kek
Need more light 

i prefer the darkness 
too much light 🤠
Show your negatives
ive noticed cd tuner is not deterministic if the values are cranked enough
((visible hands)), bad-image-v2-39000, cropped, watermark, text, error, blurry, jpeg artifacts, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, artist name, (worst quality, low quality:1.4), bad anatomy, bad_prompt_version2, bad-hands-5, bad_quality, (painting by bad-artist-anime:0.9), (painting by bad-artist:0.9), bad-image-v2-39000, cropped, watermark, text, error, blurry, jpeg artifacts, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, artist name, (worst quality, low quality:1.4), bad anatomy, bad_prompt_version2, (((bad-hands-5))), bad_quality
its a lot ik
yes,the noise is random even with the same seed u will get diff results
dk which one will work
A previous convo on hands: #🍥|anime message
Decent solutions and explanations through the following messages.
or how its resolved
i had to generate the same img 3 times until i got a similar one so maybe is not that random
i got like 4 or 5 diferent images from the same seed by changing just the brightness and contrast values aroudn slightly
but was also able to get the same image with adjusted brightness and contrast if i change them just slightly
wow i had negative_hands and im not using it
Like
forgot i dowloaded it
i started from simple and came here
negative_hand-neg, (KHFB, AuroraNegative:1.2), (Worst Quality, Low Quality:1.4), border, grayscale, (watermark:1.2)
is all I use for everything pretty much, hands are decent most of the time
wish I could hires 
would be better
(Worst Quality, Low Quality:1.4), blurry these were my negatives
theres also settings for high res fix in hr-scaling,hr-detail and hr stop step
but got spagghetti hands
low resolution,bad quality,worst quality,low quality,ugly,cropped,
FastNegativeEmbedding,BadDream,UnrealisticDream,
is my negative currently
i dont use hirez fix cuz im impatient
OH
thats what the hr settings are
that explains why they didnt do anything lol
Kek
yea,also the stop step settings are for at which samplin steps u wanna stop sending noise to the latent
ngl hirez fix is actually dope
i dont quite understand how that works
is that like ocntrolnet stop step
((visible hands)), watermark, text, (worst quality, low quality:1.4), bad anatomy, bad_prompt_version2, bad-hands-5, cropped, (((bad-hands-5))), negative_hand-neg, EasyNegative
hows this?
51/75 tokens. i still dont know the purpose and workings of token
trying to upscale a image generated with cd tuner is coming out like.. fuzzy or something
if u set stop step at 5 it will stop sending noise either brightness or contrast to the latent at the 5th step so if u have like euler a at 20 steps it will only send brightness in 5 steps then the next 15 will be normal steps without cd tuner settings
ooh
also is there a proper guide by which i can fix my weapon holding issues. swords to be specific
evenin'

SD is bad with holding weapons
"holding a sword by the hilt" is the best prompt I found to minimize the problems
If you mention sword in the prompt there's gonna be multiple xd
lemme try this
please show me the ways 🙂
I dub thee, Knights of Nubby
oh I cheat... the way is dall-e
as in only dall e or sd+dalle?
dall-e 3... it's where you go when you want real coherency 😛
nice spaghetti monster ogre guy
hey what the crap is with dall-e only giving 15 credits now? seriously?
hey ... i'd pay for more if they'd let me
but NooOOoOooo

ok ok ok.. ill fire up sd
need some new prompts... i'm uninspired 😦

do robocop vs miku
ill try. my last boost.
ahhh it did it, but it's too spicy to post
kinda odd for dalle...
😔
Lol
She angy
idk why shes holding that weapon 
i didnt even use veiny in prompt and still gave me this 
Robothicc....
I can see that being in an actual sequel.
hiyo
Heya Lewds
damn it got the font too
yeah it made a few good ones
just a no-go in SD... ugh
only way is with robocop lora
Tried pixai's animate thingy
stuff i make with cd tuner seems to come out with this noisy look or something when i img2img it
if i adjust the values in the img2img cd tuner tab i can kind of get it cleaner but then the brightness is not "right"

Scary
try with contrast instead of brightness
I think this is a fantastic time to hang up my SD mittens for the evening. G'night peeps. 
Have a nice rest
ill show u what it gives me if i zero all values out
contrast 1 adds the darkness to the latent during generation and contrast 2 adds it after the generation
ooh
is that similar for detail d2 and d1
heres with all values at 0
restart sampler 20 steps .34 denois
i should try with a cuter pic i dont wanna break tos
try with high res fix, 0.45 denoise and hr-stop step at 1
how do i create gloomy environment like this?
real miku
Let it goooo
horni bonk
Holy
ooh boy

use any anime model + cd tuner to get the ultra dark images idk if u have a similar extension on comfy tho
prob not 🤠
hmm kinda went overboard
live action miku movie?
how dare you tag me
also yes
hewo
BEAUTIFUL
there we go !

imma whack you with this lamppost yo
the hell gates are open
we live in a society
ok last one and i'm going to beeeedddD
Have you tried... Bulky armor in the prompt
ya
im smort I know all the tips
I get boobs
kek
nice model
but seriously
@stable crane "full armor, metal armor:1.3" or something like that does it for me
put negatives for skin
bare chest, bare legs, showing skin, skimpy clothes... etc
this bulky enough?
amazing I love it
almost reminds me of my spider lora, but this is mucho bulky
bulky and thicc
heres a small one 
🔫
@native halo rate this mag
thats a gun gun 
ima shoot u down
can't get em bulky enough though
bulk
damn how many layers of coats is she on
triple scope for big fire 
You should get my betterguns lora
What you mean those are bad guns I see no problem 
can't download loras for this, not using my pc atm
remember to disable dynamic prompts if u have it,it makes cd tuner buggy
O, fair :o
controlnet from a sketch my friend did
ooh that would explain some things
i was getting some weird errors earlier
i use wildcards frequently and i was getting some kind of out of index error i forget exactly what it was
yea i wasnt getting errors but i couldnt disable or remove the noise,always appeared on img regardless of settings,now its all good with dynamic prompts disabled
ooo
yeh, dynamic prompts has messed with several different extensions in the various tries I've given it. Umi-AI-Wildcards is better and has all of the Dynamic-Prompts features that I would use anyways.
bloated extension 😔
now it works for me 
Cute
@native halo okay I fixed it
I have thick nurses 
Some are too much for this christian server 
the negatives are being overpowered!
lovely
would 
good morning
Meowning
Crazy? She was crazy once, They put her in a room. A rubber room, A rubber room with rats, They put her in a rubber room with rubber rats, Rubber rats? She hates rubber rats, They make her crazy
I'm going insane
I can't get bodies to look good in this breakdomain model
To look at least in the right place 
have a link?
I did not download this myself
@lyric zealot
why so bad
hehe ... she is just big, what was the prompt?
1girl, sitting, (bikini, skirt:1.2), looking at viewer, small breast, medieval city, market stalls, houses in the background, sky, clouds, light rays, long hair, portrait
A pretty boring prompt
but I'm using it to test 
@oak spoke
hello
nice, how did you do that cyclops look? i always fail... also with leela (futurama)
a lora?
I tried using that breakdomainxl model
things broke really bad
@lyric zealot
many mikus
@lyric zealot cutefy
yea..bit of a hit or miss
blood warning maybe?
@silk narwhal this is what happened
shes painting
looks like wrong config or VAE
SDXL 1.0 yea
holy... my pc is about to take off 😄
this is what rly happened
this is what happens on my prompts 😄
my honest reaction
yea..BDXL keeps doing this orange 90s tv colorization on every generation i do with it
trashed
used the cyclops tag and strengthened it a bit, but model is very anime focused, probably couldn't do leela
@native halo is this what you like
full body 

This looks like you are using an SD1.x/2.x VAE on an XL model.
yea.. not sure how to have it change with the refiner model change
was using kakigoriv3 and refiner BDXL
guess that would be a comfyUI thing 😦
@native halo
shes coming
at least body shape isn't weird anymore

one of the 8 in the batch had one leg
she was sitting sideways so that was normal lmao
@supple raptor NUBI
come at me
tell me one touhou character prompt
Grey eyes, pink hair, pointed ears like an owl, light pink bird-like wings with purple highlights. Dresses in white and brown clothing. Hat and dress contain numerous bird-shaped ornaments.
I think if you use highres fix instead of refiner you can make it work with a similar effect. Just set upscaling to 1 if you don't actually want to upscale the base image.
Then set VAE to none/auto and make sure both models have VAE baked (Kakigori does)
You can just grab one of my images and toss it in PNG info. I'm in bed and don't have SD loaded.
but what character is that
Mystia Lorelei

mornin'
ohi
yo
what about people who aren't living in morning rn?

such discrimination
they can read it later
not arguing with you but i think that's interesting what you just said and to be fair i think thats where territorial context comes to play that are in reach
he has no way of tangibly relating to a world scenario that is miles away
He's just playin' 🙂
yep i know
of course I know people are in their afternoons and evenings 😄
i just put that thought in cause there are people you see on news that go overboard with inclusivity concept
solution "good time"
🤣
at least I've yet to see that kind of thing about time lmao
more about addressing everyone across time 🙂
What model did you use?
Gishikiv1 with refiner expmixlinev3 at 0.5
No Lora ^
I can get you the metadata tomorrow
i downloaded stable diffusion literally 5 minutes ago
Since I already turned off my computer
what's a refiner
also no clue what a metadata is...
Refiner is basically the concept of changing model in the middle of generation. If your reminder ratio is 0.1, then with 10 steps, you will generate 1 step with base model, then 9 steps with refiner model
A metadata is an embedded data that’s inside most png image people post here. It contains the information about how it was generated. Example, the prompt, the model, the negative prompt, the steps, cfg and others
However, not all image posted here will have embedded data. Most of mine, will not have them since I copy and paste them from webui, and that will not have the embedded data. However you can just ping me and ask for it







?
