#🍥|anime
1 messages · Page 198 of 1
Noyce, noyce!
yeah but thats not wha I was going for
I see. 👀 .
ok.. Need to reboot, simple one's are taking forever now. brb
Zerotwo (002) my fav ✨
Remember, we will be waiting for MORE.
Ok
play Barony!

kimono over ribbed sweater is a fun clothing gen
😃
was suppose to be bart simpson. hmmm
ahh thats better
I can hear ironman going :ooh a big boy.. or oh sheeeit
has any new version been released that can be used as fast and efficiently as 1.5 on 8gb? Cause SDXL wasn't it .😛
whats dat dress? name?
I'm on 4G 3050TI and I just upgraded to Forge branch of a1111 and holy s!
spreeeeed
peeed
is it?
yeah im using it
It's not supported in any of the UIs yet, but this was just released. The smaller stages might work on 8gb.
https://huggingface.co/stabilityai/stable-cascade
https://github.com/Stability-AI/StableCascade
There are lots of people running SDXL on 8GB now though fwiw.
You think I type more than 5 words in my prompt ?

The whole look is "Goth outfit"
You can make anything an outfit if you call it an outfit
the power of imagination is strong
👍
like this
nah its magic
there thats better
back to miku
Black WIdow Miku
now it's anime
hmm standing on building? what?
Bellybutton
Let me in!
awwww
Maybe?
Also, have a Miku.
dont type, just think loud !!
i bet everyone's out busy dating
alita vibes
maa Saraswati
yep, prompted alita
trying to create a depiction of a fallen angel with dramatic visual.... not getting there yet
the theme gets a bit off when i try to add a bit of NSFW
otherwise it could be as simple as
fallen angel, dramatic, dark vibe
but what i'm aiming for is to give it a bright sensual touch
maybe with some revealing clothing
making child safe character based art doesn't allow much room for creativity
why?
I really don't understand how the regional promter works! crazzy! or today not my day.... 😭
🤯 😠 😈
how much vram do you have and what webui are you using?
Forge WebUI on GTX 1660 Super 6Gb
ahh clever man
what i was going to suggest
what you mean?
forge webui runs lot faster than a1111
going to suggest for me using Forge? i know it... 😁 thanks!
cool
There’s basically like two different mode for regional prompter. The one most people used is to separate the region of the canvas to different regions, and put different prompts into each of them.
The other mode, is to have an auto detection of a certain tag, and add “effect” to the tag.
The way to use the second method is quite strict
in this case. I have an idea for an image. and I want to implement it more accurately using region separation.for example, this is another not entirely correct result
I'm using a mask here.
and this is prompt {__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(intricately detailed:1.15),(intricately detailed clothes:1.14),(intricately detailed background:1.15),perfect correct anatomy,perfect correct anatomy hands,colorful fill lighting,(detailed skin texture:1.1),subsurface scattering,upper body focus,((close-up)), poetic atmosphere, (brush strong),(young preteen boy hug a girl:1.4),winter sun day,snow,snowflakes,snowing,sparkle,perfect,(light romantic vintage theme:1.4),(chatacter illustration:1.12),(intricately angle shot:1.14),(intricately pose:1.15),obscure light and shadow,((soft shadow)),chromatic aberration,ambient occlusion,cinematic light,depth of field,golden ratio BREAK (1boy, black hair, open eyes, holding a valentine gift, winter coat, checkered scarf, look at the girl),boy BREAK (1girl,(glasses:1.25), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag),girl BREAK <lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>
As far as I understand, the orange area is my first part of the prompt (basic) {__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(intricately detailed:1.15),(intricately detailed clothes:1.14),(intricately detailed background:1.15),perfect correct anatomy,perfect correct anatomy hands,colorful fill lighting,(detailed skin texture:1.1),subsurface scattering,upper body focus,((close-up)), poetic atmosphere, (brush strong),(young preteen boy hug a girl:1.4),winter sun day,snow,snowflakes,snowing,sparkle,perfect,(light romantic vintage theme:1.4),(chatacter illustration:1.12),(intricately angle shot:1.14),(intricately pose:1.15),obscure light and shadow,((soft shadow)),chromatic aberration,ambient occlusion,cinematic light,depth of field,golden ratio
The left one (turquoise) should, in theory, describe the boy. like this (1boy, black hair, open eyes, holding a valentine gift, winter coat, checkered scarf, look at the girl),boy
well, the green one on the right is a description of the girl (1girl,(glasses:1.25), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag),girl
but as a result I get something a little different from what I would like according to my idea and description in the prompt
am I missing something 
in this is mode result better.
but for some reason not always
since you are using Use base prompt you will probably need to include ADDBASE
https://www.manfrotto.com/uk-en/055-aluminium-3-section-photo-tripod-with-horizontal-column-mt055xpro3/
https://www.manfrotto.com/uk-en/500-fluid-video-head-with-flat-base-mvh500ah
Wonder what I can get for those two, looks 99% new, the spirit level is 100% functional. The fluid head still functions like new.
in case of using a mask. or if I choose this particular option?
anyone here an ebay expert on photography stuff? 😛
if im not wrong, you can use ADDBASE with mask, so your quality tags at the front, the ADDBASE then your region
you can see from cmd if regional prompter is working or if there's an error
yes. it working
okay, so it's working. Meaning ADDBASE and break is the same
It turns out that if I use a mask, my prompt should look like this?{__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(intricately detailed:1.15),(intricately detailed clothes:1.14),(intricately detailed background:1.15),perfect correct anatomy,perfect correct anatomy hands,colorful fill lighting,(detailed skin texture:1.1),subsurface scattering,upper body focus,((close-up)), poetic atmosphere, (brush strong),(young preteen boy hug a girl:1.4),winter sun day,snow,snowflakes,snowing,sparkle,perfect,(light romantic vintage theme:1.4),(chatacter illustration:1.12),(intricately angle shot:1.14),(intricately pose:1.15),obscure light and shadow,((soft shadow)),chromatic aberration,ambient occlusion,cinematic light,depth of field,golden ratio ADDBASE ((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl) ,boy BREAK ((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag) ,girl, BREAK <lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>
really? why?
anything before ADDBASE is added to all region
ADDBASE
winter sun day,snow,snowflakes,snowing,sparkle,perfect,(light romantic vintage theme:1.4),(chatacter illustration:1.12),(intricately angle shot:1.14),(intricately pose:1.15),obscure light and shadow,((soft shadow)),chromatic aberration,ambient occlusion,cinematic light,depth of field,golden ratio
BREAK
((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl) ,boy
BREAK
((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag) ,girl,
<lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>```
something like that?
oki - doki trying it
damn that complicated things, what model requires that?

or is the new one just like that, when it comes to prompts
it's an extension called regional prompter
that can split the "canvas" to different regions and apply different prompt to it
not sure, it's definitely not that bad when i last used it
and you survived thrice 
and one of the most highly malignant and aggressive cancers out there
and again the result is not entirely correct
first time out people have like 30-40% survival rate, second time more like 15-20%, third time 2-10%
4th time <5%
according to the description the girl should be wearing glasses
it gets harder and harder to surgically remove it the more it comes back
but yeh, beat it three times. My body is apparently very resilient and my mind is apparently very strong willed toward survival
but uh I also might have a slow growing thyroid cancer so gonna remove half my thyroid
so yay
completely unrelated to my other cancer, it might even be a benign but aggressive benign that needs out
but after I'll just get medicine and things will still be like normal, kind of
I'm also gonna finally start learning to draw and give it 100 days and see how much I improve. If I can begin to grasp it etc. Then I can use SD to enhance it 😛
or color it
hopefully things gets better 
well things are better, I no longer have a cancer that would kill me within months
first time I got MPNST (the cancer) it grew from nothing to 42x37x28cms and 11kgs in 3 months
it took 8ish hours of explorative surgery, removal of a few organs and partial organs and a few bags of blood to get me through a 50% survival chance surgery
maybe try simpler prompt first, and also since the region and bleed into other slightly
Going through this has given me a new perspective on death for sure
probably not what people expect it to be...Rather than become "thankful for life and the time we have" (well I have but still), I've become more like "Learn to free climb mountains? I might die? Let's go!"
Yes. this is an option. but still sometimes the result is quite good. but not with a 100% hit rate. this is the case when using the second method.
SDXL?
im kinda dumb, so don't really understand how all this works, but what I do know is that, I'm just glad you are still here, and hopefully you will still be for a long time 
btw have you guys heard about the AMD related firmware/driver/software that creates their own emulation of CUDA? Fully functional emulation apparently. https://github.com/vosen/ZLUDA this one
Thank you, me too
again
maybe it's better to simplify the prompt a little first to test, before going crazy with it 
no problem. however...{__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(young preteen boy hug a girl:1.4), BREAK ((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl) ,boy BREAK ((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag) ,girl, <lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>
https://www.techpowerup.com/gpu-specs/radeon-instinct-mi25.c2983 I wonder how fast it would be to run this, or if you can run two. Even if it's old
WALL!!! 
this is second method
how is your regions setup? 
hi hi @gloomy scroll
see your cmd, there's probably an error, since for that mode, it's not used that way
hmmm
but why?
Prompt mode is something like
2girls hugging, red hair, yellow hair,
BREAK 1girl, big breasts, yellow hair
BREAK 1girl, small breasts, red hair
you type your prompt like usual, then break, and the modifying tag you wanna add to it, and the detecting tag. Example,
BREAK
glowing eyes, red eyes
but isn't that the case with me? A boy hugs a girl. Next I describe the "boy" tag and the "Girl" tag
or not?
Notice line end after each prompt
Didn't read your prompt btw
Prompt mode checks last tag after each BREAK to be target and line must end to that tag
So like no adding comma to end of the line(prompt)
BREAK
target - ((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl) tag - ,boy
Also this BREAK shouldn't be needed if you aren't using latent mode
BREAK
<lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>```
i am using Attention mode
recheck what I wrote since there was typo
Since latent is one that has some effect of separating loras but attention should just apply lora to everywhere
I'm happy that I make less typos when writing code than normal messages
It turns out that my prompt should look like this?
BREAK
((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl) ,boy
BREAK
((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag) ,girl
BREAK
<lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>```
with these settings
your last BREAK will probably make this mode not work
That last BREAK does nothing
No spaces after commas 
try with prompt-ex I don't rly know what's difference between those but I have only used it
because of your last BREAK, you are trying to detect <lora:hyperdenoiser_v090:0.65>
no last BREAK and commas {__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(young preteen boy hug a girl:1.4), BREAK ((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl),boy BREAK ((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag),girl <lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.25>,<lora:Kitazume_Hiroyuki_style_v01.08:0.45>,21thkitazume,<lora:hyperdenoiser_v090:0.65>
what is your prompt now?? 
I think it just saw empty line at the end

put your lora at the front, before the first BREAK, and try again
BREAK
((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl),boy
BREAK
((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag),girl
something like this?
This is what RP saw imo
{__wc/konica__},(masterpiece,best quality,absurdres,picturesque),(young preteen boy hug a girl:1.4),
BREAK
((boy holding a valentine gift:1.25), black hair, open eyes, winter coat, checkered scarf, look at the girl),boy
BREAK
((girl in glasses:1.35), plaid scarf,pleated skirt,school uniform,winter coat, school shoulder bag),girl,,,21thkitazume,
I think it should ignore loras
since loras aren't applied with the prompt but removed and applied separately
then you are right
But if you using latent then it should do something to apply loras to specific area but I still think they shouldn't matter for prompt mode
Alas. but neither one way nor the other, the result is not correct.
like i said, it's best to simplify your prompt first, so you can understand how it works
but this is really simple prompt...
since you are using Use common prompt then the Base ratio of 0.6 should not matters
I don't think you should be using that 
this is really simple ```(young preteen boy hug a girl:1.4)
BREAK
(boy holding a valentine gift:1.25),boy
BREAK
(girl in glasses:1.35),girl
but
can i has this elf gf?
Im on phone, I downloaded, went search for metadata found nothing so i dont think so
hmm wait lemme find it
are you using Use base prompt or Use common prompt?
pls do
`parameters
((masterpiece,best quality)),1girl, horns, pointy ears, bangs, detailed eyes, blurry background, white hair, moonlight, spider lily, flower, from side, hair ornament, japanese clothes, kimono, (spider lily:1.9), obi, outdoors, sash, solo, upper body
Negative prompt: EasyNegative, (worst quality, low quality, normal quality:1.4), lowres, skin spots, acnes, skin blemishes, age spot, glans,extra fingers,
fewer fingers,strange fingers,((bad hand)),Hand grip,(lean),Extra ears,(Four ears),Strange eyes,(three arms),Many hands,(Many arms),((watermarking)),
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1786529141, Size: 640x384, Model hash: a54c944e4c, Model: Counterfeit-V3.0_fix_fp16, VAE hash: c32b727313, VAE: kl-f8-anime2.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x_foolhardy_Remacri, Version: f0.0.12-latest-110-g15bb49e7`
this is really crazy! 🥴
in your cmd, is it working? 
Did RP prompt mode need debug enabled to give those detection areas?
I wonder if
Young woman, fair skin, long magenta hair, ice grey eyes color, dressed in peach and white colored Yakuta, regal posture and expression, holding a tessen japanese war fan in her right hand, open, dripping with blood
BREAK
A neon lit Yokocho alley behind her, several people unconscious on restaurant bar table, unconscious on street, pools of blood, raining
BREAK
Yellow lightning strikes in the sky, revealing the silhouette of a dragon behind the thick clouds
modify it and make it work with regional prompt 😛 I don't have anything installed atm
this would be more towards the A4 portrait ratio rather than landscape
I mean like RP gives you image with detection area
During generation I don't see areas like when using Adetailer
you mean this is?
two img masks
Yes
redid it a bit 😄
yeah it's not coming out right
that's the point
but
Yuri?!
not my doing
ai has it's own idea's
tries their own idea's
well hmm not there....
grabby
shes spinning them and showing off
a deep sleep is always great ! 🙂
im fiddling with a new webui called forge
and its faster than a1111 and comfy
oh yes it is! LOVE IT!
way quicker
ive only started using it
hahah
switches to XL
oh that image took 33 secs
DING done
i'll be in VC diffusers 3
😄
where does forge store your images? i dont see the default output folder that was in a1111
XL
same default folder for me
I just installed as a branch
oh damn i dont see output in mine
right... i had to delete the previous package and using git clone
took about a min, ON XL!
what size are you usually rendering on XL?
meh it's fast enough
would be faster 🙂
this IS faster!
for SDXL ?
try one
with 512x768 you wont get weird images i can assure you
is it possible to merge sd1.5 to sdxl?
nada
point taken
grinz evilly changes it and cn't share this one
huh AI said "NO!"
I bet thats xl doing that
but I admit nothing
wait. already did.. dangit
You could always get a few hentai and photorealistic nudity models, merge concepts etc into a harry potter one and then make Harry an anthropomorphic (or whatever?) Owl. With the head of an Owl and a feathery body and Hermione a inverted Mermaid
with her upper body a fish
harry got lost somehow
it gives both glasses, wtf
must be deep romance thingy
/me triesagain
yes lol i prompted for heart, but didnt come out exactly how i wanted
@nova remnant in this one too

tora walks through fire .... 👀
she likes to study, he likes to watch
little girl and little boy hugging,
BREAK
boy is holding a valentines gift,
BREAK
girl has glasses, in school, finely Detailed, elegant, sharp, focused, Gradient color, soft shaded, UHD, 4k, Hires, HQ, elegantly detailed illustration, black thin outline art style
Negative prompt: BOY WITH GLASSES, kkw-extreme-neg,
goes back to eating his Pan Dan noodles
makes it american by crumbling over some cheddar pringles
there this is good

lol.. img2img?
nice ... i was using sd model not particularly for objects
can of Pringles chips, finely Detailed, elegant, sharp, focused, Gradient color, soft shaded, UHD, 4k, Hires, HQ, elegantly detailed illustration, black thin outline art style
Negative prompt: kkw-extreme-neg
i want the meat
lick
true MAN!
how scary is she... well, she like nuclear war 🤪
shes literally me
Maggie simpson grown up
she tricked to MC if I remember correctly
crap! I think I figured out how works RP! at least for now the results are quite stable and, most importantly, predictable
Echidna would never do that shes pure
S3*
S3? oh maybe I dun rememebr now
it's not
hmm. it look like old poster for the film
why not?
😄
All the watercolor
promotion! show a bikini - get a banana(ban).. yeah.. we know... no no no....

bikinis are illegal?
Yes
You'll be in jail for bikini

Don't do the bikini!
banana and bikinis my fav combo

well that's unfortunate, i just generated a bunch..

it's frilly bikini's with naughty tats that do it
i guess i'll just put them on civit then
😄

Pony gave me a bikini shot
Without it being worn
But it's there in the image
This model is too smart

(edible bikini:2.5)
wtf, civit changed their notification thing so i have like 80 notifications
yah I just noticed too
little bit bikini
haha .. "HIghway to AD/HD.. the road to Hey look a SQUIRREL!"
no bikini 😏
UMMM
they are sinkin in the sand 💀
backs away
critic!
The degree of madness is growing! 😂
broken umbrella

ehh I cant get fangs
Try (fangs:1.8)
1.8? deym I dun go over 1.5
(fangs:1.8) 😄
where blush?
BLUSH
hands on face, blush

slight smile
😄
aliens are coming to get earthly beauties
hide all girls
?
just taking the cats for a walk 😄
TF you sayin'?
oh wait hes saying hide all girls cuz aliens are coming to get girls
this
tentacle aliens

dungeonsiege armor, hmmmmmmm?
🤷♂️ i prompted 'elaborate armor'
Ok.

You wanted yandere?
(Special measures were taken to make her look as unhinged as possible)
naurrrrrr
We posting cute girls?
100%
no no we posting a cute boy.....

reject
i can show... really!
it's the only one you get
yeah lets tame her
I mean, you asked for yandere!
yeah but thats too much
hes thorwing his umbrella
making of him a waifu?
Why is furrys so popular in this chat
havent seen a furry posted in a few days
Your profile is a furry, and you posted it just now. The cat
that's a catgirl, not furry
Ears are still furry so for #1 I'd say it's a little bit furry
you're furry
🤷♂️ you can think what you want
I'd suggest not looking through anime channels full of images if you have a problem with it though, it's pretty deep in the media and culture it comes from
My image profile is a monster of the Fendi brand (when I used to be a salesman). Your image is a cat so it's very furry. I don't mind if it is I was just surprised it's suddenly so popular
No not that but the actual cat pic guy
yeah, that's a cat, not a furry
Oh so it's like gaussian distribution curve where the tails are not furry but everything in between is. 🤔
Yes
she broke the katana
If you're in here and trying to add me as a friend but I don't know you from regular chatting, at least remind me how I know you. I get so many adds that are probably future invites or ads, trying to get into DMs with me. Hard to know who's adding for real or not, unless I know the person well 😛
I only add people as friends if I actually know them, no idea why random people send friend requests to strangers... this isn't facebook 😛
NEKO
P.O. NEKO.. runs away
DragonCubeZ
rachel mcadams as a cat furry girl
whats dis modal? albedo?
google bard?
@nova remnant I saw this one, albedo model
❤️
yes
Am I doing it right

why cry
For some reason she wants to do that lmao
I removed from the prompt but keeps happening 
dang mfer sad
@dire bridge no more sad
I changed from goth makeup to black makeup
I guess goth are all sad

'goth' is heavily over trained in some models, using 'gothic' generally has better results
same seed, goth vs gothic
ok making it more anime
The Doctore: Elizabeth Banks, and her cat
gotta say best one yet
New new Anime tron?
thats xl
1.5 , yep I like this one better
hmm i've seen that somewhre before?
in the lab
finally
ok kinda
Has anyone tried stable cascade yet?
how? couldn't get it to work
'laying in a tank of fluid'
also, tub or vat instead of tank, depending on what works best
I had : submerged on back in neon green liquid
mine keeps cutting off the legs
ahh ok
submerged for me just puts her underwater but you can't see the fluid
well not what I wanted but interesting
ummm, hoirror
lol
I mean.. she's in a tank
haha
yeah you are better than I
i wanna get eaten alive by freddy fazbear
i can iterate faster, makes it easy to tweak prompts little by little over a few minutes
that last one is cybernetic girl, robotic body suit, hooked to wires floating in a tank filled with bioluminescent green fluid, art by H.R. Giger, behind glass,
sigh, starts over
added miku to the prompt 😄
AH HA
...
haha
well then
tempted to put inside lava lamp
umm didn't go as planned but funny
I want to get in on this
hmm, turbo seems to play nicely with this model
8s gen using sdxl hires 2x with adetailer
Turbo and LCM are loras that allow fewer steps to speed up gen
this image only has 8 steps at CFG scale 1, instead of CFG 7 and 40 steps for the same model
2 hi res steps
Is this a separate LoRa or some kind of extention?
it requires a lora and the sampler, forge has Euler A Turbo and LCM samplers built in
so you just need the loras, or a model with it merged in
i just merged the turbo lora with my sugoiXL model to see how well it would work
it changes the look a little, but depending on what you're after it can speed things up a lot
will these samplers only work with those models that were made specifically for them? For regular models there won't be an increase in speed?
you can use the loras with any model
only SDXL?
LCM has a 1.5 lora, not sure about turbo
give me link please 🙏
LCM-LoRA - Acceleration Module! Tested with ComfyUI, although I hear it's working with Auto1111 now! Step 1) Download LoRA Step 2) Add LoRA alongsi...
this is?
yeah that should be it
guess i'll upload the turbo version of this, seems to be working well enough
works for what i made it for at least
a little doubtful. in a small number of steps. does not have time to accumulate a good amount of parts
works better on some models than others, but it does work
Yes. it works. because the number of steps actually decreases. and it saves time.
8 steps, 2 hires steps, CFG 1.5
SDXL model?
yeah
hm.. i will try for me
Nothing super special, just merged a few models together to get a specific look I was after for chibi characters. It does a broad variety of anime ...
in my case SDXL generation withpout HRes 4-5 min...
turbo version is currently uploading
in forge?
yes
pretty sure some of the people in here were doing XL in forge under 30 seconds with 4-6GB vram
using LCM or Turbo for XL is even better for low end cards though since every step takes so long
no real practical reason for me to reduce my 5.6s gen to 1.6s
HOW?!
they obviously don't have 1660 GTX 😂
🤷♂️
i`m sure on 3070 is faster
@vital raptor with your model what Turbo LoRa i`m can use?
you don't need a lora if you use the turbo version https://civitai.com/models/298051?modelVersionId=344308
3:15 min with Hres?!!??! but i think wrong Upscaler
without Hres better and 1:52
looks like it could use a slightly higher CFG, i'm a bit surprised it's not faster though
try the Euler A Turbo sampler and use a higher CFG
also that's not the turbo model
all those extra loras and textual embedings aren't doing you any favors for speed either
in fact nothing in the negatives matter much at all, image on the left has your negatives, one on the right has nothing in the negatives
SDXL doesn't need negatives and doesn't listen too them a lot anyway. LCM and Turbo ignore them even more because of the low steps
really should simplify your prompts in general, probably 80% of it isn't doing anything useful
illustration in 1950 style of preteen child with long black hair and green eyes, white skin, upper body, lace, closed mouth, jewelry, lace dress, outdoors, summer day
nothing in the negatives, no loras or embeddings
8 steps, CFG 1..5
Funya's wall of text prompts 
bad habits propagated by stupid youtubers doing tutorials on things they don't understand 🤷♂️
also you don't need 'preteen' and 'child', they mean the same thing to most models
girl, young girl, and little girl
this is dumb, sdxl with 768x1024 res using my turbo model is 1.2s 😆
lcm has lower quality/ less details 😔
that's not lcm, it's a turbo sampler on a normal model with no lora
well, no turbo lora anyway
yea turbo uses ADD thats better than lcm 
it depends a little on the model, but yeah, turbo seems to have better results overall
it's generating close enough to the original model that i just went ahead and merged it and released the turbo version of the model
LCM wipes so much detail that it's only good for certain things
it's interesting how turbo works with hires fix too, the default 1024zx1024 image is pretty chunky in the details, but 2 steps of 0.25 denoise cleans it up really well
<lora:add-detail-xl:1>```
Neg (low quality,worst quality,lowres,error,fault,undefined,crude:1.3)
you're still not using the turbo version of the model, but it's interesting to see how well the normal model does with 8 steps 😄
wait...
and you still dont need those negatives, sdxl for the most part doesn't have any low quality images in the training data
i`m download this is o n your link
refresh the page?
😄
the real problem is that a lot of people are just following what they saw on a video or read on reddit somewhere, but they don't have the hardware or time to actually test it for themselves so they just don't know
i wasted days on tutorials for model training and in the end none of them were actually correct, it's really annoying
but the most interesting thing is that the page with your model was NOT OPEN to me.
it was probably not at the time, it took an hour to upload it
Well, it’s 1.5 and not SDXL.. wall from text prompt must be!😂
4 min
turbo adds a lot of random crap to the image for some reason
it doesn't like empty backgrounds
(flat cell shaded:1.1), girl, chibi, petite, short black hair, (purple:0.7) eyes, cat ears, choker, black oversized hoodie, empty background, sitting, (sticker:0.8), eating donut,
yes..
pretty sure its because of how turbo works, it intentionally fills the space with extra noise so it can create more detail in fewer steps
World of donut
now this si cats
who lost a lamp?
i know i know this is Capitan? yes?
not anime char
i mena @vital raptor 😁
he he noooo this is a Tanya!!!
Good gen 
Lt. Colonel Tanya Degurechaff for you
😁

lol
m?
who enjoys pizza like that? XD




