🍥｜anime

[junior quest]

W

fangs

PERFECTION

[native halo]

https://tenor.com/view/enma-ai-anime-eat-berry-spit-gif-16294670

[nova remnant]

https://tenor.com/view/spray-bottle-cat-spray-bottle-spray-bottle-meme-loop-gif-25594440

[lyric zealot]

https://tenor.com/view/anime-dodge-gif-24234463

[keen hollow]

just some randos

[nova remnant]

cool set!

[keen hollow]

it's that lightningxl one I think you use torqx

[nova remnant]

oh?

whats the name of the model?

[keen hollow]

lightningbloomxl I think?

oops didn't mean to add an image.. but well .. there's another one 😛

[nova remnant]

oh nice

im flattered

would love it if you posted them on civitai 🙂 your prompts came out very nicely

[keen hollow]

I do'nt think I've ever posted to the site.. ill see if I can figure it out 😛

the prompts might not be in the images since I use copy/paste from my viewer, but let me know if you want any

I have lists of prompts I run randomly through models I"m messing around with

frankly I think I need some new prompts though... kinda tired of these

[nova remnant]

sure, if you want to, cause those images look quite nice

i usually play around with ideas that i come across from watching stuff online

[green temple]

nifertima best girl 🥰

[nova remnant]

its so hot give me snow

[acoustic locust]

Which model?

[acoustic locust]

@green temple

[nova remnant]

https://media.discordapp.net/attachments/931925914470711346/1061076535525785680/321666B2-CC36-467F-B254-56055DEB1C8C.gif

abe going insane from not having model

[acoustic locust]

I've been insane

[keen hollow]

cool eyes

[bitter latch]

thats what yoda is.. a Gremlin!

[keen hollow]

trying to find a good new prompt list to use to try out new models....

but everything I find is like.. massively nsfw...

just want normal prompts, come on people

[native halo]

https://github.com/AIrjen/OneButtonPrompt

[keen hollow]

ha well.. maybe... just want a consistent set of like 25-50 that I can save and run through multiple models

don't really need a 'generator'

[native halo]

id say let it run for like an hour then select and play with the prompts from the best imgs

[nova remnant]

1 hour of latex:3.5 prompts

[native halo]

already done that

[junior quest]

model?

[cerulean cosmos]

||https://civitai.com/models/50882/maturemalemix|| NSFW

[green temple]

the images i saw for the lora was on yutamixV3 but i used anythingV4 model

[green temple]

the girl with the dragon is the best but it`s expressionless , try to upscale it or smthng

[native halo]

edit the img on PS,draw something good on top of it then upscale it to 16k and post it here 4 free🤠

[keen hollow]

it was part of a large overnight run... I can create a more detailed version if interested.

simple adetailer preview...

trying to find the right denoise for the big one... 😛

[nova remnant]

Thiccc dragon neck

[keen hollow]

big version

[nova remnant]

What if I draw a kok on top of it

https://tenor.com/view/kok-cat-gif-25438805

[native halo]

i like the big 🐓

[keen hollow]

and here's a creepy expression one for Sooma

[nova remnant]

Not creepy enough

[acoustic locust]

Thanks

[keen hollow]

you should inpaint it creepier 🙂

[nova remnant]

I'll put @native halo face on it

[native halo]

https://tenor.com/view/putin-sigma-mewing-gif-13750899674712754521

[lyric zealot]

what kind of choker is this dafug

[keen hollow]

she looks annoyed

[lyric zealot]

she's probably kinda annoyed

[lyric zealot]

[nova remnant]

Love this

[green temple]

awesome ! btw what did u type for the angle ?

[grizzled moat]

Dutch angle with cowboy shot i believe

[green temple]

oh , ok thx

[wet elm]

Does anyone know the full list of artists that pony/autismmix was trained on?

[near elbow]

more brains!

[nova remnant]

Full no

Hundreds yes

Eface posted one

Not sure where it is in this chat now

@wet elm found it

Just had to search for and comment of mine tagging eface

I didn't have to tag him

[wet elm]

tysm!

[nova remnant]

[nova remnant]

the character's face looks beautiful

simplistic and cute

[nova remnant]

nice lil fangs!

[nova remnant]

[brittle jay]

can someone do a cat with a suit in a mountain setting?

[brittle jay]

need it to actually be a cat

[green temple]

@nova remnant

[nova remnant]

Skooma cat

[nova remnant]

https://tenor.com/view/headpat-cat-cute-gif-15520097

to pat

[native halo]

pat me instead

[nova remnant]

with a hat on I'll have to pat somewhere else

[native halo]

obey your superior 🫡

[nova remnant]

7

[lyric zealot]

weird hat tho

[fresh sleet]

At least for pony, I don't share the training data info, sorry about that.

[lyric zealot]

big fan of the model you created , just wanna ask, do you already have access to SD3?

[fresh sleet]

I wish, I think Emad is angry at me for stealing the next foundational model spotlight. Seriously speaking, I've tried to engage with SAI multiple times over the last year or so but they never got back to me (beyond some very generic answers). Sigh...

[lyric zealot]

well then, we can only hope that it's coming out soon rather than later

[native halo]

i've fallen and i cant get up

[nova remnant]

@native halo

innocent cat

[native halo]

lick

[native halo]

needs proper discipline

[karmic stone]

#🏞｜general-with-images Six month old baby lying on bed wearing Independence Day onesie

[nova remnant]

[gusty palm]

whats the proper use case for the dynamic threshholding thing

[gusty palm]

bit too many settings for me to feel like i understand it

[wet elm]

no problem, i just wanted to see what artstyles i could use, and someone sent a good list already, also ty for your really good model

[gusty palm]

is rabbit_v7 an sdxl model

[nova remnant]

SQUISH

[zinc flax]

https://tenor.com/view/aris-alice-blue-blue-archive-archive-gif-12082873086428133110

[gusty palm]

thats weird

i cant get it to work unless i select none vae

[near elbow]

neko gun fun

[terse ore]

Ever put that stuff through low denoise+controlnet and other things?

[junior quest]

model?

[green temple]

https://huggingface.co/zzzAI19/ChromaFT

[green temple]

chromaNeoFT v1 peruned vesion

[green temple]

wooow !

so nice

[twilit portal]

Akeno with an indian touch ✨

[merry tendon]

why do people say pony diffusion is good

[keen hollow]

" hey bill, wait up... "

[ivory pebble]

It's really good when it comes to doing NSFW/SFW concepts. I'd say it's just as good as NovelAI when it comes to that. It also knows a lot of popular characters/artists so it's easy to generate characters and mix different artists together without using a lot of loras. I personally like using AutismMix, a PonyXL based model, mainly because it improves the anatomy of the characters a lot, which is something that a lot of popular models struggle with.

[merry tendon]

very good model yes indeed

[ivory pebble]

It doesn't work like other SDXL models so if your images don't look as good or if you're just getting blank images then your settings/prompts are the issue most likely

[supple raptor]

You likely got a NaN error. Try this vae. https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors It should fix the typical NaN in VAE error. NaN in UNET is usually an A1111 error.

[quiet steeple]

Kid Baru 😄

[ivory pebble]

I do have to say tho it is not the most aesthetically pleasing model. It's to be expected tho since it's not just an anime model, but you can still get some very nice looking images if you know how to prompt/mix different artstyles together. Other than that tho it does what it is supposed to do, and it does it exceptionaly well

[vital raptor]

when AI becomes self aware..

[nova remnant]

[green temple]

Woooow !

[unique forge]

a smart dog

https://tenor.com/igr85oG7j1q.gif

[near elbow]

tank girl!

vroooooom

[gloomy scroll]

[green temple]

Nice !

[green temple]

Woooooooow !!!!!!!!

[green temple]

🔥🔥🔥🔥🔥🔥

[near elbow]

not 100% anime but how about aloy playing horizon zero dawn? 🙂

[keen hollow]

@nova remnant - I have been messing with this model for days... I basically never want to go back to non-lightning 😛

[nova remnant]

lol ... lightning models are nice .. which one are you playing with?

[keen hollow]

that same one of yours.. lightningbloomxl I think it's called?

[nova remnant]

oh cool 🙂 but i should say, there are some good 1.5 models too

i still go back and forth between sd and xl

[keen hollow]

i'm finding the xl coherence to be a good jump

[nova remnant]

yeah the coherence is good with xl models

[keen hollow]

I was doing some ultrawides

[nova remnant]

can you show them?

[keen hollow]

I dumped a whole bunch in general... one sec

[nova remnant]

im playing with sd1.5 atm lol

ok

[keen hollow]

#🏞｜general-with-images message

try that link

[nova remnant]

woops i was posting in that chanel to talk to you lol

but yeah all of those images look beautiful and the girl laying in bed with medieval settings is stunning

[keen hollow]

yeah that one took some tweaks to get the wide ratio right

[nova remnant]

awesome work!

[keen hollow]

oh yeah any prompt is free for the taking

one sec lemme grab that one

[nova remnant]

you should upload them somewhere, they are nice sets of pics

[keen hollow]

just finished running another supergirl set lol

'murica

[nova remnant]

nice pose too

are those with prompts alone, or mixing controlnet?

[keen hollow]

prompts alone

`detailed sleeping woman in a rococo bed surrounded by vines and flowers, stained glass window, gothic, chandelier, by fantasy art, trending on artstation, fairytale, art by luis royo and walter crane and kay nielsen, watercolor illustration,

Negative Prompt: (low quality, worst quality:1.2), lowres, multiple views, border, (grayscale, grainy), (watermark, signature, text:1.1), underexposed, overexposed, oversaturated, low saturation

Steps: 4,
Sampler: DPM++ SDE Karras,
CFG scale: 1.1,
Seed: 1266526755,
Size: 960x400,
Model hash: 14953e03f0,
Model: lightningbloomxl_v10, VAE hash: 63aeecb90f, VAE: sdxl_vae.safetensors,
Denoising strength: 0.4, 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: 24.1.2, Hires upscale: 2, Hires upscaler: lollypop, Mask blur: 4, Inpaint area: Only masked, Masked area padding: 32, Version: f0.0.17v1.8.0rc-latest-276-g29be1da7`

[nova remnant]

something im doing now with controlnet

[keen hollow]

well, ... oops.. didn't mean to dump all that

but there ya go lol

imma go grab some food 🙂

[nova remnant]

neat! 🙂

okay

[keen hollow]

running another large set... will take 3-4 hours

my gpu has been getting a workout this past week... probably a solid 4 days of runtime

it probably ran for a full day at one point 😐

[nova remnant]

oh wow, thats pretty intense

but not to mention the images came out looking great

creative prompts

one of the images you did with white dragon in the clouds looked really neat

civitai bot scripts need some tweaking

often flags images that shouldn't be flagged

[nova remnant]

found a cool anime xl model, called animaPencilXl , its from the author of bluepencil

[keen hollow]

xl too slow unless it's lightning 😄

[cerulean bane]

would.

[gloomy lodge]

mean to post this one but same diff

[cerulean bane]

still would.

[keen hollow]

invisible person?

[gloomy lodge]

you dont see them?

[native halo]

no,its actually a banana

[keen hollow]

wasn't sure if it was a horse

[pure oyster]

No idea what this is supposed to be, but I like its flavor.

[nova remnant]

Hello anime people

[native halo]

https://tenor.com/view/mewing-gif-14699957714023232752

[nova remnant]

interestingly that model I mentioned is a regular xl model but I had lightning model setup 4 setps and 1.5 cfg and this is what it rendered ...

was prompting cute fashion model in cute pose wearing t-shirt, pants, platform boots.

[nova remnant]

Cute black dragon

[native halo]

not anime char

[nova remnant]

and this with <lora:sdxl_lightning_4step_lora:0.8>

[nova remnant]

I don't want ugly

[native halo]

https://tenor.com/view/rin-okumura-ao-no-exorcist-blue-exorcist-punch-anime-gif-17826173

[nova remnant]

cute woman wearing t-shirt, skirt, white platform boots, bare legs, black background.

[smoky junco]

https://tenor.com/view/saignerdunez-gif-6185219

[bitter birch]

can't believe I managed to make 1920*1080 with only 4gb vrams xDDD

[silent kraken]

Haha from MHA

[latent vector]

might try and fix the eyes at point

[nova remnant]

smol

[nova remnant]

big

[green temple]

noooo😭 only kids pls no pedo 👍

[grizzled moat]

should i censor this

or

is fine

[green temple]

hmm... yes u should

[grizzled moat]

well in that case

[green temple]

same ?

[tiny cypress]

Birds are born without shackles,then what fettered my fate?

Blown away the white petals,leave me trapped in cages

[grizzled moat]

forgot to delete

[nova remnant]

https://tenor.com/view/huh-cat-gif-26460616

[lyric zealot]

just so you know, it doesnt really matter if you spoiler or not, since if it's nsfw, then it's totally not allowed (ps: your post isn't nsfw)

i usually only spoiler if it's kinda horror

[grizzled moat]

ohh

i see

[lyric zealot]

yup yup, totally no nsfw

#🍥｜anime message

[grizzled moat]

no

[green temple]

which is better ?

[vital raptor]

the left one looks like it has better detail while the right one looks like it's all blurred, i'd go for something somewhere between, clear details but not so much random particles

[nova remnant]

both are kinda blurry

did you hires fix it

[green temple]

yes

[nova remnant]

Something is weird

I think your prompt is way way over the top

And 40 hires step is insane x3

Like do 8-15, I could never tell a difference above 10 :p

[green temple]

ohh , ok

[nova remnant]

It might be "burning" the image

With that much weight

I'd first remove every weight on it

[green temple]

maybe

[nova remnant]

Then remove the fluff like (4k)

[lyric zealot]

you are using face restoration?

[nova remnant]

And try to not repeat yourself in the prompt, stuff like (white background, bad background) in the negative

[green temple]

yes

[nova remnant]

Also for things in positive

[lyric zealot]

there's probably no need for that

[green temple]

good notes , thx

[nova remnant]

Face restoration is pretty bad specially for anime

[green temple]

ok

[nova remnant]

Use after detailer

[bold ether]

face restoration is trained for realistic faces

[green temple]

ohhh

[lyric zealot]

yup, if you want, use "after detailer" aka adetailer

[green temple]

ok

[lyric zealot]

in your prompt, for example ((upperbody : 1.5)) is basically the same as (upperbody : 1.65) because a bracket without weight = 1.1 so 1.5 * 1.1 = 1.65

stick to using one way of writing, i would advise using (something:1.2) where 1.2 can be the weight you want

[green temple]

ok

[lyric zealot]

one easy way of increasing or decreasing weight is to select the word you want, and press ctrl + arrow key up or down

[nova remnant]

Negative hand embed, border, greyscale, watermark, censored, (worst quality, low quality:1.4) are my go to negatives for SD 1.5

[lyric zealot]

example, highlight upperbody then press ctrl + arrow key up

the amount increase can be configured in settings

[green temple]

really thanks guys !

[lyric zealot]

[nova remnant]

Stupid girl

[lyric zealot]

https://tenor.com/view/mayuri-steins-gate-tuturu-gif-22196871

[nova remnant]

And start prompts with (high quality, best quality)
I dunno about that model just do a x/y prompt script thing to test if it likes a lot of weight in them or other stuff like masterpiece, 4k

Fox magic

@lyric zealot 👍

[lyric zealot]

watch which server you at

[nova remnant]

I'm innocent cat

[lyric zealot]

[nova remnant]

She's just a soldier fox using magic to track the target

[lyric zealot]

surely

[nova remnant]

This one pretty cute

[green temple]

what abt this one ?

[nova remnant]

https://tenor.com/view/cat-gif-21230487

many pat

[junior quest]

info?

[grizzled moat]

Probably in the image

[junior quest]

its covered with spoiler so I guess no

[grizzled moat]

Daum

[junior quest]

yesh there is none. deym

[grizzled moat]

Its a basic prompt tho ill check history once im on computer

[green temple]

i`m dying...

[nova remnant]

looks much better

[green temple]

thx for help dude

[nova remnant]

btw omnia you can use any sdxl model along with 4step lightning lora

sdxl_lightning_4step_lora.safetensors https://huggingface.co/ByteDance/SDXL-Lightning/tree/main

[nova remnant]

the faces here look beautiful

[nova remnant]

lol i dunno why i havent yet tried 2 step lightning lora, im gonna try it out in a bit

does that mean it also works with SD1.5?

should not, kinda confusing how they wrote it

[magic wave]

Nerd~1

[magic wave]

🐖

[keen hollow]

does it work or is it hokey?

38280/38280 [14:26:26<00:00]

GPU is tired

[nova remnant]

its a lora you put along with the positive prompt

[near elbow]

have a fishy evening

[nova remnant]

i've tested the lightning lora with some of the sdxl models, they came out fairly ok in details

[supple raptor]

The wording is a bit confusing there. What they mean is if you are using the official SDXL model then you should just use the *_unet.safetensors(in diffusers) or the *step.safetensors (full model to be used in A1111 or comfy) files instead of using a lora. For all other models, you should use the *_lora.safetensors files (example: animagine, pony, etc...) They won't work on anything that's not SDXL based.

[keen hollow]

you should share some

[nova remnant]

there are bunch on my civitai page that i did last night

mostly on anime using animapencil xl

cant link it here .. nsfw

something interesting to note with all the available sd and sdxl model, you can speed up render speed with all of them using lcm lora

[keen hollow]

lcm and lightning?

ugh I generated too many images, this is taking forever to go through 😄

[nova remnant]

and we also have sdxl lightining lora .. however you could simply use lcm for both sd and sdxl

[keen hollow]

it's like 9000 images or something stupid...

3 of each prompt

[nova remnant]

wow

thats a huge set

[keen hollow]

yeah... 14 hours

[nova remnant]

faints

[keen hollow]

I run a medium res, 3 images per prompt

then curate my favorite prompts

(i use my tool to quickly view and curate)

then I can re-run those favorites alone at higher res

[nova remnant]

i never bothered to learn curating an image via inpainting lol

[keen hollow]

all kinds of random crap in here though....

curating just means I go through them and pull out my favorites lol

[nova remnant]

ohh lol

[nova remnant]

love that concept!

[keen hollow]

a lot of just generic stuff too...

[nova remnant]

maybe but they look beautiful

[keen hollow]

i ran all these in ultrawide which may have been a mistake... I dunno

but I feel like that's the only actual res I"ll use for anything

[nova remnant]

yeah would take a bit of time

[nova remnant]

@keen hollow lol #🏞｜general-with-images message

i wish we had that level of realistic beauty with sd models

[keen hollow]

don't we?

dalle is awesome though... I love messin' around with that

[nova remnant]

not exactly lol

[keen hollow]

I feel like I can get a lot more variety

[nova remnant]

i love how dalle does korean beauty so realistically

cuteness overload

[keen hollow]

all my best Tifas are from dalle

[keen hollow]

time for food

[green temple]

noooo , she dropped the cake 😢

[green temple]

nice !

[green temple]

ohhh... cooool !

[grizzled moat]

trying to remake this #🍥｜anime message

[onyx junco]

create a boy with dog

[nova remnant]

this is a pretty powerful booru tag cute & girly \(idolmaster\)

its power is in how much cute it can make a portrait

oh crap. forgot about nsfw

another interesting booru tag can't be this cute

[nova remnant]

it works great with faceID

[nova remnant]

think i need sleeeeeeeeeeep... my brain is starting to fry

[nova remnant]

It's lady gaga

[native halo]

/ban

[nova remnant]

[latent vector]

tried an sdxl model for the first time

quite nice tbf

[near elbow]

change of style

[oak spoke]

yo been a while my doodad

how's everyone doing?

[oak spoke]

what model u using?

[steady grail]

hello guys

[oak spoke]

I can't managed to get one that have hair actually covering their eyes

[green temple]

hair on eyes ? maybe

or hair on one eye ?

[supple raptor]

the tag is hair over eyes you could also add eyes visible through hair if you want to see the eyes through their hair.

[oak spoke]

doesn't work for me

it's driving me crazy

my model must be shit

recommend me some of that good stuff chief

I even use "hidden eyes" as well

doesn't help

[supple raptor]

hmm... do you have some other eye tags in the prompt, such as eye color or closed eyes, or anything else like that?

[steady grail]

i don't understand why the most cute models are the worst in following poses 😦

[steady grail]

i'm looking for more cute anime models with a good pose response 😦

[supple raptor]

probably because most of the SDXL models seem to be somehow overfitted even worse than the 1.5 models that we came from

[steady grail]

oh....

[cerulean cosmos]

getting anything other then "looking into camera" is pretty hard

[steady grail]

yep

the original pony is very good

but every derivate from it just don't respond

considering the only good reason to use xl is his response to prompts

[supple raptor]

kxl-delta seems pretty decent at poses and doesn't force looking at viewer

it also knows a ton of artist tags.

and good character knowledge.

[nova remnant]

hmm, what kxl delta?

google gives nothing relevant to txt2img

[steady grail]

this? https://civitai.com/models/249918/delta

[supple raptor]

https://civitai.com/models/332076/kohaku-xl-delta

[nova remnant]

oh sweet thanks

[steady grail]

wow, the contrast is so oversaturated

but i can see it's pretty creative

[nova remnant]

something im looking forward to trying out

[supple raptor]

wildcard list for artists on kxl-delta that I made from the actual dataset list.

[nova remnant]

i never understood how you use wildcard

[steady grail]

love you!

[nova remnant]

do i need to install any extension?

[steady grail]

guys, do you have guide to mix loras in models?

[nova remnant]

cause i mostly deal with prompts manually writing them in

[steady grail]

i can't live without dpo and turbo merged on xl

if i get the chance, i want to merge them in every model i use

[nova remnant]

you could also use lightning lora with just 4 steps and 1.5 cfg, works with all sdxl

[supple raptor]

it responds to lighting tags

[steady grail]

now we have the opposite problem XD

it lack gloss

but the result are very cool

can be used as a base and than img2img with something else

if things goes as i expect, this model is very promising

[supple raptor]

install this one. https://github.com/Tsukreya/Umi-AI-Wildcards
it comes with a bunch of mostly decent wildcard lists. and you can add more of your own if you want. if you don't already have it, install the tagcomplete extension that's listed in the readme. once both are installed, restart SD completely so that tagcomplete can grab the wildcards, now you can use them in your prompt by starting to type __(that's 2 _'s) followed by the filename and autofill will help you finish it. The full syntax for a wildcard is __filename__

[nova remnant]

ahh nice explanation, ty

[supple raptor]

I'm using it now for local hires fix + face detailer for my NAI images. it generally does well using the exact same prompt as NAI. I just have a node to convert my comfy formatted prompt into NAI format to send to the NAI api and send the unmodified version to kxl-delta.

[steady grail]

cool workflow

nai api is connected online or require extra resources?

[oak spoke]

nope made sure to removed them

[nova remnant]

@supple raptor i find it interesting you are finally using sdxl models

[supple raptor]

yep, it's connected to their service. here's the nodes for NAI stuff https://github.com/Tsukreya/ComfyUI_NAIDGenerator (the benefit of this fork is that it produces an image that can be sent to img2img directly) and the attached has my workflow.

[supple raptor]

We finally got some that are worth using

[steady grail]

so, you can't gen offline

[supple raptor]

no, since NAI is a subscription service I have to be online to use that. I just scored a free month with them so figured I might as well take advantage.

[steady grail]

oh, too bad

[supple raptor]

It's going to legitimately be difficult to go back to other models when my NAI sub runs out.

[nova remnant]

worth = tohou character feet

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

[nova remnant]

ugly

[supple raptor]

https://tenor.com/view/animepound-slam-gif-4880762

[steady grail]

i wonder if there is an xl lora to add gloss

[supple raptor]

you can probably find an artist in kxl-delta that just does that anyways. it's quite responsive to artist tags. it's trained on basically most of danbooru, so the artists are easy to look up to get an idea of how they will look.

[steady grail]

oh, good idea

ty

[green temple]

nice

[wheat falcon]

thx

[nova remnant]

no crying allowed

playing with wildcards lol its kinda fun to see how images come out

[green temple]

one hand lol

[oak spoke]

long time no see catto

[nova remnant]

tomato man is back

[oak spoke]

wan some cookie?

[green temple]

sadge

[steady grail]

charming

[vital raptor]

autismmix confetti

[oak spoke]

thanks quite an interesting name lol

[green temple]

noooo my fingers 😭

[nova remnant]

whats a good extension to merge lora with checkpoint using a1111?

trying to be specific with lora weight so i can fine tune

mmmmm supermerger sounds like a good ext

[oak spoke]

ASCENDED

[nova remnant]

if you manually rename a model filename does that break the model?

i got versions messed up a bit from merges

[nova remnant]

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

supermerger has heap ton of options, confuzzles me

[oak spoke]

open your 3rd eyes

[nova remnant]

phew made it with the lora weight merge

[oak spoke]

what with newer model unable to handle older propmt

[oak spoke]

like it kinda just messed up when I use lot of step, and high weight, or CFG

just kinda suck

[nova remnant]

sdxl should be ok to handle prompts

[nova remnant]

im eating cookies no third eye allowed

[oak spoke]

funny enough it's literally the SDXL that can't handle it

the older SDXL I use or even 1.5 can handle my stuff perfectly fine

[nova remnant]

which model are you using

some of the models might not have proper training

[oak spoke]

AAM XL (Anime Mix) the half turbo one

[nova remnant]

i kinda came across some of those that dont follow prompt

[oak spoke]

I was testing SDXL turbo

[nova remnant]

hmm

you should try the lightning model, that's built on better technology

at least thats the latest of it

[oak spoke]

doesn't seems to be any anime model yet

[nova remnant]

i have one im gonna uppload v2 soon

but its versatile for both anime and semi realism

[oak spoke]

also wtf is the pony model?

[nova remnant]

if you are looking for purely anime, you could try animapencil xl and use that with lightning 4 step lora

https://civitai.com/models/261336

https://huggingface.co/ByteDance/SDXL-Lightning

use the sdxl_lightning_4step_lora.safetensors from bytedance

and use that with anima pencil xl

[oak spoke]

is the lora a necessity?

[nova remnant]

4 steps 1.5 cfg is good for it

yes

[oak spoke]

that's some low step lmao

[nova remnant]

but the model i have up built on that lora so no need to use lora separately

[oak spoke]

all my old stuff are like 50 steps and 10 cfg

[nova remnant]

you wont need 50 steps

[oak spoke]

damn do ping me when you upload it

[nova remnant]

this is the level of detail you get with just 4 steps and 1.5 cfg

its the v2 i just worked on, gimme some time to upload it

[oak spoke]

👍

man I feels like a boomer last time I was here the anime pencil XL was still on 1.5

[nova remnant]

lol technlogy changing fast

we are gonna get sd3 soon

[nova remnant]

that wil make all of the sdxl obsolete

[oak spoke]

honestly pony model stuff look neat

[acoustic locust]

Bytedance made a model?

[oak spoke]

have yall play with it?

[nova remnant]

i couldn't settle with pony lol

[nova remnant]

yeah been a while

[lyric zealot]

pony model is kinda popular

it became the "mainstream" SDXL model for anime user

[oak spoke]

damn

I'm not surprise lol

[lyric zealot]

then there comes the pony merge that's mixed with pony model

[nova remnant]

sd3 is given to selective companies thus far so gotta wait till it comes out publicly

[lyric zealot]

it's kinda different so it has it's own tag on civitai

[nova remnant]

smol sus i pushed myself and played with it for a while lol but wasnt happy with the results

[oak spoke]

I wonder if SD3 will have higher requirement

than SDXL

[nova remnant]

sd3 will have version variants for different scales

[lyric zealot]

i liked it, but prefer the merge of it called autism

[oak spoke]

I see

[nova remnant]

i kinda got lost with getting the right output ...

admire your work tho smol sus 🙂

[lyric zealot]

In early, unoptimized inference tests on consumer hardware our largest SD3 model with 8B parameters fits into the 24GB VRAM of a RTX 4090 and takes 34 seconds to generate an image of resolution 1024x1024 when using 50 sampling steps. Additionally, there will be multiple variations of Stable Diffusion 3 during the initial release, ranging from 800m to 8B parameter models to further eliminate hardware barriers.

[lyric zealot]

did you use their trained tag?

[nova remnant]

yes all the custom prompts for it

[lyric zealot]

appreciate it

[oak spoke]

can 12gb VRAM run 8B at all?

[nova remnant]

but despite the prompts i wasnt getting results that i liked.. i dunno maybe its the style that i didnt fit with

jingo.. sd3 will likely work with 8gb gpu

there are other veriants that will require higher gpu

[lyric zealot]

i guess it's not the right style for you

[nova remnant]

@lyric zealot

[nova remnant]

sigh yes... lol i had hard time adjusting with anime itself at first

[lyric zealot]

the 8B parameters might be hard, since it's taking 4090 34 seconds to generate, but that's before the community can get their hand on optimizing it

[lyric zealot]

i woke up for Dragon's Dogma 2

[lyric zealot]

maybe exploring artist tag can be a right way? Since it knows a lot of artist

[nova remnant]

you're probably right.... cause i wasnt paying much attention to artist specific styles

[lyric zealot]

yup yup, but it's kinda a pain to find all artist tag that works. Good thing is that, there's people that already did that

search for post from eface and you should be able to find one that he made for pony specific

i think nubby also made a few for different SDXL models

[nova remnant]

yeah for sure, having the right art style would make it lot easier to get the right results

[lyric zealot]

https://tenor.com/view/dragons-dogma-dragons-dogma-2-hop-on-dragons-dogma-2-gif-16914165434528562015

[gloomy scroll]

I need to go sleep soon and then work for 7.5h

[lyric zealot]

good luck i guess

[oak spoke]

PONY bansai

[green temple]

yo ... guys are cookin` !

nice

[supple raptor]

It's basically like an LLM which means we will have all sorts of fun new ways to run it on lower end hardware. Such as quantization and CPU offloading. 8b SD3 should be similar in demands to 13b LLMs from what I understand.

[oak spoke]

it'll be fun once it got optimized

I wonder what it can make

[supple raptor]

search for images from @distant remnant in here for some SD3 anime examples.

[steady grail]

cool

meanwhile i'm merging dpo on some models lol

if this goes well, i add turbo too

guys, in your opinion? what degrade less the image, lightning or turbo?

[oak spoke]

the lighting go crazy lol

[supple raptor]

oh god. why are people merging any of those directly to models? you want to be able to control the weight of them when using the loras.

[steady grail]

to weight less on the pc and because when not merged, they usually burn the images very easily

[supple raptor]

Yeah, I'm not a big fan of the style that it has for anime, but it's better than previous base SD versions could do.

[oak spoke]

at this point I'm just very content with pony

[steady grail]

and dpo, unless is weighted at 1, looks like not doing anything

[distant remnant]

look at the quality of the fluffy tail

[steady grail]

wow, so cute

look the snow, the background

it's so cool

what happens when prompting raining dogs and cats in this prompt? XD

[supple raptor]

it's like <400mb for the lora, which you're probably going to want to have anyways for other models that don't have it. also, there is no difference between merging the lora and using it at the settings you merged it with. merging loras.

[oak spoke]

but can it do gore?

don't post it tho

just asking

[steady grail]

really? even when merged, the load time and weight on the pc is the same?

i can't wait for it to improve and be merged with other stuff

[supple raptor]

when using lora normally, load time happens once per time that you adjust it and it's so minimal that it really doesn't matter even when using a slow HDD for storage. and it takes up the same amount of vram because loading a lora is just merging it to model inside vram.

the other problem with merging a distilled model lora directly to the model is that it's relative weight will change when other loras are applied on top of it or the model is further merged.

you really should just figure out what you consider the best settings to run the DPO/Turbo/Lightning/LCM lora on your model and provide that information along with a link to download the relavent lora rather than permanently merging it.

[steady grail]

oh, ty for the suggestion

after merging the dpo, looks like the gen lost details 😦

not what i expected

no, i was wrong, the original model was of a different version. instead, looks like that after the merge, the model generate less leftover noise

[supple raptor]

distilled models will tend to produce lower levels of details when compared to a non-distilled version with both at sane settings for the type of model, they tend to do things like repeating the same or similar poses when given a somewhat vague pose prompt that would produce fairly varied poses on the non-distilled version. they have a lot of downsides, which is why a lot of people aren't using them despite reduction in generation time.

I'd much rather spend 10 minutes making 10 images to pick 1-2 good ones than spend 10 minutes making 100 images to pick 3-4 decent ones.

[steady grail]

yeah, i'm using turbo because the workload of xl on my pc is still to heavy and i want to reduce the overload to the minimum.

even if diluted, some xl models are still worth it, compared with 1.5 models

some poses or situations, are simply impossible with 1.5 models, without more effort

what? i generated something on 1.5 at 1024x1024 without mutations

wow, dpo on 1.5 is a day and night difference

[steady grail]

what is happening with forge? i'm getting less mutations with 1024x1024 than 512x512 on 1.5 models, is already the 3° i tested

[oak spoke]

I was literally about to say that 512 x 512 have so much problem now

with some SDXL model lol

[steady grail]

lol

never genned lower than 1024 on xl

[oak spoke]

it used to be perfectly fine

I just use upscaler anyway

but well that was "back in my day"

where my stuff need 50 steps to look good

🤣

[oak spoke]

but yea I need to switch up my resolution

[nova remnant]

i can't make up my mind, what i want to do with the model im working on....

im kinda indecisive about two version i came up with.... one leaning towards realism and the other semi realism (cuter)

you know what .. im gonna cut a balance between those two lol

[oak spoke]

there's already plenty of realism model just go with cuter one

[nova remnant]

lol... im thinking that too, but having difficulty deciding

[oak spoke]

well you got one vote from me

[nova remnant]

the cute variant looks somewhat like sd 1.5 anime but more crisp in quality like sdxl

what the heck.. i'll go with it

finalizing ..

[vital raptor]

Why not both?

[nova remnant]

well the version i have up atm is more realism than anime

so i wanted to be selective about v2

although v1 isnt hardcore realistic, i started with the idea of sd1.5 cuteness

[nova remnant]

wanted to keep the look perkier

[thorn crypt]

Hiii

[nova remnant]

rocks, bush, outdoors, cute woman, cute outfit.

5 mins till upload completes

[steel arrow]

Did the same prompt for fun

[nova remnant]

lol cool

@oak spoke its uploaded and live

[nova remnant]

cant link it here, page is nsfw

[nova remnant]

texture looks so good

[nova remnant]

some people are posting amazing pics on civitai with pony model

[native halo]

https://tenor.com/view/food-processing-foodie-korean-food-chinese-food-gif-23908270

[thorn crypt]

cool

[native halo]

https://tenor.com/view/azumanga-daioh-gif-16417332777048289714

[thorn crypt]

https://tenor.com/view/nervous-nervous-anime-sweat-sweating-nervous-sweating-gif-8464066390113296650

[native halo]

https://tenor.com/view/squidward-squidward-shocked-squidward-tonight-tonight-squidward-shock-gif-23057510

[nova remnant]

@oak spoke 4 steps

[steel arrow]

Is this the model that you cant share?

[nova remnant]

its up on my page

[steel arrow]

Same name as discord?

[nova remnant]

yeah there are two versions this is v2

[steel arrow]

Nice

[nova remnant]

a random pony post ... this is SD3 quality

[nova remnant]

is this sd3?

[distant remnant]

yes

[nova remnant]

no joke, looks amazing

[nova remnant]

[native halo]

https://tenor.com/view/jack-pandora-hearts-cool-shhh-jack-vessalius-gif-24202179

[nova remnant]

ugly spike hair

[native halo]

what about this hair?

[nova remnant]

messed up hand

[nova remnant]

grabby

[nova remnant]

[oak spoke]

Gib link

[steel arrow]

Handstands don't make very good faces

[green temple]

is there a difference ?

[steel arrow]

Barely

Imo

Silly me, I didn't upscale it

[steel arrow]

I'm liking these pelt armour things

[oak spoke]

the half fursuit

howdy sinners

[steel arrow]

Pretty much

I have accidentally made a robot girl death. A little depressing

[oak spoke]

come drink tomato juice with her

[steel arrow]

This one will join in

[vital raptor]

put it in a cup you heathens

[steel arrow]

Instructions unclear

[oak spoke]

why's the color of your so off? are you not using VAE?

[steel arrow]

No im not

[oak spoke]

lol figured

VAE enchanted the color by a lot

[steel arrow]

Yeah I only just redowloaded SD the other day

[oak spoke]

same I just got back yesterday literally

[oak spoke]

here's with and without VAE side by side comparison

coincidentally these rainbow color are made with lora I upload you can find it

it's the only lora I ever upload tbf

with and without VAE

even my lora won't save the VAE less stuff

[steel arrow]

Perhaps I should get one then

[vital raptor]

depends on the look you want tbh, i think the non-vae ones look better, the other ones are way over saturated

[steel arrow]

But im going to sleep for now

[vital raptor]

at least in those examples, i use a vae personally

[oak spoke]

in this specific example it's vibrant cuz of the rainbow lora which made to be as saturate as possible just to show the different

that even pushing it to limit without VAE it still can't get there

[supple raptor]

You can't make an image without a VAE. The VAE's job is to convert between latents that the UNET can use and pixels that can be displayed as an image. Checkpoints will normally have a VAE integrated in them unless the uploader was an idiot that pruned it out in which case you'll either get an error or your software will automatically use one that it has access to.

[oak spoke]

a lot of older 1.5 one function without it

and it just have weird color

[gloomy scroll]

Most of the models have VAE baked into them so when you set a1111 VAE to none it uses default baked in model

[near elbow]

have some wholesome love

[oak spoke]

married couple I see

[heavy kindle]

In which channel are you people generating images?

[umbral depot]

Most people use their local computers

[steady grail]

🔴 Bot is currently down! Bot channels are also unavailable until we are ready to return.

In the meanntime, try running Stable Diffusion locally or using Dreamstudio & Clipdrop.

[dire bridge]

they are down.

[heavy kindle]

Oh. When will it get back to normal?

[dire bridge]

Idk

[steady grail]

i'm invisible :/

i suppose i can undress myself and run around naked XD

[nova remnant]

dm'd you

[oak spoke]

me with fire extinquisher: bring it on

[steady grail]

runs around the channel totally naked, covered only by a small towel

[oak spoke]

https://tenor.com/view/fire-extinguisher-shut-up-be-quiet-fight-angry-gif-9530868

now this give me good idea for next art

also soldier in alien vs predator should have these as issued equipment

[steady grail]

damn, i'm covered in foam and not invisible any more 😦

[nova remnant]

civitai model naming convention on server end is confusing