🍥｜anime

[bitter latch]

😄

[steady grail]

very unique, almost gummy and abstract

guys, to polish images from weird things, now i use lamacleaner 1.1.2, someone know if out there there is something better?

[supple raptor]

Correct. They apply differently. I don't remember what exactly the difference is.

[supple raptor]

MUST.... PAT!

[steady grail]

thhis is justice 😎

[steady grail]

this is so old style

[south sable]

its a blend of Syd Mead, HR Giger, and Anime :p

"Come on ladies, you can't fool me... I didn't chest burst yesterday you know!"

[steady grail]

looks like i can't find raichiyo33-2184it-novae(d46bbb9c20c3) lora model

[cyan burrow]

how consistent

[gloomy scroll]

Those aren't on civitai

[cyan burrow]

it's either easily duped or just dumb

[gloomy scroll]

I can dm you place where you can find those if you want but it's nsfw list

[steady grail]

ty, honestly, the most artistic and coherent models i found were nsfw, so i'm not against using them

i just need to suppress them with nega

additional network is weird, it's hard to find and set lora with it

[steady grail]

looks like the little vampire in negima sensei

[steady grail]

ninjaaaaa!!

[runic vault]

well there is a lora for her already but i wanted to see far one could get without one. (reference image)

[cold talon]

yas

[steady grail]

ohhhhhh, Rachel Alucard

i love bb

[steady grail]

how do i make this 2 token m_akram, m_ouroboros affect the prompt while using aog model?

[supple raptor]

You'd have to merge those models into it. Since Vil was just using regular AOM2 and not a merge with those models, I assume they just grabbed that prompt from a previous image that was genned on a merge with those models and didn't bother/forgot to remove the activation tokens.

[steady grail]

yeah, probably different gens of inpainting

[cyan burrow]

testing my latest private build, based on the model that I won with this week and another I have stashed away (neither is dark gemini)

[gusty palm]

its sad seeing the corporate AI's struggle with their guidelines limiting

"i have no biases or anything but im unable to comment on that or engage in this topic for these obviously preprogrammed reasons"

i was going through conversational models today is there any llms that run on like standard hardware

https://github.com/imoneoi/openchat

the ones i saw needed like 24gb vram

i need mnore like this

[south sable]

they all seem to need a ton, I can run H2OAI but its slow as molasses in january

[gusty palm]

maybe since its like 100F here in florida itll be faster

;]

[south sable]

https://tenor.com/view/kel-omori-redacted-gif-26497143

[gusty palm]

i wasthced star trek earlier so i lold

sponk got turned into a human

[south sable]

SPONK

https://tenor.com/view/laughing-dying-gif-24782560

I almost fell out of my chair

[gusty palm]

;D

[south sable]

He is forever Sponk now LMAO

https://tenor.com/view/out-of-shape-out-of-breath-too-hot-the-studio-handsome-filipe-gif-17315062

me after laughing so hard

[gritty spruce]

Just a friendly reminder, we need everyone to go and vote on the bot channels, please select what image you like best: "A" or "B", this is for helping improve and finetune what will be the gold ⭐ SDXL 1.0 model.

[gusty palm]

that was fun

[cyan burrow]

awww what an adorable undead family

[supple raptor]

I would recommend looking into 4bit quantized GGML models. These can be run on CPU using system ram, optionally with parts offloaded to the GPU.
This guy converts an insane number of models to GGML. (For people with a bit more vram, he also does 4 bit quantized GPTQ models, which run primarily on GPU in VRAM and parts can be offloaded to CPU.)
https://huggingface.co/TheBloke

Both GGML and GPTQ quantized models can be run using Oobabooga
https://github.com/oobabooga/text-generation-webui

I run Oobabooga in WSL fwiw.

[gusty palm]

sweet

[shy sedge]

anyone has sd on google colab?

[gusty palm]

saw the bloke earlier I didn't understand what half of it meant though

[gusty palm]

saving these thanks

[steady grail]

how do we vote?

[gritty spruce]

go to any of the bots https://discord.com/channels/1002292111942635562/1100170312106127410 to https://discord.com/channels/1002292111942635562/1101178553900478464 and please vote on the generated images you see your preference A or B

[gusty palm]

that stylepile extension is kinda neat

[gusty palm]

i was totally voting in the wrong channel

[runic vault]

well it kinda works

[gusty palm]

first attempt at loopback scaler

very improve

[gritty sundial]

do you use any LORA? because my skeleton is all deformed on my face

[gritty sundial]

which model do you use?

[drowsy dove]

Every picture takes 6 minutes, hopefully I'll find out what is causing this weirdness. The problem was the VAE
(removed the SDXL pictures I sent here because they weren't good at all)

[cyan burrow]

private one

[drowsy dove]

This is so ironic. I can't use the desktop to generate images with SDXL because it has not enough RAM to use while having 12 GB VRAM. Meanwhile the laptop with its glorious GTX 1650 with its 4 GB VRAM can generate pictures with SDXL

[supple raptor]

SDXL runs fine on 12GB vram. Which WebUI are you trying to use?
Also, I saw you mention a problem with the VAE, the official SDXL VAE is the only one that's compatible with the model. Embeddings from SD1.5 based models sort of work since the small text encoder is the same, but they really aren't worth using due to not being compatible with the larger text encoder.

[cyan burrow]

Their daughter turned human as she aged, prompting daddy to become the man we know as Skeletor when he realized his wife cheated on him with He-Man

[drowsy dove]

I'm using Automatic1111, 1.5.0-rc1

[drowsy dove]

I'm impressed it still works with 4 GB VRAM

[cyan burrow]

Im just waiting for my dreamstudio credits so I can try sdxl on it

[supple raptor]

It must be running on CPU to work on 4GB vram because even if you converted to fp16, the model alone is 6.46GB

[supple raptor]

Just use the bot for free. I don't think DreamStudio even has the 1.0 candidates yet.

[drowsy dove]

Well, nvtop show 100% GPU load and nearly 4 GB of VRAM, so the video card is doing something

[supple raptor]

Hmm... Maybe it's swapping 1 part at a time in. It could probably hold either one of the text encoders or the unet, just not at the same time.

I'm looking to see what I can find about 1.5.0-RC's SDXL support right now. I've only used XL in Vlad's fork and ComfyUI so far.

[drowsy dove]

SDXL seems to understand nighttime better
And it seems to understand the cube is glowing too

[cyan burrow]

Uh maybe tomorrow then, don't normally like public prompting

[supple raptor]

Looks like Auto1111's SD-XL support is lacking refiner model support which is odd because he does have the settings needed for the refiner's aesthetic score (which Vlad is missing, despite him having support for the refiner model).

[supple raptor]

Ok, so it does support refiner, the implementation is just kind of dumb. To use it, you have to send to img2img, switch to the refiner model, and just do normal img2img.

[untold glacier]

...Imoen??

[pale monolith]

yes!

[untold glacier]

That's only my favorite game of ALL TIME

[pale monolith]

ME TOO!!!

literally why my go to is "heya" to this day

[untold glacier]

[untold glacier]

I'm currently on my nth playthrough!! Modded nearly to death. I've never talked with someone who loved the game as much as me, lol.

[pale monolith]

same! First time i'm using the ascension mod, no clue what im in for xD

[untold glacier]

Oh boiiiiiii

Let's take it to the DMs

[cyan burrow]

Have we taken over this channel now or something?

[supple raptor]

Yes, the anime channel has taken over the anime channel.

[drowsy dove]

Take two because the skirt didn't make much sense before. A hint I should be sleeping now

[steady grail]

what's the game?

[untold glacier]

(Almost realistic anime)

[bitter latch]

sneaks back in

hmm little out of focus

hmmm

makes a new anime merge

now let's test it

any idea's for prompt theme?

besides just anime

interesting

removes the lora's

goes with "miku sailormoon"

waits

[bitter latch]

o.o

hmm works well... I think

[pale monolith]

astronaut?

[bitter latch]

interesting shading

suppose to be 80's drawn she-ra

[bitter latch]

dats creepy

[steady grail]

wow, very clean, how did you do it?

[bitter latch]

marged 3 anime toon models

ok two anime and 1 toon

[steady grail]

i love it, can you share?

if you feel it

[bitter latch]

tho the last few where img2img which prob helped

[steady grail]

it's really fantastic

[bitter latch]

thanks

someone recommened using Heun instaed of euler and it seems to work better

takes longer tho

used this one as the config file which is what i think helped. https://civitai.com/models/84586?modelVersionId=89927

[bitter latch]

this is a good one too https://civitai.com/models/83231

and mistoonsanime model . thats the three I used to merge

yawns I should be in bed

hmm the coco one , same seed and prompt.

blah

one last one then bed

goes to bed.. thud

[steady grail]

night

[young gulch]

#🍥｜anime dog

Сообщение

[supple raptor]

There isn't a bot in this channel. Go to #1100170312106127410 to generate images.

[wooden coral]

SDXL

[nova remnant]

Have I missed any exciting new models, extensions, loras, etc in the last month or so?

[ocean reef]

depends what you like, but there are a lot of new ones

I always filter all models/lora by newest on Civitai

[runic pine]

What model are you using? Sdxl?

[ocean reef]

Camelliamix

[slate narwhal]

Some Sdxl (bot):

No bokeh prompt:

[zinc flax]

won that hand gacha

[untold glacier]

Anaglyphic eyes. I love it.

[ocean reef]

😍

so pretty!

[lyric zealot]

the magic of this lora https://civitai.com/models/112808/chinese-art

[ocean reef]

[lyric zealot]

rare lora that's good at weight 1

at least for expmixlineV3 model

[ocean reef]

yeah, depends on model, some loras are hardly visible for some I use

[lyric zealot]

[untold glacier]

Then you have some like the Yoshitoshi ABe LoRA which can still be too strong at 0.5 with many models

[ocean reef]

sure, can find loras like this 😄 I usually use only add_detail one

[untold glacier]

I have that LoRA but still haven't found a use for it. Usually when I want more detail I switch to something like SushiMix or Breakdomain

Often times my problem is too much detail, heh.

[ocean reef]

oh, I have one model where add_detail even with -2 doesn't work 😄

it generates waaaay too many details

[untold glacier]

Oh right, I didn't even think to use it as a negative. I'll have to try that.

[untold glacier]

Which model?

[ocean reef]

Nabimix v2

v1 allows to remove some details, v2 doesn't

I did a looot of tests on those two 😄

[supple raptor]

there are at least 2 add details loras btw
https://civitai.com/models/58390/detail-tweaker-lora-lora

https://civitai.com/models/82098?modelVersionId=87153

[ocean reef]

Yeah, I use 1st one

[untold glacier]

Yeeeah there's been a few I've seen.

[supple raptor]

1st one is the original. 2nd is also quite good. Seems like it might be compatible with more models from what I've seen.

[ocean reef]

daaaamn

that's so pretty

meanwhile I'm trying to get this hairstyle more often - not an easy task to do xD

[ocean reef]

[ocean reef]

0 details look cool

sometimes less is better 😄

[lyric zealot]

true, but the fingers are fucked

[ocean reef]

I wish we will live to the day when fingers are generated perfectly

[untold glacier]

[steady grail]

fingers are hard to draw even by people

[lyric zealot]

weight -1

[steady grail]

she mistaken sugar with salt and is trying hard to hide it 🙂

[untold glacier]

Even by established mangaka. I was reading a manga just the other day where there were some mighty wonky fingers.

[steady grail]

it's... very... tasty... anata

[supple raptor]

waifus just like sticking their hands in blenders. they can't help themselves

[ocean reef]

1st one is better 😄

[untold glacier]

[steady grail]

lol

[steady grail]

don't make me think about it.... i studied drawing for many years...

fingers are weird, i can't just grasp them well

[supple raptor]

hands are complex af. there are so many different shapes that they can make.

[steady grail]

that's because i cheat and i just use my hand as a model lol

[lyric zealot]

kinda weird style at weight 1

[steady grail]

i chose the position and make a photo, fu*k yeah

[ocean reef]

as long as you know anathomy, everything is easy 😄 but anathomy is not easy!

[ocean reef]

I rather draw animals than humans! They are easier!

[ocean reef]

kinda overweighed

[steady grail]

anatomy totally another subject of study

[lyric zealot]

totally, but kinda cool nonetheless

[steady grail]

it totally helps, but it's so hard core

you need a life to learn it

[ocean reef]

you just need to learn it early and study until you understand all muscles work, bones, etc. So yeah, takes some time

But training makes you the master

[steady grail]

why when i try to refine something, i get some form of eldritch monstrosity?

[supple raptor]

https://cdn.discordapp.com/emojis/1091075010074001489.webp?size=64&name=mikupanic&quality=lossless

[steady grail]

i don't know if i should call it just a mistake or directily Gore!

it's scary

[steady grail]

oh, cute, but something is on her air.... tandrils?

[wispy canopy]

just me prompting abstract stuff 😄

SDXL is gonna be so good 😮

[steady grail]

i suppose control net and change of model don't relate too much

eface, do you think i can run sdxl with tiled vae and 4g vram?

[untold glacier]

This is SDXL?

[wispy canopy]

that's sdxl 0.9

[untold glacier]

[wispy canopy]

you might, but i don't thinkyou'll be able to get highres. SDXL is outside the ballpark

[steady grail]

i could try to make the shape with sdxl and than refine it with some other model

[untold glacier]

That begs the question as to how well it does at lower resolutions than it was trained on. We already know 1.x and 2.x don't do too well below their threshold.

[steady grail]

i can see his coerence is pretty high

[wispy canopy]

sdxl simply fails hard below treshhold

[steady grail]

oh

[wispy canopy]

you go under 768x768, it will fail.

[steady grail]

what's his standard of res?

[steady grail]

768?

[untold glacier]

Hmm. Looks like we'll need to use 1.5 for low res gens then.

[wispy canopy]

normal sd = 512

sdxl = 1024

[steady grail]

well, it's not too high, i can generate at 1024 already

[steady grail]

i'm i like to push the boundaries of stable diffusion

[untold glacier]

https://tenor.com/view/this-is-the-way-this-is-the-way-mandalorian-mandalorian-i-have-spoken-baby-yoda-gif-24159898

[steady grail]

ok, i admit the driver of my gcard crush pretty often, but, usually it runs

[supple raptor]

I would turn the denoising strength way down, like 0.4-0.5 probably.

[true basalt]

Damn that’s super clean and detailed. SDXL 0.9?

[wispy canopy]

yes

[steady grail]

is not worth it, the 2 models are just not compatible

[steady grail]

when refining, i noticed, the compatibility of the the models is very important

they must think alike to get something decent

looks so much matrix

[supple raptor]

refining just needs lower denoise if the models aren't compatible enough. There are a ton of people using various 1.5 based models as their refiner for SDXL models currently.

[steady grail]

now i'm doing the last experiment to see if i can refine this image how i want it. if not, i just change metod

[steady grail]

giving an example, with this model, at denoise 1.5 i got this

it's just weird

[gloomy scroll]

#1055215351731851354

[steady grail]

i love it

[wispy canopy]

is that with refining?

[steady grail]

it's just not worth it.

[wispy canopy]

i'm using refining

but using sytan's method

[steady grail]

how?

[wispy canopy]

do about 60% of the steps on the base model, and then hand over the latent noise to the refiner model to finish it

is a workflow in comfyui

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[steady grail]

oh, i'm not using sdxl

[wispy canopy]

this one saved SDXL local for me

but you can apply this workflow to regular models too

[steady grail]

she is a yandere

[zinc flax]

Maybe

[ocean reef]

kuudere

[wispy canopy]

i'm pretty sure there's people in this channel that have a degree in dere-ology

[wispy canopy]

you could say she has a lot on her mind

[zinc flax]

beware of the hand

[supple raptor]

If you want to try using a 1.5 based model as the refiner, you can do it by adding a vae decode using sdxl vae, then vae encode using a sd1.5 vae to the latents between first pass and second pass. if you do that, make sure to also switch the vae connected to the final vae decode over to an sd1.5 vae.

[supple raptor]

wdxl made sad miku

[bitter latch]

how you do that?

[wispy canopy]

uhh... i prompt? 😮

[bitter latch]

https://giphy.com/gifs/StayBK-george-smarty-pants-a-bear-named-KHJSKOsvc5glNvl85w

[wispy canopy]

it's just a txt2img prompt in sdxl

but, throw in the tag "double exposure" in sdxl, and magic happens

[bitter latch]

that is alot of info to dig thru . wow

[zinc flax]

hapi

[wispy canopy]

ooooooh ❤️

[bitter latch]

hmmm

[wispy canopy]

she's seeing the whole galaxy at once

[bitter latch]

could be

[zinc flax]

she's chilling

[untold glacier]

Love this one

[bitter latch]

Good to see other anime style that isn't the same ole same ole

looks at this piece sideways.. huh?

[bitter latch]

tho the art came out wierd i'm loving the stylish anime it's giving me with the merge I did

[untold glacier]

How's she gonna get off of there?

[bitter latch]

takes out double exposure, not doing much with the normal models

[wispy canopy]

anime models won't be able to use it

that's why i kinda stopped using anime models allthogether

[untold glacier]

This is really nice. But it's not anime 😅 You may wanna put that in #1072015001029722182

[wispy canopy]

(darkalfa is an anime model, but i made sure it does know other things outside anime)

[brittle rampart]

oki-doki

[bitter latch]

oooh

[bitter latch]

I use non anime models to make anime art too

[slate night]

Nia my beloved

[slate night]

hugs her gently

[bitter latch]

ah HA!

I would so put that on my wall!

[untold glacier]

Btw, how does XL do with generating upside down people? Something still beyond the grasp of current gen SD.

[wispy canopy]

i suppose that's still the area of "don't bother trying "XD

[untold glacier]

Lol

[bitter latch]

adds miku as part of the prompt....

[bitter latch]

she has no feets

[untold glacier]

Chibi-style characters often don't have feet.

[slate night]

Don't need feet to be cute

[ocean reef]

she has boobies tho

all what she needs

[bitter latch]

ok time to watch some star trek

[ocean reef]

angry man 😄

[bitter latch]

just realized I had very little neg prompt on that.. huh

[gusty palm]

can she swim

[bitter latch]

switches back to the anime merge. let's see how this goes and that didn't go well (hides that one)

[gusty palm]

they should do an episode of star trek with human characters from lower decks

[untold glacier]

...not really

[gusty palm]

looool

[untold glacier]

I have the same problem with such poses. I haven't looked to see if there's a lora to fix it yet but I'm tired of having 3-4 knees.

[spring lion]

i would like to try lora! so far i know its kind of like a pre-set controlNET pose with a pre-loaded character

[gusty palm]

theres pose loras and theme loras too like aSthetic

theres a hugging knees lora

[supple raptor]

LoRA is basically a miniature model. It's trained similarly to the way a model is trained and it can learn different poses, styles, characters, objects, or other concepts. They're much more efficient (faster) to train than a full model and result in a file that is only a few hundred MB in size or smaller rather than a full checkpoint which would be ~2GB or larger.

They are effectively merged into the model that you're using inside vram so they are also much more flexible than a fully trained model.

[spring lion]

thanks 🙂

[supple raptor]

Depends on the hardware you have available. Even with a 3090/4090 I would probably stick with training LoRAs. If you have datacenter hardware available then finetuning could be worthwhile.

your goal for training would also matter, if you just want to train a character/concept or train on a style that you don't have a few thousand images for then LoRA would probably be better.

[brittle rampart]

very disappointed that I can not pull out the details

[untold glacier]

This is some sort of glitch, I'd try restarting SD and/or slightly altering the CFG by 0.1 in either direction. I've encountered it myself but still don't know what causes it exactly.

[brittle rampart]

but the essence is not only in these places. When the size is increased, the details are not specified, but distorted. this is frustrating

[untold glacier]

I'm talking about the general desaturation and hazy look from the glitched gen. As for the issue you're talking about, it could be how you're upscaling it. Unless you're not upscaling it--in which case I'd recommend using hires fix with an upscaling model such as fatal_anime or HugePaint at a denoise of 0.4-0.5.

[brittle rampart]

ah yes. I'm using hires fix with an upscaling model on this is before. now i try upscaling in Img-to-img

I'm using rather weak hardware and I can't immediately perform the scaling process quite strongly

[brittle rampart]

i`m using hires fix with an upscaling model such as 4x_foolhardy_Remacri at a denoise of 0.2-0.3. then try upscaling using img-to-img changed bit denoise and CFG

[untold glacier]

SD likely needs more to work with when you're using Remacri to upscale with hires fix. I'd increase to 0.45 or 0.55-0.6 with the denoise so SD can do its thing to add detail in.

[brittle rampart]

oki-doki... now try it

[gusty palm]

idk what this is but i likey

[brittle rampart]

and where i can get hugepaint upscaler?

[untold glacier]

https://1drv.ms/u/s!Aip-EMByJHY26wU-wOqApp7Hu7tt?e=4ihnBr

[gusty palm]

is there a trick to those super wide images or do u just set the resolution that way

[untold glacier]

Lol

[untold glacier]

Depends on the model and prompt, really. Part of it is luck, and much of it is understanding how a model works to get it to do what you want.

I've found that a sweet spot is often 440x1024, before hires fix or upscaling.

[brittle rampart]

and not envy is from the model?

[untold glacier]

Sorry, what?

[brittle rampart]

oh it not for me. i`m sorry

[gusty palm]

had no idea queercore art was a thing

[supple raptor]

You need to enable a VAE.

• Download this: https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.ckpt
• Put it in .\stable-diffusion-webui\models\VAE
• In WebUI go to Settings Tab > Show all Pages > Press CTRL+F > Search "VAE"
• Click the blue refresh button next to the vae dropdown
• Select the file you downloaded in the first step from the VAE dropdown and hit apply settings.

[brittle rampart]

tell me what tod o VAE?

[supple raptor]

The VAE is the part of a model that is responsible for turning latents (a numerical representation of the average of 8 pixels) into actual pixels that can form an image. Without one at all, you would get no image. They are supposed to be built into models, but many people remove, or "prune", them to make the file size a little smaller. When you use a model with a pruned VAE, you get images that look mushy and lack details like the ones you posted.

[brittle rampart]

and how much the change style of the images of the selected model if the VAE does not match it

[supple raptor]

VAEs are universal between all SD1.4, 1.5, 2.0, and 2.1 based models. Some can work slightly better with certain models than others, but in all cases they will look way better than just using a model that has a pruned VAE without loading one.

SD-XL uses different latents and needs it's own VAE.

[runic vault]

fumo satan

[brittle rampart]

thanks a lot for the clarifications. I will try to regenerate images already obtained earlier using VAE

[noble lion]

good fumo.

[brittle rampart]

VAE really gives amazing results

[supple raptor]

Much better! That scythe is outstanding!

[lyric zealot]

[bitter latch]

oh and werewolfs are not in flockes....

[bitter latch]

restarts auto1111 and tries again

listens to some rag'n'bone man

hmm

[bitter latch]

nice if a bit overloaded

little red grew up and she's not taking crap from no one

tries to make it more anime

[gusty palm]

def moving to Liecoe one day

[untold glacier]

Knees are the new bane of my existence.

[brittle rampart]

heh... i `m just using controlNET for my regeneration..

[gusty palm]

do u need mega hardware for that

[brittle rampart]

yes... But, Ruby herself has matured brutally!

[supple raptor]

inpaint only masked on the face should take care of that.

[brittle rampart]

huh! this is another unplowed field in my work with SD! I don't understand the whole process..

[supple raptor]

• Put your image into the inpainting tab, draw in the WebUI over the area you want to inpaint (in this case it would be the entire face)
• Set mask blur to ~4. This is the number of pixels to blur around the edge of the mask in order to blend it into the rest of the image
• Set mask mode to inpaint masked,. This means it will inpaint the area that the mask covered, the other option would inpaint everything except the masked area.
• Set masked content to original. This means it will use the original image as context for the inpainting.
• Set inpaint area to only masked. This will resize the masked area up so that the inpainting happens at a higher resolution, then resize it back down to fit into the original image for you. This is the most important part for getting more details in your inpainting.
• Set Only masked padding, pixels to approximately 8. This is the minimum number of pixels around your mask that the setting above will use for the inpainting.
• Set denoising to approximately 0.4 to start with. Increase this if the image isn't being changed enough, decrease it if it is being changed too much.

I usually leave the resolution set to the same resolution as the image for this step. For the prompt, you want to mainly prompt for how her face should look, and keep any style tags or descriptions that you used in the original image. Try to avoid including things that will not be within the area that is used for inpainting, such as the scythe or the wolf.

[brittle rampart]

i`am sorry "need for what"?

[gusty palm]

controlnet

[gusty palm]

thats so concise and helpful

[brittle rampart]

Wow! Lot of Thanks! I am must try it. but first i am need a really big coffee cup!

[brittle rampart]

why?

[gusty palm]

jw if i can use controlnet

i saw a extension for pulling poses from koikatsu party into controlnet

[supple raptor]

No, controlnet isn't too bad for hardware demands, just make sure to enable the Low VRAM setting in the Controlnet extension and use the pruned versions of the controlnet models. I can dig up the links for the pruned models in a few minutes.

[brittle rampart]

For the prompt, you want to mainly prompt for how her face should look, and keep any style tags or descriptions that you used in the original image. - Should I change the Prompt string so that it only describes what I want for the face?

[gusty palm]

ok

im not sure yet if id get use out of it im not sure what the uses are

koikatsu is a lot of work lol

[supple raptor]

Yep, because the inpaint only masked settings make the image that the AI actually sees be a cropped version of the original image.

[supple raptor]

openpose is a really nice controlnet option that is pretty simple to use. Basically find an image of a person with the pose you want and it does some AI magic to make a stick figure skeleton of the pose then uses that pose to influence the noise used to generate the image which results in the image giving the pose of the skeleton. There are even extensions to manually edit the skeletons right in WebUI.

You could even use a screenshot from koikatsu to make the skeleton.

[brittle rampart]

i am love use controlNet sometimes for poses..(exp: for Ruby i using Preproccesing "softedge" and yes just set Low Vram )

[supple raptor]

for example, if the first image was your original image with the black area being the mask, then the second image would be approximately what the AI would actually see for the inpainting, with only the black area in the first image actually being changed.

[gusty palm]

i wouldnt mind posing a skeleton that would be useful

[supple raptor]

The skeletons can also be used to precisely pose hands, which can make getting good hands much easier.

[brittle rampart]

Understood! Thus, it is possible to "pull up" parts of the image that initially turned out to be bad?

[supple raptor]

Yep, exactly. Usually problems like the face in your image are due to the resolution of the face just being too low for it to draw nice details. But this method performs the inpainting on that small area at a much higher resolution so it solves that problem.

[brittle rampart]

It looks damn tempting, creatively!

especially if you combine it with controlNET

[supple raptor]

Yep, if you use controlnet, just don't add an image to the controlnet image box, it will use the one that's in the img2img/inpainting box automatically and adding it to the controlnet box can make it do the controlnet magic based on the entire image instead of just he inpainting part which can screw things up.

[brittle rampart]

controlNET i`am use only at the initial stage of obtaining an image

[supple raptor]

@gusty palm here are the controlnet models
Controlnet v1 models (smaller but not as good) - https://huggingface.co/Jemnite/ControlnetPrunedfp16/tree/main
Controlnet v1.1 models (about twice the size but generally give better results) - https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

[brittle rampart]

What a 100500% damn cool tool! just compare the face in two pictures!

[gusty palm]

controlnet sounds very useful

[steady grail]

is maybe aurora the definitive chibi farmer?

[brittle rampart]

I completely agree!

[steady grail]

yeah, too bad it don't do too good with images not generated by the same model

[untold glacier]

Miles ahead of what you posted earlier, well done! Looks great

[brittle rampart]

What do you have in mind

[gusty palm]

does this do anything if i dont have controlnet, it looks like it is bsaed on the console but idk

[untold glacier]

Was referring to your earlier post: #🍥｜anime message
Big difference. Thought you were already using a VAE at the time, which is the reason behind the advice I gave--but yeah. VAEs make all the difference in cohesion and proper contrast with models, among other things.

[brittle rampart]

oh yeees! but I still can't fix her right hand using the method inpaint...

[brittle rampart]

in AUTOMATIC1111 webui it's like an extension

[untold glacier]

Hands are a still a crapshoot at the moment. I've had gens where I generated dozens of inpaintings and all of the hands were still bad. Sadly inpainting won't help much with that. The best combo is a halfway decent gen, a little bit (or a lot) of photoshop, and then throwing it back into image2image or inpainting.

[brittle rampart]

and if you use the same controlNET with Preprocessor openPose?

[zinc flax]

i didn't expect heart pupils to show up but I'll take it

[untold glacier]

Openpose can help some yes, but a canny/scribble/lineart control type with reference image may help more as there's a more direct reference for controlnet to "see".

[brittle rampart]

sounds like a good idea to "try"!

[gusty palm]

i inpainted her facec that worked nicely

[untold glacier]

Beautiful face. I look forward to the day we don't have horror hands though, heh.

[untold glacier]

Also, you can use the Adetailer extension to do that automatically with all of your gens. It's a lifesaver.

[supple raptor]

Maybe I'll put together a little guide for inpainting with the Krita extension tomorrow if I have time. It makes inpainting so much more enjoyable.

[brittle rampart]

😭 I want a new cool video card!!!!!
I'm tired of receiving these messages!
OutOfMemoryError: CUDA out of memory.

[supple raptor]

Do you use the Tiled VAE feature from the multi-diffusion extension?

[brittle rampart]

oh!! dont know really. where i can see it?

[steady grail]

attention, it increase the generation time very much

[brittle rampart]

in my extention tab i not see that "multi-diffusion"

[supple raptor]

Here's the extension link- https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git

Those settings are good for 12GB VRAM. If you tell me how much VRAM you have, I can suggest appropriate settings for you.

[steady grail]

but it's a very cool extension when you want to push the quality and have time

[supple raptor]

If you're nearly running out of VRAM, you'll be swapping to system ram anyways so Tiled VAE will be a time saver in that case.

[untold glacier]

Also, general upscaling guide, using multidiffusion: #🍥｜anime message

[steady grail]

pretty true

i sometimes i push a 2400x2400 on a 4 g

[supple raptor]

The trick for tiled VAE is to find the highest tile sizes you can use without OoMing so that it will only activate when you really need it.

[brittle rampart]

in my case, everything is really sad. I work for GF 1660 GTX 6Gb

[brittle rampart]

Wow.. it look very cool!

[steady grail]

she is living her life, for a girl with 6 fingers

go for you happiness, girl

[untold glacier]

[steady grail]

eheheh

i usually redraw them manually

there for example

so cute lol

forgot i had this lol

[supple raptor]

In that case I think 976 decoder tile size and 80 encoder tile size should be good for you. Try that and see how much VRAM it says you're using on a gen and adjust up or down as needed to try to get the VRAM used to be as close to the max as possible without going over. Decoder handles turning the latents into an image and Encoder turns an image into latents (this happens whenever you do any type of img2img), if you get an OoM error you can try to watch and see at which point it happens to know which one needs adjusted. In txt2img, if you aren't using Hires fix or Adetailer, the decoder will be the most important.

[steady grail]

sometimes i can push to 1360 with my 4g vram

but probably, i'm pushing outside the ram

[brittle rampart]

oh! thanks! for help! now i try all this!

[crystal meadow]

Does anyone know how can I achieve similar results?

[steady grail]

with a good model focused on bg

[supple raptor]

1344 decoder with batch size 2 will put me just below 12GB when I'm running SDP-no-mem optimization. Xformers buys me some more headroom, but I don't bother to. Hange for it.

[crystal meadow]

Any recommendation here?

[steady grail]

and abusing control net

i would need one too

[supple raptor]

The colors and texture are similar to my model, but I don't typically get that much detail in scenery.

[steady grail]

for such details, you need to construct the image piece by piece, use controlnet and set the prompt well

[supple raptor]

https://civitai.com/models/100505/kakigori

[untold glacier]

What Anime Model Should I Use??

• **DarkAlfa **is your best bet if you want the ability to follow the prompt at top quality, but it can have a somewhat distinct style with a cartoon-edge. It works very well with LoRAs though, and gives great results while sticking closely to your prompt.
  https://civitai.com/models/64843/darkalfa-effortless-high-quality
• **Aurora ** Is my favorite premiere danbooru-tag based model, as it gives fantastic, cohesive results with lots of versatility when using creative prompting. It works great with LoRAs and can replicate many styles.
  https://civitai.com/models/40199/aurora
• Breakdomain Anime has a great natural lightly painted art style. It works less well than others with LoRAs, but the amount of detail and cinematic presentation you get with it naturally means that many times style-LoRAs aren't needed. The author removed it from Civitai when some people complained on its potential similarities to other models, but it's still available on huggingface.
  https://huggingface.co/nyxia/breakdomain-anime/tree/main
• Darksushimix Colorful is another detailed, style-filled model if you're less concerned with cohesion (as is often the case with models that give lots of detail.) "2.25D" is less flat, whereas "Colorful" works best for flat styles. It's what I used for such things before using Breakdomain.
  https://civitai.com/models/24779/dark-sushi-mix-mix

All of the above can give great flat color, traditional anime styles. All of the models work best at clip-skip 2 aside from DarkAlfa, which is best at clipskip 1. Additionally, DarkAlfa works well with natural language prompting, whereas the others often work better with danbooru tags.

[supple raptor]

10/10 would pat

[zinc flax]

Same

[gusty palm]

intuatively tileddiffusion splits the image into tiles for rendering to let u upscale to larger values but im not sure if thats what it does

[crystal meadow]

This is quite helpful. Thankyou

[brittle rampart]

@supple raptor @untold glacier btw you know extrntion for restart all SD server without having to constantly interrupt the execution of the batch file and run it again manually. maybe an extension that adds the simplest button on the interface for this?

[gusty palm]

Time taken: 4m 49.92sTorch active/reserved: 2715/3226 MiB, Sys VRAM: 4966/6144 MiB (80.83%)

[supple raptor]

Yes that's basically right, if you use the region prompt control part of it, it instead splits it into zones that you define in the interface.

[gusty palm]

i dont mind the time as long as it works

does it do nothing if i dont use the region control

[supple raptor]

As long as you don't need new dependencies to be installed, you can go to the extensions tab and just click the Apply and Restart UI button.

[supple raptor]

Without region control, it splits the image into tiles like you thought.

[gusty palm]

oh ok

oh ic what u mean now

[brittle rampart]

OMG! it just be all time here!

[supple raptor]

Yep, if you install or update an extension and get any weird behavior or errors afterwards, you'll want to re-run the batch file as it likely means there are dependencies it needs to install. But most updates and many new extensions don't have any new dependencies.

Sitting on pillows so she looks taller 😅

Time for me to do some studying and then get some sleep. Night all 👋

[untold glacier]

G'night Nubbs!

[zinc flax]

Have a nice rest

[gusty palm]

hoonigth

[untold glacier]

And with that I'm out for the night. G'night peeps.

[gusty palm]

oooo

[real nexus]

Made a genshin style Lora

Not too sure if I should release it

[zinc flax]

Looks clean

[gusty palm]

im so glad i tried genshin

[gusty palm]

not sure what i broke but the settings>defaults view/apply changes doesnt work anymore

[real nexus]

I might release it idk

[glacial bane]

Ima need the data

Pwease

[lyric zealot]

there you go, metadata

[glacial bane]

Thank yow

Very much

[steady grail]

today i will test regional prompt, some theme to suggest? or i let google suggest it for me?

[glacial bane]

What kind of loras/embeddings do you use? Your eyes are perfect how do you do it do you use your own mix?

[lyric zealot]

most of the time, I'll use ADetailer for fixing eyes

[glacial bane]

ADetailer a negative embedding?

[lyric zealot]

it's an extension

[glacial bane]

wow i really love your work

I better experiment then byee

[lyric zealot]

appreciate it, but the extension is actually just one click to activate

[steady grail]

earing you talk about eyes made me think about the opening dna^2

i found my argument

the title of the song is BLURRY EYES

[glacial bane]

Song? You makin a song?

[steady grail]

a song of when i was a child

for now i'm still not able to compose well

[glacial bane]

Darn welp hope you can make it

[steady grail]

one day, when i master the piano a little more

i can't wait too

[glacial bane]

Im here struggling to experiment cuz im on a laptop that uses my cpu

Cuz it has no gpu

Takes more time haha

The agony

[steady grail]

eheheh

[glacial bane]

Yet i can make good photos

[steady grail]

i'm pushing my hadware too, i'm on a 4g i'm pretty slow too

[glacial bane]

Fr

I wanna get a good pc for SD

Yet inflation and money problems

[steady grail]

they are pretty pricy

[glacial bane]

Exactly

Now ill be off to experiment bye

[lyric zealot]

have fun and good luck

[glacial bane]

Hope i can get good generated pics like yours

[steady grail]

i made another horrible mutant lol

[gritty sundial]

Does anyone know a good realistic model that looks very realistic. I'm needing to do a job

[steady grail]

not a fan of photorealistic

let me se what i have...

https://civitai.com/models/4834/visiongen-realism-reborn captured my interest for some reason

https://civitai.com/models/4201?modelVersionId=114367

https://civitai.com/models/4384?modelVersionId=109123

i know about these but they are pretty old, maybe now there something better out there?

[gritty sundial]

Do you know why it's blurry like that?

[steady grail]

low res?

[gritty sundial]

yes

Is it because of VAE?

[steady grail]

the vae usually influence the colors and a little the details

but in this case, i think the prompt is too small, not enough steps and the low resolution

try to use this photo as a base on controlnet as canny and set highres fix at 2

[gritty sundial]

I don't know how to use controlnet

[steady grail]

but honestly, i'm not an expert with realism, i sold my soul to chibi moe >.<

[supple raptor]

LoRA weights are too high, CFG is at the upper limit of where images tend to look decent, the negative prompt has way too much emphasis on the first part, and lastly the resolution is too low.

[steady grail]

cfg is good between 6 and 9, right?

[supple raptor]

Yeah, that's usually a good range. 9 can be too high on some models/loras, but usually that's a sign that they are overtrained.

[steady grail]

it's me, or working with realism gives the creeps?

[supple raptor]

Genning actual people, like celebrities or public figures, gives me the creeps. I don't mind realism otherwise.

[gritty sundial]

aaaaaa, why does it come out so bad

[steady grail]

when is realism, something hits me inside. like shaping something fake

i suggest to download some negative TI, they helps very much gens

[supple raptor]

(Asian, clone faces, duplicates, tiling, NFSW, shirtless, deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, painting, anime:1.5) this part of your negative prompt is likely hurting your images a lot more than it's helping.

• clone faces isn't something the AI would have been trained. it's a symptom of the AI attempting to create symmetry.
• duplicates duplicates of what?
• tiling sd doesn't try to make tiled images usually
• deformed iris likely understood as deformed and iris separately. same for pupils
• semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, painting, anime in generally you should simply use a model and your possitive prompt to get the style that you want. adding all of this to the negative is not very helpful.
• all of this being at 1.5 emphasis is either evidence that it isn't working or it's part of the reason for poor quality images.

[steady grail]

more than 1.3 usually give weird results

[gritty sundial]

now it's better

[supple raptor]

lastly, try enabling Hires Fix. If you have enough vram, 2x upscaling would be very helpful. Select one of the ESRGAN upscalers with 0.51 denoising. If you're short on VRAM, try 1.5x upscaling with the same settings.

[steady grail]

once i put (drinking water:2) and ended with generating a person under a waterfall lol

looks like i added too much water lol

[supple raptor]

waterboarding your waifus

[steady grail]

lol

that was an epic fail XD

[supple raptor]

I've done similar on many occasions. Loras at 8 strength instead of 0.8 is always a fun surprise 🤣

[gusty palm]

https://cdn.discordapp.com/attachments/548263195135967234/1131996871125115083/00366.png

in love w this t bbh

[steady grail]

doing stuff in kitchen?

[gusty palm]

nu but randomizing settings in stylepile

[steady grail]

how works style? it's empty to me

[lyric zealot]

you meant this?

[steady grail]

yep

i usually set styles in prompt

[lyric zealot]

that's basically for saving prompts

[steady grail]

so, the only way to randomize styles is a wildcard

[lyric zealot]

like you can type a basic prompt that you usually use for your generation, then press save style

randomise, probably

[steady grail]

i should find a wildcard with many artist names

[lyric zealot]

you can create your own wildcard as well

[gusty palm]

i downloaded a bunch of wildcards ><

im not creative

[lyric zealot]

same, but i seldom use wildcards now

maybe more dynamic prompting

[gusty palm]

i didnt quite understand that is that where u {option1|option2}

[lyric zealot]

yup yup, that's what i meant

[pliant forge]

Imagine/ dog

[lyric zealot]

if you are interested in using SDXL bot, go to #1100170312106127410

[bitter latch]

hmmm....

this comfyUI isn't easy

hmm something isn't right

[cyan burrow]

ComfyUI is pretty intuitive and when working on a singular image, you can achieve amazing results

[bitter latch]

yeah if i figure it out. need to watch some vids I think

[supple raptor]

ComfyUI could really benefit in the new user experience category by including a few well made workflows in the repo. with a few notes nodes to explain what some of the nodes are doing and why they are connected the way they are. If you don't have a pretty good idea of the process that SD uses then building a workflow from scratch is way beyond the average user's capabilities.

[bitter latch]

nods

[supple raptor]

@bitter latch do you normally use Hiresfix in Auto1111?

[bitter latch]

no

used to

[supple raptor]

Well, this might come in handy anyways. Here's my Hires Fix workflow for Comfy. You'll need to select your own models. Comfy handles upscaling a bit differently than Auto. If you use a 4x upscaler, this will give 2x upscaling and uses tiled vae for the Hires pass to save vram. It will also save both high and low res versions of the image.

[bitter latch]

still figureing out

tho the nodes are making more sense

just a different thought process

[bitter latch]

first successful image

[gusty palm]

im gonna see how big i can upscale before something freezes or ooms

[gusty palm]

i dont think it ooms it just takes a long time lol

[gusty palm]

realistic miku

[bitter latch]

thta one was done with just regular a1111

[gusty palm]

oo ur doin comfy

[bitter latch]

yeah tried it out, back to a1111

😄

more used to it

[supple raptor]

I'd much rather use a1111 as well. Comfy is neat, but inconvenient.

[nova remnant]

Hi, nice picture.

[bitter latch]

thanks

[nova remnant]

Hello, the picture is really great, I really like it.

[lyric zealot]

thank you, really appreciate it. Here's the same pic with metadata.

[bitter latch]

even darker

[nova remnant]

That's a really mega anime character, I really like it.

[bitter latch]

thanks

[gusty palm]

i love flat colors and thos big lines

reminds me of like those cartoons on mtv or something i think, drawn together or ugly americans

[bitter latch]

I have the style Modern European ink painting in there and it sorta changes it

hmm sometimes the huen model does ok then it does that

oi

hmm lets take out hte ink style and switch back to DPM__ 2mSDE

hmm some of these models just don't do good anime

[gusty palm]

im all over the place with sampler, i sortof switch randomly and sometimes its nice and sharp and sometimes its not

[bitter latch]

yeah

[gusty palm]

the ac wont shutuppppp

[bitter latch]

that means it's working

[gusty palm]

true i shouldnt complain

[bitter latch]

ac's are like kids if they go silent.. worry

[gusty palm]

its 95F here ><

[gloomy scroll]

https://github.com/Tsukreya/Umi-AI-debloat

[gusty palm]

the more words in a sampling method name the better the sampler

[bitter latch]

shrugs

[gusty palm]

i ask gpt occasionally and it just gaslights me into thinking theyre different

[bitter latch]

they are

different methods

[untold glacier]

ᴡʜᴇɴ ᴀʀᴇ ᴍʏ ʜᴀɴᴅs ɴᴏᴛ ɢᴏɪɴɢ ᴛᴏ ʙᴇ ɴᴏᴏᴅʟᴇs

[swift lava]

Are any of the major anime models capable of doing ink wash painting? Like in the style of Josetsu?

[untold glacier]

For something like that you'd probably want a non-anime model? But lemme look for a LoRA.

[bitter latch]

just put in in style of Josetsu , ink wash or hydro ink wash

shrugs

[swift lava]

Not worth looking for a LoRA. I have just been experimenting with a prompt and putting it in different styles. Couldn't find a keyword that would work for this one, so I thought I would ask.

[untold glacier]

For styles that aren't represented well in a model, I have to use LoRAs or creative prompting, perhaps with controlnet on a good source.
Using composable diffusion when prompting can also help get you there. It's how I achieved this using Aurora:

[swift lava]

Nice. Sounds like that is another tool I need to learn. That ink drawing style can be finicky to get. At least I have been able to get decent watercolor and pencil drawings.

[untold glacier]

For when you wish to learn: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#composable-diffusion

[bitter latch]

if you are using auto1111 look for the extension style pile

hmm let's see what this does, I added hydro ink washed in style of Josetsu to the prompt 😄

[gusty palm]

been using style pile all day its so fun

[bitter latch]

hmm might have to tweak it

[swift lava]

This is closer than what I was getting. It seems to have broken Regional Prompter, but that is ok.
I thought it was another extension rather than built in.

[untold glacier]

Nice!

[mighty saffron]

does anyone who has played around with the API know how to make the anime style have more range? we've been experimenting with it in this app and it works ok, but all the art style converges a bit generically: https://apps.apple.com/us/app/cereal-anime-ai/id6450517877

[untold glacier]

I don't use the API. You may have more luck with such questions in the more technically minded channels.

[mighty saffron]

thanks!

[bitter latch]

got rest of day off so i'll be here 😄

ponders different subject, think i've run this one in to the ground

[swift lava]

Well, not exactly what I was going for, but I have got some interesting results. Likewise, I think it is time to try something different.

[bitter latch]

miku's from diff dimenions?

[untold glacier]

I've found my issue. Don't be like me and forget a foundational rule, people. Always prompt for the model. Don't expect cohesion from natural language prompting with booru-based models that work best with clip skip 2.

[bitter latch]

oh she got angry

[gusty palm]

those are cool

[gusty palm]

can u elaborate a little ive been changing clip skip from 1/2 randomly but dont really know what changes to expect

[bitter latch]

"I didn't need that hair anyway!"

[untold glacier]

[bitter latch]

without syle pile

[supple raptor]

CLIP processing is handled in 11 layers. There's technical reasons for that, which I don't know off the top of my head. The relevant part, is that the final, or penultimate layer can be very chaotic. NovelAI and EluetherAI found that by removing, or skipping, the final layer results became much more coherent. Fewer instances of extra limbs being added, less color bleeding, and things like that.

TLDR: Clip Skip 2 will typically make your results more coherent on most models, but may make generations a little more samey on some of them.

[untold glacier]

So am I completely wrong in what I said?

[gusty palm]

ah ok

[supple raptor]

You're right that basically every Anime model should be used with CLIP skip 2, but many (probably most) realistic models benefit from it as well if coherency is what you're after.

[untold glacier]

Ah, coherency vs following the prompt?

[steady grail]

so, what's the point of the last 2 steps?

to encrease creativity?

[supple raptor]

The thing is that the penultimate layer is often not very good at following the prompt and will just start doing random shit. Sometimes that random shit can be interesting, often times it just makes things look worse.

[untold glacier]

That's strange, as it's not been my experience much of the time. Here's a recent study I did on clip skip 1 & 2, on a variety of popular models.

Warning, the image is very large (though it's already been scaled down by half from the original 😅 )

Perhaps I'm not viewing it objectively enough. Or not interpreting the evidence correctly.

[supple raptor]

CLIP is a text (or more accurately "token") processing model. It exists outside of Stable Diffusion. In some cases the behavior of the penultimate layer can be desirable. Also, the term "CLIP Skip 2" is a misnomer, what it is actually doing is stopping at layer 2, so CLIP Skip 2 is actually only skipping 1 layer.

[steady grail]

oh

[bitter latch]

well that got choatic

chaotic?

[supple raptor]

It's also worth noting that clip skip is only a thing in SD 1.x based models. 2.x based models use OpenCLIP which is a different architecture and it ignores the clip skip setting. SDXL uses both CLIP and OpenCLIP together and I'm not sure if clip skip will have any effect on it.

[steady grail]

in that comparison clip skip 1 does better than clip skip 2

in most models

[untold glacier]

I dunno. I'm second guessing what I'm actually seeing. There are elements that are better yes, but is that just from added noise (chaos) to the gens? It's hard to tell. @supple raptor, if you get a chance to view the Clipskip comparison I uploaded, let me know your thoughts when you can. I'd love to hear them.

[steady grail]

well, i see less people fused with matter and less extra limbs in 1 skip, unique exception is dark alpha that looks to do better with skip 2

but 1 thing is certain, this comparison is crazy and eye opener

[untold glacier]

With the spacesuits, a lot of them are fusing themselves to the cockpit in CS one rather than two. Driving scenes are also often more chaotic and less focused of a presentation. The same with the gothic lolita matrix sword fights. -- ᴛʜᴇʀᴇ's ᴀ sᴇɴᴛᴇɴᴄᴇ ɪ ɴᴇᴠᴇʀ ᴛʜᴏᴜɢʜᴛ ɪ'ᴅ sᴀʏ

[bitter latch]

did some layered effects 😄

[cyan burrow]

Heh found some old images that I made on MJ back in last September, the last one was remade in SD with img2img etc

well I have a lot more but...not really the place for flooding MJ pics

for awhile though, MJ went crazy obsessive with rendering n00ds

[bitter latch]

origional, 1, 2

[runic vault]

oh this one is perfect

[steady grail]

nighty night

[cyan burrow]

ah finding old renders from january/feburary that I made with DG3, or an early version of it

and this was from Dark Gemini v1

[bitter latch]

Did some layered processing effects

😄

[brittle rampart]

WHY!? 🤣

[bitter latch]

just to see what I could do

shrugs

[brittle rampart]

huh.... oki-doki

[bitter latch]

ok. bbl

[gusty palm]

ooo incomprehensible horrors

[drowsy dove]

Even SD thinks I've generated enough 1girl with robotic arms

[wet elm]

whats the best like neutral style anime model?

[runic vault]

like this? that would be anything 4.5/based66 (image is based66)

[wet elm]

yeah but i kinda fell they are outdated rn

[runic vault]

outdated to what exactly? sdxl is brand new it's going to take a while for any pure anime focused models based on it to come out.

[cyan burrow]

🎶 Look at this PHOTOGRAPH🎶

[wet elm]

idk, when i use those i tend to get some bad results when comparing to the newer models, maybe its just me tho

[supple raptor]

https://cdn.discordapp.com/emojis/1097119300386562138.gif?size=64&name=AyameDespair&quality=lossless

[drowsy dove]

SDXL seems to have more potential than 1.5 (2.0 and 2.1 didn't run when I tried them), at least I was able to get results closer to what I expect: pictures at night are darker, glowing objects do illuminate the environment around them, a general better understanding of the prompts, but without the 1.0 being public, it's too early

[cyan burrow]

That would make it like offset noise loras

[cyan burrow]

Nickelback don't want you to talk to the hand, he wants you to LOOK AT THIS PHOTOGRAPH

[drowsy dove]

If the base models starts like this, then LoRAs would improve it further (hopefully)

[cyan burrow]

it'll take awhile before it catches up

[supple raptor]

Yes, they fixed the problem that required us to use offset_noise on previous SD versions.

[stable crane]

anyone tried waifu diffusion XL?

[bitter latch]

The miku you do not mess with

[slow needle]

yes and its kinda meh compared to sdxl 1.0

[bitter latch]

hhe

haha

[bitter latch]

hmm interesting

Cat Girl Roman Miku

😄

[bitter latch]

there we go. anime