🍥｜anime

[swift lava]

Asking here because I saw some discussion of it yesterday, but when using Multidiffusion upscaler, is it possible to combine it with X/Y/Z plot?

I want to use the plot to automate changing some parts of the prompt, but it doesn't seem to work for that.

[bitter latch]

yeets himself out a window

on the first floor.. what i'm not crazy 😄

much....

[untold glacier]

I haven't tried that, let me see.

[bitter latch]

goes to play games and waves

[untold glacier]

Yep. Works without issue.

[swift lava]

Are you able to use prompt S/R to change things in the tiles? That is what I can't figure out. Like, I want to do a batch substituting out elements of the prompt in every region simultaneously while keeping the rest the same.

It doesn't work by default because I tried it, but I am wondering if there is a setting I can adjust.

[untold glacier]

Yep. That's what I tested. Works flawlessly.

Wait.

I misread what you asked.

[swift lava]

Oh, how did you do that? It is at least similar to what I am attempting.

[untold glacier]

That's just your standard prompt s/r.
Now as for region prompt control, I don't believe you can change things within multidiffusion using prompt s/r, as it would need to be listed as an option within the x/y/z plot. It only looks at the main prompt.

Now, it does look like Adetailer has that to a certain extent. So we know that that particular extension has support for it--but it'd have to be within the list, and I don't see that for multidiffusion.

[swift lava]

Ah, that is unfortunate. I might need to try regional prompter instead because it seems to work entirely from the main prompt. Or this might be the impetus to learn python coding and make an extension to the extension.

I will also take a look at Adetailer.

I don't think many people use prompt S/R heavily, so it makes sense that there is not that much support for it.

[untold glacier]

Do know that Adetailer is specifically for inpainting faces, hands, and people. So the models that it uses will automatically inpaint those, depending on what you choose.

[swift lava]

I am guessing that you also cannot read region prompts from a file? That is the other way that I tend to do large batches.

[untold glacier]

I'd use it more myself, if it had more options.

[untold glacier]

I haven't seen the option to do that, no.

You might have better luck reading the Multidiffusion documentation, as there may be things that I'm missing in that regard that aren't well known.

[swift lava]

I guess that is my call to be the change I want to see in the world. I tend to use it for changing large parts of the prompt to get pseudo timelapse effects.

[swift lava]

Read through that and didn't find anything, and a google search also didn't turn up too much.

[cyan burrow]

is adetailer something requiring command prompts/script?

or is that just a log?

[fickle pasture]

made with revanimated model

[bitter latch]

cat cafe sign?

and thatrs a totally diff type of business

[supple raptor]

Sounds to me like you should just be using wildcards unless there's a specific reason that you want to have them all arranged in a grid to compare.

[supple raptor]

adetailer is an extension. It installs just like any other normal extension. A full SD restart may be required after adding the extension for the first time to allow it to download the detection models that it uses. https://github.com/Bing-su/adetailer.git

[cyan burrow]

Yeah but even some extensions require you to write a commandlist and whatnot to make it do stuff

I guess adetailer is mostly a menu and prompts and values?

[swift lava]

What are wildcards? Is that another term for prompt matrix or something that I am unaware of?

Generally the grid is important, though, or at least the ability to input a series of specific prompts from file.

[stable crane]

https://github.com/AUTOMATIC1111/stable-diffusion-webui-wildcards

[supple raptor]

For just inpainting faces, all you really need to do is turn it on and it sill just do it's thing with default settings, Leaving the prompt blank means it will just use your main prompt, same goes for the negative. Here's the full interface which is integrated into the txt2img and img2img tabs.

[cyan burrow]

separate steps and noise multiplier etc is what gives it a highly beneficial use right?

[supple raptor]

wildcards is a term that refers to a list of things that can be dynamicallyand/or randomly inserted into your prompt. There are a few methods to use them. UmiAI makes it easy. I use this fork of UmiAI which has extra wildcard lists and has all of the embeddings that the main fork includes removed from it. https://github.com/Tsukreya/Umi-AI-debloat.git
There are more detailed instructions in the repo that you can check out, but the simplest way to use them in your prompt is to just type __FilenameOfTheWildcardList__ in your prompt and it will choose a random thing from that list for every generation. So, if you want a bunch of random generations, you would just increase your batch size and/or batch count to get the number of images that you want and each one would have a randomized set of wildcard things.

[supple raptor]

No, what makes it highly beneficial is that it automatically inpaints the detected area at full resolution which means far greater level of detail. It's particularly good for faces and especially when it isn't a close up portrait.

[swift lava]

Thanks. I am not sure if this will do what I am looking for, but it will probably be helpful in other ways.

[cyan burrow]

How does it accomplish this better than others? Cause we can select the resolution of inpaints now

[quiet bramble]

Looking for this style of LORA

[supple raptor]

It happens during the txt2img generation process as an added step, similar to how Hires fix works.

[cyan burrow]

but faster and more precision?

[supple raptor]

I mean, ideally you use both. So when you click generate on txt2img the process becomes:

• low res generation
• upscale
• img2img at new resolution
• automatic face detection
• inpaint at full res on detected face

[cyan burrow]

I thought multidiffusion was better for upscaling

[supple raptor]

For actual upscaling, sure. That doesn't mean you don't want to use Hires Fix and adetailer on your images anyways so that you have a good image to feed to multidiffusion.

Multidiffusion is also completely unnecessary for most images and is a way slower workflow than simply letting Hires Fix and ADetailer work their magic first.

[cyan burrow]

I suppose, any suggestion for upscale model?

[supple raptor]

Depends on the image. fatal_Anime, AnimeSharp, and DigiPaint are the ones I use the most.

[pale monolith]

try Aetherion style, cogecha, motomurabito, and/or pyramid ghibli , all on civitai

nothing "exactly" like that, but flat colors and hard outlines are in the ones i mentioned

the pose dynamic, fisheye, and stuff will all be prompt work

[cyan burrow]

Cheers mate

[pale monolith]

hugepaint8x if you want to keep glow effects, gradients, and semi-realism/ realism, otherwise fatal_anime, reksanime, lolipop, or yandere300k , reks will apply almost like a multiply to the color tint, the other three are just pretty good at upscaling without much change

[cyan burrow]

thank you, is it possible to run multiple upscalers at the same time on different unet lanes?

[supple raptor]

Definitely not with pixel upscalers. unet works in latents.

[cyan burrow]

mmm

It just came to mind how great it would be to have a different upscaler for style, and another for concept

[stable crane]

man

the realesrgan anime upscaler can be quite good

[cyan burrow]

Cursed blonde character...CURSED!

[unreal ridge]

does easynegative change the art style?

[untold glacier]

Slightly, yes.

[unreal ridge]

I'm tweaking with my neg tags

haven't changed this shit in months

[untold glacier]

Always exciting to change formulas.

[untold glacier]

But only very slightly. I've found it to be very clean of a negative embed. I use it or Auroranegative when I want to use one.

[unreal ridge]

I was using both

[untold glacier]

Uh.... both auroranegative *and * easynegative?

[unreal ridge]

yes

[untold glacier]

Yeah, they work best as one *or * the other.

[pale monolith]

i just use auroraneg and khfb

[untold glacier]

Yeah, khfb is designed to work alongside auroranegative, whereas both easynegative and auroranegative are set to do the same type of things, but differently.

I'd imagine you could also use khfb alongside easynegative just as easily.

negative_hand can be used with anything, really. Such a light presence that it has, with a very specific focus.

[supple raptor]

@unreal ridge

All of those have negative_hand included as well.

[quiet bramble]

Thank you very much!

[unreal ridge]

removed a few

[pale monolith]

[untold glacier]

New WLOP LoRA is fire.

But that's not how you read a book.

[untold glacier]

Why do they all have skirt wedgies?

[south sable]

https://tenor.com/view/judge-gif-9081447

[untold glacier]

https://tenor.com/view/86anime-frederica-eighty-six-derica-squint-gif-25166025

[cyan burrow]

This is its voice and how it introduces itself

[cyan burrow]

or this one

or this

[cyan burrow]

which do you think fits the character of his best?

[untold glacier]

That wood shine tho...

[silver tree]

how was this done, if i am allowed to ask?

[cyan burrow]

he felt very targetted by this. I mean imagine if this was Jack the Ripper and he sounded like this and spoke about these things

[untold glacier]

That's just a gif from an actual anime, no SD involved,

[silver tree]

hmm, okay. i was quite impressed

[untold glacier]

Hey, I'd be too if it were SD. Lol.

[silver tree]

still new to this so i am wondering the animation potential here

[untold glacier]

We're not quite there yet, but I've seen significant SD animation innovations on the horizon in regards to temporal stability.

[silver tree]

definitely seems possible in the future, if not possible now with extremely time consuming processes.

i wonder really how to recreate the same character with convincing poses, relative to the same scene

[silver tree]

i can imagine there’d be techniques of rendering the background separately, using advanced inpainting techniques to somehow create proper enough keyframes, and then somehow getting cohesive animation out of that that doesn’t look super deep-dreamy. i was really impressed by that gif haha

[zinc flax]

Huge

[zinc flax]

Also sideboob + pits

Nice

[stable crane]

for me it's more about the shoulder

dude I gotta just update my SD so I can try the hands lycoris

[stable crane]

so cute

[zinc flax]

Very thic

[supple raptor]

New version was just released so you might as well.

[stable crane]

damn cool

latent coupling? luck?

[supple raptor]

Just trying out one of this dude's LoRAs. https://civitai.com/user/konyconi/models

[stable crane]

he's got so many

I'll have to find & try those negs too

[brittle jay]

animation potential at least from stable diffusion in my opinion is very very small

[untold glacier]

But what are they talking about?

[brittle jay]

would be easier to use different technology

very hard to solve temporal something something forgot the name

[wispy canopy]

sorry, i'm not fluent in stablediffusion-ish, but probably how you should invest into waifus

this would make for a fun ingame menu like that xD

[untold glacier]

[silver tree]

my anime girls are ugly and weird and semi 3d n shi

you guys all have dat cute aesthetic with the glossy haair

[brittle jay]

use different models

[wispy canopy]

yeah, use different model 🙂

[silver tree]

which models are recommended for cute anime girls

[pale monolith]

or loras

[untold glacier]

Part of that is the model, the other part (since many of the popular models lean toward that 3d look) is prompting against it.

[wispy canopy]

my model is pretty easy to use, but it's an acquired taste, i suppose

[silver tree]

this my example of weirdly 3d vs 2d n cute

only difference here being seed

[supple raptor]

What model are you using?

[silver tree]

only protogen and waifu diffusion

not together

[pale monolith]

try aurora or flat animerge

or darksushimix

this is flat animerge

[wispy canopy]

depending on prompt, you can steer the output too: this is all the same seed, difference being 2d/3d/realistic/flat shading

[supple raptor]

Waifu Diffusion is sadly still pretty bad and protogen as a series is all over the place. I'd second the Aurora or darksushi recommendation. Haven't used flat animemerge personally.

[silver tree]

i need to delete some fo these jawns fr

ima delete waifu, it didnt do much for me trying to make old schoo, style anime shit

[wispy canopy]

protogen, like most non-anime models work very well when used with [anime lineart:0.2]

[pale monolith]

hmmm i should try that on my model

you sure 0.2? that doesn't feel like it does anything xD

[wispy canopy]

are you using a generic model?

anime models generally do very little, if anything with those tags

[pale monolith]

nah my buttons and thread merge

[wispy canopy]

you could go for 0.1

[supple raptor]

It's adding it 20% of the way through the gen. It should have plenty of an effect.

[wispy canopy]

earlier makes the anime influence too stronk

[supple raptor]

I'm surprised buttons & thread worked this well with all of these LoRAs.

[pale monolith]

:3

tbh i think i should just release it on civ at this point

[supple raptor]

It seems good to me. I don't know what I would want to change about it.

[pure oyster]

oops, set my lora to 10.7 instead of 0.7, that explains a lot.

[pale monolith]

i guess it is a bit noticeable

[silver tree]

im not making thesse images for nefarious purposes. simply to master the software

[pale monolith]

i'd actually want to merge in the lighting lora i have on like 99% of the time but i dont want to deal with kohya stuff xD

[wispy canopy]

porn gets boring fast, art is eternal!

although, i do miss being able to express artful nude in this channel, but i suppose that's only fair

[silver tree]

so the anime lineart tag is not useful on aurora model?

[supple raptor]

Just use SuperMerger.

[supple raptor]

Not needed because Aurora is anime style by default.

[wispy canopy]

yeah, not needed

same as with darkalfa, doesn't need it, although lineart does change it a lot

[supple raptor]

Ctd released a couple new lighting LoRAs tonight BTW. I haven't tried them yet but her gens with them were 🔥 as usual.

[silver tree]

still a little bit too hd i suppose with the variety of shadows and shining light rays whatnot

[pale monolith]

ghostlight? cause i got that one xD

[untold glacier]

I do wish there was another way to influence DarkAlfa into pure 2D anime style other than "flat color"--it seems like it tends more towards western animation stylistically in general, and with flat color it goes further in that direction.

[supple raptor]

Now you need a VAE https://huggingface.co/nubby/kl-f8-anime2-blessed

[wispy canopy]

i suppose that's the generic model shining through

[supple raptor]

https://huggingface.co/closertodeath/ctdlora/blob/main/lightingv1-000005.safetensors
and https://huggingface.co/closertodeath/ctdlora/blob/main/lightingv2-000010.safetensors

[pure oyster]

oh, a anime VAE with less contrast? Yes please

[pale monolith]

night time lighting is real nice

[supple raptor]

Oh that is nice!

[untold glacier]

Nubbssssss all the embed previews lol

[supple raptor]

I'm on mobile. No easy way to remove them.

[untold glacier]

Discord needs to get on the ball

[supple raptor]

Complain to Discord cause my ass is in bed

[untold glacier]

You don't type with your ass do you?

[silver tree]

okay, now we're talking. good shit, nubby

[supple raptor]

Yep VAE is very important. That's what turns numbers into pixels so you can have an actual image.

[silver tree]

yeah i dont know what any of that means but hey

[pale monolith]

important but also not much variation, lora's is the fun part

[untold glacier]

Darksushi looks interesting.

[pale monolith]

it makes pretty anime

[untold glacier]

Which one do you use?

[pale monolith]

this is my buttons and thread base style

[supple raptor]

I like colorful personally, but they're all good.

[pale monolith]

colorful as well

[untold glacier]

From the preview images, it looks like a more stable version of the Niji ripoffs.

[pale monolith]

that one image is probably all ive done with it, mostly got it as a candidate for more merging

i really like wwcolor mix

[silver tree]

im fecklessly downloading

[untold glacier]

Is this booru-language or natural based? Probably booru, right?

[pure oyster]

Good luck remembering trigger words.

[pale monolith]

idk how it "should" be prompted i just prompt via booru tags xD

[silver tree]

my computer has been fucked to its last gig

[untold glacier]

Gotta keep notes

[untold glacier]

Gotcha.

[pale monolith]

despair

[untold glacier]

And my LoRAs only take up less than 1/4th of the space of Eppinette-Chi's

[supple raptor]

You know most LoRAs have the tags list in the metadata right? You can view them by hovering on the preview tile and click the icon in the top right if it appears. Trigger words will be the most common tags used.

[silver tree]

im rookie still

[wispy canopy]

just dedicate a disk to it 😄

[pure oyster]

I'd say that I usually don't see that when I look, but then again my sample size is only like 4 lol.

[pale monolith]

[silver tree]

my other ssd still has my old os on it, imna fry this cookie

does adding more sampling steps do shit? i see people making hyper realistic cyborg shit with 20 steps still

i suppose, would doing 30 steps make a difference for anime style

[pale monolith]

kinda, there's a point where an image isn't formed so there's a minimum

but after a point only small variations take place

[wispy canopy]

adding steps usually increases image coherency, up to a maximum point

some models like higher steps than others

[pale monolith]

the actual step count depends on the sampler used

[wispy canopy]

cfg, model, sampler, are all factors into stepcount

i generally go for 35, but that's just a personal preference

[pale monolith]

i stick with 27 cause im impatient :3

[silver tree]

was using 27 earlier and reverted back to 20

hmmm

[supple raptor]

I don't remember the last time I used anything other than 20 steps except for some upscaling.

[silver tree]

that was in my quest to get more simple anime girls which ive now solved

[untold glacier]

I've looked at a few that have tags, and so far it's less than 50% that have that info in the metadata sadly. That and many of the LoRAs require additional steps, optimal weights, or models that they're best used with. It's not something that I could begin to use without my notetaking system.

[supple raptor]

I use all of the DPM++ samplers.

[pale monolith]

i find 20-26 steps quality varies a lot between steps for Euler A but 27+ is when it stop really changing much

[untold glacier]

For Euler A, you definitely get more detail at higher steps, such as 40.

[supple raptor]

If you want more notes you should create a .txt file for each one with the same filename as the LoRA. Place it in the same folder as the LoRA and it will appear at the bottom of the preview tile.

[untold glacier]

After 50 though, your returns aren't worth it.

[wispy canopy]

I only ever use UniPC

[untold glacier]

How do you even upscale??

[pale monolith]

not detailed?!

[untold glacier]

I can't use DDIM in a million years for that stuff.

[untold glacier]

I didn't say it wasn't detailed, just that there is more detail at the higher steps.

[wispy canopy]

set 2nd sampler to dpm++ 2m

[untold glacier]

Wait, what second sampler?

Where do you select that?

[silver tree]

the sampler on the img2img page?

[pure oyster]

2048 detail in a 640 sized image. Pretty good

[wispy canopy]

ah, it's a vladmandic setting

unipc will revert to ddim in 2nd pass

[supple raptor]

In settings you can enable setting a different sampler and prompt/neg for Hires fix.

[supple raptor]

It's an Auto setting. Vlad just defaulted it to on.

[untold glacier]

There's not a lot of detail in these areas--it's rather smooth. Now, that's one of the benefits of Euler A and it helps to give focus where it needs it--that was one of the reasons that Euler A was great for so long.

[wispy canopy]

oh right... this one

[untold glacier]

Something like DPM++ 2M Karras will have a lot more light detail and slight noise or art markings all around, but that's often something that you may not want depending on what you're trying to gen.

[untold glacier]

Yeah, I don't have that option.

[wispy canopy]

git pull 🙂

[pale monolith]

never! I'm already 4 months out of date

[untold glacier]

You think I mess with that stuff? I use WINDOWS. I download and plop it into place when I wanna upgrade.

[silver tree]

upscale seemed to have worked i guess

[supple raptor]

Unless you haven't updated in 3+ months, you do. You just need to turn it on

[untold glacier]

This is why I wish there was an easier way to know the upgrades. I've even kept up with most of them manually and didn't know about this one.

Thanks for the heads up.

[supple raptor]

There's been a changelog for like 6 months now

[pale monolith]

yup 3+ months out of date not there for me

not that i like ever use upscale besides SD upscale script verrryyy rarely

[supple raptor]

Update your shit

[pale monolith]

when you point out a feature i use that is better i will !

[supple raptor]

Inpainting has built in zoom as of tonight

[untold glacier]

AAAAAAAAAAaaaaaaaaagh
I've been reading the commits MANUALLY

https://tenor.com/view/anime-head-bang-wall-yikes-gif-15778741

[pure oyster]

I never recognize anything you guys ever screenshot of your webapp lol.

[pale monolith]

as of tonight as in its totally broken something else and wait a couple days for it to settle not tricking me with that

[pure oyster]

10 forks and 200 mods deep

[untold glacier]

You're gonna love mine then, it's a whole different UI

[supple raptor]

Should be good. I was using the RC that became the stable release for a month. I only reverted a couple of days ago to troubleshoot a problem that was unrelated.

[pure oyster]

Borders between the tools are a lot more clear, so that is nice.

[supple raptor]

Oh, you have Vlad, that's why you don't update because he's incapable of not breaking it literally every day 🤣

[pale monolith]

ehhhh idkkk im pretty precise with my tablet on default i haven't really needed to zoom on anything tho xD

[untold glacier]

Nope, never. I have standard 1111. I just have CozyNest as the UI extension.

[pale monolith]

.> basic b for mee

[untold glacier]

It's a classic

[supple raptor]

Besides that, there have been noticeable startup speed increases in 1.3 and now 1.4

[pure oyster]

look how ugly mine is

[untold glacier]

First impressions are ɴ ɪ ᴄ ᴇ for DarkSushi

[pale monolith]

does this count lora refreshing speed?

cause that.... that is tragic

webui boots faster than a lora refresh for me now xD

[supple raptor]

Unfortunately no, I have the same problem.

[pale monolith]

[silver tree]

does seed hold its effects between models

[wispy canopy]

only minor

[pale monolith]

somewhat

[silver tree]

this is aurora and darksushi same seed

[pale monolith]

its a way to kinda see the difference between models and and pinpoint what a single change does

[silver tree]

seems like darksushi overall adds more background detail

[untold glacier]

Traditional hand problems, but good imagination

[pale monolith]

aurora has an incredible clip so it listens to what you say a lot, but this also means if you don't specify stuff in the background it ends up being kinda empty a lot

[supple raptor]

You can add mksks style, detailed background on Aurora to influence it towards more background details. Those are activation tokens for one of the models that was merged into it.

[silver tree]

hmm, ill try to add the background stuff. i strayed away from trying that cause how shit it is by default

[pure oyster]

any hand looks good when you hide 3 fingers

[wispy canopy]

it will be a while before we can get real ranni's as a 1-shot generation

[pale monolith]

i messed up on the prompt to get 6 arms xD

this version 2 is realllyy close

but the ghost face is still elusive

[wispy canopy]

yeah, the ghost face being the hard part

😦

[pale monolith]

i get hints sometimes

[pale monolith]

closest one i've gotten so far

[wispy canopy]

yeah, the hint is there

[pure oyster]

That baby face

[wispy canopy]

still, i coulnd't keep playing elden ring

[silver tree]

never thought i’d have these powers to generate such waifus

[wispy canopy]

post your best generations here!

[pale monolith]

everyone always expects the blade to be the part stuck in the stone smh

[silver tree]

you guys have only just barely opened my eyes to 1% of the possibilities here

[supple raptor]

I just didn't feel the need to any more after beating it 8 times.

[wispy canopy]

i'd also be exhausted if i whack my blade that way into a stone

[untold glacier]

Here we go.

[wispy canopy]

don't worry about posting stuff, i have no idea what i'm doing either!

[wispy canopy]

did she bend the blade on her biceps?!

[supple raptor]

All 8 were full completions too. No cave, catacomb, quest, or world boss was safe from me.

[wispy canopy]

as ever, i have no idea how i got to this, but me likey!

[untold glacier]

I feel like I'm getting close to DA:I art territory.

[wispy canopy]

can't unsee the 3rd hand 😄

[gloomy lodge]

your luck it aint a 3 leg

[wispy canopy]

😄

[silver tree]

some of yall gotta be high

[silver tree]

im tlakin bout whoever made this

[silver tree]

like what do u even type do achieve this lmao

[silver tree]

that style is sick Eface

[pale monolith]

143279853915981432980378-170

[silver tree]

you guys get crazy detailed clothing, you look like fully upgraded characters. necklaces, rings, jacket patches, headgear, hella shit. is that all specified, or general randomness achieved via some setting?

[wispy canopy]

random

[silver tree]

you look like you solo horntail in .02 seconds

[untold glacier]

Nearly 4 here. I gotta sleep. G'night peeps.

[wispy canopy]

unless specified, the model just "does it's thing"

[pale monolith]

i should bed too is like 2

[wispy canopy]

dude. xD

[silver tree]

hmm i see

[wispy canopy]

dude. XD

gnight america ^^

[silver tree]

does higher denoising on upscale give the thingamajig a better chance to produce better quality hands and such?

or lower denoising if its already decent on the source img

[pale monolith]

denoise determines how close it stays to source image

[silver tree]

thats why im wondering, if the hands are shit will it produce high quality alien hands

or if u wanna up the denoising to reduce that

[pale monolith]

0.1-0.3 is minor detail shift, 0.4-0.5 is moderate, 0.6 and above starts to completely shift image

[silver tree]

okay i see what u mean

[wispy canopy]

try to keep denoise as low as possible

[silver tree]

id prob want it at .3 for this

[wispy canopy]

some samplers/upscale methods can get away with lower denoise

[silver tree]

still fixes hands and such but doesnt change pose and allat

[wispy canopy]

latent upscale, with 2x increase, need ~0.58 denoise

looks at @supple raptor weren't you from US too?

[supple raptor]

If you're talking about Hires fix, then not usually. The reason for using that is that models struggle with high resolutions so we set the composition of the image at a low resolution then upscale and img2img so it doesn't have to do too much at the high res.

[silver tree]

yup, ive lost my waifu to the denoising. ggs

[pale monolith]

im on my first satyrday tho

[wispy canopy]

satyrday?

[supple raptor]

[silver tree]

ive not used hires fix yet

[pale monolith]

[silver tree]

wait, this higher denoising gave me, like, a montage of waifus lol

[wispy canopy]

the original SD is trained in 512x512 pixel images

[silver tree]

this shit is hilarious honestly

[wispy canopy]

so when you go higher than that resolution, you end up with duplicates and the like

[supple raptor]

Oh, you're using one of the tiled options, yeah don't use high denoising with those.

[wispy canopy]

so, the highres fix first makes a lowres image, and then upscales it using SD again

however, with high denoise, it tends to become... creative

so you want to keep it as low as you can

[silver tree]

should i not be using a script at the bottom of the img2img panel?

[supple raptor]

I go sleep now. Night all.

[silver tree]

ive got sd upscale selected at the bottom but ive got literally no idea if any of it is doing anything

[supple raptor]

SD Upscale is good, but you want to keep denoising at 0.25 or lower.

[silver tree]

i see

gg nubby

u are goated

[supple raptor]

Use that when you want more detail on an image that's already pretty good.

[pale monolith]

wheat

[wispy canopy]

pretty nice detail 😮

[wispy canopy]

old gen i had, on wheat 😄

same model lol 😄

[pale monolith]

back in the olden times before custom models

[wispy canopy]

sd1.5?

or shudder 1.3?

[pale monolith]

i think 1.3 tbh

am i realllllly tired or is there like an optical illusion think with this one?

when is full screen i feel like the highway blur where it stretches

[wooden coral]

Breakdomain models are gone

[atomic stratus]

Who did you got beautiful like this ?

[pale monolith]

[unreal ridge]

rip chat 🪦

[south sable]

Good Morning

[south sable]

tfw you and your friend are about to be sucked into the digital realm (possibly for all eternity).

[grim portal]

this one sparks happiness

[zinc flax]

i got jumpscared by my own gen

[unreal ridge]

Not inside the glass, but still decent

[bitter latch]

same prompt same seed, but standalone , non merge model

[untold glacier]

If we still used CRT monitors, I'd make a remark about how she is technically still behind the glass...

[bitter latch]

in the jar to long?

hmm 2 jars or one cut in half?

[bitter latch]

非常によくやりました

文法が悪い場合は申し訳ありません。翻訳者を使用しています

😄 apologies if bad grammer, using translator

[untold glacier]

I wish I could find out a specific tag or keyword that does this. I love the look where it's sort of painted in from white space.

[bitter latch]

challenge accepted

[bitter latch]

almost , white canvas, multicolor pastel paints, cloudy art style, anime angel warrior girl, art by makoto shinkai
Negative prompt: an6, kkw-extreme-neg
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 102022262, Size: 960x640, Model hash: c0986c2136, Model: ElegantRPGAnimeMagick, Clip skip: 2, Version: v1.3.2

[supple raptor]

Just view it on your phone and you can still make the joke.

[untold glacier]

ᴛʀᴜᴇ

[supple raptor]

unless you have a foldable phone in which case, why would you spend so much on a phone?

[unreal ridge]

He forgot about me 😔

There's no grammar errors on this one. Nice translator, whichever you're using.

[bitter latch]

using google

[unreal ridge]

Idk the specific tags either

it just does that sometimes to me

[untold glacier]

Same.

[bitter latch]

could be the model too

[unreal ridge]

weird one but still nice

[supple raptor]

Unfortunately there isn't a specific tag for it afaik. Sometimes you can force it with something like [white background::5]

[unreal ridge]

I'm using ink watercolor art as tag

give it a try

[untold glacier]

Sweet. That sounds like a great starting point.

[supple raptor]

you can also try watercolor \(medium\) which is the booru tag for watercolor

I like the effect you're already getting though

[unreal ridge]

What's up with the \(something\) in some tags?

[unreal ridge]

My model can't do it very well without increasing it to 1.2~5. Try increasing it if your model struggles with it

[supple raptor]

booru uses parenthesis in some tags, but WebUI uses parenthesis to increase emphasis on the word so you have to escape the parenthesis with a backslash before each so that it will interpret them as text.

[bitter latch]

switches from the merged to antoher one

[unreal ridge]

ooh

Nubby the 🐐

[supple raptor]

https://cdn.discordapp.com/emojis/1093686592071282739.gif?size=48&name=ASdiscodoge&quality=lossless

[untold glacier]

Nice. I think this was the key to the right direction. Thanks.

[untold glacier]

That and possibly the prompt editing from Nubby.

[unreal ridge]

Is there any list of art styles that work without loras and shit? Like
ink watercolor art
ukiyoe
art nouveau?

[zinc flax]

retro style anime screencap pixel art flat color

[untold glacier]

Yes and no. Yes there are lists, but I haven't seen any that have assured stylistic effects. Just total lists of art styles and the like. There are limited studies that I've found on the web, but nothing with more than a few styles and examples.

[unreal ridge]

I'll add these to my list

ty ❤️

[unreal ridge]

[untold glacier]

[unreal ridge]

art nouveau my beloved

[unreal ridge]

is it on the booru tags thingy?

[supple raptor]

some booru tags that work to varying degrees are painting \(medium\), marker \(medium\), watercolor \(medium\), graphite \(medium\), colored pencil \(medium\), millipen \(medium\), nib pen \(medium\), acrylic paint \(medium\)

[bitter latch]

now lets see if clyde allows it....

[unreal ridge]

I doubt clyde is still up at this point

I've seen way worse being posted here

[bitter latch]

we broke him?

[untold glacier]

I can't remember specific sites at the moment, but there are wildcards. And as I said earlier, it's not assured that they'll affect things--especially in anime models--in the way that you expect.

[unreal ridge]

I need to figure out how wildcards work

[untold glacier]

Here's a link to where you can download a gigantic list of wildcards (each of which is a list). Do know that there are tags that are not SFW in the least in there. https://www.sdcompendium.com/doku.php?id=wilcards:wildcards_0108

[unreal ridge]

I downloaded two zip files of wildcards alrd

I'm pretty sure you sent one of them

[untold glacier]

Very likely.

[unreal ridge]

her ears 😭

[supple raptor]

I like this extension for wildcards. It comes pre-loaded with some wildcard lists that cover most things that you might want. I used to have the giant list of them from sdcompendium and at that point there's just too many lists to find new things to try. Using them is super easy, you just use __wildcardlistfilename__ in your prompt to get a single random item from the selected list. More in depth uses are explained in the readme for the extension. https://github.com/Tsukreya/Umi-AI-debloat

It also integrates with the sd-webui-tagcomplete extension so it can autofill wildcards for you so you don't have to try to remember the filenames of them.

[unreal ridge]

👍

[untold glacier]

Wildcards are extremely useful. It's what makes prompts like random D&D character generators possible. I myself use the extension Dynamic Prompts to use wildcards. I love the toolset that it is.

[silver tree]

is this relative prompt for multiple character or more shìt

[south sable]

this is the power of "Everyone" as a tag and then describing everyone as best you can. Although, there are such things as Latent Couple that could also get you more characters.

[untold glacier]

How abouts we keep things nice and chill here, eh?

🧊 🧊

[south sable]

https://tenor.com/view/chil-squad-thums-up-agree-okay-ok-gif-15155162

[unreal ridge]

almost sent something extremely sus by accident

thank god image preview exists

[south sable]

https://tenor.com/view/montana-jones-alfred-melissa-sone-thorn-anime-gif-18073822

[untold glacier]

Niiiiiiiiice

[unreal ridge]

This one looks nice

[silver tree]

my bad didnt realize curse words were unaccepted

[untold glacier]

Attitude, bud. We like a light environment here and that sounded like a direct attack.

[zinc flax]

Stylish

[silver tree]

i was just asking if he was using relative prompt or more complicated of setup. sorry if ti seemed like something more

definitely wouldn't come in here and attack anyone lol

[supple raptor]

That's how I read it. You're good. Just a misunderstanding.

[silver tree]

well i will counteract the negative energy and emphasize that you are all goated and have helped me immensely with ideas and actual technical aspects 🙂

[untold glacier]

[supple raptor]

https://cdn.discordapp.com/emojis/1107715323009966092.webp?size=48&name=saffronheart&quality=lossless

[unreal ridge]

I sometimes forget that cow and pow exists

[unreal ridge]

Holy shit, I need to start doing this

I thought you had nitro for a sec lmao

ayo

[supple raptor]

owo

[unreal ridge]

No one saw the misspell, noice

[supple raptor]

https://cdn.discordapp.com/emojis/1069076345260343318.webp?size=48&name=anyasmug&quality=lossless

[supple raptor]

Oh I did

[unreal ridge]

https://tenor.com/view/memes-meme-dead-death-bad-ending-gif-26049237

[supple raptor]

https://media.discordapp.net/stickers/967094889332498454.png?size=160&name=AYO

[unreal ridge]

what do y'all get if you prompt nothing? Just the neg tags

[supple raptor]

Most of us will get 1girl cause she's overfitted in almost all of the anime models XD

[south sable]

https://tenor.com/view/anime-love-after-world-domination-coffee-tea-shaking-gif-25869781

[unreal ridge]

I tend to get 🐯 40% of the times for some reason

[supple raptor]

this model is adorable

[supple raptor]

no prompt on toonyou-JP

[south sable]

ill try it!

[unreal ridge]

I'm getting some cool stuff for some reason

[south sable]

weird lol

no prompt animepunk

[stone onyx]

does this have meta data? 👁️

[unreal ridge]

No. I remove metadata cuz I find the idea of people seeing my prompts without me knowing uncomfortable for some reason. Also, I sometimes put some sus prompts and I don't want to be judged

[unreal ridge]

It was posted by @wispy canopy a few days ago

[supple raptor]

watdafuq how this model so damn cute?

[unreal ridge]

I'll try to find it

[bitter latch]

https://giphy.com/gifs/simpsons-homer-bushes-3ohs7KViF6rA4aan5u

[unreal ridge]

This

[south sable]

no prompt again

[unreal ridge]

Why is no prompts giving me better stuff than when I try prompting 😭

[south sable]

https://tenor.com/view/hmm-think-i-see-stressed-yuri-gif-13973217

what have we been doing... lol

[drowsy dove]

Do you use nsfw on the negatives on the images without a positive prompt?
Asking because I don't think I can send what I'm getting with Anything v5

[wispy canopy]

oh that one ^^

[unreal ridge]

This one is perfection

Wasn't expecting it to work well with my model

[wispy canopy]

even with the spelling error 😄

[unreal ridge]

yes

[unreal ridge]

Fucked hands but looks cool

[wispy canopy]

lets see if i can improve upon it 😄

[south sable]

just the negs (mistoon)

[silver tree]

damn so what you’re saying is no prompt is better than prompts i be workin on all night

[south sable]

https://tenor.com/view/nodding-crossed-arms-pikachu-eyes-closed-pokemon-gif-15194868

[silver tree]

the images speak for themselves

fugg it

[unreal ridge]

bro 😭

I'm switching teams. Fuck prompts.

[drowsy dove]

The power of negativity is overwhelming

[silver tree]

she got a wooden musket

[bitter latch]

no don't do that.. that makes nsfw content

😄

[wispy canopy]

dudes, fuck prompts indeed -> throw in song lyrics again!

[silver tree]

that model looks nice with no prompt for sure

[south sable]

toot ghost

[unreal ridge]

I can recognize this prompt just because of the ghost lol

[wispy canopy]

lightened up on the negative prompt, looks quite better actually ^^

[unreal ridge]

@wispy canopy your prompt is spreading

[south sable]

#justhenegs

[unreal ridge]

oo

this looks nice

[south sable]

darksushimixmix

[drowsy dove]

I got a not 1girl

[unreal ridge]

I sometimes just get a room

😔

[wispy canopy]

well... darkalfa IS an ex-hentai model ... so there's that...

no pos, no neg, this model sucks LOL 😄

[bitter latch]

with the one girl?

😄

[unreal ridge]

nope

just a room

[bitter latch]

https://giphy.com/gifs/travisband-travis-fran-healy-jhaFMx06pRuE9M0qOP

[south sable]

https://tenor.com/view/hachimitsu-shikimoris-not-just-a-cutie-shikimoris-not-just-a-cutie-characters-shikimoris-not-just-a-cutie-anime-shikimoris-not-just-cute-gif-26000313

[south sable]

#justhenegs

lol

[unreal ridge]

id:browse nopromptsupremacy

[south sable]

its slowly going insane

[unreal ridge]

it failed

Don't wanna fix it

[bitter latch]

gives it a try, no prompt no neg or pos , haha

CAT

[unreal ridge]

at least it looks like a cat

[south sable]

its not unlike a cat

[bitter latch]

no eyes

[south sable]

looks like a failed taxidermy

[bitter latch]

we need to build a wall - trump

[unreal ridge]

weed wump

[south sable]

#justhenegs

[wispy canopy]

#nonegatives

oh wait. there are

FUUU

stupid toggle

[south sable]

https://tenor.com/view/the-goon-dodgeball-movie-ben-stiller-white-goodman-gif-14923367

[wispy canopy]

that's why it didn't want to do the thing i wanted

[bitter latch]

umm I can't post the one I did with just negs.. nope

[silver tree]

is that explosion one negatives only too?

such random cool stuff

[bitter latch]

put clothes in regular prompt area, well she has clothes but the buttons burst (facepalms)

put in nsfw for neg

😄

[wispy canopy]

i can't live without a lot of anti-nsfw in my neg: #justhornymodelthings

[unreal ridge]

what about masterpiece on prompts?

[bitter latch]

I think I broke it

[south sable]

https://tenor.com/view/serial-experiments-lain-lain-stare-into-the-abyss-y2k-2000s-gif-24846261

[wispy canopy]

https://tenor.com/view/kratos-boy-kratos-boy-boi-god-of-war-gif-12278930

yeah. these are cool ❤️

lets add masterpiece, and see what happens

[south sable]

"Positive Prompt" in the pos and "Negative Prompt" in the neg

lol

[wispy canopy]

masterpiece editions ... eh

[unreal ridge]

Can't see the difference

[wispy canopy]

less creative use of colors, bit more washed out, the hair has less blending into background

masterpiece restricts it a bit in this regard

it does increase image coherency though, but well.. me and image coherency

LETS PUT MASTERPIECE IN THE NEGATIVE!

[south sable]

https://tenor.com/view/explosion-anime-gif-24705382

[wispy canopy]

seeing this, i should remind you to watch trinity and beyond if you never did

[south sable]

best quality/bad anatomy only

masterpiece/bad anatomy only

[unreal ridge]

got some weird images when I searched by images

[untold glacier]

The machines have already replaced you

[wispy canopy]

i suppose i just killed masterpiece in my model 😄

[south sable]

https://tenor.com/view/morpheus-free-your-mind-freethinking-enlightenment-the-matrix-gif-14970364

[wispy canopy]

https://www.imdb.com/title/tt0114728/?ref_=nv_sr_srsg_0_tt_8_nm_0_q_trinity%2520and%2520 is movie about nukes

[unreal ridge]

Haven't watched a movie in like

4 years

[wispy canopy]

this one very worth it imho

should be mandatory viewing for anyone

so they know that they should shoot anyone who even considers nukes weapons to actually use

[bitter latch]

now to change to the pure anime model

ok maybe not pure

[bitter latch]

mistoonanime

[bitter latch]

welll

[supple raptor]

Nah, that's one of my old prompts. Just trying some random stuff on this model I found.

[bitter latch]

BINGO- Giant goddess girl silhouette in the rainbow nebulas strands merging into universe, cinematic lighting, reflections, concept elegance, 16k, hi-res, RTX, cute and adorable, fluffy abstract , wet ink watercolor art, universe merging, ethereal holographic paint style, clouds, cloud merge, art by makoto shinkai
Negative prompt: an6, kkw-extreme-neg
Steps: 20, Sampler: UniPC, CFG scale: 6, Seed: 2140876664, Size: 960x640, Model hash: c7751e6108, Model: aZovyaRPGArtistTools_v3, Clip skip: 2, Version: v1.4.0

Used embeddings: an6 [40f4], kkw-Extreme-Neg [5dd3]

[bitter latch]

changed sampler

[wispy canopy]

aight, imma go to bed ^^ sleep well everyone

[bitter latch]

😄 cuts out

interesing mix

[bitter latch]

and back around to this...

ponders a cat nap

sets it back to 680 x 960 sets up batch of 8 and takes a nap

[bitter latch]

sadly time to go listen to ppl btch abt orders....

[silver tree]

no prompt loves this white cat

[untold glacier]

This model is pretty nice, if a very specific style.

[silver tree]

the white background look with cascading watercolor stuff really is a cool style

what does controlnet even do? if i put an image here does it do anything?

[pure oyster]

oh my, alot of discord servers just dropped

[untold glacier]

We're back!

[south sable]

https://tenor.com/view/confused-anime-gif-25086783

me thinking i had a stroke

[untold glacier]

When the icon straight disappeared I thought I was somehow banned for a sec, heh.

[pure oyster]

Same, my stuff is only mildy lewd

[pure oyster]

control net is one of those extensions than need models that don't come with it pre-installed, they can all be downloaded here https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

all the details on what each control net model does is here https://stable-diffusion-art.com/controlnet/#Preprocessors_and_models

[silver tree]

clicks profusely

[unreal ridge]

I'll save this one

haven't touched control net yet

[untold glacier]

Yeah I guess now is as good a time as any to get controlnet 1.1 installed.

[silver tree]

im trying to turn a picture into anime

[untold glacier]

Major FFVII vibes.

[silver tree]

presumably control net is what i need although i just cant tell if im affecting anything

i have a preprocessor selected and image preview showing the little canny outline but no model selected so probably nothing happening

[pure oyster]

yep, you need to download the model type you need to use then put them into "stable-diffusion-webui\extensions\sd-webui-controlnet\models"

[silver tree]

just did it

im hucking files into folders like no other

not sure if canny is what id want for this, the previews seem to be fairly similar

[pure oyster]

canny would probably work

[silver tree]

alright it seems to have made a significant difference, this controlnet shiz allows me to turn up denoising while maintaining the scene and not having flying anime mermaids appearing

[pure oyster]

Tried Darksushi for the first time using my normal workflow, I like it.

[unreal ridge]

darksushi is nice

one of the first models I used when I got into sd

[south sable]

i have mix mix and reg darksushi. both very fun

[pure oyster]

Here is the unedited text2img.

[pale monolith]

which model?

[untold glacier]

https://civitai.com/models/92834?modelVersionId=51593

[pale monolith]

oh i have the toonyou extract lora!

[untold glacier]

Nubby was testing it out and it looked good to me, so I decided to try it too.

Yeah, it's likely better suited to a LoRA--maybe used with DarkAlfa, but it'll be harder to get a flatness that the model gives. Without coloring it with keywords in a slightly different direction.

[pale monolith]

my lora results don't look nearly as cute as that style tho

[south sable]

https://tenor.com/view/error-wait-computer-download-cat-gif-10480536

[untold glacier]

Lol

[south sable]

lol def checking it out now

[untold glacier]

https://tenor.com/view/spock-computer-gif-24940930

[untold glacier]

Yeah, I haven't seen what all the fuss was about when I tried XL. But, I'm used to the amenities of neg prompting, hires fix, upscaler models, etc. etc.

[pure oyster]

god damn it it's another booru model, I got to learn how to properly prompt for these models.

[untold glacier]

Yeah...

They're not easy, especially when you're used to a natural-language based prompting style.

[pure oyster]

welp time to go onto Safebooru and pick out tags on their website

[untold glacier]

You don't have autocomplete?

[pure oyster]

there is a mod for everything huh?

[untold glacier]

If only my wants didn't outweigh availability, heh.

Just search tagcomplete

Or here's the direct https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

[pure oyster]

yea, I see it in the list on the extensions tab.

[pure oyster]

real men use purses.

[supple raptor]

Yeah, it's a fun one. Hands are kinda messy on it, but I've seen worse.

[pure oyster]

Now that I got autocomplete installed I wonder If I should try aurora once more.

[slow needle]

maybe wait for sdxl

[pure oyster]

I can confirm Aurora is now working

[silver tree]

meinamix is in the gutter, you cant no prompt this thing

[untold glacier]

I've finally found it. A fantastic, sharp, detailed upscaling model midway between anime and real. It's smooth when it needs to be, and sharp with the details.
A new go-to for me when I don't want to use fatal_anime to darken the image further, or ultrasharp that may give slight graininess or patterns. All examples (aside from the first image) don't have Adetailer or Multidiffusion on, so as to ensure same seeds.

HugePaint

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

[pale monolith]

yeah hugepaint8x is my go to usually xD

[untold glacier]

It may have been you who clued me into this, I tried searching the channel history but couldn't find the message I originally saw. If so, then THANK YOU.

[pale monolith]

yeah all the other upscalers want to remove my depth of field and blurry foreground and glow and stuff

[stable crane]

very cute

just recently learned there's a counterfeit 3

[pure oyster]

damn, I don't think img2img can go past 4000x4000

darn, I wanted to make some dank shit

[supple raptor]

Pretty sure you can make it go as high as you want. Just have to modify ui-config.json to allow it. 4000x4000 is going to take a shit load of vram if you aren't using tiled VAE though.

[lyric zealot]

help guys

is this worth it

2.8k sgd

[runic vault]

if you're going for a 40 card go for the 4090

[trim current]

uhhhhh can always go cheaper by building it yourself tbh

that too

4090

[next thorn]

need that vram

[lyric zealot]

trueeeee

[pure oyster]

tried inpainting a small portion of a 7052x4700, SD forced the image back to 4000x2667, can't go above that in extra's tab either.

[untold glacier]

Change this:

It's in the "Saving images/grids" tab.

Sorry, didn't see the "inpainting" part until after I posted that. My above solution may not work.

[pure oyster]

still both worth the try, and the lag!

[supple raptor]

That converts to roughly $2073.40 USD. I just threw together a build in pcpp with the same specs and low but not necessarily the lowest priced parts and it came out to $2034.31 without an OS. Not sure what component prices are like in Singapore, but based on US pricing it looks like a solid deal. https://pcpartpicker.com/list/Q48FFg

[untold glacier]

Usually what I do in such situations is take the image into photoshop, crop to around the section I want inpainted, bring it back into SD, then inpaint and put it back in the original image in photoshop. This saves time, your vram, and your sanity.

[lyric zealot]

well, i'm almost certain if i build this pc with component bought in Sg, it's gonna be almost the same price

[supple raptor]

I would always recommend inpainting at lower resolutions than that. It's better to do your fixes and inpainting after the initial hires fix gen, and upscale after which will allow the upscaling process to help smooth any visible transitions between the inpaint areas.

[lyric zealot]

which is why, this is an actual solid deal

[pure oyster]

to be specific , I'm inpainting a 7952x4700 image at 1200x1200 one little piece at a time.

[supple raptor]

Yeah I would think the prices should be pretty similar after currency conversions. One consideration that might be worth thinking about is that you may be able to get a used 3090/3090ti for less than the 4080 if you were to build it yourself which would get you more VRAM and possibly save money.

[supple raptor]

Have you considered using the Krita extension for your inpainting? It'll give you much better control over both the masking and the area total inpaint area. I never inpaint in WebUI anymore. https://github.com/Interpause/auto-sd-paint-ext.git

[pure oyster]

Eww krita

as a colorist I must not

There is a photoshop one I think, I just have to sail the seven seas and update mine from 2021 to 2023 for it to work

[lyric zealot]

true true, imma have to think a little more

[untold glacier]

One of these things is not like the other

[grim portal]

thats a solid deal at 2k usd

not to scammy

[supple raptor]

Nothing stops you from doing your quick fixes and inpainting in Krita and switching to PS for everything else. Krita is immeasurably better than the WebUI experience for inpainting.

[pure oyster]

I'll try it

[lyric zealot]

totally agreed, just need to think a little more if i should get 4090 or even wait for 5000 series

[grim portal]

u gonna be waiting at least a year

minimum

[lyric zealot]

well, that's gonna be okay, not in a rush tbh

[pure oyster]

I'll upgrade whenever my computer no longer turns on.

[untold glacier]

Shortcut #🍥｜anime message

[silver tree]

i am still rocking a 1080Ti lmfaoo

[grim portal]

[pure oyster]

oh my, possibly the worst video guide ever for this kirita stuff

[pure oyster]

yo, do you know what the error "backend can't be reached" implies?

[supple raptor]

Do you have --api in your COMMANDLINE_ARGS?

[pure oyster]

maybe, do I put a space after "set COMMANDLINE_ARGS="?

[supple raptor]

It shouldn't hurt if you do, but you don't have to. set COMMANDLINE_ARGS=--api will work fine.

if you have multiple args you do want a space between them though. For example: set COMMANDLINE_ARGS=--listen --opt-sdp-no-mem-attention --enable-insecure-extension-access --api

[pure oyster]

Then I definitely have it set, I actually have all the windows running in kirita and it clearly is reading my SD files.

[supple raptor]

I would recommend fully restarting WebUI and Krita.

[pure oyster]

Mine looks like this

Might be a misuse of layers, the video guide wasn't exactly clear

[supple raptor]

Yep that all looks right. I'd recommend using picking the Airbrush brush for drawing your mask so that you get soft edges on the mask so that it will blend into the rest of the image better.

Here's how I have mine set up with an example of a mask drawn with the airbrush

Here's my SD Plugin Config tab

[pure oyster]

I was able to generate an inpainted layer just now

[pure oyster]

lines are a little harsh even with air brush so I'll check more settings

[pure oyster]

will try to add gaussian blur to my mask

[supple raptor]

The easiest fixes are manually blurring the edges even more with the Filter Blur brush or adding gaussian blur, increasing the size of the airbrush (which will also blur the edges more), or lowering denoising (not always practical when you need to change something significantly).

You can also manipulate the layer that the inpaint creates to make it blend better. Eraser Soft with a light run around the edges can often be a quick and easy way to do that.

[pure oyster]

Oh boy, this generation time is possibly double if not more than usual

[supple raptor]

Check your base and max size and make sure they're set to resolutions that you can gen without running out of VRAM. It shouldn't be any slower than inpainting in WebUI.

[pure oyster]

I set base to 512 and max 1024, I'll see how long it takes

[pure oyster]

Harold smile

[zinc flax]

Gotta love paw pose

Try claw pose

[south sable]

[grim portal]

i be generating and generating but still cant get decent hands

[lyric zealot]

(Decent_hands:999)

Well, generating hands is still a problem for most models

[south sable]

https://tenor.com/view/david-byrne-same-as-it-ever-was-always-has-been-same-gif-22266189

finally something

[zinc flax]

There we go

claw pose

[south sable]

hard to get both hands

[grim portal]

i think i got something good! i mean her eyes are a lil funky but what can i do

[supple raptor]

Looks good, eyes are likely either due to not enough resolution or not using an anime VAE. The Adetailer extension can help with the resolution problem without having to increase the size of the gen. For VAE, I'd recommend the one here that ends in 09: https://huggingface.co/nubby/kl-f8-anime2-blessed

[grim portal]

i used hinata hyuuga pictures, they definetly arent of best quality , also threw humans into the mix lol

[supple raptor]

Oh is it a LoRA that you trained?

[grim portal]

yeap

took a good 2 hrs 52 pictures but ima retrain it again apparently had some doubles, i removed some lower quality ones

when i only ask for it to generate hinata hyuuga and nothing else it does pretty good

[supple raptor]

It could help to take a few images, make copies of them and crop them to be close-ups of the face.

[grim portal]

does adding more desciptive text to the txt help ?

close upps of face got it i can do that rn since i brought them with me

im also going to seperate the humans out too

[supple raptor]

The captions in LoRA training are sort of like a negative prompt. You want to caption the things that you don't want to be inherent to the LoRA. For a style LoRA, that generally means basically everything. Most people use the WD tagging tool to automatically caption their images which really helps with consistency. https://github.com/toriato/stable-diffusion-webui-wd14-tagger.git

[zinc flax]

She really said 🐱

[supple raptor]

It made me think of the sad cat dance meme

[grim portal]

is there a point in doing more than 150 sample steps for training ?

[supple raptor]

Depends on learning rate, number of images and batch size. On its own number of sample steps doesn't really mean anything.

[supple raptor]

Wtf 🤣

[grim portal]

constipated 4 eyes demon with super hearing