🍥｜anime

[untold glacier]

Thanks, and that's fine. Now it's my hope that the process is easy enough to glean using the extension for someone who's just now dipping his toe in the water.

[sweet bramble]

have some fun mixing models too, I've been doing that a lot the last couple months

use the block weight merge technique

[untold glacier]

Just what I need, another worthy SD endeavor to lose my time in, lol. Thanks. I'll add it to the list of awesome things I still need to delve into.

[untold glacier]

This is nice. Though the hands give me major creepy factor. 😅

Love the lighting though.

[bitter latch]

yeah the hands go wonky

[untold glacier]

You using a negative embedding?

[bitter latch]

yep

usually does good

hmms

[untold glacier]

Gotta be the prompt or model then. I hate it when I can't break the bad hands.

Also, you can try using Adetailer with the hand dropdown and a slightly higher Adetailer denoise.

I've had some luck with that, but it's SD after all--it's hard to gen good ones with a bad prompt/model.

[pure oyster]

I crudely add and remove fingers and other stuff in photoshop then put it in inpaint to blend it.

[bitter latch]

yawns. gonna crash here shortly

[untold glacier]

Yeah that can work, I've done that on occasion.

[supple raptor]

Yup this is my method. Inpainting a bad hand has hardly ever worked for me without manual edits first.

[bitter latch]

thats better

[bitter latch]

this is addictive

[untold glacier]

ᴊᴜsᴛ ᴏɴᴇ ᴍᴏʀᴇ ɢᴇɴ... ᴊᴜsᴛ ᴏɴᴇ ᴍᴏʀᴇ

[bitter latch]

grabs a anime prompt from a site

[south sable]

https://tenor.com/view/one-more-time-kiss-anime-toradora-gif-24219176

on the trail of the biblically acurate waifu

[bitter latch]

... and 9 months later...

[supple raptor]

Just up the width several notches and you'll be there in no time XD

[bitter latch]

last one then bed

[bitter latch]

there have some anime

thicc theighs...

[untold glacier]

Lies

[bitter latch]

brain slowing down

[pure oyster]

These days I almost exclusively try to find good seeds so I can inpaint an upscaled 3000x3000 image for 6 hours.

[south sable]

https://tenor.com/view/akudama-drive-anime-cyberpunk-hacker-intro-gif-19161157

[untold glacier]

A classic

[bitter latch]

there last one anime furry 😄 uwu

[south sable]

https://tenor.com/view/chips-eating-anime-lunch-break-gif-11881516

[untold glacier]

Major jig-jig street vibes, without the futuristic look. Love it.

[untold glacier]

G'night Kumoa!

[bitter latch]

thuds

https://giphy.com/gifs/cat-funny-tOwkgfESVDua4

[south sable]

look like they are saying, "Bro, how much you even inpaint?"

[pure oyster]

Open in browser and zoom to see my power level.

[untold glacier]

[south sable]

https://tenor.com/view/black-clover-anime-boy-chico-finral-gif-20172253

[untold glacier]

ᴍᴜsᴛ... ɴᴏᴛ... sᴀᴠᴇ... ᴛᴏ ʜᴀʟʟ ᴏғ ғᴀᴍᴇ ғᴏʟᴅᴇʀ...

ᴍᴜsᴛ... ɴᴏᴛ...

https://tenor.com/view/anime-boy-its-so-hot-blush-sweat-gif-17477945

[untold glacier]

Been a while since I've seen this prompt. Love the contrast.

[south sable]

oldie but goodie. always fun to check in on it once in a while

元 素 法 典 The Code of Quintessence still can deliver

I think these are some kind of finger trap chains...

time for another classic from the Code of Quintessence...

[pure oyster]

edit

[untold glacier]

I'd say the ComplexLineart LoRA extraction was relatively successful! Left original, right with LoRA and keyword.

[sweet bramble]

nice!

[south sable]

[untold glacier]

[zinc flax]

i didn't expect this

[cyan burrow]

I gotta say, my time in here since like September of last year has been really positive. I like you guys and no I'm not drunk or high. I'm saying you're a good community

[cyan burrow]

a future form of Sepuku

[magic wave]

So smol :O

[cyan burrow]

Trying to get control net to put puzzle pieces together and get an archer to shoot a victim :p but I think I need layer divider

[karmic sentinel]

hello can anyone help me

i want to install stable difussion

anime version

[cyan burrow]

anime isn't a been so much as a model(trained file) that you load up and it has been trained to make it specialize in anime

I'm not home so can't give you links but there's a webui client (local) called automatic 1111. it's more or less plug and play for the installation part as long as you have an Nvidia card 8gb vram or more otherwise there's a few more steps that I don't know

can't guide you now than that, going to the doctor's

[karmic sentinel]

hey its okay man

everything okay ?

[cyan burrow]

eh.... keeping that to myself for now. I'm sure all regulars know what's been going on though

I'm alright though

[karmic sentinel]

glad to hear that . Take care buddy

[cyan burrow]

take care and good luck

[nova remnant]

This channel is based af

[nova remnant]

one more then I'm done. I love making funny faces with AI lol

[bitter latch]

cant' post nsfw

dude delete it

can't

not here

another server

[nova remnant]

can go to civitai lol

[bitter latch]

oh dude (in a surfer voice) you wanna get banned don't ya

[nova remnant]

damn 23.40 MB malware

[bitter latch]

riiiight

just don't post it here at all they do not allow it

quit being a moron

goes back to eating breakfast

and Jane Jetson

And Judy

[nova remnant]

that's some cool retro anime art

[bitter latch]

and .. leroy.. no

[bitter latch]

haha

[bitter latch]

hmm no pants?

oops

there

[slate night]

@dull jackal Might wanna take a look at this (the messages, not the attachment for obvious reasons lol)

[cyan burrow]

how so?

[nova remnant]

lots of lovely anime art for my eyes to enjoy

[cyan burrow]

You can just post the images here, it's really easy. You drag and drop, or copy and paste

[cyan burrow]

Ah of course

[bitter latch]

here have a bunny. i'm off to work

[cyan burrow]

I actually like that 3d generate I managed to get. The 2.5d and 2d too

[nova remnant]

There's a ball jointed doll style I found on Civit that I really love

[cyan burrow]

and I dunno what this art style is

and a slightly more different. Somewhere between 3d and 2.5d. Kinda like pastel on 3d?

[nova remnant]

ball jointed doll style

[finite herald]

[finite herald]

nico yazawa bjd wen 0:

[cyan burrow]

yaml.scanner.ScannerError: while scanning for the next token found character '\t' that cannot start any token in "D:\AI_Apps\Automatic_1111\models\ControlNet\control_v11p_sd15_lineart.yaml", line 3, column 1 I think my controlnet yamls are broken....or should I just delete them?

cyberpunk cybergoth

[zinc flax]

Very cute

[cyan burrow]

are there any control net models and pre-loaders that aren't on the official download page of HF?

[cyan burrow]

alrighty then, got some new control nets, installing a few things to speed cuda up and trying to resist the temptation to buy a 3090

[wet elm]

may i have the parameters for that?

[summer shuttle]

cant tell whether if it's an explosion or not

[craggy turret]

That’s AI?

It looks so much like the real thing

[craggy turret]

Like you an tell when a “photograph” of a person is AI generated,

But if you just look at that normally with no context of it being AI,

It really looks like the real thing

[remote vine]

what’s the most popular/best model now?

[compact warren]

Look around, test, and find out lol

[cyan burrow]

Mine of course

[compact warren]

Thats such a general question that really cant be applied to this AI stuff

[cyan burrow]

Dark Gemini 3

[magic wave]

I'd say Aurora

BUT

Sis is working on a V2 of Aurora

[cyan burrow]

and soon to be released Arch Nova

[compact warren]

If u ask me a lot of them have same capacity to output similarly styled images to each other

[magic wave]

I mean, its all down to the LoRAs, what LoRAs are the best

[compact warren]

The ones that can blend together without screwing over everything 😭

[magic wave]

Mine

[cyan burrow]

https://civitai.com/models/6209/dark-gemini-v3 this be Dark Gemini, the strength is in its versatility and natural language prompts

[zinc flax]

My unreleased personal merge is the best model fr

[compact warren]

Yeah for a lot of people their best model is actually their own lol

[cyan burrow]

I'm making a new model to fit a more "pseudorealistic" style in a sci-fi concept

Kind of like this, but I'm trying to move it more towards a drawn style

(model is still in alpha or beta, it's a merge. I merge with block weights and use those preview extensions and DAAM to see if I get what I wanted)

[supple raptor]

TBH, trying to make a full model just to get a particular style seems archaic to me at this point. LoRAs are just so much better for that purpose. The important thing for models is that they just have a decent default artstyle, good LoRA compatibility, and listens well to the prompt style it's intended for.

[compact warren]

^

[magic wave]

Yeah, I mean thats why I use Aurora

It reacts well to style change

[cyan burrow]

Most models out there use booru tags and I don't like em 😛

[supple raptor]

Same. I use Aurora 90% of the time, but rarely without LoRAs.

[cyan burrow]

and merging towards a concept model isn't anywhere close to as resource and time heavy as training one is

I also prefer to use my own model, even if it is a merge. Cause I feel I contributed towards what's there

Eventually I'll tire of merging I suppose

and just pick different bases and then make lora's instead

[supple raptor]

Training a LoRA is like a 2 hour task including gathering and tagging the dataset. That's maybe enough time to properly test 2 models that you think you might want to merge together.

[cyan burrow]

I test them much faster than that, but I'm not that thorough. Also my GPU isn't fast enough to train a high quality lora that fast

or well not enough vram

and gathering takes me quite awhile longer, unless I'm after something common

[supple raptor]

Everybody has access to free Colab.

[cyan burrow]

uh you can't use that for SD anymore

[supple raptor]

Google doesn't want you to, but you very much can and many people do.

[cyan burrow]

If you don't care about getting banned...also free colab is a major pain in the ass for me. Since I can't automate more than 90 minutes before it shuts down

so no...not colab

and gpu vram not high enough

so no I can't do it in 2 hours. I can do it in maybe 4

[cyan burrow]

a lighttoy

[supple raptor]

The actual training is insanely fast. Out of that 2 hours I quoted, maybe half an hour at most would be actual training. Most of the time is just gathering the dataset and prepping it. Even on 8GB the training should be at worst double that amount of time.

[wild fern]

what's the best model for anime?

[cyan burrow]

Yeah it's about 45-90 minutes to train. Gathering takes me longer cause I'm picky and I don't have any software to autoscrape

[supple raptor]

I mean, even if it takes a full day, that's still a lot faster than the weeks you spent on the last model https://cdn.discordapp.com/emojis/1017116778784096329.gif?size=48&name=shrugIdk&quality=lossless

[cyan burrow]

I spent weeks cause I was sick with cancer

and didn't really do a whole lot of merging or whatnot

most of the time

[cyan burrow]

So yeah

[supple raptor]

Honestly, most of them are pretty similar at this point. Personally, I use Aurora because it has high compatibility with LoRAs. You can't go too wrong by browsing the popular models on CivitAI and picking one that looks good to you.

[cyan burrow]

If I had like a 3090 I would train a SDXL model though 😛 not merge, train

I would probably also spend more time focusing on LLMs and voice or music AI

[supple raptor]

SDXL isn't even released yet, but I doubt it will be able to be trained on a 3090.

[cyan burrow]

It should be according to Emad

They're working to get its requirements down

[supple raptor]

According to Emad, it will be able to be run on consumer hardware. He didn't say anything about training requirements AFAIK.

[cyan burrow]

Might be right, but we'll see a mass exodus or just people sticking with 1.5 and not giving a damn about the new ones until they can use them

and train on them with consumer hardware

though I'm sure dreambooth and other extensions will find a way to train sdxl

eventually

[pure oyster]

Thought I was being too waifu pilled, so I made a husbando this time. 3000x2000

[zinc flax]

is it husbando time

[untold glacier]

[shy sedge]

what is this model called

[zinc flax]

personal merge unreleased

still doesn't look good to me

[shy sedge]

hm

[wet elm]

what model do you use it on?

[zinc flax]

no clue what he meant by this hand gesture

[supple raptor]

https://cdn.discordapp.com/emojis/1107522740652621874.webp?size=48&name=KikiLaugh&quality=lossless

[pure oyster]

@untold glacier since you liked my OC yesterday, Here is a portrait of her.

zoom in and feel those skin pores

[zinc flax]

don't look at the feet

[quiet bramble]

Any models that are like NijiV5?

[bitter latch]

anime bots

[pure oyster]

At a quick glance https://civitai.com/models/12262/night-sky-yozora-style-model seems to be somewhat close

[quiet bramble]

Thank you!

[bitter latch]

model i'm using https://civitai.com/models/8124/a-zovya-rpg-artist-tools

[cyan burrow]

Did you all know that Mushoku Tensei season 3 starts in a month?

no 2.5 weeks actually

[bitter latch]

ok no more of those, gonna eat 😄

[slate night]

Sorry for the late reply! I use it on my own model, I feel like any "basic" anime model would be good with it... could maybe try pastelmix, darksushimix, maybe aurora and see what happens

Setting it at .8 strength works well for me too

[supple raptor]

I use it on Aurora, but you could also try it on one of Horrid's own models. I think one of these two are what were used to make the previews:
https://civitai.com/models/52346/animepunk
https://civitai.com/models/47792/horridmegamixv1

[slate night]

T

It got my hair (kinda) right on that one lol

[keen hollow]

I need to get some good prompts that render well at high res

I hate how I lose so much detail when I expand to higher resolutions.

Example:

Perhaps a better example.

Lower denoise keeps more of the original design but quality suffers significantly. 😦

[wild fern]

why does it always struggle with the hands

[untold glacier]

Ah, the drawbacks of (relatively) early AI image generation.

Welcome in.

[wild fern]

then it's best to promt it to not generate them xD

actually struggles with faces as well, but after a few generations it gets better and better

[untold glacier]

Well, you can get around it to a certain amount by using good models, prompting well, using negative embeds, and inpainting... as well as just photoshopping and then inpainting. I'm sure Nubby is writing up a detailed explanation as to why it happens though, heh.

[supple raptor]

To give you a real (though somewhat simplified) answer, it's because hands are extremely complicated. Think of all of the different shapes your hands can make as you manipulate your fingers. The AI doesn't have understanding of how hands work, it's been shown many different positions that hands can be in, so it is easy for it to mix up multiple different positions together when it tries to draw them. It can help when you define a specific position that it has likely been well defined durring training for the hands. waving and hands on hips are two that come to mind that have a much better success rate than just letting the AI guess.

[untold glacier]

As for real world usage of my previous statements:
Good anime models: Aurora, DarkAlfa.
Prompting well: Use the language that works best with the model. Aurora often works best with booru tags, whereas models like DarkAlfa work better with natural language.
Negative embeds such as easynegative, or auroranegative help in general, whereas negative_hand can help somewhat with hands in general.
Inpainting: If you're using Auto1111, you can use the Adetailer extension to automatically inpaint hands to a certain extent (it's about 50/50 though.)

Read about the model that you download. Understand what works best with it. Research and learn best prompting methods, as well as negative prompts.

This and more can help you on your road to generating fantastic images with SD.

[wild fern]

yeah i figured out that when i prompt it to generate hands in the pockets or away from sight it manages better

seems like it all boils down to how well you prompt, thanks a lot for the explanation and the tips! I'm having so much fun learning to work with sd

[untold glacier]

Glad you're having fun, and good to see ya here. Always great to see fresh faces in the anime channel.

[cyan burrow]

DarkAlfa, the model that took my pure innocent Dark Gemini 3 and...made her unclean

^^

I gave Eface my blessing to use it though, so it's fine

[untold glacier]

Tbf it's the easiest anime model I've ever used. Super responsive to prompts, and is extremely clean when used correctly.

You had a big part in bringing it to life, heh.

[cyan burrow]

yeah got the natural language prompting from my dark gemini and the conceptual stuff. Then cleaned up the art quality slightly so there was fewer poor renders

the alfa cleaned the art

[untold glacier]

Indeed! Your dark gemini was good, but I shied away from using it much because of the slight dirt/paint/splotches that it would sometimes add into my more... inventive prompts.

[cyan burrow]

I'm continuing the natural language in dark gemini though for Arch Gemini (Dark Gemini 4)

well what dark gemini 3 requires is to know an art style, otherwise it has a higher risk of creating that

[cyan burrow]

so like dark fantasy art style, gothic aesthetic, etc

and maybe 2.5 digital illustration

[untold glacier]

Wow. This must've taken a fair amount of work.

[cyan burrow]

I've got 2 models in the works now, Arch Gemini and Arch Nova

Arch Gemini taking the art further towards 2d and Arch Nova being a from the ground up 2.5d-3d or pseudo/hyperrealism but not realism. More like DAZ3D renders

[wet elm]

np, thank you, i will try aurora because everyone in this chat seems to like it lol

[cyan burrow]

pr sp,etjomg öole tjat

[wet elm]

dayum

[untold glacier]

ᴘssssᴛ... ᴛʜᴀᴛ ᴀɴᴅ ᴅᴀʀᴋᴀʟғᴀ

[wet elm]

is it bad to use a lot of embeddings in the negative prompt? ive been using like 7

[wet elm]

im gonna try it too then

[untold glacier]

I haven't seen that many before, lol. Usually I use one or two at most.

For the best models, they work best with one or two. They have a lot of the crap trained out of them.

[wet elm]

i once saw a good image with a lot of them and i just copied it all

[cyan burrow]

@untold glacier examples from Arch Nova

[untold glacier]

Cool beans. Always a fan of cyberpunk.

[untold glacier]

My go to negative prompt for DarkAlfa is

BREAK
(disfigured, unclear, indistinct:1.3), ugly hands, extra arm, extra hand, split arm, missing finger, extra finger, three fingers, four fingers, six fingers, merged fingers, (bad anatomy:1.3) misplaced hand, misplaced foot, (text:1.3), (signature:1.3), (title:1.3)```

It's an edited version of Eface's.

[cyan burrow]

@untold glacier and this is from current state of Arch Gemini

[untold glacier]

It's also a good starting place for negative prompts with anime in general. Just try to add additional prompt-specific text before the break, to help keep things evened out with CLIP.

[slate night]

Actually got a legible word that I didn't even prompt for

[untold glacier]

Sweet! Looking good.

[cyan burrow]

I'm hoping to make an easy to use version with natural language that also accepts conceptuals and style tags and produces artwork free from artifacts 90% of the time at least. You can never escape artifacts or mistakes

but can minimize them

[wet elm]

thanks!

[wet elm]

i didnt think AI could do words that well LOL

seiba

[wispy canopy]

just got back from company trip (teambuilding /getting drunk with colleagues)

and i see dark alfa? 😮

[cyan burrow]

so if someone wants to say make a prompt like this In the center of the eye of the storm where the wind stays still but devastation brought a village to ruins, a man in a black velvet robe with vertical front stripes in gold sits on a stone, his head shaven with an oroborous tattoo'd on his scalp, anime art style, type seinen, hard light, dim lighting, low contrast, low saturation, dutch angle, medium close-up shot, hdr, 8k and have it create it pretty perfectly. Now I'm gonna have to test it myself to see what it currently does. But still

[untold glacier]

You betcha

[wispy canopy]

😄

[untold glacier]

And welcome back to the land of the living

[cyan burrow]

Eface, yes...the bastard child of the hentai villain that dirtied my beautiful innocent dark gemini 3

[untold glacier]

[cyan burrow]

ok lets see what Arch Nova does with it first

it shits the bed with it

[wispy canopy]

imma update my shit first ^^

have been playing a lot with LLM models though 🙂

those things are evolving crazy fast

[untold glacier]

DarkAlfa, 3rd seed. Mostly gets it. With Adetailer on face and hands.

Oh, and I had to take out close-up.

[cyan burrow]

no I'm afraid it doesn't mostly get it

that's not the eye of the storm, a dutch angle, dim hard light etc

My issue is I don't have an LLM that can write the perfect syntax prompts for "natural language" (generated by an AI)

[untold glacier]

☑️ devastation brought a village to ruins
☑️ a man in a black velvet robe with vertical front stripes in gold sits on a stone
☑️ anime art style, type seinen, hard light, dim lighting,
That's more than half the prompt bud, it mostly gets it.

[cyan burrow]

half is not getting it, I'm a lot more strict on what I pass. My current models didn't fare any better either

[quiet bramble]

Any good negative embeddings I should get?

[untold glacier]

There are some things that SD will never understand without a significant increase in our processing power to access a larger latent space or ways to compress said latent information. Saying "medium" won't affect "close-up". These models are trained far more on woman than men, let alone with bald heads so it'll automatically be harder to send it in that direction. A specific tattoo on a small specific area of a subject that's any less than super close to our viewpoint will be nigh impossible to gen. I don't think we'll see that within a year without a significant increase in processing power or compression. Don't expect the impossible from current gen methods, lol.

[untold glacier]

Easynegative, Auroranegative. Aurora, though made specifically for its namesake's model works great on just about all the anime models I've tried.

[quiet bramble]

Thank you. Would you recommend badhandv4?

[untold glacier]

I've used it. It's not bad but I prefer negative_hand. Lemme see if I can find that.

Nevermind I'll just give it here.

[cyan burrow]

Not expecting the impossible, but I'm never going to lower the quality I strive for

cause that'll affect my performance

[untold glacier]

I respect the ideal, but it is impossible with the current tech. Don't stop aiming for the stars, but don't expect them to get any closer.

[quiet bramble]

Appreciated

[cyan burrow]

Still think the render can be done better, probably by altering the prompt

[quiet bramble]

Do you use these all at once?

[untold glacier]

Only ever Easynegative OR Auroranegative. Sometimes I use negative_hand by itself, as is the default with my DarkAlfa negative prompt, but it can be used with either of the others without issue in most cases.

[wet elm]

is bad_prompt_version2 any good? im trying to lower my amount of embeddings on my negative prompt

[quiet bramble]

That’s good to know. You saved me from smashing them all together.

[untold glacier]

It's been a while since I've tried it. I've moved on to easynegative for non-anime/some anime when I don't use auroranegative.

[untold glacier]

Glad to save your gens! Haha.

[untold glacier]

Ughhhh. My heart.

[wispy canopy]

too sweet?

[untold glacier]

If I want to see too sweet I'll look at Vil's chibi gens, lol. I'm a sucker for romance stories and the like. Great gen.

[wet elm]

i smashed them a lot together lol, now i have to un-smash some

[cyan burrow]

it got one thing right 😛

and the crazy guy. Now I'm gonna slap em together

[bitter latch]

and with that monstrosity of a image.. work!

blah

[wispy canopy]

needs more hands tho

[cyan burrow]

wonder if it can solve this control net

damn slow

[cyan burrow]

it better show something good

and why do I recognize her?

from some anime

[untold glacier]

Very nice. Looks straight out of a VN.

[cyan burrow]

where though

which VN

[untold glacier]

Haven't played enough to know.

[cyan burrow]

good good 7 seconds per it

3 control nets and a prompt :<

[cyan burrow]

yeah not what I was looking for

looks nice

but not what I wanted

[cosmic sleet]

I used 4 Lora's for this.
mushoku tensei, hoyo/genshin art style (i think the two together gave me this art style), Anime screen cap and the detail tweaker thing.

[cosmic sleet]

Also thank you -v-)/

Not sure I did share it here

But this Roxy i got

[cyan burrow]

ROOOOXXXXYYYYYYY

hmmm it came up with this

[nova remnant]

It's a not finished Lora project I'm working on to gather every funny face in anime. I have a dream to create a visual novel using stable diffusion art since I have zero talent as an artist. Stable Diffusion by default can't really do that much with the more comical anime faces so I wanted to make a lora to fix this.

When it's released I'll post it to Civitai

[nova remnant]

https://civitai.com/models/26477?modelVersionId=33642

Using this lora. It's really really good

[wet elm]

nice

[untold glacier]

Reading up on the new EU passed law "AI Act", and LOL. I didn't know they were that concerned Psycho-Pass would be a real thing:

[nova remnant]

EU is forbidding it for a good reason but you can bet the US will be like AWW YEAH MORE WAYS TO ABUSE PEOPLE'S RIGHTS

lol

[untold glacier]

"for a good reason"
That's the excuse people always use to justify their actions. The more the western world limits their ability to innovate, the closer China and other like minded governments get to having control over everything.

[nova remnant]

I just don't have any concerns regarding AI regulation. AI was thrown at us already they can't stop what's out there

[nova remnant]

I don't care if the US government comes out and says "AI art is illegal, punishable by death" I'm not uninstalling my stable diffusion

[untold glacier]

Not everybody will have that choice. If people like us don't vote or communicate in the direction of freedom, the choice will be taken from us too.

[nova remnant]

Well I do get somewhat involved in the politics of AI, not like a political commentator but I am spreading the love and joy of AI on my own community and in various other communities I'm helping with tech support, generating interest and stopping the feer mongering

it's not much but if we all do our part it will add up to a big deal

I waited my whole life for this tech I'm not letting some corpses who shouldn't even still be alive tell me I can't use it

[untold glacier]

I doubt that's enough, but I appreciate your spirit.

It's those propping up the 'corpses' that concern me.

[nova remnant]

there's many hobbyist programmers doing good work too. Neuro-sama is a great example of this. They can regulate it like they canr egulate drugs, it'll stop it form being openly available but it'll be underground instead

[brittle jay]

very nice image, unfortunately jpeg🤮

[brittle jay]

crazy guy hehe

😳

[nova remnant]

I'm also a programmer I have an ebook on designing a neuro network from the ground up as well as 2 udemy courses on applications of data training. Nobody is stopping me

[brittle jay]

except to trainn a model you need a ton of gpu power which costs a lot of money

that is if you are training from scratch

[nova remnant]

depends on what you are training the model to do

[brittle jay]

though no worries, some guys on 4chan or somethinng will always keep saves of everything

so it's fine

[nova remnant]

you can create much smaller neuro nets if you focus it on a specific task

[brittle jay]

even if it becomes totally illegal

which is unlikely to be the case

probably only commercial limitations

[brittle jay]

no

not a diffusion denoising one

that is good enough to be usable or comparable in quality

maybe a GAN

but it would be worse most likely

and less control and stuff

[nova remnant]

only reason I haven't started yet is cuz I hate python

even looking at it makes me want to puke

[brittle jay]

python is very easy

too easy in fact

[nova remnant]

it does too much for you

I like to have more control

[brittle jay]

yes

nah

just let it

less work for you

[nova remnant]

Why isn't this python shit working?
Tries to check Python code
No idea! Cuz the code has a thesaurus full of depedency hierarchies to trace

oh well

guess I'll try shit until it works!

[brittle jay]

ask chatgpt

haha

[nova remnant]

chatgpt can be quite useful yeah

but when it's wrong it still sounds right

lol

[brittle jay]

you could probably also connect it to the docs of all the things you are using through privategpt or something

omg nubby is now named nubby???????

damn

[supple raptor]

to be fair, so do people in various internet forums most of the time which is where people would have been asking before.

[brittle jay]

new pfp as well??

wow

[supple raptor]

the pfp came first. I asked for thumbs up and SD gave me double middle fingers instead. It was too good to pass it up.

[brittle jay]

haha

sd knew what you really wanted deep down

[supple raptor]

I think SD was just giving me the finger XD

[brittle jay]

maybe

since you abused it so much

lol

[brittle jay]

omg lol

this is too funny

[untold glacier]

Hey Kazoo. Ikr?

[brittle jay]

hello

ye

[brittle jay]

lower the cfg perhaps? are you on 12?

[untold glacier]

It's straight outta Psycho-Pass.

[brittle jay]

I should actually watch it

only watched like one episode once then dropped

[untold glacier]

First season is really, really good. Second is pretty good. Movie was meh, and I don't think I made it through the third season.

It's really sad when the quality of writing just drops off.

It had so much promise

[brittle jay]

damn

[brittle jay]

man I have too many anime to watch

and I just finished all the big tests and practically everything at school today

so I do have time

[untold glacier]

USE IT. I miss those days.

[brittle jay]

but I also plan on working my ass off to fly to japan in sept

[untold glacier]

Wow

[brittle jay]

only me and step brother

I need to save up like 5000$+

[brittle jay]

should I watch sub or dub?

[untold glacier]

I believe they're both pretty good, but I'm sub all the way.

[brittle jay]

then sub it is

[untold glacier]

Enjoy it! It's a blast.

[brittle jay]

oh shit it's already 3am

well

I'll go to sleep later

[brittle jay]

until now I never knew that kcal stands for kilocalories

[cyan burrow]

now I just need img2img to clean it up

[brittle jay]

bruh the girl is like the definition of naivete

[untold glacier]

So was I my first day of work

[brittle jay]

oh rly

[untold glacier]

We all start out green man

[brittle jay]

is that how it goes for most people?

like she thought she would find her life's purpose in her work

[untold glacier]

I'd imagine that's what a lot of people think coming out of college.

[brittle jay]

only because she would be the only one to be able to contribute

[brittle jay]

damn

[untold glacier]

[brittle jay]

everyone is an existentialist these days

[untold glacier]

Have you seen the times we're living in? Lol

[brittle jay]

existentialism is like softcore nihilism

people are clinging to their last hope

"philosophical suicide"

anyways

https://media.discordapp.net/attachments/546763235051700314/1112692343808737361/tSoyMtM.gif

[terse ore]

Anyone use the Photoshop SD plugin?

[brittle jay]

isn't there the new generative fill now on photoshop?

[terse ore]

There is but it isn't as good

[cyan burrow]

There are a few, which one?

[terse ore]

Al Faraj one. It was working for me but then A1111 stopped working, so I switched to Vlad. Which did show up in PS but it isn't showing any change to the images when I hit generate

[cyan burrow]

https://www.flyingdog.de/sd/ there's this one that costs money and is on the adobe store, then there is https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin

[terse ore]

Maybe I need to add --api to Vlad. Not sure how because it's a little different than a1111

[cyan burrow]

I dunno, maybe it's in the settings

I prefer auto cause Vlad is just vomiting on new stuff and no wiki to explain them

[terse ore]

Hmm, I'll try to figure it out. I use ComfyUI because it's so much quicker and I don't need new stuff, but I just use auto/Vlad for the Photoshop ability

[cyan burrow]

There's also InvokeAI, I guess

if you want options

[terse ore]

I also got the "free" generative ai and firefly so I'd like to have everything good in one spot.

[keen hollow]

evenin'

[brittle jay]

ok I go sleep

just finished ep 3 of psychopass

ig it's alright

for now

[keen hollow]

sleep? psh. nah

[brittle jay]

it's like 4 and a half AM for me

also did you guys see the new steam interface?

until now it was only on beta

but now it's out of beta

[untold glacier]

I've been on beta so I've seen it for a while, it's nice. Anyway, good night to you!

[keen hollow]

I saw it

[lyric zealot]

thanks to whichever mod that did that

[keen hollow]

did what?

[lyric zealot]

some scammer posted stuff here

[keen hollow]

ah

[bitter latch]

haha

[keen hollow]

trying some new Tifa styles...

[keen hollow]

quiet tonight

[keen hollow]

used the pose 😛

[lyric zealot]

ohhh, using controlnet?

[keen hollow]

yeah

just fartin' around

[arctic stratus]

a girl with white skin, black hair, honey-colored eyes, very attractive and seductive

[untold glacier]

The darn faq command isn't working for me again... sigh

[lyric zealot]

same

[lyric zealot]

you can try using going #1100170312106127410 for bot usage. Most of us here uses local installation of Stable Diffusion

[untold glacier]

Hey bud, the bot channels are further up the list under "SDXL Beta Bot". As for anime image generation though, we're all running it locally (I believe all of us are anyway.) There's also the paid service NovelAI that works great as well. For more info you can check the #1072229020520947753 or #1072220168534642768

[untold glacier]

Wow. You got me. It's been a while since I've been successfully rick rolled.

[polar arch]

😆

[keen hollow]

never gonna give you up 🎵

[quiet bramble]

Any good So-Bin models out there?

[supple raptor]

Cirno has seen some shit

[nova remnant]

CIrno after her 100th revival from being killed by Flandre

[cyan burrow]

Yeah I'd have a messed up face too

[nova remnant]

hahah wtf

[bitter latch]

hehe

[cyan burrow]

she a bit sexy no?

[nova remnant]

the funny part is this describes the anime pretty well too

[cyan burrow]

When your girlfriend wakes up one morning and discovers your interest in anime

[nova remnant]

nice snow waifu

[cyan burrow]

thank you

[magic wave]

weird sandwich

[cyan burrow]

In Shrine City the people live peacefully, are fed, sleep well, stay in good health and have a roof over their head for as long as the people in The Gunk work and slave away with the promise of receiving youth restoration and healing after 40 years of work. Should they survive this. They don't. This is Clara, she lives in Shrine City. She will live to be 260 years old without aging a day. She's thankful to the Gunks for being slaves

[magic wave]

Those are really cool, also I just realized the other day that your PFP is Yareli xD

[magic wave]

O damn those are dope :O

[wooden coral]

Sure is
My most played frame by far

[magic wave]

I don't have her

But she looks fun :O

[wooden coral]

I finished warframe the other week
Like I now have leveled every item there is

Only 3000 hours in

[magic wave]

Oh damn :O

I mean, that sounds pretty fast :o

[wooden coral]

Eh I took my time, i was MR30 in under 1k hours tho

I finished it on my 800th login day, thought that was kinda fitting

[magic wave]

I'm like, 12 I think and I got 500 hours xD

[wooden coral]

Yeah there is a lot to do

I think being a new player nowadays is very overwhelming

[magic wave]

There is yeah xD

[magic wave]

I mean, I got 500 hours, but I'd say I'm still newish, I feel newish at least cause there is so much stuff xD

[wooden coral]

Yeah the amount of content is crazy, especially for a F2P game

But its also its biggest flaw

[magic wave]

Yeee, and thats true too, its really really really big

xD

[sweet bramble]

evening

[magic wave]

Evening :o

[wispy canopy]

imma rerun this batch with adetailer on

oh. not installed. oh well, next batch then 😄

[sweet bramble]

nice sword

[wispy canopy]

Imma cheating a bit tho, as i always do. just keep spamming lyrics in a LLM, making it convert into an image description 😄

Imma run this prompt again tho, this is cool ❤️

[granite fog]

huh didnt know you can get the cards to look that good.

[wispy canopy]

seems my model recalls what a deck of cards should be

well... actually, it seems it failed quite horribly

[granite fog]

well, it's pretty good, these model are bad at understanding text and the like anyways.

[nova remnant]

takes hand

[wispy canopy]

perhaps the term deck of cards actually did make it actual playing cards, instead of "anything related to cards"

[nova remnant]

lol I tried training a lora on the word "cat" written in 50 fonts each font is a separate training image. It couldn't write cat no matter what I did or how many times I changed the training settings and retrained.

[wispy canopy]

there's simply too little parameters in stable diffusion to get it coherent atm, unless you're overtraining a specific model like mad

[nova remnant]

was a fun experiment but damn it didn't even come close once

I think the model probably influences that more than anything

but who wants to add text in training data on a model

that'd be bad

but there is a lora that can write FBI on a hat. He disguised the FBI letters in a really golden textured

[wispy canopy]

🖤

[nova remnant]

ain't bad

[wispy canopy]

i suppose you'd have to manipulate the tagging so it'd doesn't write "fbi" just applies the "image in the lora"

[nova remnant]

yeah teach it that the word "fbi" is a requirement to make the hat complete rather than letters slapped onto the hat

[cosmic sleet]

"arson approved" -her

[gritty patrol]

wow this is beautiful

[unkempt cape]

Is there any model or lora for anime that can make fake anime screenshot

Like something u would see in anime tv

[cyan burrow]

I just realized I won't have to pay for Starfield, Lies of P, Cities Skyline 2. Still got my "kinda bought" gamepass ultimate. I wonder if xbox live gold conversion trick still works? Once my sub is up in Feb 2024. I bought 2.5 years of Game Pass Ultimate for 65 euro or so :P. Thanks to them converting it in a 1:1 ratio after adding it and converting them in a specific order

[wooden coral]

Ill buy it because of some circumstances, on this note, I hate Microsofts player support

[cyan burrow]

player support? You mean support lines?

I never had issues with support, in fact when I had forgotten to cancel a sub during a trial (for F76) and wanted a refund. The support agent said she was unable to do anything about it right now but said she would call me tomorrow and fix it for me then after finding out how to do it. She did just that, called me the next day

I've never had a support agent go out of their way like that to REFUND me

[obsidian willow]

Making politicians

[unkempt cape]

Anyone know a model or lora for making fake anime screenshot

[supple raptor]

Who could have guessed the reddit protests would do nothing against the thing they were mad about? https://www.nbcnews.com/tech/tech-news/reddit-protest-blackout-ceo-steve-huffman-moderators-rcna89544

In other news the trolls with armies of bots at their disposal will have the power to vote out moderators, so that seems productive.

[unkempt cape]

Oh wow its almost like the protest lasted very long and not 2 days

[wispy canopy]

all subreddits should've gone indefinite

[molten shadow]

I always search for SD knowledge on reddit, and now that things happend

[wispy canopy]

oh yeah, just replace mods and reopen -> that's why they're re-opening the subreddits

[molten shadow]

is there any side that provide anything about SD knowledge like that

[supple raptor]

https://www.sdcompendium.com/doku.php?id=start has a fair bit. CivitAI also has a bunch of guides but with how shit their search is, good luck finding them. Much of it is and always has been stored within discord servers as well. Wait a week or two and the subreddits will probably basically all be back.

[autumn violet]

If they don't change anything I will probably try and use reddit less. I always use a third party app on mobile.

[molten shadow]

many thanks, there're a lot of tutorial in youtube, but non of them can tell me what's differ between a lora and textual inversion, I really need a place to just not provide guide

[supple raptor]

It wouldn't have really made a difference. Reddit would have just implemented the policy they're now talking about anyways and the users would vote out the mods because the majority of users don't know or care what an API is, they just want to see their memes/thots/cats, or argue with strangers on the internet. Nothing is going to stop Reddit from wanting to become a profitable company.

[wispy canopy]

i don't want to agree, but that's a realistic take yeah

[supple raptor]

You can think of textual inversion as sort of like a prompt that's been refined by the AI to more closely match the images that it was trained on. It can't add new data to UNET (the image part of the model) it simply finds the parts of TENC/CLIP (the text part of the model) that most closely matches the training data. These "embeddings" then get loaded by WebUI on top of CLIP so that they can be called in the prompt by using the filename as a word.

LoRA is like a miniature model. It includes both UNET and TENC/CLIP components and can therefore add completely new concepts. In the implementation we have, these miniature models are effectively merged with the actual SD model you have loaded when you generate an image.

[unreal ridge]

My man is the whole SD Wikipedia

[molten shadow]

many tks for your details

[supple raptor]

I really can't blame them either. They aren't and never have been a non-profit business. 1 billion in revenue sounds like a lot, but really it's basically nothing at their scale. That doesn't mean I like it, but I do understand that it isn't going away. I hate the Reddit app and paid for RIF so that I wouldn't have to use it. I probably won't visit reddit again except when linked directly to something that happens to be there.

[magic wave]

Nubby is the GOAT

[wispy canopy]

same. although reddit has done some seriously scummy moves this time (although they've been scummy all this time)

[unreal ridge]

I'm starting to miss NAI prompt mixing 😔

[supple raptor]

I haven't used NAI.. well, not their site anyways. How does their prompt mixing work, because we have a few methods to mix prompts.

[supple raptor]

The main issue is just the amount of money they want to charge. It's completely reasonable for them to charge something for their API access, but it shouldn't cost an app like RIF ~$12m/y to function.

[unreal ridge]

I've only heard of wildcards, but I'm pretty sure that it's not the same thing

I have 0 clue of how it works either 😔

[untold glacier]

https://docs.novelai.net/image/promptmixing.html?highlight=mix#advanced-prompt-mixing

[wispy canopy]

agreed, i'm also not adverse to a small subscription either, but the high price and just inane remarks really are sending me into looking for alternatives

[untold glacier]

Just your standard alternate each step.

[unreal ridge]

Oh, I remember seeing this on the doc nubby linked me a while ago

[supple raptor]

@wispy canopy can probably explain wildcards and how to use them better than I can. I don't use them much. Basically they take a list of words or groups of words and allow you to randomly add 1 or more of them to your prompt each batch.

[wispy canopy]

yeah, i can tell about wildcards 🙂

@unreal ridge need a rundown on how to use wildcards in A1111?

[unreal ridge]

🙏

[wispy canopy]

most imporant extensions would be dynamic prompts (which includes wildcard support), and autotag completion

[supple raptor]

@unreal ridge as for the type of prompt mixing that NAI uses (thanks @untold glacier for the link) we have a few options, the simplest is the built in one
Alternating words: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words
Edit, fixed the link

[unreal ridge]

I'll give it a try

[untold glacier]

Here's a mega list of wildcards you can use that people have put together. Be forewarned, there are some nsfw wildcards in there, as I haven't touched what people combined.

Just be sure to put them in the extension's wildcards folder to have access to them.

[wispy canopy]

don't forget the collections folder inside dynamic prompts

[untold glacier]

snip Oh, wait you were talking about something completely different.

[wispy canopy]

especially since i notice that @untold glacier's wildcard zip has files that reference those specific files

[untold glacier]

Yeah. Wildcards can be amazing if used right. It took me months of refinement but I put together a great prompting/wildcard system for generating high quality and varied random D&D character portraits.

[wispy canopy]

but then i grew tired of wildcards, and switched to LLM's

biggest mistake of my life

[supple raptor]

We also have
Prompt Editing (change all or some of the prompt from one thing to another partway through generation. Can also be used to add or remove parts of a prompt partway through generation): https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing
**Prompt Blending ** (similar to the above alternating words, but instead of alternating the prompt every other step of the generation, the prompts are averaged together. This often gives better results than alternating words): https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#prompt-blending
Prompt-fusion-extension (another one similar to alternating words and prompt blending, but even more complex and often gives even better results): https://github.com/ljleb/prompt-fusion-extension

Or you can always do things such as using one prompt for the initial gen with Hires fix and specify a different prompt for Hires fix, or use controlnet, but have it activate later in the generation or turn off partway through.

[wispy canopy]

nah... not THE biggest mistake

biggest SD mistake

[untold glacier]

Not to mention my personal favorite, composable diffusion. And the great extension that goes with it, neutral prompt. More info here: #1087494110358798547 message

[wispy canopy]

in other news, i've gone down yet again with gas usage, prompting an investigation of my energy company, as my pc is now basically my house-heating

[supple raptor]

Oh yeah, I forgot the methods to just straight up assign different prompts to different areas of the generation.
Multidiffusion/tiled diffusion: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git
Regional Prompter: https://github.com/hako-mikan/sd-webui-regional-prompter.git
Two-shot/Latent Couple: https://github.com/ashen-sensored/stable-diffusion-webui-two-shot.git
Composable Lora: https://github.com/opparco/stable-diffusion-webui-composable-lora.git

[unreal ridge]

oh yeah, prompt blending seems to work way better

alternate is just giving me a random portrait 💀

[untold glacier]

I need to get and understand composable lora. I've seen it for a while but never looked into it.

[supple raptor]

Yeah, I find that alternating words can work well when the two things are pretty similar to begin with, but can take a lot of experimenting to find the right strength so that one doesn't completely overpower the other. Prompt Blending will just work without much fuss, but can also benefit from adjusting the strength of the different things.

[supple raptor]

I find that it's mostly useful for using two character loras at the same time. Basically it just allows you to use different loras in each AND prompt. It also allows you to change how loras are applied so that they don't get applied to the negative prompt which can sometimes be beneficial.

[untold glacier]

Gotcha, thanks.

[unreal ridge]

Two shot looks pretty cool

[wispy canopy]

i REALLY shouldn't do SD in the summer on my own pc 😦

having a A+ isolated house doesn't help

[untold glacier]

Tell me about it. My house has bad insulation from the outside, and here in Texas it gets a bit hot. It's been inches away from 80° in the house for the past few days in the afternoon.

[wispy canopy]

well, if i didn't use my pc it would remain pretty cool here... but i've not used my heating since february (with some nights freezing here), home isolation and pc being enough

[supple raptor]

Just downgrade the cooling in your pc so the heat is contained https://cdn.discordapp.com/emojis/1093687011765923850.gif?size=48&name=Kanna_fire&quality=lossless

[wispy canopy]

https://tenor.com/view/sam-wilson-hes-out-of-line-but-hes-right-just-saying-gif-24518939

[supple raptor]

A side benefit is that you'll be able to cook bacon on the back of your GPU

https://tenor.com/view/bacon-pig-heater-closer-piggly-gif-24115615

[wispy canopy]

right now i get partially cooked cat at times

[untold glacier]

We need more people using DarkAlfa. They must join the cult

[wet elm]

whenever i generate an image its giving me the error: TypeError: 'float' object cannot be interpreted as an integer
what should i do?

[supple raptor]

Post a screenshot of the full error message in the CMD window. It should say what value is being given as a float instead of an integer.

At least the last several lines of the error if it's a long one.

[keen hollow]

I'm frustrated... generated a few hundred high-res images overnight and today, and just... nothing. Nothing I like. I have total "prompt block" and can't get a good new wallpaper. 😛

[supple raptor]

When I'm struggling to come up with a new prompt I dig up a prompt that didn't work well in the past and just start changing things until I manage to force it to give me something I like.

[wispy canopy]

which of the 30000 already existing prompts you have? XD

[keen hollow]

I need to narrow in on a theme or model or something.

I feel like a lot of the models I have loaded right now are just.. too similar.

sorta miss the days of the original SD model 😛

[supple raptor]

I just start scrolling up in my outputs folder and stop somewhere randomly

[wispy canopy]

you have not used darkalfa 😄

[supple raptor]

Usually it doesn't take long from whereever I stop to find one that didn't work.

[keen hollow]

what's that one?

[untold glacier]

Only one of the best

Great with natural language, and gens clean.

[wispy canopy]

it's my model mix

[untold glacier]

https://civitai.com/models/64843/darkalfa-effortless-high-quality

[keen hollow]

hmm

I'll scope it out

[wispy canopy]

be warned, your usual booru shenanigans will not work well

good part: no need for masterpiece highres 😄

[keen hollow]

I also tend to run into the issue where the hires re-render loses a lot of the interesting details from the smaller version. That's tough to work around. 😐

I don't really use the booru stuff much.

[wispy canopy]

you'll love it in that case 🙂

[untold glacier]

I haven't had too much of an issue with this. Are you using multidiffusion+tiled vae?

That's made the biggest difference for me in keeping the small details when upscaling.

[keen hollow]

this is sorta interesting...

[keen hollow]

I dunno what that means, so probably not. 🙂 I can try to find an example.

[untold glacier]

https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

It's an extension.

And it's, in my opinion, absolutely necessary for upscaling.

It replaces all other upscaling methods, so don't try to use this in conjunction. You can use whatever upscaler model you want with it though.

[keen hollow]

I am not using any upscaling. Just hires, which is an img2img full-res render.

[wispy canopy]

i just run with highresfix 640 -> 1280

[keen hollow]

the img2img transfer is tough... lower denoise retains some of the original interesting details, but quality and clarity obviously suffer

[wispy canopy]

sometimes i don't get my model. i'm not complaining. i just don't get it.

[supple raptor]

Which upscaler do you use in Hires fix?

[keen hollow]

SwinIR, lollypop, and ultrasharp

those are my go-tos

[wispy canopy]

i'm using ultrasharp only

[untold glacier]

This will work in conjunction with the hires fix. After a point in the size of the image, the multidiffusion tiles will kick in. It won't if it's too small, so you won't need to worry about it with smaller images.

[keen hollow]

lemme create an example to show what I mean.... too many images to sort through to find one i might have kept 😄

[supple raptor]

Try using latent since you're wanting more details. You'll need to increase your normal denoising by ~0.1 so if you normally use 0.5 go up to 0.6.

[keen hollow]

latent... old school... i can try it

[untold glacier]

Animesharp, fatal_anime, ultrasharp, are my staples.

[wispy canopy]

lowest denoise for latent & dpm2 types is 0.58

any lower and you get that hazy look

[untold glacier]

I can never get latent to give me what I want with in a cohesive whole. In my testing upscalers have worked much better in the long run.

[supple raptor]

When latent works well, it is by far the best option. Some models and prompts just don't work well for it though so the GAN upscalers are far more reliable.

4x-AnimeSharp, fatal_Anime, and Digipaint35000 are my go to GAN options.

[untold glacier]

Digipaint35000? What do you use this one for?

[keen hollow]

Example:

So this isn't the best example, and hard to see if you don't zoom in.

But if you pull them up, check the buildings on the right.

The hi-res version is very flat and boring.

The lights down at street level just all get blurred together

[untold glacier]

Yeah, you'd want a sharper upscaling model for that. Something like ultrasharp or fatal.

[keen hollow]

lots of straight lines and almost glass-like planes...

you lose a lot of the dots and details and grain and things that make it interesting

[supple raptor]

Anything that's closer to a painted or hand drawn style.

[untold glacier]

That and you'd want multidiffusion's tiles, giving more creative presence to the gen.

[keen hollow]

This is RevAnimated, simple prompt like 'cyberpunk city' or something

@untold glacier , can you post an example?

before I go get a PhD in how to use that plugin, would love to see what it does 😄

[untold glacier]

If you want, you can give me the prompt info/settings and I'll gen an upscaled version.

[keen hollow]

sure

prompt: a dark cyberpunk city, intricate details Negative Prompt: negative_hand-neg, bad_prompt_version2, (worst quality, low quality:1.4), Steps: 23, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 655095045, Size: 1200x500, Model hash: d725be5d18, Model: revAnimated_v11, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 2, Hires upscaler: lollypop, Version: v1.3.2

[untold glacier]

Btw what card do you have?

[keen hollow]

RTX 3060 12GB

[untold glacier]

Ok. The seed might be slightly different for me, either way I'll gen a normal sized and upscaled version.

[supple raptor]

more digipaint here

[wispy canopy]

my default upscale settings on darkalfa

[untold glacier]

Nice. I'm definitely going to have to give that a look.

[untold glacier]

My internet is slightly slower than it usually is so it'll take me a second to get the model.

[supple raptor]

Jemnite, from the Touhou server, made a model inspired by the mix of LoRAs that I usually use and I'm really enjoying it.

[untold glacier]

Neat

[keen hollow]

ah no worries

wait.. you really don't have revanimated already?

[untold glacier]

Dude I have tons of models and very little hard drive space, lol.

Oh. And one thing to remember when using DarkAlfa is that it often works best at clip skip 1, believe it or not.

[keen hollow]

RevAnimated is really popular, just surprised. 😄

[untold glacier]

I'm slow to upgrade, and deep in experimentation. Once I saturated my hard drive with models, I find it difficult to give up space for another.

It certainly looks interesting though

[wispy canopy]

noooo yumeshiro, i cannot lose you!!!! 😄

[untold glacier]

[wispy canopy]

as one of the last darkalfa defenders 😄

[untold glacier]

Oh... No, I'm definitely sticking with DarkAlfa. I haven't encountered a model yet that has the ease of use that this one does.

[keen hollow]

Can DarkAlfa do more creative stuff and not just portraits/landscapes we've seen billions of?

Like "A cat with a body made of cheese swordfighting with popsicles"

[wispy canopy]

only one way i can find out!

prompting that atm 😄

[keen hollow]

Here is another example. Not talking about the changes to the image, but moreso, the loss of details in the textures. Look at the castle walls and see how the higher-res generation lacks the character in the texture.

[untold glacier]

It's quite good at sticking with the prompt that you put in, but if you're looking for insanity you're never going to get great cohesion with any model.

[keen hollow]

LOL

Well I exaggerate. 😄

[wispy canopy]

i dunno what i was expecting though 😄

[untold glacier]

Oh, what VAE did you use? And I assume clip skip 2?

[wispy canopy]

clip skip 2 is really bad for background details imho 😦

[untold glacier]

So you used clip skip 1 then?

[wispy canopy]

i'm always at clip skip 1

[untold glacier]

And your VAE?

Standard anime2?

[wispy canopy]

anime2

yeah

putting clip skip is a "weakness" because you're laxing the control, because the model doesn't understand your prompt

but that's just like. my oppinion man

[untold glacier]

I agree completely, though it's needed for many anime models to work properly.

[wispy canopy]

@keen hollow is this the kind of stuff you want?

[keen hollow]

the bottom of this is very interesting. The main focus point on the portrait .. meh. 🙂

[untold glacier]

https://tenor.com/view/psychedelic-cat-whoa-glasses-sunglasses-gif-24905370

[wispy canopy]

well, it's just the song lyrics i was listening to whatever i was liysting to, not even bothering cleaning them up

[keen hollow]

Superman with Iron Man's doohickey

[wispy canopy]

imma start up my verser again 😄

[wet elm]

i managed to fix it

[wispy canopy]

song lyrics are the best 😄

[keen hollow]

the owakita style ones are kinda fun to do..

[wispy canopy]

@keen hollow here it is. the eye of the storm!

[keen hollow]

the purple in the center is neat

[untold glacier]

Omnia... it is clip skip 2, lol. It wasn't set to 1. This is with 2.

This is 1.

The top is very close to your original image.

[keen hollow]

I use clip skip 2

ugh.. those lines are so straight.. just makes interesting images so boring 😦

[untold glacier]

Ohhhhhhhhh. I see what I did. I asked you a question earlier but Eface responded, lol. I didn't look at the username carefully and that it was you. My bad.

Ok. Now to test the upscaling.

These are my settings

Aaaaand here's the beautiful upscale

[wispy canopy]

copy paste the image in controlnet, do tile resample

that might help

my image description had mona lisa... oh well

[untold glacier]

It's definitely different, looking at the two side by side. But with the denoise comes more freedom for SD to gen detail.

Oops, forgot to tag @keen hollow : #🍥｜anime message

[wispy canopy]

oohh... i like this ❤️

[keen hollow]

This is actually at a HIGHER denoise.

[untold glacier]

Yep. 0.4

[wispy canopy]

@untold glacier the song used for those images is "The Master of Illusion", funny you respond with the spades 😄

[untold glacier]

Hah

[keen hollow]

This seems interesting.... what exactly is this doing?

[untold glacier]

It's generating the image, then upscaling with ultrasharp, then splitting the image into two with an idea of how it looks as a whole, and finally generating each section and seamlessly blending them back together for the end result. To put it simply.

[keen hollow]

how is that different than SD Upscale?

[untold glacier]

I believe the tiled VAE makes the difference. Now don't ask me how that one works, as that's beyond my knowledge.

That and you can set the latent tile dimensions if you need to (I usually keep it at 128x128)

You can also have region prompt control, and specifically prompt for certain things at certain areas of the image. That one has helped me a lot with people in larger images.

[keen hollow]

ok i'll give to scope it out

[untold glacier]

Cool beans. It's just "multidiffusion upscaler for automatic1111" in the extensions tab.

[keen hollow]

Someone needs to render "cool beans"

[untold glacier]

@wispy canopy Back when you were still waiting to hear back from Cryo about DarkAlfa, I made a couple merges using your model name as a guide--which got me DarkAlfa early to experiment with. One of those merges ended up being what I call DarkAlfa-Strong. Which is sometimes slightly less cohesive, but often stronger in it's focus. For instance, it may not gen hands as well, but the cinematic presentation might be slightly better. It's often not worth my switching from DarkAlfa, but it acts as a great alternative if I have a great prompt but need slightly more oomph out of it and am not concerned with cohesion as much.

DarkAlfa Strong

DarkAlfa

[wispy canopy]

looks pretty damn nice, actually!

[untold glacier]

It can be great, but I often prefer the quality that comes from regular DarkAlfa. If you want the model though, let me know.

[wispy canopy]

don't think i do SD regular enough fo more models though, waiting for the next big model XD (more parameters)

[supple raptor]

Unless they learned how to properly train text encoder I don't think that XL is going to be nearly as exciting as people think. OpenCLIP's lackluster training is a big part of the problem with 2.x and why it's so much harder for people to train. CLIP had a really good initial training which is why the technically inferior 1.x models are so much easier to work with.

[wispy canopy]

lets hope it's not as bad 🙂

[untold glacier]

Exquisite, Nubby. Great work.

[supple raptor]

I give credit to the beta model I'm using https://cdn.discordapp.com/emojis/978879333248667718.webp?size=48&name=AYAYAYA&quality=lossless

[unreal ridge]

The reflection looks really fucking nice on this one