🍥｜anime

[acoustic locust]

where do merge models go in a1111?

why happen

it still generate

[gloomy scroll]

Model folder

[acoustic locust]

I just made

the best merge

it's so consistent

and the hands are great

that's the whole prompt

[steady grail]

wow

this iss xl?

[acoustic locust]

no

1.5

it's the best possible merge imo

[steady grail]

you got such a precision with 2 words?

[acoustic locust]

yes

[steady grail]

damn

[acoustic locust]

look at the hands

they're perfect

[steady grail]

works only with cirno?

[acoustic locust]

no

[steady grail]

yep

[acoustic locust]

I"m just using cirno

see

it's literally the prefect merge

[steady grail]

try sailor moon usagi

[acoustic locust]

ok

first try

[steady grail]

whoooooooooooooo

[acoustic locust]

I've never watched sailor moon so idk if that's right or not

[steady grail]

nah, usagi it's the blonde protagonist

but it's pretty close

maybe he will get it with Usagi Tsukino

but she have pigtails

and the school drees is the same

[acoustic locust]

I might need a lora

[steady grail]

you will share it?

it's so damn cool

it's so responsive

what's the mix? i'm very curious

[acoustic locust]

Should I upload it to civitai?

[steady grail]

how do you prefer, the result is very cool

would be a waste to let it vanish

the last one need a fix to the hand lol

[acoustic locust]

that's a 1% though

others are really good

it's really consistent

[steady grail]

yep, they are just small mutations to the number of finger

it's easy to fix

i even know how to do it manually

[acoustic locust]

regardless of an extra fingers that hand still looks like a hand

[steady grail]

yeah, really

never saw such response with such a simple prompt

toh, i just fixed an image, with simply a img2img

[acoustic locust]

It's going up

[steady grail]

i'm still surprised of what ai can do

it's simply mind blowing

i have a certain idea of what's going on behind the code, but still, my mand can't grasp how it is possible

[acoustic locust]

It's really impressive.

[steady grail]

so many times i made art out of my lunacy i feed him lol

what happens if connect these nodes? wow cool, and if i concat this 2 results and than feed the noise in this other result? should be an absolute caotic mess, but no, it gives me a master piece lol

why? XD

till you follow some basic rules, the ai nowadays are so stable, to be able to make art with anything

even feeding it drawings from paint

[acoustic locust]

https://civitai.com/models/268705?modelVersionId=302948

it's up

[steady grail]

yoh, thanks

i hope comfy didn't crush

i usually don't run gens at such high res

yeah, crush

[acoustic locust]

post some gens when you get it working

[steady grail]

yep

[acoustic locust]

@untold glacier

I made a great merge

👍

[steady grail]

by the name, i can say one of the components is soushiki lol

[acoustic locust]

lmao I know

It was my favorite model

[steady grail]

got you lol,

[acoustic locust]

the only reason it was the best model

is beacuse mine wasn't made yet

https://tenor.com/view/gojo-gojo-smile-gojo-maniac-gojo-laugh-gif-13257859273272055083

[steady grail]

uahahaha

[acoustic locust]

https://tenor.com/view/fnaf-gif-25599710

[steady grail]

some lowres stuff

chi-chi from dragon ball

[nova remnant]

[steady grail]

pretty close

[acoustic locust]

is that with my model?

[steady grail]

yep

[acoustic locust]

they look like hands

and not a grotesque amalgamate approximation of what hands look like

[steady grail]

i agree

[steady grail]

when i add some style, the air became red

weird

[acoustic locust]

can i see?

[steady grail]

upscale a little do add some details

for such a simple prompt, the result is very cute

and the mutations are almost none

interesting, just by inserting black hair, change even the expression, and the character looks closer and closer to chi-chi

seriously, i don't know what have you done, but the model is very responsive

very cool

[untold glacier]

Is it SDXL or 1.5?

[steady grail]

1.5

impressive, right?

[untold glacier]

[native halo]

yes,when u realize some 1.5 models can do good hands

[untold glacier]

I just can't go back. I need the greater awareness that SDXL has in regards to natural language prompting and understanding shapes and structure. It's why efforts like PonyDiff are so important.

[acoustic locust]

I know, and you're right. But SD 1.5 is also really good.

[steady grail]

and that's when you try to make an artistic model and instead you get a porn generator lol

[acoustic locust]

On god, I hate getting NSFW results

[steady grail]

they are just boring and enbarassing

[native halo]

me when the sex

https://tenor.com/view/chaos-head-chaos-world-takumi-nishijou-nishijou-takumi-crying-gif-25314258

[steady grail]

uahahaha

[acoustic locust]

It so sed

[steady grail]

about sex, i remember when i was 16 and they taught me just 1 thing.
sex is not about the stuff you can see everywhere, but the feelings you share with the other person

[native halo]

thats cool is that a quote from a lifetime movie?

[steady grail]

no idea, it's just something they said and I always carry with me

[untold glacier]

So... I spent some of the last few hours attempting to verify what I had thought was a good solid foundation for generating with Pony, and in fact, it verified that I had a bias towards the model. I exclusively tested without artist names, as that should be doable for a base model, and it showed me that the SFW prompt adherence just isn't where it needs to be.

Image #1 is supposed to be a maid. #2 is supposed to be a girl in goth lolita at a convention. #3 is supposed to be a girl dancing in her bedroom. #4 is supposed to be a girl in a wheel chair with a robot pushing her. All of these I tried were prompt and settings optimized FOR PonyDiffusion, so theoretically they should be the best they can with the model. But they're just not great at all.

For comparison purposes, here are some of the same images but with HiwamariMixv2 or Hiwamari+BluePencil Upres--usually the same or adjacent prompt but without the keyword modifiers.

[untold glacier]

Oops, there's the convention one.

[steady grail]

looking for the right person and focusing on the feelings and not the superficial acts, that's the vision of love i matured

[acoustic locust]

how do i make obvious box go bye bye when using regional prompting

?

[steady grail]

and sex is just one of the way we express love

[acoustic locust]

I like that.

[native halo]

ima try that on tinder

[keen hollow]

so @untold glacier ... what's the leading contender these days?

[native halo]

oh there was an official competition?

https://tenor.com/view/whaaat-what-shto-emotiguy-shocked-gif-27492547

[keen hollow]

well

[native halo]

how big is the prize

[acoustic locust]

You're never getting laid.

[keen hollow]

isn't yumeshiro the one that does those comparisons and finds the winner?

🙂

I rely on those!

ya'll, I don't have time to test 300 models with 30000 iterations each

[native halo]

xyz plot and leave the computer on for 3 days 🤠

[untold glacier]

All in all, with my testing, it's still HiwamariMixV2. Though I will say I'm very impressed with the latest Animagine... and that's not something I ever thought I'd say. It's a fantastic flat color 2d model out of the box.

[keen hollow]

I have done this quite a few times lol

[keen hollow]

oh wow, so things have changed a lot

[acoustic locust]

the 1.5 or the xl model of hiwamari?

[untold glacier]

I'll let you guess.

[keen hollow]

is that on civit? 😐

[native halo]

the one that can create anime characters

[acoustic locust]

Oh man, I wonder which one does that.

[untold glacier]

BluePencil3.1 is right up there with it, but it has large subject cohesion issues comparatively. Large as in size, so bigger things like anatomy, subject recognition. Things where they should/shouldn't be, etc.

[acoustic locust]

I'm not seeing a hiwamari mix v2

[untold glacier]

What I've found often works wonders is Hiwamari on the base gen and BP3.1 on the upscale.

[native halo]

try to create mikasa on both versions

should be easy enough right

[keen hollow]

yeah I remember bluepencil was one of your leaders in the past

[untold glacier]

https://civitai.com/models/131611?modelVersionId=257436

[acoustic locust]

thanks

[untold glacier]

Yeah. It's still great, but it lost some cohesion with all of the additions the creator has made over time.

[keen hollow]

I still have v2... what is he up to?

3.1 is the latest?

[untold glacier]

Yep

Last I checked

[untold glacier]

Feelings will always change or fade with time. I've heard it once described as thus: Love is an unflinching, unwavering commitment to the welfare of someone else at your own expense.

[steady grail]

i need 18 min to generate something with a 4gvram XD

[native halo]

but its new

[steady grail]

and i can't use lora or controlnet

or i'll crush everything lol

[native halo]

you can use loras there u can choose from the massive list of 160 loras on civitai

[steady grail]

like the water in the river, things never stay the same, but they naturally flow in their destined direction

[steady grail]

on 4 gvram, even a single lora will block everything

it's already a miracle i can run it, but it's not worth it

i attempted it for fun, but overworks my pc for an unworthy result

[native halo]

its ok im sure they will soon release another trimmed version of turbo with just 700mb params and 0.5 steps

[steady grail]

i get better results with a good comfy workflow

[native halo]

or we could also wait until the a100 drops to 200 usd,easy imgs and easy training

it took the tesla p40 8 years to drop price to 150 usd im sure we can wait for it

[steady grail]

when this will happens, we will get ai able to generate a moon 1:1

and our plays will become some childish joke

[native halo]

we will live in the moon by then so no need to gen one🌚

[steady grail]

uahahah

[untold glacier]

So what you're saying is... "If you wait by the river long enough, the bodies of your enemies will float by."?

[acoustic locust]

what card?

[steady grail]

gtx 1650

[steady grail]

something more romantic. good things with grow and bad things....... will die <.<

damn, the last part is not very romantic lol

upscaled

these hands are insane

[native halo]

try with a harder hand pose like this one

[nova remnant]

Tinder is a nice way to find people with cats to follow on instagram and then have cute cats to look at in the instagram

👍

thats what tinder is for, right ?

[vital raptor]

yes, for cute 'cats'

[nova remnant]

(fingers, fingers together, fingers in front, showing fingers, heart on fingers, finger heart, making a heart with fingers:1.3)

[native halo]

its great for one night cat viewing adventures,no compromises

[steady grail]

i'll try it

[nova remnant]

[untold glacier]

Now I want to see this

[vital raptor]

.

[nova remnant]

I'm innocent

[native halo]

tried that hand pose on the kakarot model and it had a 12% chance of success so 88 pics came out bad and 12 good ones

[steady grail]

almost

this is a low res

[nova remnant]

food

[native halo]

the 6 finger

[steady grail]

yep

[vital raptor]

it's trying

[steady grail]

i should prommpt to prevent it

oh, i made it

[native halo]

try with this hand pose now

[steady grail]

how is this even done? lol

[native halo]

thats one of the 12 imgs from kakarot that came out good

[nova remnant]

What even pose is that

[vital raptor]

close enough

[native halo]

i just typed (heart hands, own hands together,hands up) 100 img batch

[nova remnant]

She stole a finger

[native halo]

me lookin for good fingers

[nova remnant]

I'd rather look for elves

Reading books with them

[untold glacier]

Excellent

[native halo]

here u go🫡

[nova remnant]

anime character

[native halo]

https://tenor.com/view/shinobu-oshino-shinobuoshino-monogatari-gif-8451549

[nova remnant]

https://tenor.com/view/stranger-things-fire-flame-gif-14920648

[native halo]

https://tenor.com/view/shinobu-oshino-tongue-out-tease-blush-gif-9753291

[nova remnant]

[acoustic locust]

🤐 🧏‍♂️

well that was fast.

what do you think?

[native halo]

how much is that in 💸

[acoustic locust]

3 ruples

[native halo]

https://tenor.com/view/neco-arc-gif-25900194

[acoustic locust]

Me after spending fiddy bands on a mediocre animation

(I’m verbalase)

[native halo]

verbased

[acoustic locust]

Verbankrupt

[acoustic locust]

https://www.youtube.com/watch?v=3j6W0RpxzEg

[silent kraken]

Where is this from

[native halo]

Kizumonogatari movies,the first one

[silent kraken]

Thanks

[silent kraken]

Thank again

[stable crane]

https://www.youtube.com/watch?v=fM4GuLgWi4U

[acoustic locust]

gendump time

[acoustic locust]

Model?

[brittle rampart]

yours

[acoustic locust]

Thanks

[acoustic locust]

Thoughts so far?

[brittle rampart]

very well. The first generations did a good job with legs and arms.

[acoustic locust]

It’s really consistent and I like it alot.

I’m thinking about renaming it to BorealisMix

It sounds better

Souoroa or BorealisMix?

[brittle rampart]

I'm not the creator. I rather judge by the result than by the name😁

[acoustic locust]

Ok

Well I am and now it’s BorealisMix

[brittle rampart]

Borealis - what is this for you?

[acoustic locust]

?

[dawn barn]

lowvram is for 2GB VRAM, medvram is for 4GB VRAM and will also perform similar to just leaving it out

Can you send prompt

Ok my A41 opened just now

[dawn barn]

Yeah just put (green:0.1) eyes

And adjust up as needed

Like (green:0.1) eyes

They have a professional card that you can just buy from ebay

A400

0

A4000

16 GB VRAM, you can buy it for about 600€

What UI are you using?

Even then, you will still get different outputs

[brittle rampart]

A1111

[brittle rampart]

right. but not so different if the seed was completely random

[dawn barn]

No, it's very different. But that doesn't matter anyhow, what user interface are you using

[brittle rampart]

I would like to know if it is possible to get a truly random sequence of seeds when use a "Batch count"

[dawn barn]

Can u pls share how you get that art style

[brittle rampart]

@acoustic locust i think her eyes, little big 😂

[dawn barn]

Tried opening in png info, no prompt

Could be, though once it's loaded into GPU it should be fast

So each step should be fast, but from steps to image, and from prompt to steps should take longer

[native halo]

a 3070 doesnt fill the pcie 3.0 bandwidth

[dawn barn]

I'M STILL CATCHING UP

YOU ARE ALL TYPING SO FAST

[dawn barn]

Uh should be somewhere in settings

[native halo]

no one is typin right now its actually kinda dead

[dawn barn]

Like

It shouldn't do that

[dawn barn]

Oh no

that's intentional

[brittle rampart]

I’ll be glad to see where THIS is!?

[dawn barn]

But

Wait

Okay from what I saw

I can't find it anymore

I know it used to be there, so it prolly just got removed somewhere down the line

What you can do tho is

Make sure that this here is enabled

[brittle rampart]

"produce same images in a batch!?""

[dawn barn]

It means that

When you create it in a batch

That it will be the same

As if you had created it alone

[brittle rampart]

and this guarantees me a random row of seeds?! how?

[dawn barn]

It doesn't

I can't find the setting anymore

I know it used to be there

But like again, it doesn't really matter

The pics between each seed should be substancially different

[brittle rampart]

i want random seeds in batch

[dawn barn]

You can switch to ComfyUI tho if you really care about this, but as I said, it's not really necessary

[native halo]

it doesnt matter,just by changing a single number at the end it already makes the seed different than last one

[brittle rampart]

yes, but as I noticed, a simple change by 1 often gives fairly close results.

[dawn barn]

It really shouldn't

If you can document it

Then submit it

Then it'd be great for them to know as well

Because it shouldn't

[native halo]

ok then try this, instead of doing a batch generation,try to right click the big orange generate button and click where it says "generate forever"

[dawn barn]

You have variation seed enabled

That's why

[dawn barn]

Yeah, disable variation seed

Just disable it entirely

It keeps the same variation seed

Across the batch

I think

[nova remnant]

Forever is a long time

[native halo]

generate forever then come back in a month

[nova remnant]

(hands above head:2)

The true way of prompting

[native halo]

(pov:2.5)

[dawn barn]

I need to start deleting some models

It's getting too much again

[dawn barn]

From brief testing, I don't think hands perform better for me

Hands on head does perform decent

Like 60% of the time or 70%

But hands folded together for example, which is my go to, sucks, and at times sucks worse

The fingers just end up misshapen

What VAE were you using?

I think in this case VAE can affect finger quality

[brittle rampart]

" VAE can affect finger quality" Rally? i think VAE more for colors, bright...

[native halo]

no vae has nothing to do with fingers

[brittle rampart]

+1

[dawn barn]

VAE can affect some shapes and textures

You can look at some things in detail, try it out without VAE and with VAE

Or maybe the model I tested that on just was particularly unstable

Could also be the case

[brittle rampart]

and which VAE`s do you think are good for fingers?

[dawn barn]

Come to think of it, I was forced to run xformers back then

[native halo]

these are the type of vae artifacts u can get,nothing to do with fingers

[dawn barn]

Lemme test again

Here an example

Just now genned

[brittle rampart]

but where fingers?

[dawn barn]

.

[brittle rampart]

i think this is colorcorrection and brightness

[dawn barn]

Look at her dress

Or whatever that is

Look at her head

Just use arrow keys

To move back and forth

Use left and right arrow keys

It's just sometimes when it's more ambiguous from the looks of it

[brittle rampart]

in fact, most of the image is not related to changes in shape or texture

[native halo]

https://tenor.com/view/funny-gif-27451299

[dawn barn]

That being said, is the SDXL VAE 1.0 still artifacting like that

[brittle rampart]

It would be nice to see all this in relation to the fingers.

[dawn barn]

The head position is different, there is a nub there, and the wall is different

[dawn barn]

I don't know and haven't tested yet whether it affects fingers too when they are misshapen, but I am not excluding it

[dawn barn]

The changes are typically much smaller

E.g. here the difference

Wait I think I grabbed the wrong pic

Can be way smaller

That's kinda cuz the VAE is trained to be similar to each other, it is a decoder after all

But different VAEs can decode things differently

[nova remnant]

https://tenor.com/view/ken-jeong-cant-see-this-funny-gif-14446155

[native halo]

https://tenor.com/view/yapping-gif-3664239514826816233

[dawn barn]

From a bit of searching, apparently it just uses the "original" VAE if None is selected

And other VAEs are a finetune of that particular one

[nova remnant]

No

Post elf

[native halo]

something wrong with my model i think its using a pirate vae ( non original)

[brittle rampart]

😂

[dawn barn]

Huhhh

[native halo]

[nova remnant]

Is this

Anime character?

[native halo]

https://tenor.com/view/anime-krul-serap-of-the-end-dance-gif-26790118

[nova remnant]

[nova remnant]

[native halo]

https://tenor.com/view/redanime-gif-25752591

[nova remnant]

anime not allowed

[nova remnant]

o:

[native halo]

https://tenor.com/view/full-metal-panic-anime-fumoffu-gif-13091153

[dawn barn]

Another new way to hide hands

[brittle rampart]

yes. like this is

[lyric zealot]

are you not using hires fix?

[dawn barn]

Not really

Unless I have a pic I want

[lyric zealot]

ahhh, alright

[dawn barn]

For this one I fixed hands in inpaint

That's why no metadata

[dawn barn]

This here was just good gen

[vital raptor]

an old used card for a discounted price doesn't interst me

I already have a 4090

[dawn barn]

I mean it has video output and gaming features

[vital raptor]

I was talking about a new product that's affordable to hobbiest AI users

[dawn barn]

Think that most hobby AI users would want their monitor output not on a separate GPU tho that they'd also need to buy separately

[nova remnant]

(stealing your card:3.2)

[vital raptor]

maybe, but most AI users have the cheapest GPU they could get because gaming doesn't need a 4090

having a cheaper AI only option would suit a lot of people

also useful for a headless system that's just for AI

[nova remnant]

I see you're not a path of exile player

[native halo]

the intel 16gb could be that option ( only if intel could fix their driver issues)

[nova remnant]

Its not bad for AI anymore

[vital raptor]

the point is that vram is cheap, it's a lot of the other part of the cards that are more expensive to produce, they could easily make a non-gaming 48gb AI card for under $1k

they wont, but they could, and it's what i want 😛

the only reason i upgraded to a 4090 was the extra vram for AI stuff, and tbh it wasn't worth it in the end, it's not enough

[native halo]

yes i remember huawei had a few of those chips,they called them NPU's ( neural processing units) they were made specifically for AI but idk if they will ever release them to the public,why would they,more money if they sell them expensive

[vital raptor]

I doubled my vram but still can't double the resolution because 4K is 4x 2K

[nova remnant]

Won't models like bug out with that much res

[vital raptor]

i could have just rented some gpu time for a few hundred hours and still saved most of the cost of the gpu

[vital raptor]

not if you render at a normal res and use hires fix

every image i've posted here is 1024x1024 with hires fix to 2048x2048

[native halo]

yea either high res fix or tiles upscale can do them

[vital raptor]

but you can't hires fix 4K unless you have the vram for it

and the other upscaler options are generally inferior

[slate narwhal]

This is from the bot; the art style is by Russell Mills (just put Russ Mills on the prompt). The Russell Mills prompts are a blurriness fest, ruining SDXL gens easily, so the test in this one is to get maximum sharpness through upscaling.

[lyric zealot]

angy

super angy

[dawn barn]

Thanks

[native halo]

she can punish me anytime

[lyric zealot]

she will have to punish me first

[vital raptor]

mornings..

[nova remnant]

huh

[lyric zealot]

didn't i told you about downloading that extension, and then suddenly finding i have tons of embeddings, im trying some of them

[dawn barn]

Didn't get much results beyond variations of this, which is already pretty good results. But it doesn't really look blurry

[valid bridge]

Does anyone here have experience with training anime loras? Just curious.

[dawn barn]

Like don't get me wrong, I'm looking for something similar to that

I'm just confused

[native halo]

depends,is there any money involved

[valid bridge]

Well, okay, so I can train the lora myself, I just want to be able to know the best configuration for it so that it isn't stiff, or muted, or anything like that.

Because I can tell there's a style change with the loras I'm training.

[slate narwhal]

The eyes in your gen are blurry for a portrait close shot, and the lines are dim, even though the bot is much better these days. This happens with most of the illustration/watercolor/cartoonish type prompts.

If you wish to get the similar lighting to what I got in that gen, use a photographer in the prompt. Then upscale.

[dawn barn]

I want the blurriness and smudgyness, just in a more anime way

I was hoping that I could use this to replicate

The style here

Didn't really work but I have found an alternate method to get closer to i t

[slate narwhal]

Try IPadapter using the image with the style you want and a model close to this style, and if that fails, try creating a Lora for the artist, or train a checkpoint.

[dawn barn]

I cannot find artists

This is AI generated

But the prompt

Is really messed up

[nova remnant]

Are you using embeds

They often lock your style more than not

Like easy negatives and such

Negatives make images too clean for what you want

[dawn barn]

How does IPadapter work?

I don't use negatives

At best I used to use (bad quality, worst quality:1.3), then adjust negatives based on results I get

But for this style

I found out that bad quality, worst quality interferes

[nova remnant]

Yeah

[dawn barn]

Like found out today, so I'm putting colours in there instead

And that works pretty well

[nova remnant]

That exactly makes images too clean

For what you want

Low quality and worst quality are the trained tags

At least on the models I like

They do the most

Watercolor is a good prompt too I like

[dawn barn]

Might need to bump watercolour up

[slate narwhal]

Ipadapter tries to imitate the style using the current model you are using without putting in new content. It can fail depending on the model.

Easiest way to do this would be to take the main style image, put it in the image prompt tab on Fooocus, try the closest you can to the prompt, and try different anime models. You could also try the "variation" feature in Fooocus to create more images close to the original style. Check the detailed instructions @:
https://github.com/lllyasviel/Fooocus/discussions/557

[dawn barn]

Another webui

Okay, why not give it a try

Tho it is already p decent now

With all the adjustments, maybe not really

[slate narwhal]

If you use Fooocus for this, don't forget to turn off all the default SDXL/Fooocus styles in the advanced/styles tab

[dawn barn]

This is still A41

Still downloading Fo3cus and scanning

[slate narwhal]

Yeah, it's pretty nice already.

[dawn barn]

I'll see how resistant it is to more modifications

[broken osprey]

Tried to replicate your first prompt only using deepboru and a pose, I think somewhere along the lines sd went a little nuts on the adetailer phase

[supple raptor]

@finite herald when she gets to swing the banhammer:

[finite herald]

Little ol me?

[supple raptor]

https://cdn.discordapp.com/emojis/1192365419936350218.webp?size=64&name=miku&quality=lossless

[finite herald]

[gloomy scroll]

Wild Fruit spotted

[steady pumice]

interesting

[wheat falcon]

hey

[timber crown]

@vital raptor, what parameters are you using for fastest gens?

[nova remnant]

Help channel man

[gloomy scroll]

@finite herald Smile for the camera 📸

[nova remnant]

Would press

The buttons

[nova remnant]

This image is very random now that I look at it

[vital raptor]

i'm only using --xformers and --opt-channelslast in my webui-start.bat if that's what you mean

I also use TensorRT on my main models to speed it up a little

[timber crown]

Did you try --opt-sdp-attention or --opt-sdp-no-mem-attention

[vital raptor]

at one point, before torch 2.0 i got a little more performance from sdp-attention, but since torch updated it's always been slower for me

[nova remnant]

seconds before the fall

hello captain

[vital raptor]

I haven't tried in the last few months though, ~11 seconds for a 2048x2048 image is fast enough that i don't really mess with tweaking settings

[vital raptor]

👋

[nova remnant]

gib promppt

[supple raptor]

--opt-sdp-attention should be roughly the same speed as --xformers, but uses around 20% more vram on most systems while also being non-deteministic (slight differences everytime you generate even with the exact same prompt, seed, and parameters).
--opt-sdp-no-mem-attention is both slower and even less memory efficient but at least it's deterministic (like xformers is since xformers 0.0.19).

[timber crown]

Why is it the same?
When you say deterministic, will ALL generations be separate, even with the same seed, or is it like "oh, same gpu, same seed, xformers, you get the same result for yourself but no replication can be made from other users?"

[supple raptor]

Even with the non-deterministic optimizations the differences are slight enough that you might not even notice if you aren't paying much attention, but it's enough to potentially get mangled hands on one attempt then get decent hands if you generate the exact same image again. That part is not much of a downside, but when combined with no increase in speed and you're using more vram then it doesn't do it any favor.

The speed is roughly the same because as I understand it anyways, sdp-attention is very similar to xformers, just not as memory efficient. The main benefit of sdp-attention is that it's available on GPUs other than Nvidia.

[nova remnant]

huh

[vital raptor]

It’s in the metadata

[nova remnant]

I tried to look for it and it really was

magic

[keen hollow]

evenin'

[terse ore]

AAM XL is out, another anime model.

[gloomy lodge]

ignore the fucked up cat dog

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1197789322984837320/00184-3356903518.png?ex=65bc8b3c&is=65aa163c&hm=d442afbda38a49950891c4c6d53ada802f757923fe2aff6a0e809e2feeadc1dc&=&format=webp&quality=lossless&width=397&height=662

[untold glacier]

Miku dump before bed

[bitter latch]

welp i'm messing around with wildcards and well.. 😄

and wow

totally random

[acoustic locust]

@wooden coral @dull jackal

Mods

Strike him down

[wheat falcon]

User warned, message deleted

[lyric zealot]

HUGE

[zinc flax]

Huge smile

:D

[misty jungle]

如果知道会在这场同学聚会上再见到邹宴，我是不会来的。 张雅楠大概也知道，所以她叫我来的时候给我的名单上并没有邹宴。 又或者，她就是想看看这种热闹。 “邹宴你可以呀，几百万粉丝的大美女网红都被你骗到了！” “哎呀，人家这叫郎才女貌！佳偶天成！“ “那可不，咱们邹总现在可是互联网新秀，名人排行榜上的人物，老牛逼了！” “邹宴，不介意我和你未婚妻合个照吧？咱这第一次见这么大的网红，完了我好拿出去吹牛啊。

[native halo]

你说什么?

[brittle rampart]

一个人写一本小说

[misty jungle]

第一次使用，有很多看不懂而且不会使用

[dull jackal]

Thanks for the ping!

[bitter latch]

made using wildcards 😄

[solemn cairn]

very good image

[nova remnant]

I want her to smile on top of me

Get covered in happiness

Boba

[zinc flax]

Giving you the big happy

[nova remnant]

@zinc flax sleepi

[zinc flax]

eepy

[steady pumice]

migu

[gloomy scroll]

Did you guys know that this is what Fruit looks irl?

[native halo]

no

like this

[sour shadow]

migu migu migu migu mi

[solemn cairn]

what todo if the loras that i put in the paste doesnt appears ?

[native halo]

dance

https://tenor.com/view/dance-marge-simpson-bart-simpson-the-simpsons-dancing-gif-16685690

[gloomy scroll]

[brittle rampart]

Meruru Meruru Meruru ruuu😝

[sour shadow]

[sour shadow]

Sorry but I'm not sure what you mean

[gloomy scroll]

Could you explain little bit more

[solemn cairn]

if i use a model XL , some Loras doesnt appears?

this could happen?

[gloomy scroll]

Is that lora for XL?

[solemn cairn]

no its a normal lora

[gloomy scroll]

sd1.5 loras don't work on XL

[solemn cairn]

for sure?

[gloomy scroll]

Not rly since I haven't actually tried but to my understanding they should be incompatible

[obtuse trellis]

Yea they don't work well that's the one downside to xl, can't get all the 1.5 stuff

[solemn cairn]

but they doesnt not appears?

[gloomy scroll]

Appear where?

[native halo]

there

[solemn cairn]

on SD

SD has a paste name Loras

[gloomy scroll]

[zinc flax]

I don't get ir

[gloomy scroll]

What does SD mean in this context

[native halo]

bogos binted

[solemn cairn]

stable difusion

[zinc flax]

I feel old now

[obtuse trellis]

[gloomy scroll]

Do you mean like user interface (UI)?

[solemn cairn]

yes

[gloomy scroll]

Then that isn't SD

Are you using Automatic1111 webui or ComfyUI or some thing else?

[solemn cairn]

yes automatic1111

the loras doesnt appears

[gloomy scroll]

1. So you are trying to paste lora name to prompt input area
2. You opened lora tab and can't find your lora there
3. You copied lora file and trying to paste that into prompt area

is any of those correct?

[solemn cairn]

no , do you know , to use the lora on stable difusion you have to put the lora on a paste call "loras" , then you turn on the stable difusion and the Loras should be there , but only 1 lora "add details " appears

[native halo]

are you on colab?

[solemn cairn]

yes

[gloomy scroll]

You mean this is all you see?

[solemn cairn]

its other lora

but yes

[native halo]

sorry this is not the colab server

[gloomy scroll]

Don't rly know how colab works but try that button

[solemn cairn]

i i already try so it never happen to you

[obtuse trellis]

I use local idk colab since they made that restriction for free accounts

[gloomy scroll]

No since I don't use colab

[vital raptor]

sometimes they don't show up, just type it in like <lora:the_filename:1>

[nova remnant]

@native halo she has personality

[native halo]

https://tenor.com/view/fnaf-gif-25599710

[nova remnant]

[vital raptor]

i need an extension that adds a button just to toggle a group of settings on/off, i'm lazy 😄

[cyan burrow]

meh 4070 Super is just 12gb

wonder what 4070 Ti Super will cost

[vital raptor]

likely more than it's worth 😛

[nova remnant]

sounds very weird

I don't think they will launch a 40 super series

[vital raptor]

the 4070 super launched 2 days ago

[nova remnant]

4070 ti super

[cyan burrow]

theyve already began the launch, 4070 super was on the 17th, the 24th is 4070 ti super and 31st is 4080 super

4070 ti Super will apparently be about 1000 usd

too much

[vital raptor]

i like how their page compares the 40 super cards to 30 series cards

apparently not too confident in the upgrade from the original 40 series versions

[cyan burrow]

that's cause it's the same chipset, they just added more pipes etc

[twilit portal]

bride 👰‍♀️

[dusk seal]

thought Ai images would take over the world by a year

shamefur display

[nova remnant]

ai images will come out of your pc and be president

[steady pumice]

big if true

[obtuse trellis]

Ai image will take over ur computer

[nova remnant]

AnimagineXL v3 is pretty cool for more classical anime styles

[supple raptor]

That tab is only supposed to show loras that are compatible with your currently loaded model by default. I think there's an option to make it show them whether they're compatible or not somewhere in the settings but a lora made for SD1.5 still won't work on an SDXL model, nor will an SDXL lora on a 1.5 model.

[supple raptor]

What? No. 40 series does not use the same silicon as 30 series.

[cyan burrow]

I was referring to the 4000 series vs the 4000 super series, as to why they're not comparing them

because they're almost identical under the hood, only spiced up a little bit. Something that could've been done from the start

[supple raptor]

The charts should have included both 3000 series and existing 4000 series so that they would actually be useful for people to see where the different cards land in relative performance, but ultimately 1st party performance numbers are not trustworthy anyways so they're irrelevant

[nova remnant]

me seeing a new anime model come out and having to learn new settings/prompt styling for it

[native halo]

me seeing a new anime model that cant promp anime characters

[steady grail]

just found out another bug in my workflow -.-'

a very stupid one at that

i mistakenly connected ipadapter and depth controlnet to the upscaler and made the upscaling 10 times slower XD

[nova remnant]

girl sitting on a super tall rock when

10000x512 image

[supple raptor]

10000x512?

pshh....

I don't gen anything at 512

https://cdn.discordapp.com/emojis/930965738632671283.webp?size=64&name=downey&quality=lossless

[nova remnant]

Then double it!

[nova remnant]

no feet

[supple raptor]

my starting res is 960x544 or 544x960 for almost all of my gens, then they get 2x upscaled to leave me with almost exactly 1080p images.

[nova remnant]

is this from txt2img

Models I use wouldn't do a rock this long

[supple raptor]

yep, pure txt2img the mashup of loras I'm using does some funny stuff. It's the same setup I was using for these ones #🍥｜anime message

The tall rocks weren't intentional, but when it happened I thought it looked kinda neat so I didn't fight it.

[nova remnant]

mushroom hat girl

[supple raptor]

remilia? dats a mob cap

[nova remnant]

This one is mushroom hat girl

lmao

[supple raptor]

nyo, dats Marisa Kirisame

and it's a witch hat

[nova remnant]

mushroom hat

[supple raptor]

yesh, dats mushroom hat

[nova remnant]

Kinda funny how it really tries to make a witch hat

[nova remnant]

What does our local expert think @native halo

[native halo]

licks

[nova remnant]

this is a family friendly discord

[native halo]

just lickin my ice cream

[supple raptor]

https://media.discordapp.net/attachments/718529989301108846/1170233610184441867/QQ20231104132925.gif?ex=65584bfa&is=6545d6fa&hm=d7145b4f3a56d0dbe73f34995a120ae5bacfcbf000e6543fd5c926a995753f88&

[nova remnant]

[nova remnant]

Hands behind head

In other words

Armpit

Hewwoooo sleepi

[zinc flax]

Hello there

[supple raptor]

You've said the word and summoned Reimu.

[lyric zealot]

i just realised starboard is removed?

[native halo]

its gonna be like XL,they will soon release a trimmed down version of starboard called "starboard turbo"

[lyric zealot]

[supple raptor]

every image that gets a reaction goes on it, but they have a smoothing filter applied.

[solemn cairn]

hi are you using th SD?

[supple raptor]

Yeah, pretty much all of the images I post are SD gens. Mostly txt2img, sometimes with some manual touchups and/or inpainting.

[zinc flax]

Huuuge

[supple raptor]

https://cdn.discordapp.com/emojis/930962898761695232.webp?size=64&name=tewi&quality=lossless

[solemn cairn]

me to text-image

too

[supple raptor]

beeg

smile

[zinc flax]

Huge smile

[cyan burrow]

everytime someone I don't know adds me or starts a DM with me I always expect them to drop a channel/server invite. It always happens -.-

[supple raptor]

yeah, I just delete any friend requests from people who's names I don't immediately recognize from one of my active servers. If they send me a DM I delete that too most of the time. I'll take DM without the friend request, but I have no interest in being friends with some random ass person even if they aren't running some scam (which they probably are).

[native halo]

i always keep my DMs open in case a nft millionare who is also on the brink of death is lookin for a strugglin poor hard working citizen like myself so he can donate all of his fortune to me as a last sign of goodwill.

[nova remnant]

Or some bodysuit pics

[native halo]

i prefer the ones where the bodysuit is gone

[supple raptor]

https://cdn.discordapp.com/emojis/1028509125044359259.webp?size=64&name=MomijiWide&quality=lossless

[nova remnant]

Sleepy bodysuit

Why does yawning look like something else

Am I imagining things

[native halo]

shes just hungry

[nova remnant]

I am very hungry

[keen hollow]

Cool, nice to see some different stuff. 🙂

[nova remnant]

Miku punch this burger

hello burger

@native halo SD focusing on the wrong things 👍

[native halo]

https://tenor.com/view/lick-anime-gif-26198065

[nova remnant]

[vital raptor]

is it though?

[native halo]

her hip is broken lemme fix it

https://tenor.com/view/inori-yuzuriha-guilty-crown-aim-gun-gif-12793602

[nova remnant]

[nova remnant]

I have too many broken hips gens that I can't post

[native halo]

dw i got an extended magazine ready 🤠

[keen hollow]

do ya'll have that live lcm thing set up to test prompts?

I feel like I should look into that

[vital raptor]

live lcm?

[native halo]

live?

[keen hollow]

yeah

where it continually updates live based on prompt changes

extremely fast generation... not great quality but would be good to look at prompt impact

I always get frustrated with comfy and leave but maybe I should use it for things like that 😛

[vital raptor]

based on how far off the lcm lora is normally, i wonder if a live updating thing is useful at all

[nova remnant]

There is one that updates as you prompt

I've seen a post of it

[keen hollow]

i'm messing with a 'turbo' model right now that sorta helps.. I find that I dont have a lot of prompt control/creativity with these though

[native halo]

even if theres one,the results will always be less detailed

[vital raptor]

but i guess it depends on what exactly you're trying to 'test' with your prompts, just changing the CFG and adding a few steps can drastically alter an image, using a completely different model/lora is even worse

[keen hollow]

yeah it's less detailed but it could be useful for testing prompts for various styles and such

[vital raptor]

lcm uses less steps, so it 'listens' less

if it's just fast generation you're after, lowering the steps (depending on your model) might be a similar, but more accurate, test

[native halo]

yea and even with the same ammount of steps ( 6 or 8) euler a still gives more details so just try it out with that

but i heard sdxl turbo works with 1 or 2 steps

havent tried it yet

[vital raptor]

for awhile I was using tensorRT and lcm at the same time, had my 2048x2048 gens (hires x2) generating in ~7 seconds, and while the quality was just as good, the detailing was lacking

if you're doing stuff with no backgrounds and simple characters, it's great

[native halo]

yea its great for 2d flat styles

like this

[vital raptor]

yeah

this on the other hand, not as likely

[keen hollow]

I stumbled onto those zhmix models a while back and can't decide which one I like the most.

[nova remnant]

@native halo do you know this character

[vital raptor]

ahri

[nova remnant]

https://tenor.com/view/cat-pets-kitty-pussycat-animals-gif-14851744

[keen hollow]

Today's Tifas

back to pretty straightforward images, but trying to use LCM to create... so that's what is new i guess

[nova remnant]

[native halo]

https://tenor.com/view/kda-imaginetears-gif-19424813

[nova remnant]

[gloomy scroll]

||You have been visited by dog||

[nova remnant]

very needed spoiler

[native halo]

https://tenor.com/view/inori-yuzuriha-guilty-crown-aim-gun-gif-12793602

[nova remnant]

why do that to plant girl

[native halo]

im allergic to plants

[nova remnant]

you sure

[gloomy scroll]

Here have a leek

[nova remnant]

Very tight jungle

[native halo]

[terse ore]

Is there a model that can turn images into monochrome manga lineart?

[dawn barn]

https://civitai.com/models/16014/anime-lineart-manga-like-style There is a LoRA for that

[terse ore]

Yes I tried that but the input image has color and the output also ends up having color

[dawn barn]

And you added "lineart, monochrome" to your prompt?

[terse ore]

Yes

[dawn barn]

What model?

Send over prompt while you are at it

[terse ore]

I looked up some workflows and there are some but it's in a1111 (I use comfy). It's using an image but also a plain blank white image as the img2img as well and generate lineart from that

[native halo]

for img2img u need controlnet + high denoising otherwise it wont work

[terse ore]

I am using animelineart model (which was made from that Lora) but I tried other models and same thing. Prompt is lineart, monochrome

I used controlnet but I realize it could be denoise, yes. I used .5 only. Let me try again

[dawn barn]

Why are you using a blank white page

Just use txt2img

[terse ore]

I didn't, but just that I saw other workflows so that but with img2img with the original image. I'm.nit sure how they did it

Ok I did it. Kinda worked but some things like T-shirt is very strange. I'll experiment. Thank you

[zinc flax]

Try this lora
https://huggingface.co/2vXpSwA7/iroiro-lora/blob/main/test/style-lineart.safetensors

[supple raptor]

For something like this you shouldn't be using img2img at all. Just use controlnet in txt2img.

[glacial bane]

Bing makes good stuff

Somehow

SD better

[winter fable]

help

how do i increase resolution?

images very grainy

[glacial bane]

Increase the size of the picture

512x512 is default

[winter fable]

i think i'm using 512x700

lemme check

yeah, it's 512x700

[glacial bane]

Increase it dawg

It may take more time yet it's worth the wait@winter fable

You get killer pics like this

[winter fable]

what size you using?

and i don't think i can do 1024 with 1060 lol

[terse ore]

I make renders in blender and then use the image as input to make characters consistent. To perhaps make a manga. But the plain areas usually has a lot of lines and stuff where it's supposed to be flat white.

[pure oyster]

Model? I like the hair.