🍥｜anime

[nova remnant]

ahh

[wispy canopy]

no way of getting that merged

[nova remnant]

yeay i know

[wispy canopy]

and generally, anime sdxl models are... not really well trained

as in, not leveraging the potential sdxl has

[nova remnant]

is that even possible to acheive that in sdxl or do you feel stability ai should release something else beyond sdxl 1.0?

[wispy canopy]

i've been playing a lot with pixart as of late, and while the base model is exceptionally poor at anime, it's comprehensional levels are off the chart

[nova remnant]

i hope you keep enough time for sdxl too lol

[wispy canopy]

this one done in pixart

i mean, the image quality is not great, that's something that's easy to see

[nova remnant]

nice but also its more towards painterly effect

[wispy canopy]

but the image composition -> good luck getting that on sd1.5 or sdxdl

i've tried making such images in both, but can't get that "joy of life" feeling

[nova remnant]

sd hasn't been able to achieve midjourney like oil painting

something like this impossible with sd

they can get close but not quite

[wispy canopy]

pixart is awesome tho, i love my crappy birds

[nova remnant]

the best i've seen ppl do with sd for oil painting effect is smudge some areas, but overrall image composition is not right

[wispy canopy]

and it does painterly REALLY good

and i mean, REALLY good -> this is just the base model

[nova remnant]

this looks brilliant

[wispy canopy]

yeah, the compositions the model makes are out of this world

[nova remnant]

i see it

its like top tier

cant belive thats from base model

[wispy canopy]

just grabbed a random non-character prompt from civitai, a single generation, and stuff like this just comes rolling out

[nova remnant]

i wonder how accessible that will be for general users

[wispy canopy]

if you have 20+ gb ram & 6gb vram you can run the model on comfyui

[nova remnant]

bit heavny on ram

[wispy canopy]

hey, be glad it can be offloaded!

[nova remnant]

but i should upgrade my 16gb to 32 anyway

[wispy canopy]

the OG spec had the t5 encoder in vram too

[nova remnant]

im glad its not eating up vram too much

[wispy canopy]

t5 encoder is what pixart uses to tokenize your prompt, sd1.5/sdxl use clip

[vital raptor]

I think for what it is, bpn does pretty good 🤷‍♂️

[cyan burrow]

ah tis Clara, she survived the merger

[wispy canopy]

actually, i was working on that prompt in bpn

and threw the prompt in pixart, and it... just was better

[cyan burrow]

daaaamn she aiming for the guinness record. Look at all them fingers

[nova remnant]

with sdxl there is lot less of those anomaly

[vital raptor]

i guess it depends on what you're after, i didnt try to match the style but the 'feeling' i get from the image is similar

[cyan burrow]

she took it back a bit far, overcorrecting

[nova remnant]

but be honest... her face looks so good

[cyan burrow]

I'm not gonna complain to her about her new longfingers robot hands

yes I worked awhile to get that face

consistent through all my models now

[nova remnant]

that's the thing about faces i look for in sdxl

could be even better

[cyan burrow]

this is sd 1.5

not sdxl

[nova remnant]

i know

[wispy canopy]

image quality worse, composition/color -> it's not a contest

[nova remnant]

thats the point im making

you can pretty much look at sdxl and sd1.5 faces and can tell which is which

[cyan burrow]

5 fingers again...malformed, a little large....maybe she....a ladyboy?

sdxl doesn't do it better as such, just different

there are styles SD 1.5 is better at and those SDXL surpass it by far at

[nova remnant]

sdxl has distinctly different texture , but what i dont like is the gloss

[cyan burrow]

she ready to do battle

what gloss? what do you mean?

[nova remnant]

you dont see it?

[cyan burrow]

well I just don't understand how gloss is on an entire image

I dont know what you mean

Im probably seeing it without knowing what I'm looking at

[nova remnant]

there is clear metalic or plastic feel to it but some recent models have improvved

[cyan burrow]

oh, well that's cause I put 2.5d in the prompt

[vital raptor]

I think it's a matter of expectation. the models themselves are not fundamentally different in capabilities, it's the training and merging that's been done, it's no different than changing checkpoints and expecting it to be the same. You also have to change your prompts and weights to fit the model

[cyan burrow]

time to test this merge-not-yet-merge on fallen angels

[nova remnant]

if you have tried bpn you wont see it, and some other recent models that ive tried fixed that too

mind you that's in last 1 mo or so

[cyan burrow]

My models tend to stand apart in versatility, but theyre not newbie friendly

[vital raptor]

I was struggling for a bit moving to SDXL because i had a "perfect" 1.5 model that I spent months on, but once I learned to prompt better for the models I was merging and training on, the quality keeps improving without changing the model i'm using

[wispy canopy]

@nova remnant i know what you mean, it's really noticable in the early sdxl models

[nova remnant]

im hopeful sdxl can get lot better

heck .. do you see when you can render sdxl images with 512x768 WITHOUT weird glitches that becomes accessble for low power pcs too

and the traiing mechanism can be much better to enocurage devs to get involved

[cyan burrow]

different variants of Clara

[native halo]

Best anime tbh

[cyan burrow]

I killed them all oni-san, give me headpat!

[wispy canopy]

the prompt that makes you feel a bit better about life ^^

[vital raptor]

tbh, the more i try to replicate what other people post, the more i'm impressed with this sdxl model

the prompting style is completely different from what i'm used to with 1.5 but the more i figure it out the better it gets

[cyan burrow]

is she a bit more drawn now?

even though not entirely

[nova remnant]

nice cinematic vibe

it gets boring to watch a character just stand lol

need more poses 🙂

dynamic pose adds some interesting pose

[cyan burrow]

sure but right now I'm checking to see how this merge recipe does

[nova remnant]

looking good so far

drop me a link so i can try it out

[cyan burrow]

far from done but sure. Right now my old model from April is up on civit

[nova remnant]

link?

i'll add you to follow list so i can get updates

[cyan burrow]

trying...but civit is so slow

[nova remnant]

whats your handle there?

or name a model that's up lol

[vital raptor]

dark gemini?

[cyan burrow]

https://civitai.com/user/Cryonicus https://civitai.com/models/6209/dark-gemini-v3

[nova remnant]

got it

i like the model you have up

looks good

are you about to release v4?

[cyan burrow]

no, this is another one

[nova remnant]

oh

how different are they

[cyan burrow]

that one is very versatile and can do just about anything (DG3), this one I'm only taking blocks towards building a cyberpunk/magepunk style

like a world where magic is used to power technology and they live in a cyberpunkish dystopia

[nova remnant]

ok cool, gemini is what i will get since its got more broad use case

not a big fan of cyberpunk style

[cyan burrow]

sure, it has some learning curve but people who reviewed it did just fine with it so

I guess it's good

[nova remnant]

arent you going to prune that model

[cyan burrow]

aww she almost had it

there's no point to prune fp16

it ends up with same size

its on the file menu on the side

[nova remnant]

its 5+gb

[cyan burrow]

then wrong one I said the file list on the side

[nova remnant]

isn't this the right link https://civitai.com/models/6209/dark-gemini-v3

[cyan burrow]

is she ok?

yeah it is.....

[vital raptor]

she doesnt look ok

[cyan burrow]

I told you, the file menu

there's an fp16 file there

only way there isn't, is if you're missing it

her resulting trauma has made her put her hair in a suspicious way

[nova remnant]

oh yeah i see it now

but even the smallest size is 3+ gb lol

anyway i'll try it out

[cyan burrow]

yeah, I can't get it any smaller

though I havent tried since April but Im not about to try

[nova remnant]

your model cover images look good

[cyan burrow]

thanks

my problem is that I don't feel there is much I can do to improve the model further, all I could do are LORAs but I just get annoyed trying to make it work on 8gb

so meh

brb

[cyan burrow]

gonna configure quest 3 in a bit

I got them to give me 20 dollar off on the quest 3 and 70 dollar off on the elite strap

almost getting it for free 😛 the strap

otherwise I wouldnt have bought it

[cyan burrow]

saving early draft merge of 0.51

XCell_Obfuscate 0.51 done

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1183443555667169341/00239-80694551.png?ex=65885ab2&is=6575e5b2&hm=967338b2ffc5897a992a011200e7ef21ace2089d93a3928e8f438dc8a94ca011&=&format=webp&quality=lossless&width=441&height=662

[bitter latch]

Lilo of the Von Sweets 😄

Hi-res effects

goes back to bed... thud

[nova remnant]

sad that you spoilered that sweet girl

just cause she is beautiful

[cyan burrow]

https://store.steampowered.com/app/479010/Kodon/ cool, vr sculpting

[nova remnant]

https://civitai.com/models/226443 pixart fine tuned model, pretty sick imagery

its 1.15gb

@vital raptor you were able to run pixart on a1111 before?

[cyan burrow]

Heh yeah....quest 3 is a ultra-drain on battery

gonna have to get them backups

[cyan burrow]

good god I created something rather disturbing

wish I could show it

[cyan burrow]

it escalated

[nova remnant]

BPN (sdxl)

[vital raptor]

I have not tried Pixart yet

[nova remnant]

someone put a fine tuned model up on civitai

and its only 1.15gb

and i checked the hugging fac page, but looks like they dont have vae for it yet

to run on a1111

https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/sd-vae-ft-ema

but soon to come i hope

[supple raptor]

No special tricks on that one just a fairly simple prompt. The model is Soushiki which you can find on Civitai. You can put that image into PNG info to get my prompt and all of my generation parameters.

[gloomy scroll]

Excuse me, where did I arrive at?

[nova remnant]

welcome to the other side of the black hole!

[gloomy scroll]

Well time to open another portal to home

[vital raptor]

turn back! the environment isn't friendly

[nova remnant]

been searching for you all over the universe

[untold glacier]

Looks like a Class Y to me, very unsuitable for life of any kind. 😄

[untold glacier]

Top left is 👌🏻
It's been a while since I've added an image to my hall of fame folder

[nova remnant]

thanks 🙂

[untold glacier]

If you have a civitai account, you should feature that one in your image showcase

[nova remnant]

sure, i will upload the entire set

[nova remnant]

needs more goggles to innovate

[nova remnant]

did you know they are working on age reversal technique that is getting closer to reality? by 2030 this will have likely chance of going mainstream

[nova remnant]

@untold glacier this model looks good from the author of bluepencil and more anime focused https://civitai.com/models/221379/kawaiipencil-xl?modelVersionId=249687

[untold glacier]

But this is the cover art?? Lol.

I'm getting

vibes

[nova remnant]

yes

will take me 10 mins to download it, gonna try it out

[vital raptor]

blue pencil already makes cute girls, what's the difference?

[nova remnant]

with metadata

this one is more tuned for anime i think

blue pencil is bit more versatile

[untold glacier]

Hmm. I'll have to try it out. I need to test the latest Hiwamari and Anime Illust anyway

[nova remnant]

i tried the anime illustration one if i have it right, the tone was bit washed out unless im confusing it

hmm i like this kawaii pencil

very good form and image tone

btw im generating images at 512x768 then upscaling by 2x with hires fix.. and yes thats with SDXL

i wonder why some of the authors are releasing lcm specific model

you can pretty much run lcm lora with every models

at least that's what i do

[terse ore]

Is there a way to use prompts to pose characters in SDXL? I can just do basic stuff but not drastic side poses for example.

[nova remnant]

describe it in spoken language

[native halo]

some poses require controlnet

[terse ore]

Maybe I need to be more detailed is what I want out of it

[nova remnant]

detail lora for sdxl https://civitai.com/models/122359/detail-tweaker-xl

[bitter latch]

not what I was going for but.. ok

[bitter latch]

hmm guess since it's a anime model it's wanting to put in a anime character

like this

[nova remnant]

@vital raptor you may want to take a look at https://civitai.com/models/229002/icbinp-xl?modelVersionId=258447

[untold glacier]

Yeah, the naming is a problem. The gap between what it's purported to do what what it actually does is quite large.

[terse ore]

There are a ton of SDXL controlnet models on the download page. How do I know what to get?

[nova remnant]

each models are specific to what they do

have a look at what those models do https://stable-diffusion-art.com/controlnet/

[untold glacier]

I think I have new favorites when it comes to creativity and pristine quality when it comes to SDXL anime models. This is good stuff.

Himawarimix V2. Fantastic quality of gens. Very sharp and precise.

KawaiiPencil. Really great aesthetic and creativity with this one. Great at following the prompt too.

[nova remnant]

btw this TI makes a decent amount of difference for anime https://civitai.com/models/119032?modelVersionId=245812

[untold glacier]

Yeah, the relatively new hk1 is fantastic.

A needed part of everyone's workflow.

[nova remnant]

yeah i like it, been using it quite a lot

[vital raptor]

I try not to use any negatives

[untold glacier]

For proper anime style, it's often needed with many models to get that extra mile.

[nova remnant]

dont think use of negative prompts with SDXL is completely meant to be excluded, just that they vary a lot from sd1.5 practice

[vital raptor]

it's not required, but I prefer to not use them if I don't need too

the more I can get a model to do without the negatives the better

[nova remnant]

you can have a visible diffence from using them

something positive prompting alone cant achieve

[vital raptor]

yeah, but the same can be said about just changing your prompts, everything changed has an affect

and if you don't need it, then why bother

[nova remnant]

positive and negative have different influence on the output

[untold glacier]

This is a good way of thinking for prompting in general. The more you can do with the least amount of tokens, the less potential there is for something to go wrong.

[nova remnant]

you do need it depending what you want to settle with

[vital raptor]

well then I guess I haven't needed it because i'm satisfied with my results 🤷‍♂️

[nova remnant]

yeah its taste dependant not meant to be generally excluded

[vital raptor]

I have tried them, I just found I didn't need them

[nova remnant]

i have tried them too they are nice

with distinct visible difference 🙂

you can run a plot xyz on them to see the changes

[iron star]

its just a negative ti it cant change style specially in XL

[vital raptor]

was just doing that actually

[untold glacier]

This is incorrect.

[nova remnant]

not just style, but the composition of the image

[iron star]

oh yea ? post 2 images with different style here

[nova remnant]

you can test it locally to see the changes

[untold glacier]

Textural inversions are just another form of network extension. Just because they aren't as powerful as something like a hypernetwork or a LoRA, doesn't mean they don't affect the output.

[iron star]

sorry i dont use XL i dont like downloading 6gb checkpoints just to change style like in early 1.5 days when you had 4gb ckpts each with a different artist

[nova remnant]

sdxl models are far more versatile than sd1.5 btw

[iron star]

not with style i like fuzichoco , ramdayo styles

[nova remnant]

yeah thats the thing, they are not strictly narrowed down to a set style

unless you train them in special cases

[native halo]

😔

[untold glacier]

Versatility doesn't equal style reproduction, lol. No one said it did.

[vital raptor]

super basic prompt, feel free to suggest a different one and i can run it again

far right image is no negatives

[iron star]

i dont care about versatility either i just like different styles without changing 6gb checkpoints every time

[nova remnant]

you really dont see the subtle difference in them?

[vital raptor]

@iron star I don't change checkpoints

[iron star]

thats cool

[vital raptor]

I didn't say I can't see a difference, I said I don't care about them

[nova remnant]

well some do care about them 🙂

like i said its more of a prefernce

[vital raptor]

what is "wrong" with the far right image that needs correcting that these negatives corrected?

exactly, it's just taste

[iron star]

i also dont like anime checkpoints that cant prompt anime characters that i like,since theres no loras for XL i really dont use it

[nova remnant]

doesnt matter its your choice not to use them and some prefer to use them, there is obvious difference that matters to some

what im saying is negatives are not meant to be generally excluded

[untold glacier]

Allow me to help you. Base gen, and then with sky. Take a look at the differences between the two.

[vital raptor]

all i said was that i try not to use negatives, i didnt say everyone else had too

[untold glacier]

There's a clear difference in end quality.

[native halo]

i like the one with teeth

[nova remnant]

yeah same thing i was saying..

[untold glacier]

There's less blur and more detail in that one. (Edit: Blur not blue)

But granted, this isn't showing a style change.

[nova remnant]

you can see the difference easily with a portrait render

[native halo]

since we talkin about styles take this chinese style miku 💅

[nova remnant]

but i have to agree with sdxl its not necessary to use rigorous negative prompts like 1.5

[vital raptor]

the real problem is when the TI starts restricting or preventing something you do want in the image and you fight with it for hours before realizing it was the embedding

[untold glacier]

That's user error though, not a fault of the network extension.

[vital raptor]

and I'm not saying it does that exactly, but I just have a problem using something that i don't know what is in it

[nova remnant]

yeah, that applies to sd1.5 models as well and that's important

[vital raptor]

if I'm trying to generate something specific, I like to know what's affecting it

maybe i'll add one after to try and improve it, but i don't use it by default

[nova remnant]

[native halo]

💅

[nova remnant]

Where is the famous Chinese tight bodysuit

[native halo]

soon

[nova remnant]

sure that's a good practice to have good control over the intended image, but aside that when you know what you want to achieve or prefer with a partcular TI then its fine to use it, and they do make a difference in a good way knowing you apply a ti properly

[untold glacier]

Indeed. It is hard to see how what you're changing affects the end result. That's the biggest downside to AI image generation. Things like CFG, prompting, network extensions, steps, and samplers are all amorphous ideas until one spends time with them and tests them thoroughly. Only then can one understand how they affect the end result.

It's like seeing storms and understanding that what makes them is a combination of pressure, wind, water, and heat/cold, and how they all interact. It's quite difficult, and very easy to get wrong.

[nova remnant]

btw @vital raptor im not saying using bunch of TI w/o knowing what they do is a good practice incase you got me wrong

[vital raptor]

yeah, that's why I try to replicate some things people post in here, it gives me something to try and achieve with my model and my oiwn prompting to see if I can make something similar. Practice with the tools is worth it

[vital raptor]

I think we were just arguing the same thing, use them if you want them, they can help certain things, but they're not required

[untold glacier]

Yep. This is one of the best ways to learn the ins and outs, replicating images by prompting/changing settings. Taking apart others' prompts. Testing and verification--always with more than one seed and usually at least 4, etc.

[nova remnant]

the wording "they are not required" is not justifying the use case

[vital raptor]

that said, I have seen models that almost require a negative TI almost always just to get a good render, but I consider that a bad model

[nova remnant]

you might wanna claim whole SD is a bad project on itself

just for allowing the use of negative prompts

[untold glacier]

This

[vital raptor]

except that I'm not saying it's bad to use negative prompts

[nova remnant]

they are required as necessary

but you can choose to use them or not

[untold glacier]

We're going in circles here torqx. I think we all understand that.

[vital raptor]

yes

[native halo]

idk what happened here tbh

[nova remnant]

the thing i was trying to rectify is that you first claimed they make no visible diference as well as claiming they are plain not required 🙂

[vital raptor]

I never claimed they make no visible difference, that's where you're misunderstanding something

Also, they are NOT required. Python is REQUIRED, negatives are not

[nova remnant]

...........

[vital raptor]

what?

that doesn't saying anything about the visible differences and I posted an image that shows the differences

I don't see the problem

[nova remnant]

then why say negative are not required lol

[vital raptor]

What I said was that I don't need to change my images

because they're NOT required

they're optional and up to the person

[nova remnant]

how do you define require and not required... something by choice or something that makes applicable differnce ?

ok thats the thing i was saying too, taste dependent

[vital raptor]

i define it with the dictionary

essential, compulsory, not optional

[nova remnant]

when you say not required that implies that has no effect

[vital raptor]

no, it implies I can make the choice

[nova remnant]

so call it preferential

[vital raptor]

because the effect may not be what i want

[nova remnant]

because it is required to those who want it

[vital raptor]

semantics

[untold glacier]

Goodness torqx. Is this really worth the argument? Let's make some waifus, and leave the bickering alone, shall we?

I saw your uploads in Civitai

Nice stuff

[nova remnant]

lol im not arguing just clearing the confusion

[untold glacier]

Love the softness to the images.

Are you prompting something special for that or is it just the model?

[nova remnant]

just regular prompts ... minimal negative too lol

[untold glacier]

What's your CFG? Is it pretty low?

[nova remnant]

the last few images were experiemental with bit anatomical negative but they werent necessary with sdxl

yeah 2

[untold glacier]

Ah, that explains it. Great cohesion for that low of a CFG though.

[nova remnant]

generating all of the images with LCM LoRA

[untold glacier]

I take it back, lol

That's a regular CFG then, haha

[native halo]

take em all

[untold glacier]

I guess the model is a bit light and soft overall.

[nova remnant]

most of them have cfg 2 except those that i posted before lcm thingy came out

[untold glacier]

This one?

[nova remnant]

definitely 2

i think ... lol pretty sure i was using LCM

[nova remnant]

yep

[untold glacier]

Ah. Didn't realize it was there. I'm viewing the image on civitai in a squished window, so it thinks that it's on mobile--not showing me the deets by default. Cool beans.

[nova remnant]

lol yeah im on pc so might have been different gui on phone

[untold glacier]

Really wish I could get LCM to work with anything I try, really.

[nova remnant]

i've been using lcm with every models

they have decent outputs

[untold glacier]

Yeah. You and Eface using it up a storm and things looking great, but I've tried it many times with many models and even SD implementations (Comfy and Auto) and it never works like I want it to.

[vital raptor]

same, I haven't really had any models it doesnt work with

[nova remnant]

exactly

[vital raptor]

some models don't looks as good, some look better, but they all seem to function

[untold glacier]

https://tenor.com/view/kanokari-anime-fall-kazuya-kinoshita-gif-26402084

[nova remnant]

you should be able to get it working, there are two versions sd1.5 and sdxl .. and you should use LCM sampler with it, the lora weight also matters inbetween 0.6-0.8 should be ok

[untold glacier]

I define function as a good end result, heh. In this instance, I never got that.

[vital raptor]

fair enough

[untold glacier]

Not 1?

[nova remnant]

1 could be bit high

[untold glacier]

Aaarrgh

[nova remnant]

and if you were using other samplers instead of LCM then 1 is fried

[untold glacier]

I thought that was what it was supposed to be at

[untold glacier]

Nah, I was using LCM sampler.

[nova remnant]

ahh ok

so you used LCM sampler, sampling steps at 8, cfg 2 and you didnt get decent output?

[untold glacier]

Yes, but if the LoRA wasn't supposed to be at a 1 weight, then that was probably my issue.

[nova remnant]

try with 0.6 weight

[untold glacier]

First I gotta get this gen right

[nova remnant]

oooh epic

[vital raptor]

1 weight works with my model, just a little more contrast, most of mine are 0.8 this one is 0.9

[nova remnant]

you got this uncanny cool touch with most of your images

[vital raptor]

it's like anything else, tweaking to fit your model and wants

[untold glacier]

Lol. Thanks?

[nova remnant]

lol yeah didnt want to flatter you too much but i saw your images

[untold glacier]

I hear ya, but that only works if you understand what you're working with. LCM is still new to me.

[untold glacier]

Oh the civitai ones?

[nova remnant]

yeah

[untold glacier]

A lot of those are rather simple prompts, surprisingly. Some just testing out artist name recognition with BPN, haha.

[nova remnant]

i was wondering to myself how you came up with the concepts cause those are not directly relatable with text prompts

[vital raptor]

BPN is a good model to test LCM with btw, it handles it really well

[untold glacier]

It's the cohesion. It's top of the line, model-wise. That's what LCM lacks the most.

I've been told by Eface that the best way to use it is with FreeU, to get what you lose with LCM back.

[nova remnant]

btw @untold glacier LCM possibly adds a bit more cohesion to the render

[untold glacier]

I find this hard to believe, based on what I've heard of the testing of others, and what I've seen.

[native halo]

this true without freeu i get 💩 results

[nova remnant]

i can't give technical evidence but i was surprised to see how a lot of the images on sd1.5 came out surprisingly well after using LCM LoRA

i see many of devs are recently uploading versions of their models with LCM implemented

i find that strange cause ive run almost all of the sd 1.5 models in my collection with lcm lora with great results

now only running sdxl

[nova remnant]

take a look at dreamshaper, they have a model up just for lcm

and some other devs are doing that too

[vital raptor]

blue pencil did it too

[nova remnant]

im sure they may have technical reasons but as a user all i had to do was apply the lora

[vital raptor]

I think it's a performance thing, at least for the first gen it's faster if it doesn't have to load a lora

[nova remnant]

could be

but even then for initial speed boost thats kinda redundant to download a 6 gig file

[vital raptor]

I've tried to merge it with my model just for less hassle, but my model won't merge with some things for some reason

but yeah, downloading another full model doesn't make sense

[nova remnant]

yeah cause lcm lora is lot smaller in size

[untold glacier]

Not at all what I asked for but I like it.

[nova remnant]

came out beautiful

[untold glacier]

No fat people allowed on that plane. You see the size of that seat?

[nova remnant]

lol im more absorbed by the scenic view

and her sitting there looks very high tech and simple

[untold glacier]

FINALLY

[nova remnant]

she is sitting outside on the wing?

[untold glacier]

Issue with the non straight wing, but still, it's what I wanted.

[vital raptor]

"There's something on the wing!"

[untold glacier]

https://tenor.com/view/william-shatner-twilight-zone-gif-5445543

[nova remnant]

this one is pleasing to the eyes

and mind

[untold glacier]

IKR? I'd fly on that plane in an instant

[nova remnant]

but watching her sitting on the wing triggers anxiety

lol you should work on the first image

altho its perfect as it is

how the heck did you even prompt it, are you using controlnet? 🙂

[untold glacier]

heheheheheheh

no

It took 30 minutes to prompt correctly.

[nova remnant]

wow damn

[untold glacier]

These days, I usually don't spend so much time on one thing

POW being a thing of the past and all...

[nova remnant]

well this is impressive

[untold glacier]

Thanks

ʏᴏᴜ ᴄᴀɴ'ᴛ ᴘᴜᴛ ʏᴏᴜʀsᴇʟғ ɪɴ ʜᴏʀɴʏ ᴊᴀɪʟ

[untold glacier]

But that's the plane she takes to school every morning.

[nova remnant]

my adetailer extension is gltiching out

[untold glacier]

https://tenor.com/view/natsuki-subaru-funny-re-zero-subaru-cold-gif-6609270426476203740

FINALLY GOT LCM WORKING! KOHYA DEEPSHRINK AS WELL!!

FEAST THINE EYES ON THIS 3456x3456 IMAGE

[steady grail]

i remember that film lol

🤣

[untold glacier]

https://tenor.com/view/float-the-twilight-zone-stars-sparkle-gif-17642561

[steady grail]

yep, that was it

the sabotaging gremlin, very cool

[untold glacier]

They aren't broken for me? 🤷🏻‍♂️

Dunno. Have you tried restarting Comfy?

Now if only I had access to multidiffusion, these tiling artifacts would be a thing of the past

Concatenated:
flat colors. a stunning cinematic 2d artwork. high contrast. black and white. hazy. chromatic aberration. surreal. dark and shadow
simple anime artwork. top down view of a girl with a beautiful face lying on her back with her arms spread on a cold stone floor. view from directly above. (album art)
Neg:
(photo. photograph:1.3). 3d. deformed. bad quality. amateur. clone. repeating

But it's a multi step process: the later upscale prompt is changed from "simple" to "detailed"

I also add the neg embed in at the upscale.

Wait...

That's after I changed it. No "even bangs".

Don't use a VAE? Lol

Try super low CFGs.

<2

"Edging"???

Less VRAM usage.

[vital raptor]

because your computer is a potato? 😄

[untold glacier]

I use it myself, and turn off the hardware acceleration to ensure it doesn't try and use my GPU. If you aren't doing that, you should.

[vital raptor]

I used edge once... to download chrome

[untold glacier]

In fact, for any program that you have open that could use it, like Spotify, you should turn off hardware acceleration.

[vital raptor]

discord can use a bit of vram as well, depending on what channels you have showing

[untold glacier]

https://tenor.com/view/albedo-albedo-overlord-albedo-smug-smug-smug-anime-gif-23800405

[untold glacier]

Oh yeah, definitely that one. That's a big one.

Enjoy this 4096x4096 image.

[vital raptor]

chatting in this channel while watching a youtube video at 1440p resolution uses about ~1.5gb of vram for me

[untold glacier]

That's horrifying.

[vital raptor]

but I don't usually use more than about 20-22GB of vram for SD, so it doesn't bother me

I also spent an absurd amount of money on my computer

yeah

got it a couple months ago to replace my 3080TI

1024x1024 using hires fix x2 and LCM is about ~8-9 seconds

1024x1024 on sdxl with no upscaling or lcm is about 5 seconds, 2.2s with LCM

when I have it working I can use TensorRT and LCM at the same time which makes a base 1024x1024 gen about 1-1.5s

TensorRT should work on a 3070

[nova remnant]

a

[untold glacier]

But that's a pain to set up

[vital raptor]

yeah, it's a pain

[nova remnant]

i just use xformers

[vital raptor]

yes

[untold glacier]

It does. Not gonna be a ton for 3070, but you'll see an increase for sure.

[vital raptor]

I get almost twice the speed

[untold glacier]

That's because you have a 4090 lol

[vital raptor]

but i have the vram for that

[nova remnant]

with wut

[vital raptor]

I don't think a lower card could do TRT and LCM at the same time, it takes a lot of vram

[nova remnant]

ohh

[untold glacier]

I don't know. I've only ever used it on Auto

[nova remnant]

tensorrt can make it faster but more vram?

[untold glacier]

From memory, it's not too much--it's just an added unet on top of everything else, which is under 400 mb I believe.

[nova remnant]

cus when i use high res 4x it still on like a 512 base i dont even use half my vram

so im willing to trade vram for speed

rtx 6000 ada

i use xformers and pretty unscientific

but ive seen 20 it/s

[vital raptor]

you should be getting well over 35it/s with that card

[nova remnant]

yea i dont do much sd

more llm stuff

so im pretty sure i dont rly know how to optimize sd well

[vital raptor]

but yeah if you want more speed, look into LCM and maybe TensorRT

[untold glacier]

Aight, I'm heading out for the night. G'night folks.

[vital raptor]

LCM is the easiest

[untold glacier]

https://tenor.com/view/sleep-anime-love-gif-22569103

[vital raptor]

yeah, i'm off to work, later

[nova remnant]

what about the x formers?

[vital raptor]

xformers works with LCM

but TRT replaces xformers

[nova remnant]

large language model

no its my personal rig

really good

like a bit less than 4090 but much better power consumption

cus the slower memory, less core voltage but more cuda cores

yes but games dont care about that

nah

not for a single card

7950x

with 96gb ram

uwu

thats not a sd workload

maybe sdxl with 4x upscale something dumb

but ive been happy with sd 1.5 models

with a heavy upscale

seed hunting is quite quick

networking job

and not expensive girlfriend

tech jobs that require a degree yes

yes but tech and degree

and throw in some certs

anyways i will sleep, its monday now

i will look into the lcm and trt and tensorrt

wow get some sleep, im in ca so its only 12

[eager sedge]

hi

[steady grail]

i can feel you 😦

i meditate exactly for that reason

[bitter latch]

sleeping thyme

sleeping thyme 2

sleeping thyme 3?

[west atlas]

you can
but you're gonna need an anime sdxl version... the stock sdxl makes these way too rarely

[bitter latch]

😄

yep i'm being random artsey

artsy?

it's that thyme again!

ha ha , all my saved styles with no extra prompt

hmmm something wrong with that leg?

[west atlas]

vampire cats v human rabbits...

man, chat gpt prompting is weird

[bitter latch]

I see you and raise

better yet

cat herding?

was suppose to be... Battle between Vampire cats and Human bunny monsters in the heavens

but noooo

gives up

[west atlas]

best is to do it in stages

make one, then img2img, and crop the original into the picture back up and redo again and again

or if you're using a1111 then img2img in an area far away from the 1st successful catfighter

[bitter latch]

meh to complicated

ok not really but I should be in bed

thuds

[nova remnant]

I went to sleep

unzips

[native halo]

feelin kinda cozy rn

[nova remnant]

@lyric zealot get one of these too

[lyric zealot]

surely i can get that

[steady grail]

@gritty spruce @wooden coral @dull jackal scammer there

[gritty spruce]

thanks!

[west atlas]

f ya wanna do something funny then add semi transparent balls/swimming aids to the pictures

doesn't matter if it is a dragon/cat/mouse etc... just make them wear it

[vital raptor]

@gritty spruce @wooden coral

[gritty spruce]

thank you

[nova remnant]

[bitter latch]

one punch miku

haha

[bitter latch]

well.. she said don't make her angry but no one listened...

hehe

[vital raptor]

lol

[bitter latch]

and can't post the next one cause well that shirt went umm.. gone

ok just got the hulk. hmmmm ok

[vital raptor]

😄

[bitter latch]

with a chibi lora

[vital raptor]

pink fox girl miku 😄

another one @gritty spruce

also @wooden coral or @finite herald or whoever is around 😄

[steady grail]

or @dull jackal ?

hello

[vital raptor]

lol

[wooden coral]

btw you can just report the messages, no need to ping us all each time

[vital raptor]

i did that too, wasnt sure how it worked 😛

now i know

[wooden coral]

you can react with ⚠️ to the message or right click->apps-> report to staff

[steady grail]

oh, didn't know worked like this

[bitter latch]

wasn't me this time

ponders.. bed or starfield?

[vital raptor]

not what i was going for, but ok..

🤷‍♂️

[bitter latch]

t-rex arms -rawr

[vital raptor]

lol, this one is cute

[untold glacier]

How strange. The "Add Detail" LoRA doesn't appear to be on Civitai anymore. Dunno why. I'll upload mine to a google drive and get you a download link.

Wait, it was just a different name than I remembered. Here you go: https://civitai.com/models/122359/detail-tweaker-xl

It's not a huge effect though, but it does add some detail.

If you're referring to the link that I sent above, that's the only one I've ever used. I don't have any further info on others.

[cyan burrow]

hmmm anyone know a good cheap joystick for space combat games and such?

Elite Dangerous, Star Wars Squadrons, No Mans Sky

etc

for use in vr

I need to make a bigger play area for VR stuff too, there isn't much room in here...I can't take half a step without reaching the end zone

def gonna get an apartnent with 3 rooms next and dedicate one to hobby room and create a 3x3m sandbox. Maybe take some kind of track and attach it to something. Or no maybe just get a KatVR if I get that invested

[untold glacier]

@lyric zealot You're now being used as part of Discord advertising lol

[lyric zealot]

they should pay me for that

[untold glacier]

They really should lol

https://tenor.com/view/anime-money-gif-11773610

[cyan burrow]

when did EA stop saying "EA. IT'S IN THE GAME" or any other variation?

[stable crane]

many years ago

I miss it

[untold glacier]

Yep

Apple? Permitting sideloading? That'd be a cold day in hell

Well good luck to ya then

[nova remnant]

[untold glacier]

https://tenor.com/view/kuroo-tetsurou-haikyuu-anime-laugh-lol-gif-17746244

Sorry... but I got a chuckle outta that

Apple really loves their walled environments

Yeah

[keen hollow]

evenin'

wow people are here?

waves

[native halo]

ok

[terse ore]

I was reading on the stable diffusion ai site about what each SDXL controlnet does, and it says he couldn't get the openpose one to work. But I saw in this chat that people use it. So, which openpose is best for SDXL? I have a 3060

[lyric zealot]

[keen hollow]

well it's quiet again 😛

[dull jackal]

Was there a p i n g?

[native halo]

/ping

[bitter latch]

-sneaks over and saves-

darn no info? what? prompt please!

https://tenor.com/2TXT.gif

tries to figure it out

[keen hollow]

oh come on, you know i cheat at this...

i did post some actual SD images the other day...

gotta keep people guessin' 😛

These are Dalle3

[native halo]

where are the Dalle2 imgs tho

[keen hollow]

hmm lemme see...

[bitter latch]

punk rocker wearing black top , skirt and silk hosiery, long black-pink hair , sharp focus, epic detail, elegantly drawn with colored pencils , art by Stan Lee

[keen hollow]

there ya go

[bitter latch]

thats one way to come off the page

took out the artist, added pov chest, and replaced pencil with pen

[magic wave]

Tomboy Supremacy

[native halo]

needs more muscles

[magic wave]

Its girly tomboy

[bitter latch]

ok.. .replaced rocker with nerd so it's punk nerd, and for some reason it comes out ..ummm.. I can't post these

😄

[keen hollow]

try "colored pencil drawing"

[bitter latch]

no more punk

[native halo]

cry about it 😢

[steady grail]

wow

what model is this?

[keen hollow]

😐 the dalle model 😐

[steady grail]

oh

[keen hollow]

give it a try in SD though! I like trying to get as close as I can.

[untold glacier]

Love these

[native halo]

she can kick me anytime

[cyan burrow]

she innocent, didnt do a thing,

[steady grail]

dale is a fantastic model. too bad can only be used online, and so, the mixes are limited

[cyan burrow]

Dale?

[steady grail]

dall-e

[cyan burrow]

oh

yeah dall-e is alright, dall-e's strength is responsiveness to prompts

but it's not so good at making fantastic art, like breathtaking art

[steady grail]

i would love to model a comfy workflow around it

[cyan burrow]

god damn it, my gens are semi-nsfw, dunno if I can even show em

[steady grail]

ok, i'm melting down

see ya

[cyan burrow]

wha?

[steady grail]

very tired

like a nuclear power plant in melt down lol

[random valley]

Guys

Got any new places where I can generate SD images for free?

And I mean the ones with good customization like controlnet and Loras

[vital raptor]

really liking this anime merge I made, really clean

[vital raptor]

what do you use in your prompt to get the little speech bubble emojis?

I haven't been able to get my model to do that

[zinc flax]

spoken heart

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1184056686500139058/00041-2148096997.png?ex=658a95b8&is=657820b8&hm=7c1ee7235fcf7760406d6f292ccaf30a9e72fd4469ba6f07ae0eb5c4d8f93b8a&=&format=webp&quality=lossless&width=441&height=662

[vital raptor]

i guess my model can't do it, i get a heart but no speech bubble 😛

hmm heart in a speech bubble worked, but it's adding random hearts everywhere.. i think i fried the text encoder in this merge

8/8 [00:01<00:00, 5.69it/s]

SDXL is slow 😄

512x512 SD1.5 is a bit faster 30/30 [00:01<00:00, 32.99it/s]

[vital raptor]

a little faster with TRT 30/30 [00:00<00:00, 42.64it/s]

I think I can go faster with a different sampler, but it usually looks bad

lol, PLMS sampler 30/30 [00:00<00:00, 52.37it/s]

[vital raptor]

that's a lot better

[nova remnant]

would hug

and push it down more

[zinc flax]

Lewd

[bitter latch]

hhe

3d Anime?

anime 3d?

[dull jackal]

I love these

[bitter latch]

she had a long night

😄

[maiden rose]

https://cdn.discordapp.com/attachments/1100842198972497932/1184057794450702406/image.png?ex=658a96c0&is=657821c0&hm=2d963d76229874fb0b8828989d35665789aee0ed50528024e13c0e65d1cb6b3a&

all anime xl kinda sucks

[bitter latch]

yeah it's not the best

wow

[vital raptor]

What part of it sucks? Each model is different but if you want something specific you might need a custom merge or training

There’s just not very many and most are based on the same one or two base models

[native halo]

just train bro, get an h100 and train

https://tenor.com/view/enforma-sfm-mom-died-sad-bro-tip-number80-gif-15420166

[vital raptor]

A 4090 would be cheaper. But also renting a service for a few hours is cheap

[native halo]

training a checkpoint with a 4090

[vital raptor]

I did

[native halo]

😮

[vital raptor]

Takes about an hour or two depending on what you’re training

[native halo]

damn so thats why i couldnt do it with 200k images

https://tenor.com/view/drake-notebook-gif-20708336

[vital raptor]

lol

I was training a face, i used 15 images

To be fair, most information and tutorials are mostly wrong or very specific to something most people aren’t training

[native halo]

to get a good checkpoint like counterfit or neko u need a lot of images that wont do it on a 4090

[vital raptor]

You can, it’ll just take a while

[native halo]

yea like 6 months

[vital raptor]

In most cases you don’t need to train an entire model though

Something like counterfeit already knows anime, so just train enough to change the style

Or merge models. The pictures I posted earlier are from a model I merged including counterfeit and kawaiipencil

[native halo]

to change the style you would need more images unless u want it to come out in same style

[vital raptor]

A style shouldn’t take more than 30-50 images, but it depends on the style too

Although if it’s drastically different from the base model used it may take more

Back to work..

[native halo]

ill be patiently waitin for that xl anime checkpoint release

[maiden rose]

you can crowdfund the cloudpower to train your model, $300 should do it

getting the dataset and labeling it is 90% of the effort

[native halo]

how long would it take to tag if i have a varied set of 200k curated images

[chrome moss]

Does anyone know how to merge this model with the face shape of another model? This model is beautiful but the face shape is too cute, I want to merge it with a model with a more mature face shape.

https://civitai.com/models/209997/maidmix

[vital raptor]

does it not change the face if you prompt it for mature/adult or a certain age?

if not, you can try this method to extract a lora then merge it with a different model https://youtu.be/CnoXyIcXQV4?si=esxQvin-zZiSeyfF

the video is a little old, but the method is still valid

[chrome moss]

it changes very little

thank you, i'll try

[stable crane]

nice

[zinc flax]

Merged multiple style loras for that gen. Kinda love how it turned out

[supple raptor]

Try MBW with these starting values:
Character Transfer
0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,1,1,1,0,1,0,0,0,0

Expect to need to tweak the values until you get it working well. You want MaidMix as model a and your mature face model as model b.

[vital raptor]

🥰

[nova remnant]

https://blog.segmind.com/segmind-vega/

[vital raptor]

and they use a A100 for that speed

seems like pretty much the same as the existing LCM

[nova remnant]

high end scaling, but the comparison is scalable

i think i will take a break from lcm lora uses

specially after having some quality loss in images

[vital raptor]

i'm playing with some new merges, doing all my testing with LCM on so I know exactly how it looks

[nova remnant]

cool

there is image quality loss tho

with lcm

[vital raptor]

what kind of loss

[nova remnant]

quality

[vital raptor]

ok, but what do you mean by quality? is it leaving noise in the image, low resolution, lack of detail..

[nova remnant]

not as crisp

compared to using say 30 steps or more with dpm samplers

the lcm image tone misses some detail

kind of like bit washed out

in general you could say they both deliver a level of aesthetics but in comparison lcm lacks a bit

[vital raptor]

i can't even replicate my above image without LCM, it changes a lot having 30 steps and a higher CFG

but, the above image is what I was trying to get from the model, so I'm good with that

[nova remnant]

not asking you to stop using lcm lol

just saying that has some loss

and also wont handle complex prompt too well

[vital raptor]

yeah, i was just testing to see what difference it made. It's pretty drastic

[stable crane]

I think LCM kinda sucks

[nova remnant]

nah not suck

[stable crane]

kohya deepshrink is much better

[nova remnant]

its going to evolve

[stable crane]

suck -> not worth using

[nova remnant]

lcm is very fresh

you can get decent images with lcm but im making a comparison

the detail are bit restrictive with complex visuals

but hopefully that will change in future

[vital raptor]

I think it's more the prompt coherence when it comes to complexity, if the model can do complex imagery with a simple prompt, it will

[nova remnant]

it adds some coherence but the overall image tone misses some detail

but they are ok for general viewing purpose, again im making a comparison

not using any lora or detail, just basic generic prompt, the image has some depth to it

btw i love it that i can render images at 512x768 on sdxl and hires it by 2x

[vital raptor]

what's the prompt for those?

[nova remnant]

didnt save the prompt, this has the metadata

[vital raptor]

this models isn't very creative

[nova remnant]

is that with lcm lora?

[vital raptor]

yeah

[nova remnant]

familiar

[vital raptor]

but that's not why

[nova remnant]

content came out ok, but texture is bit flushed out

[vital raptor]

at least not completely, that model is supposed to look like that

[nova remnant]

with soft touch?

i was noticing the color and texture depth

[vital raptor]

same model, with LCM

it looks exaclty the way i made it to 🙂

[nova remnant]

neat

i like the anime better

i mean the 2nd

[vital raptor]

but, it's just a 'style' at this point, it's not very creative or detailed

[nova remnant]

i think the 2nd image came out more pronounced

regardless of image complexity

[vital raptor]

like, the image i posted earlier with the underwater city or whatever, that was my BPN based model, this is with my new anime model, same seed/prompt/etc

massive difference in 'filler' in the image

[nova remnant]

image looks cool but texture is missing depth

you can tell by the buildings on the forefront and back

[vital raptor]

well this model is specifically for lineart, it knows almost nothing else at this point

i've mostly been making stupid cutsie stuff with it

[nova remnant]

ahh ok, thats also relatable to lcm

[vital raptor]

because that's what it's for

[untold glacier]

This is nice. I particularly love the lack of tiling issues.

[vital raptor]

it's not tiled

rendered at 1920x800 with hires x2

sorry, read that wrong, thought you meant like tiling artifacts

[untold glacier]

You misunderstand me. When an image is generated beyond what is best for a model, you can get "tiling" or "repetition" issues where some "ideas" are repeated throughout the image to make up for the lack of data.

Such as the repeating elements in the BPN gen.

The fact that there's none with your recent image is impressive.

[vital raptor]

i hadnt even paid attention to that actually

that's the first oversized render i did with it, just to compare

[nova remnant]

i was surprised with that too before when i was rendering images at 512x2048 on sdxl, i think most sdxl can do well w/o needing deep shrink

[vital raptor]

yeah, it's really not tiling elements in the image much... interesting

[vital raptor]

even stretching it wider, one of the models i used must have had some massive source material

[nova remnant]

looks cool

[vital raptor]

stupid UI wont let me go over 2048 😄

[nova remnant]

what happens if you try to enlarge it more from Extras tab?

[vital raptor]

that just changes the resolution, it won't add more to the image

[nova remnant]

A spirited girl with fiery red hair races through a field of sunflowers. She holds a basket of sunflowers in her hand, her smile as bright as the sun itself.

ahh right

[vital raptor]

editing the ui config to change the slider maximums...

[nova remnant]

she has a big smile lol

[vital raptor]

8192x1024 is too big for my vram 😩

[vital raptor]

7280x1024 works though, now we can see the tiling from the model 😄

[nova remnant]

lol

[vital raptor]

lol, looking at your images it's like she was happy and enjoying life, then her sunflower farmer partner dumped her and she burned it all down

[nova remnant]

haha

didnt think of it that way lol just playing with random ideas

[vital raptor]

are these blocks specific to certains aspects of a model? If so is there a list somewhere describing them?

I've been trying to wrap my head around how the blocks affect things but finding details is proving difficult

[supple raptor]

There are several guides out there that give a general idea of the things that each block generally effects most prominently. The problem with those guides is that all of the blocks are interconnected and can vary quite a bit between some models. There really is no replacement for exploring each block independently to reach your goal. To do so, lock the seed and set a prompt that is pretty reliably produces good images. Generate an image on model A, then model B, then begin runtime merging test each layer at 1 with all other layers at 0 and take notes on what changed from the baseline model A image. Then you can use your notes to make an educated guess at how you should merge the models to get your desired outcome. A good rule of thumb for your final merge is that you generally want to try to avoid very sharp differences between any given block and the ones next to it. The values should transition somewhat smoothly in most cases.

[vital raptor]

I found part of my problem was that the supermerger extension is broken.. the Add difference mode doesn't work at all

probably a large part of my confusion when I was trying to figure it all out

[supple raptor]

Oh, yeah I basically never do normal weighted sum or add difference merging any more. I pretty much only use MBW.

[vital raptor]

which mode though? weighted sum?

[supple raptor]

MBW is only weighted sum AFAIK.

[vital raptor]

i mean, if not weighted sum