218 xl

[floral mulch]

I like it

[wide beacon]

But it’s not good

Artiwaifu is a much better model is you wanna play with styles replication

[floral mulch]

it looks great imo

[wide beacon]

Also I was hoping 4th Tail actually made Pony good when they said Nai3 at home

It wasn’t what they advertised at all

[floral mulch]

never lliked 4thh tail

guy said he removed score tags but it just looked llike pony without score tags

[wide beacon]

imo with all the knowledge KohakuXL has if it got trained to actually use what it learned it could be great

It’s literally the open source SD model with the best Touhou knowledge

[floral mulch]

u just take all of this

but most ppl dont care cuz to them its some old shit

I like some of the touohu songs, but what the fuck does he know about images

https://tenor.com/view/kanye-west-meme-laugh-shrug-gif-25843815

[frail zealot]

owo what's this

[floral mulch]

omo

scuffed lora

[jaunty vortex]

I'm hoping flux TIs will be possible to make on low vram gpus 😔

[floral mulch]

gonna trai this I think

[jaunty vortex]

looks cool what this

[floral mulch]

maronrice

[full lynx]

6009 messages

[plucky dove]

:)

[floral mulch]

this clandestine model is pretty good

[olive onyx]

[plucky dove]

I didnt like the baked in AI art look, but it seemed okay besides that

[wide beacon]

Are you telling me it’s all shupogaki?

[floral mulch]

https://x.com/Saki_Vtoob this vtuber cute, me want lora

[floral mulch]

almost done

[plucky dove]

[plucky dove]

https://civitai.com/models/435207?modelVersionId=782125

[heady echo]

[plucky dove]

lets see if its any better

[heady echo]

i can test after work

oh those from few weeks ago

[floral mulch]

nioec

[frail zealot]

Well the recommended parameters are at least a lot better now

25 steps Euler a

[nocturne cosmos]

feels like the same as awa1 for me

[plucky dove]

Me too

Anatomy still super jank

I think neta is better

[floral mulch]

sounds like it''s worse

[wide beacon]

he probably should rebase onto Kohaku or Animagine and remove the "by" tag

[jaunty vortex]

artiwaifu v2, pretty similar to v1

they changed the tagging though compared to v1 I'll try to adjust it to match what they said to do in the description

[jaunty vortex]

cfg pp euler a, 1.5, even this is too strong

[plucky dove]

crazy hands

i also never know what to do with his negatives

he doesnt even use the same negatives as he suggests

[jaunty vortex]

idk I just use the same negatives everywhere

this model does good clouds lol

[floral mulch]

Guess I'll use it for my cloud merge

I just need to make schizo methods work in bayesmerger then I'm taking a break and actually merging

[jaunty vortex]

it's better than v1 although still makes duds, more knowledge and styles

[floral mulch]

wtf is this lookism lookin aah

[frail zealot]

@floral mulch saw some cool 4th tail gens on the vtuber ai art thread

it looks pretty sick

[floral mulch]

which one

[frail zealot]

the one on the 4chins

[floral mulch]

the latest evr?

ver

tried ur last model

was good I think but I still prefer the ovverall look of qromeq

need to do aa good merge

[frail zealot]

0.4.5

[frail zealot]

qromeq had a big problem with double navels which is why i stopped using it primarily

[floral mulch]

don't recall it much

it had a lot of ribs and back creases

but that wasnt too bad

I found a few pony models I want to merge now

[jaunty vortex]

Yeah that looks bad but v1 was also kinda bad on this prompt

https://danbooru.donmai.us/posts/5263078?q=blackbox_(blackbox9158) the actual artist's style is better

Less 3D more anime proportion

[floral mulch]

I'm looking through my favs from 2012

it's a bunch of furry and league shit

[jaunty vortex]

Vampire knight

[floral mulch]

that manga ending was so ass

[jaunty vortex]

I never read the ending

[floral mulch]

saved

[jaunty vortex]

I got filtered by the anime was too boring ig

[floral mulch]

I watched so many boring shits my sister had at one point cuz had no internet

[jaunty vortex]

Lol I watched kirarin revolution

[floral mulch]

[plucky dove]

where plz

just latest threads?

oh i see em

[frail zealot]

I prompted a lot of abs...

[floral mulch]

I didnt cuz it still gave me

[floral mulch]

should be possible to run 200 iterations now for merges in reasonable times, not manual scoring though

[jaunty vortex]

https://civitai.com/models/564727?modelVersionId=820512 leekZ dropped

[wide beacon]

Same model but kohakuxl zeta instead of delta

[jaunty vortex]

Kohaku Epsilon Rev2 actually not delta

[wide beacon]

His typo then

[jaunty vortex]

actually not sure bc the bottom of the page says epsilon

doesn't matter tbh

[plucky dove]

https://civitai.com/models/739071?modelVersionId=826511

https://civitai.com/models/282341?modelVersionId=826474

New version of 4th tail

Minthy recommends tofu for complex sfw

Havent tested it myself

[floral mulch]

ye trying it already

[plucky dove]

Any good?

[hidden matrix]

Hmm, for some reason, 4th Tail 0.5 (left) does worse on the fuzichoco style compared to Tofu 0.5 (right). It doesn't look close at all.

tried without by and with artist: and that didn't improve things so it shouldn't be the prompting style.

[plucky dove]

Yea, it seems that some artists were lost while some artists were gained

Can you try the tofu model?

Oh wait

I didn't read ur message properly

[hidden matrix]

That's this image.

[plucky dove]

Yea tofu seems better

Because it doesnt have to conflict with pony

Maybe some weird pony fuckery is messing with 4th tail?

[hidden matrix]

Pretty good naga u style from Tofu without prompting for chibi.

Compared to NAI3LikeDiffusion v1.0, which may be closer, but often misses the flat eyes.

[plucky dove]

What negatives do you use?

edit: nvm i can check with bot

I didnt really like the one recommended

[languid cliff]

Silly mofus hab been purring here on da warm gpus

[wide beacon]

https://civitai.com/models/758322/iris-xl @floral mulch @frail zealot might interest you lol

[floral mulch]

@ruby thistle sup

[floral mulch]

hm

still havent tried to get by artist name merged into animagine type models

I want to have ziyun and stuff

[wide beacon]

Do a better Leek by merging Animagine Kohaku and Arti

[floral mulch]

I'll merge all of the good non pony models together

[wide beacon]

Here's a list then

• KohakuXL
  • KohakuXL Zeta
  • SanaeXL
  • Heart of Apple 3
• Animagine
  • UrangDiffusion
  • SwampMachine
  • Leek
  • Heart of Apple 2
  • Umbra
  • Naimagine
• ArtiWaifu
• Tofu

[full lynx]

you could try the ppa "correction" merge too

[floral mulch]

First I'm doing arti60 kohaku zeta kohaku alpha Delta and sanae maybe and my current umbra which is a weaker version of what I want to do. What's urang diffusion

[floral mulch]

Link

I can merge loras too at the same time

If there are any cool ones, I never check

[wide beacon]

no

nyo

[olive onyx]

we're cooking in here ?

[olive onyx]

Some of these don't do well for merges

Better list would be:

KohakuXL
SanaeXL
Animagine
NAI3LikeDiffusion
Umbra
Artiwaifu

[wide beacon]

he's gonna use a license that forces you to release weights if you share images made with a modified model

[wide beacon]

you have no clue what umbra is then

[olive onyx]

I mentioned umbra in my list

[wide beacon]

imi has done merges with models that are incompatible since umbra it doesn't matter if some models are harder to merge or not

[olive onyx]

he is the goat afterall

I couldn't possibly wrap my head around his merge techniques

[wide beacon]

urang is an Animagine finetune that adds a few games and is trained by a member of cagliostro (although it's not trained on the original dataset so styles are weaker)

[olive onyx]

@wide beacon best kohaku model ?

is delta yes ?

[wide beacon]

zeta

[floral mulch]

Umbra doesn't really have incompatible models in it, it's just a merge of the best animagine finetunes at the time with some layers manually scaled. I only started bleeding parts of pony and artiwaifu into it after, with limited success the models after have some unique qualities but most of them did stuff I didn't like at the same time

[floral mulch]

I just want to try out new merge options that weren't doable before and see how it goes

[plucky dove]

https://tenor.com/view/making-potion-baba-yaga-the-witch-of-the-woods-smite-mixing-gif-17218647

[floral mulch]

0 merging today again

[wide beacon]

ngl what I have in my head would be
merge sanaexl with heart of apple 3 (get sanae's knowledge with apple's aesthetic)
merge urang with swamp machine (take urang's character knowledge and swamp's aesthetic)
merge urangswamp with naimagine (add some animagine back in + aesthetic)
merge sanapple with nairangswamp (combine them to a middle ground)
merge arti with tofu (uuuh dunno)
merge artofu with abomination created above (yay)

I'm going insane

[jaunty vortex]

you should do it

[olive onyx]

yeah but like if you merge them all at default 0.5 weights for all blocks it's not going to end up looking good

[wide beacon]

that's why I'm waiting to see what imi will come up with using his cursed methods

this is just my headcanon and I assume it wouldn't be 0.5 for every block

[jaunty vortex]

just dare ties all of them

[wide beacon]

@olive onyx here you go added details on what I'm thinking

[olive onyx]

we will see if his sanity prevails or the merges

[wide beacon]

imi is already insane wdym

who else would spam Gemini to make new algo's

[jaunty vortex]

we breeding models here 😂

[wide beacon]

back to the aom days

[jaunty vortex]

why have designer dogs when you can have designer models

[olive onyx]

I feel training LoRA's for like better anatomy and stuff and merging them is a better way to do it

but I'm too lazy to that ofc

[olive onyx]

for context makers of juggernautXL do this to

[wide beacon]

you could diff extract LoRAs and then block merge them yesh

[jaunty vortex]

I don't merge or train
however I do have my funny hive beating denoiser

[wide beacon]

getting a balanced dataset to make a LoRA like that is just a task that none of us wants to tackle

Rundiffusion has a whole lot of money

[jaunty vortex]

https://x.com/nearcyan/status/1836598231415054705

near (@nearcyan) on X

DisTrO formula released

Twitter · Sep 19, 2024, 2:48 AM

maybe with this we can hivemind train

[wide beacon]

Speed 📉

[jaunty vortex]

there must be enough gpus within big cities

[wide beacon]

combining compute over the internet is a flawed idea

[olive onyx]

dude

1gb VRAM from like 500 users

that's like 500gb's of VRAM

ofc this isn't how it works tho

[jaunty vortex]

instead of crypto miners

funny image generators

[wide beacon]

Think about latency in games

200 ms is annoying

But 200 ms + having to calculate everything is just stupid

[jaunty vortex]

I wonder if 800x less latency is enough for wifi training

[wide beacon]

The main server that coordinates everything would kill itself

[jaunty vortex]

if I can combine me and my sister's gpus that's already not bad

two 3060s

[wide beacon]

50ms for everyone in the swarm would be necessary to even make it viable

would you be ok with not being able to use your card for 12 months?

[jaunty vortex]

can't I just train a flux lora

why big model

[floral mulch]

Definitely is when it comes to trying to use pony

[olive onyx]

maybe we could like find some pony LoRAs that are trained on natural language captions and not schizo and try and merge them

[floral mulch]

don't think that exists but who knows

[olive onyx]

We'll make one

https://civitai.com/models/376450/duchaiten-pony-xl-no-score?modelVersionId=817261

@floral mulch

[floral mulch]

I'm not using a ducchaiten model

just start building datasets for shit u want

better sooner than later

will be useful in the future as well

[olive onyx]

I guess I'm going to be a regular at e621 and danbooru now

[misty siren]

Can tofu model generate lineart?

[floral mulch]

probably, most models can

spent the last week messing around with samplelrs

was a good side quest but failed

or got tiring

[floral mulch]

are there any

90000 ovvercclcockced

juicerian

loras

giga herz heinstein

[plucky dove]

Anything nice or just still good ol euler a

[floral mulch]

#1279347749149605939

[wide beacon]

new model is very cool

[floral mulch]

yeah

already on the list

was fixing lora merging in bayesian merger all day

was getting trolled

[floral mulch]

nvm guess I'm still getting trolled, time to spend the entire tomorrow on it

[plucky dove]

what new model

?

@wide beacon

[olive onyx]

[plucky dove]

does the name start with "i"

cough cough

[wide beacon]

it's da angel model

[olive onyx]

@floral mulch have you tried correcting 💢 Illustrious ?

I don't think I can do it with my abysmally low knowledge of merge methods

It just gives noisy images as outputs

if you just transfer the CLIP of another model to it, it just completely generates noise

[floral mulch]

no

I'm busy with something else

I'll try it when I'm done with this aids shit

[olive onyx]

good luck with your AIDS I hope they get better

[floral mulch]

mayb finally unfucked it

didn't enable loras tho, if it doesn't work I'm prob spending tomorrow again on it

[wide beacon]

holy crap

[floral mulch]

sadly can't merge boft and other lycoris probably like this

theyre kinda quirky

informative error messages

[floral mulch]

that's some of the most cursed shit I've ever seen, thanks

I'm so fucckcing retarded

I put umbra and sdxl as normal models and a random model as the base model so I did sdxl-random model and umbra-random model should add an anti ritard check like notice ur not using sdxl-base

[floral mulch]

Wao not fuzzy

[olive onyx]

did your AIDS fix work?

[floral mulch]

Kinda, more to do

And I can't merge lycoris

Anyways this model is insane

Makes me not want to merge just from how much it levels up when you throw it dt some dogshit and ik pretty sure it's doing all the weight lifting, literally

As far as I'm aware it's usable, but it has some pedantic shit like not being able to specify a base model as the second or first model in the list cuz it will slice off the 3rd model if the method requires only 2 (most of them)

[floral mulch]

Yeah this isn't easy to handle tbh

[olive onyx]

just do like a few OUT and IN Blocks for the merge and the CLIP

as far as I've seen the one of the things the model excels at is character knowledge

besides that, there are models better than it in other aspects

would be amazing if we could just get the characters

I usually block merge IN05, IN06 and OUT00 with a strength of like 0.33 and it seems to work

[jaunty vortex]

maybe find some characters that it knows but other models don't and see what happens if the layers are replaced one by one?

[olive onyx]

I'm afraid that "unalive" might be a valid solution for my mental state trying to merge this

[jaunty vortex]

you should ask ljleb about mecha

[olive onyx]

any progress

[floral mulch]

imo it's way better at pretty much everything compared to when I try merging with other models, it's more than just the chars, and it's new type of knowledge, different than what I'm used to with animagine

like poses are different

the same errors don't happen

etc

no cuz my dogshit crashed

adding safeguards against it now

https://discord.gg/EkxQ5WQd ✅

[olive onyx]

I see so it's like completely different from like any other XL models that are available publicly ?

[floral mulch]

I haven't tried all of them

but to me it's different, I've been running the same prompts for merging for months and different poses and the errors I usually encounter weren't there

[olive onyx]

[floral mulch]

1seccc

[olive onyx]

did you get any good results from the merge models ?

I tried doing ComfyUI Dare/Ties and tinkering with supermerger but I couldn't get anything that wasn't NoiseDiffusionXL

[floral mulch]

can't show the ones where anatomy and poses are better cuz nsfw but

just to show that it's different

not neccessarily better to my taste but it's a good sign

on that note the "usual" is already to my taste aesthetically, so I'm only really looking for new interesting stuff that's not necessarily a backbone style

[olive onyx]

did merging help with gaining higher precision on character tokens the model didn't know previously ?

[floral mulch]

didn't check for that depends on what charactcers you'd need to test for yourself

[olive onyx]

yeah but I don't know how to do that without making modern noise art

my merge output

[floral mulch]

if u have accccess to xyz start with low values and raise them sequentally and see whhicch one starts introduccing the worst noise

not really something you can do manually

[olive onyx]

[floral mulch]

I think I used to do that in supermerger

I think it was also trained with a lokr method but don't quote me on that, might be just misinfo, so you could try treating it as a lora

[olive onyx]

treat the model like a lora

https://tenor.com/view/kurt-angle-stare-100-yard-stare-meme-gif-251436459491629941

ok I can do that

[wide beacon]

KohakuXL is trained using the lokr algo

afaik Illustrious is a KohakuXL finetune but not using lokr

[floral mulch]

hm

guess that's why it's a bit peculiar in ccocntext of the other models

there haven't been any big finetunes on koihakui

[wide beacon]

Sanae and Heart of Apple 3 are KohakuXL finetunes

[jaunty vortex]

Maybe we can try illus - kohaku

[floral mulch]

theyre smaller

[wide beacon]

Yeah their dataset are more curated

[hidden matrix]

Is this the naimagine you mentioned?
https://civitai.com/models/716416?modelVersionId=801166

[wide beacon]

yes

[hidden matrix]

k

[olive onyx]

any idea how I could like not have this happen ?

[wide beacon]

What did you try to do

looks like you're not loading a LoRA?

[olive onyx]

LoRA extract of Illustrious

from Kohaku Beta

made it work

so I tried merging it

but supermerger was like "nah"

[wide beacon]

When was the last supermerger commit?

[olive onyx]

5 months ago

[wide beacon]

yeah that's probably it

[jaunty vortex]

wow very cool

[olive onyx]

but I don't know any other LoRA merge extensions

[wide beacon]

there's probably some custom comfy nodes you could use to do the same thing

[olive onyx]

yeah I wanted to do that

but like

My ComfyUI backend in swarm wasn't loading up properly

btw

I feel like

a 512 or 1024 extract from Illustrious using KohakuBeta as base

would get better results from LoRA merging

but built in LoRA extractor in swarm only allows upto 320

[wide beacon]

did you update comfy and swarm?

[olive onyx]

yesh

[wide beacon]

huh

[olive onyx]

i'll try again

hol on

I believe that after merging in this LoRA the UNET won't be a bitch when trying to merge blocks so that's why I had the idea

[jaunty vortex]

what is the difference between merging in this lora extract and merging in the real deal

in terms of not getting noise...

maybe there's something that's getting filtered out 🤔

[olive onyx]

who knows

https://tenor.com/view/shrug-gif-25269231

I'll find out after I merge it in and then try merging in CLIP/UNET of illustrious

@wide beacon do you have a link for a ComfyUI LoRA merger

the backend works after updating

[wide beacon]

it's slightly less accurate

so if you extract even lower dim

[olive onyx]

nvm I found it

[wide beacon]

you could maybe

[wide beacon]

there's like that custom node that adds a custom node shop

[olive onyx]

how do I search for a node to add

disclaimer: I don't know ComfyUI

[jaunty vortex]

double click

type in

[olive onyx]

thank you

okay soo.....

there isn't a node that let's me merge a LoRA INTO a checkpoint

@floral mulch help pls

[full lynx]

lora loader ...

[olive onyx]

pls explain mr yeggles

[full lynx]

LoraLoader

the node

[olive onyx]

wut now

I'm sorry I dun know comfy

[full lynx]

pass the model into the lora loader input (left)

[olive onyx]

how

what node

[full lynx]

checkpoint loader

you should read my comfy guide

[floral mulch]

?

use sd-mecha, if u have a lora u can just do add_difference to a model I think

[olive onyx]

next

?

[full lynx]

do you want to save the model?

[olive onyx]

yesh

[full lynx]

checkpoint saver

or save checkpoint

[olive onyx]

is it ?

[full lynx]

and the vae...

[olive onyx]

okau dun

how do I run the nodes now ?

[full lynx]

queue prompt

[olive onyx]

[olive onyx]

okay thank you sm it works ❤️

[wide beacon]

nodes are easy just need a bit to learn

don't hesitate to bother yoinked some more if you need more help

[wide beacon]

I'm not lying it's just a different logic but it's easy

[jaunty vortex]

I use comfy for everything since sdxl got good

Once in a while it's a pain because I forget to connect something

[floral mulch]

I don't like it, I don't like being presented with too many options, if there are too many options I will get stuck doing everything for 5% of the time instead of doing something for 100% of the time, and it looks less appealing than traditional ui

[jaunty vortex]

I just use the same workflow 99% of the time

Swarm would be ok too

[olive onyx]

okay new idea

block merge LoRA into model

so not everything goes bye bye

[floral mulch]

Ah yeah, could be easier than block merge whole model

Too lazy to get a Lora of it tho

I'll just stick to subtract

And it's 10 am

Fucking dogshit

[olive onyx]

I can upload it to huggingface if you want

[floral mulch]

Sure, I'll try it tmrw

Haven't been able to test lora merging cuz all of my later ones are oft

[wide beacon]

Just install swarm

it's built in

[olive onyx]

anyways uhhhh

wanna merge illustrious ? turn it into a LoRA

someone's gonna have to edit the LoRA extractor code to allow for higher dim

and somehow get a result closer to the original model

@gritty lynx

also need a Comfy Node for LoRA block merging into a checkpoint

to actually be able to use it

[full lynx]

lora block merging

[floral mulch]

they can be directcly translated back to normal bloocks

[olive onyx]

supermerger allows for it but supermerger is outdated soo

It doesn't work

[gritty lynx]

what's illustrious

[hidden matrix]

So I tried a ties merge of all that in the order given.

Next attempt will be a straight up ties merge of everything.

[gritty lynx]

I think illustrious is a sdxl/pony model?
if so, the nodepack I made has a feature for this

not really super user friendly atm ig

[hidden matrix]

rip lmao

[jaunty vortex]

illustrious is finetuned off of kohaku beta

[gritty lynx]

oh okay

that explains why it's hard to merge I suppose

[jaunty vortex]

it was trained on like 20m pics, lot more than other finetunes

[gritty lynx]

yes. kxl is a full rank finetune using lokr so that doesn't help model compatibility

neither

[full lynx]

if you're gonna do a full rank lora why use lokr

[gritty lynx]

lora stands for low rank :> why would you use a lora if you're gonna do a full rank finetune

lokr is just capable of doing a full rank finetune while using less parameters than an actual finetune

[full lynx]

sorta? if model dim =< lora dim then its just a finetune

[gritty lynx]

but with lora dim = model dim you end up with like 2x the param count you would have if you just straight up finetuned without lora

[floral mulch]

https://cdn.discordapp.com/emojis/930962869401571348.webp?size=44&quality=lossless

[gritty lynx]

why not?

[floral mulch]

does things only for money so he doesn't have anything of substance to offer for me

[full lynx]

[frail zealot]

whay are you remoting in via a PHONE

[floral mulch]

cuz I was in bed

[frail zealot]

imagine if clandestineXL was trained ona better style mix instead of that blsuhyspicy one...

[floral mulch]

dunno if I just got lucky but I haven't got a run like this since then and I'm 150 iterations into a merge now lol

I mean the process is kinda working but it just got stuck at blurry images

prob shit scorers

[frail zealot]

mfers will bwm 50 checkpoints for 20 hours instead of just training 60 different loras for every possible need...

[wide beacon]

funny insect jar to create poison thingy

[floral mulch]

https://cdn.discordapp.com/attachments/1271923652694380574/1288202102354477148/convergence_plot.png?ex=66f45332&is=66f301b2&hm=d3ea6b355f6ac5b6762c3e3666aaa4ba7d117aaf633c038a435669abc3440d99&

yeah even the broken scorer works

just sucks ass

[full lynx]

scale it

[floral mulch]

it's the choice of scorers

I'm putting this in now https://github.com/city96/CityClassifiers

[floral mulch]

finally, shit scores for shit images

[floral mulch]

havent turned off my pc for like 4 days

[floral mulch]

just realized I'm retarded

the reason I've had a tough luck suddenly going fro m1 day to another merging illust is that

I changed my code

and it forced the range of the text encoder to .4-.69

so basically always forced to be a 0.5 interpolation between an already rough to mereg model

[jaunty vortex]

oof

[floral mulch]

that's why the first attempts I did were like "wtf are u talking abotu this is easy to merge

then from that on just getting fucked over and over

poor 200 iteration merge tried looking better but it just couldn't with that handicap

[floral mulch]

need to come up with a better way to define the objective

average score doesnt do shit when a shit quality image can show elements of an artstyle or character or the opposite

[jaunty vortex]

multiple scoring objectives? but then you have to weight them

[floral mulch]

yeah

I'm trying to think about whether or not multiple score types averaged are more meaningful than just overall image scores averaged

[jaunty vortex]

maybe if you weight them by squaring

or something

square root

[full lynx]

did i hear math

or

mayb even

[floral mulch]

hmm

atm I'd need at least 4 separate metrics, quality, style, character and anatomy

but I don't want to add a 4x multiplier to the optimization space

or wait

nah

it's the parameters of the merging that is used to optimize towards a higher score

not sure if that restricts how many scores there can be

[full lynx]

weighted sum + some scaling func could work well

[floral mulch]

nice

[full lynx]

given a,b,c,d are the scores and aw,bw,cw,dw are the weights, you could use something like this to calculate the score (w being a max of 100, a,b,c,d being max 10)

(this func is extremely generous btw, if a=10 and the rest are 0 the final score is 50)

[floral mulch]

hmmm

thanks

[floral mulch]

doesnt exhibit any difference after replacing the position_ids so that didn't really doanything

[floral mulch]

last 2-3 days

[floral mulch]

damn they dont like this prompt

[plucky dove]

euge looking for help

[floral mulch]

512 h100 gpus?

the fuq

@wide beacon new aid wont suck

[wide beacon]

ping Angel

[floral mulch]

dunno if ppl in orgs want to work together

[wide beacon]

@low berry interested?

could be worth talking to them

[plucky dove]

kek

[low berry]

I'd definitely want to, however now I have things to process first so time is the problem

I'll arrange my time to help 🫠

[wide beacon]

https://civitai.com/models/812489/badapplexl

😭

new Animagine finetune just dropped

[floral mulch]

hmmm

I'll give it a gander

always more stuff to add

[floral mulch]

https://civitai.com/models/820208

[floral mulch]

[floral mulch]

getting lost in the sauce

[floral mulch]

time to make incestousxl

[plucky dove]

any good?

[plucky dove]

make WHAT

[plucky dove]

i was wondering, does it erase styles?

gonna try it out

[floral mulch]

dont think so

it looked one of the best so far

[plucky dove]

from limited testing

it is pretty good

[wide beacon]

I'm gonna do coil whine again

[feral elbow]

current best anime model for nsfw? anything good as NAI yet? ive been using pony but havent been active past month

[wide beacon]

we don't do nsfw here

go ask somewhere that does

[feral elbow]

where do you suggest i do that

kind of stupid to imply that anything less than 99% of people using anime ai models dont use them primarily for nsfw

but anyway the question can just be reframed to "whats the current best anime model in general"

because unless it blocks gens then the best model overall usually does the best nsfw

for example nai

[wide beacon]

bruh

[full lynx]

pony is prob best for Sex, illust/wdv01 for sfw

[floral mulch]

arti for purely stylistic, leek/neta/clandestine~ for general (as in not sfw but everything), pony for coomer/furry adjacent, illustrious if you want to try the new model. it's new so it hasn't been fully established where it shines the best and has some rough edges still, maybe check the finetunes of it, I had some fun with the one made by counterfeit guy but just an hour or 2

[plucky dove]

Using obsession illustrious finetune, its pretty ok for nsfw, can dm if u want

[wide beacon]

https://civitai.com/models/823566/t-illunai3

[hidden matrix]

Doesn't seem to keep the original styles that well. Guided more towards that NAI style.

[wide beacon]

It’s a test with 100 images

[floral mulch]

so many bullshit model uploads that look the same and dont say what they did to the model

[plucky dove]

https://huggingface.co/PotatoBox/Momo-XL

new model?

[floral mulch]

havent tried yet

so many models coming out to merge

[plucky dove]

ill try

anything specific u look for in models

?

[floral mulch]

not really currently

I like what i have now, but I want to merge illustrious

to it

or the other way around alternatively

which I havent tried yet I guess

illustrious cant do some artstyles I use and some characters

[plucky dove]

styles are nice

anatomy is worse than illustrious imo

[plucky dove]

after some more testing

very unstable for certain prompts

[olive onyx]

Thread died?

[plucky dove]

Yes

[floral mulch]

don't feel like doing anything

busy working on dogshit code rather than merging

[hidden matrix]

Why not turn the dogshit into catshit? It's still shit, but with a higher percentage of weebshit.

Probably.

[floral mulch]

it's less dogshit now

technically it could work for sd3 or flux or other types of model merges too now (as long as ljleb implemented them in mecha )

[jaunty vortex]

I'm a 🐒, where are the merge sample images

[floral mulch]

[plucky dove]

tbh that looks pretty cool

[floral mulch]

I've been using the same prompts for months for merges, seen some interesting variations for sure

counted 600 images manually to decide which iterations to check and still ended up with this many

[wide beacon]

new version

This version is based on the early access version of NOOBAI for training, using a variety of filtered 522 images for additional art style and foot training, strengthening some of NOOBAI's art style issues and some of the flat images. Compared to v1, the art style effect of Nai3 is not as rigid because the training set is large enough for me to include more content suitable for neutral art styles. This version was trained for 200 epochs and 36 hours on a 3090 graphics card

[floral mulch]

foot training

[wide beacon]

Tianliang duhoe fandongye fan be like:

[floral mulch]

https://media.discordapp.net/attachments/1230387093793013883/1296653186894921798/image0.png?ex=671311e2&is=6711c062&hm=e0495084b2c1fdf7840ca6345cfc47c69f5573bdcb849bc2404b567917f532e5&

Brightness the merge

[nocturne cosmos]

is there any good anime finetuned flux now?

[wide beacon]

haven't heard any news around that part yet

but there's some cool thing happening in the SDXL anime sphere

so at least there's fun to be had

[nocturne cosmos]

you mean the angle's model?

[floral mulch]

yeah

and euge/lax models

the latest public early release is already better than pony at pony's own things

and it was already a good anime model

[olive onyx]

it got delivered via an inherited train

[floral mulch]

I think it merges kinda well?

better than illustrious 0.1

but

I don't really see the point of merging at the current moment

like what with?

this is the best we have

at least they helped solve the how to merge pony question

I think I'd make loras and merge those atm

[wide beacon]

new update for this

only change in the changelog is base model being noob eps 0.5

[jaunty vortex]

looks a bit like animagine or is it just me

also looks like tponynai

[wide beacon]

check the previous changelog

-# dataset

[jaunty vortex]

I tried it, he trained it to be not vpred anymore, what is the point

[wide beacon]

what

the base model is the 0.5 eps model

it was never vpred

[jaunty vortex]

oh nvm

[floral mulch]

nice, swarm has nice api documentation for every feature I need like load unload and gen images

[floral mulch]

bruh

how the fuq do II change folders on swarm

[floral mulch]

used symlinks lul

[wide beacon]

server settings

[feral elbow]

noobai based on illustrious currently my go too

still not as good as NAIv3 in my opinion though

[languid cliff]

m-mofu is here as spoon warrior!

[floral mulch]

en garde 🥄

[languid cliff]

m-mofu will hab da ice cream!

[olive onyx]

is merging dead ?

[floral mulch]

yep

this is how we merge now

mergepilled by ljleb

[olive onyx]

https://tenor.com/view/yakuza-kazuma-kiryu-pissed-off-gif-14586175

[gritty lynx]

we had a good run

[wide frost]

What the fuck

[floral mulch]

gonna see if this works compared to advokats https://civitai.com/models/922689

guy said pony loras woork

[heady echo]

looks rly clean

[floral mulch]

(they didn't work)

[heady echo]

[floral mulch]

we need to gather all the artists from old aid

most of them are dog shit on these newer models

like not even comparable

ceey, arutera, powa (36iii), zumizumi, wanke, shiun (siun 5513), funabashi (iwzry), yogisya, narue, solipsist, bebe (bebe pp), suzukasuraimu, agm, haru57928031, 99 yen (tsubura)

who else

[heady echo]

narue

[floral mulch]

right

[nocturne cosmos]

aid still has the best artist style imo

[floral mulch]

Yeah we need to recreate then but on the new models

M

[wide frost]

donate to euge so they make a new model

oh they're on noobai team nvm

[floral mulch]

fun times

[wide beacon]

soon 218xl anniversary

[floral mulch]

https://cdn.discordapp.com/emojis/967097962167697418.webp?size=44

[wide frost]

the artist combo? you can still get that style with noobai rather easy i think

[floral mulch]

not the same

weaker

will need style strengthener loras or finetune

[wide frost]

yeah shiun didnt really work

[wide frost]

trying to "copy" this style with other artist combos

ill let u know if i find one 💀

[floral mulch]

thanks

[wide frost]

unlikely tbh

[wide frost]

feels like u could use solipsist and some others

[floral mulch]

yeah that one works alright

[wide frost]

schizo

[ruriri:shion \(mirudakemann\):15], [diathorn:mdf an:0.6], (solipsist:0.8), (sweetonedollar:0.1)
[yogisya:1.5], [as109:loundraw:0.4], (honnryou hanaru:0.6), [potg \(piotegu\):shion \(mirudakemann\):0.7], (solipsist:0.5), (diathorn:0.2)

not very similar tho tbh

train a lora on ur rella, ziyun

[wide frost]

imagine u could train vision models on all of danbooru and u were like "give me artist recommendations based on X artist"

[floral mulch]

yeah

[wide beacon]

deepghs has tried to make that

[hidden matrix]

I'm just wondering how to get the mofu style in noobai.

[wide frost]

i have found the best mofu style

but i dont use original noobai

but here u go anyway leilin, [blvefo9, fumihiko \(fu mihi ko\):15], (hideousbeing, yogisya), (wanke:0.8), [potg \(piotegu\):haguhagu \(rinjuu circus\):15], (otohime \(youngest princess\):0.65), (honnryou hanaru:0.8), (sweetonedollar:0.65), ([hiroki \(yyqw7151\):lam \(ramdayo\):15]:0.5)

its a mess

[hidden matrix]

I don't think this is the intended style, but it's still interesting.

Also, how did you even figure out that combination of artists?

[jaunty vortex]

just add random artists until it looks good

[wide frost]

wasnt home so couldnt answer

uh i find artists on danbooru and just ask chatgpt to make combinations based on a1111's prompt editing and weighting features

then make xyz grids

[floral mulch]

TagDescriptionvery awaTop 5% of images in terms of aesthetic score by waifu-scorerworst aestheticAll the bottom 5% of images in terms of aesthetic score by waifu-scorer and aesthetic-shadow-v2......

[floral mulch]

using gemini to tag my ai gens for these for datasets

can't get the api to take nsfw though

on ai studio it can tag nsfw but I guess api is stricter

original prompt: best quality, aesthetic, detailed, by solipsist, solo, 1girl, young girl, anime screencap, 1980s \(style\), backpacker hiking up a mountain trail, it's gonna be dark soon

[floral mulch]

@asht

@wide beacon

found poire

https://www.pixiv.net/en/users/6735502

[wide beacon]

🍐 ?

[wide frost]

@floral mulch @wide beacon check out https://gelbooru.com/index.php?page=post&s=list&tags=aogisa

[floral mulch]

@heady echo gonna make a lora of this

[floral mulch]

not bad

yoinkies

[heady echo]

i forgot about poire

[floral mulch]

I'm thinking of doing a combined lora with separate sets for each tag, and also different subsets for mixed tags

will see how it goes

[wide frost]

train lora on ur umbra mecha

[floral mulch]

nah, want it to work on noob

[wide frost]

i mean a noob lora with the style of umbra mecha (or 218xl)

[floral mulch]

umbra doesn't really have a style other than adding more noise during denoise could turn any model into that just haven't got around to experimenting because supermerger is for 90 year olds that still use a1111 and doesnt work on forge

and 218xl is a lot of styles, gathering datasets for the ones i like now

[wide frost]

supermerger works on forge

dynamic_modules_utils.py

changes:
original: from huggingface_hub import cached_download, hf_hub_download, model_info
new: from huggingface_hub import hf_hub_download, model_info

then on line: 287: "resolved_module_file = cached_download(" i just changed it to "resolved_module_file = hf_hub_download(" and itworked

[floral mulch]

that is an aids noise pattern

seems like glaze or whatever attempt

gonna train on it anyways

we don't do soy models

got this so far

• the ai gens

@heady echo is this bebefate https://danbooru.donmai.us/posts?tags=bebe_(bebe_pp)

[heady echo]

hmm

i think illustrious does support bebefate

[floral mulch]

fro mthe eyes it looks like it is

[heady echo]

google shows

so it does show

[floral mulch]

I'll get it either way

pretty sure all tags were changed to danbooru equivalents

for noob

so bebefate became bebe (bebe pp)

let me check

yep

[floral mulch]

huh

cant find anything about this artist

not this

this

wtf

I read it wrong all along

goat

[wide beacon]

they did

the dataset is completely different from the euge's arti/aid

[plucky dove]

did arti/aid have stuff from other sources?

[floral mulch]

Probably, a lot of artist names had the ones they used on social media etc

[olive onyx]

wow this chanell is actib

[floral mulch]

added bomhat

[floral mulch]

already pretty good like 30% in training

gonna try bundling more artists since this is just an enhancer

still going

[wide frost]

@floral mulch what are u training

a lora?

or trying to finetune

[floral mulch]

lora

pretty good app for dupes

czkawka

[wide frost]

i use https://dupeguru.voltaicideas.net/

works the same pretty much

i should find one that doesnt use python cuz thats hella slow

[floral mulch]

narue has so much cursed shit

[wide frost]

looks dope af

[floral mulch]

this one is pretty fast actually

[wide frost]

probably why

pogchamp

i should practice rust

i dont like the syntax

[wide frost]

but its a good lang

[floral mulch]

vs earlier stuff

[wide frost]

@floral mulch https://danbooru.donmai.us/posts?tags=chiori_(_chio_ri)&z=1

xd

[floral mulch]

multitask

[floral mulch]

actual goat

[jaunty vortex]

finally aesthetic gens

bebefate reminds me a bit of ohisashiburi

[jaunty vortex]

there's enough copycats that there's even a ohisashiburi_(style) tag

[floral mulch]

LMAO

wtf it looks almost the same

other than texture I guess

its the eyes

[wide frost]

@floral mulch hiroki_(yyqw7151)

madana_(xesa7885)

[jaunty vortex]

these are just nsfw

[wide frost]

nice eyes+film grain pretty pog

[wide beacon]

Not that simple

[wide frost]

👍

[wide beacon]

Oh wait I know czkawka

It’s originally a Linux app lol

[jaunty vortex]

I used czkawka too

good for finding dups

[wide frost]

why do you have similar pfp

[floral mulch]

[floral mulch]

[floral mulch]

this feels too hard

might have to not train 15 styles in 1 lora

[floral mulch]

I think it's working so far

Except yogisya cuz ass dataset and poireau 13 I think wasn't trained at all in noob so it's not picking up much

[floral mulch]

50 minute grid time

[floral mulch]

lmao wtf is that wave

[heady echo]

[floral mulch]

got 3-4 epochs that are good

or better than the rest anyways

this is like halfway trained for ziyun I think

kinda same for narue

the rest are kinda iffy

gonna split the datasets into separate loras for the completely untrained styles

[floral mulch]

which one to keep for training?

or fuck it and both

[wide frost]

i think it might be easier to train without film grain

tho idk

[jaunty vortex]

both

[floral mulch]

yep

curated a bunch of them down

not like artists have static styles for 10 years

[floral mulch]

finally found poire real tag

https://gelbooru.com/index.php?page=post&s=list&tags=powa_(36iii)

@wide beacon @heady echo

[heady echo]

the plot thickens

honestly i think euge made it hard to find the actual artists

somehow

🫠

[floral mulch]

well, got them now at least

training agane

[floral mulch]

https://www.artstation.com/aoiogata @jaunty vortex

[jaunty vortex]

inconsistent

they're kinda shit at anatomy

[floral mulch]

looks good tho

[jaunty vortex]

they're just tricking you with the low saturation

I could draw this in 2-3 hours

[floral mulch]

dont like this one

but some of the other ones I think are nice

like these

[jaunty vortex]

it's good but not top

[heady echo]

pretty style imo

[wide beacon]

imi if you find the artist use https://saucenao.com to find the danbooru tag

or also https://huggingface.co/spaces/deepghs/search_image_by_image

[floral mulch]

right

I'll do that next time

hope the fingers wont be this bad in actual use

[wide beacon]

maybe could've been worth adding cleaner styles like the ones people use on nai3 into the dataset

of course tagging them properly too

[floral mulch]

Wtf the fingers are recovering

[floral mulch]

Not training in ai images currently, maybe I'll do it in a separate Lora or just take someone's

Will try to merge into models either way

It's also learning the brand new styles surprisingly

And this is 15 styles in one 16 dim lora

[wide beacon]

nono I mean adding artists like ciloranko and ask in the dataset

[floral mulch]

Well I'm doing subsets so they don't affect each other

But maybe

Don't wanna mess up the scuffedness of ziyun iwzry etc

[floral mulch]

think it needs more training

https://files.catbox.moe/dyzfxa.png

[jaunty vortex]

6 looks ok

[lethal sierra]

omg cute samples

[floral mulch]

continuing

[floral mulch]

https://www.pixiv.net/en/users/158253

[jaunty vortex]

nice texture

[floral mulch]

cant find an artist

had a pic with yellow background, a chair, and a paper plane

and city scape abstract images

nvm

wasnt yellow

and not paper plane

and wasnt the cityscape one

[jaunty vortex]

wow abstract

[floral mulch]

https://media.discordapp.net/attachments/696442309578981400/894228358324420669/93193402_p0.png?ex=6749ce8e&is=67487d0e&hm=68a7c4892e75e4c417e19ba53fa6b6cd65f4881d26d33ea78426e94edeccb4df&=&format=webp&quality=lossless&width=545&height=702

[jaunty vortex]

they need to wipe their glasses

[floral mulch]

this is how u keep looking at images for hours

[jaunty vortex]

https://danbooru.donmai.us/posts?tags=ibuki_satsuki+&z=5

what about

[floral mulch]

nice, and lot of images

couldn't find the cityscape one, but either way its similar to this vibe, maybe more simplistic

[jaunty vortex]

there's a bunch of ibuki satsuki pics that aren't on danbooru

too bad I can't find highres

https://danbooru.donmai.us/posts/6508672?q=ibuki_satsuki+ lmao there's random fabric on this one

before AI: cool pic
after AI: wow AI artifacts

[floral mulch]

welp

[jaunty vortex]

🚨🚓 OVERTRAINED 🚓🚨

[wide beacon]

lmfao

[wide beacon]

How the training going?

[floral mulch]

stopped it after that ^

[floral mulch]

might redo this continue but only with the undertrained artists

makes more sense imo

[floral mulch]

last 3 checkpoints are the above pic before loss started going up, I think I'll redo with 32 16 instead of 16 8, ziyun for example didn't pick up any of the background style I like

samples looked a bit better

I 'Ll try both removing datasets and increasing dim

or keeping them just lowering their repeats of the learned ones

[wide beacon]

Avoid doing more than 64 dim

[floral mulch]

yeah

16 8 already too strong for some styles, just want to see if it helps pick up more stuff

[wide beacon]

Try having more repeats for styles that are too weak

[floral mulch]

yeah doing that atm

moved the rest to single

oops

guess cant change dim on existing save state

[wide beacon]

Changing dim is like doing 4d chess

You won’t win

[floral mulch]

I'll just change alpha instead

I'm curious what will happen

with the already learned stuff

[floral mulch]

ok that changed a lot lol

think it got worse tho

too strong at 16 16

[floral mulch]

green is previous 1000 steps

[floral mulch]

f, huggignface broke my cool page

[wide beacon]

rip

at least it's still slightly functional

[floral mulch]

now just to try merging vpred

[floral mulch]

maybe got it

it's only on model differences, not full models, but I can take a vpred and eps model and by doing the opposite of what you'd expect I can end up with either a working eps or vpred model, oh and ztsnr doesnt work on the vpred results, so that would need more tinkering

[jaunty vortex]

Vpred without zsnr

[floral mulch]

[olive onyx]

what is zsnr

I don't know

[floral mulch]

zero terminal signal to noise ratio, starts from 0 noise or some shit, look it up

[full lynx]

ends at zero

[floral mulch]

ah

yeah terminate

pretty significant differences between some of them

[floral mulch]

gonna need to crank up the noise if i stick to an eps model, kinda underwhelmed with the lack of detail compared to vpred gens

also wonder if nyaflow has mergeable parts

prob similar like vpred x eps

[nocturne cosmos]

noob won't update now

[full lynx]

after vp1.0

[nocturne cosmos]

ah, i see

[nocturne cosmos]

Lax says next version for noob won't include the cosplay dataset.

[floral mulch]

great, new model to merge

[floral mulch]

merge of random 1.1 models and 0.9

[heady echo]

what does noob cosplay dataset do

[floral mulch]

Cidolays

Cosplays

Realistic

[heady echo]

[languid cliff]

Mofu will hab da ice cream

[floral mulch]

u were 6969

@nocturne cosmos do you know any good vpred or noob 1~ ish loras?

[nocturne cosmos]

maybe these 3 art stlye lora? I dont use a lot lora on noob
https://civitai.com/models/1049532
https://civitai.com/models/1027557
https://civitai.com/models/823860

[floral mulch]

its so hard to find a vpred model that does what I want

but I'll get there

[floral mulch]

maybe if I keep distilling them I'Ll get an inbetween

luckily I can use cpu to run tests