218 xl

[floral mulch]

oof

thats a pretty big difference from vpred 1.0

[floral mulch]

maybe not a big issue, none of the other vpred models retain how 1.0 looks like

maybe cyberrealistic is closest?

idk

I'll run a merge between 5 and 8

https://huggingface.co/Enferlain/sdtests/tree/main/noob/epstovpred here's everything so far, the out8 and 1bbwm ones are the ones from the xyzs above

[jaunty vortex]

none of them can pass the fence test

[floral mulch]

let's keep it realistic

[jaunty vortex]

noob doesn't make me a sandwich either

[nocturne cosmos]

whys that 404?

[nocturne cosmos]

maybe include the OG vpred v1 in the xyz as well

[plucky dove]

its there, 4th from left

[nocturne cosmos]

mb,

[plucky dove]

its ok lol

[floral mulch]

It's there, 4th image

Also the repo is private, oops

[floral mulch]

made it public

the first 7 models from these maybe worth to check to see which ones are the best, hard to see from the tests I've done

[nocturne cosmos]

still 404

[floral mulch]

I'll upload to a new repo

[floral mulch]

https://huggingface.co/Enferlain/juice/tree/main/noob

should be good

[floral mulch]

Spent the entire day trying to get my merge to make ffffff background

Still not there

[nocturne cosmos]

205out8wai just doesnt work for me

is it because forge doesn't know this is a vpred model?

[floral mulch]

Yeah none of them have the keys to be automatically recognized

https://github.com/DenOfEquity/forgeFlux_dualPrompt

U need this on forge

Only way to manually enable vpred edm etc

[floral mulch]

spoon varia

[nocturne cosmos]

oh wow, I think its a lot better than the og vpred v1 check point( I only tested the 205out8wai)

[floral mulch]

the og vpred model is really bad compared to the other vpred models

except for colors

for those it's still one of the best

and of these imo obsession does the best for texture and color combo

so that's what I'm trying to merge atm

rest results in undesirable levels of saturation

[floral mulch]

see 1 and 3

[floral mulch]

I think the cat ending one is the "best", or one of the bbwm ones of the ones in that repo

actually it's probably easyfluff but working to fix the colors in that atm

[olive onyx]

what got you back into merging? @floral mulch

good results also

[floral mulch]

haven't really stopped

[plucky dove]

Obsession is rlly nice, it has that AI shine to it tho since thats what its trained on

Which is subjective good or bad

[floral mulch]

well, if ur looking at the xyz above, it's the least shiny/metallic of the vpred options

https://cdn.discordapp.com/attachments/1019446913268973689/1324306468870033458/xyz_grid-0041-2367640929.png?ex=6777ac02&is=67765a82&hm=8c32487efbf5723c4c1a15df6d1fd978bd704c1f985c9d15ad675b638d7721c5&

maybe last one worth to fix

5 lmao

I like that bad fingers are like the rarity now

2 years only

[nocturne cosmos]

try negative prompt with only "worst quality, low quality"

[floral mulch]

I use worst quality, low quality, lowres, artist name, signature, bad anatomy

[floral mulch]

1st half is my eps-vpred merge x panchovix's eps-vred merge 0.5 (both stuck being grey) each one's out8 block replaced from an unfucked v-pred model to get rid of greyness, but results in overly saturated images

middle model is out8 from an older noob based v-pred model/merge/finetune, so replacing it didn't fix the greyness

2nd half is the middle grey model, but the out blocks, or the out.2 layer remerged from un-fucked v-pred models. still not entirely satisfied but so far 21 and 24 are close to what I want (first 2 models on the catgrill xyzs above)

basically going for 000000 adn fffffff backgroud ability, and avoiding washed out colors while retaining eps like results

got more work to do

some of the results I have can already do full white or black bgs but depending on the subject they don't work well whereas on the pure vpred versions the effect is a lot more drastic, especially on 1.0 here, that might not be entirely dependent on out layers but I don't want to mess with my eps knowledge

[plucky dove]

Do you use any quality tags in positives?

[plucky dove]

Cat tower looks cool here

Tho i just saw from chaos lab ur right

Black background varies a lot

[floral mulch]

Yeah cattower is probably the best vpred model from the ones I tried regardless of color differences

But I saw some more models on civ it lately so maybe they do better, will check later

[plucky dove]

https://civitai.com/models/1095750/neta-noob?modelVersionId=1230787

Thinkge

[floral mulch]

yeah saw

how is it?

there's 2-3 models i wanna test

[wide beacon]

wtf

[jaunty vortex]

tf is flexible deformity

flexible tag is 50% jack-o' challenge

[plucky dove]

schizo phrenia negs

[plucky dove]

havent tried yet

[nocturne cosmos]

i dont like it

[floral mulch]

f

not gonna bother exploring other models until I'm done with my current thing

[plucky dove]

I also do not like it

[floral mulch]

f, even though 2 can do 000000 black background, it looks miserable when both subject and bg has black in prompt

[floral mulch]

schizo

[floral mulch]

well

shit

now I have to figure out if any of them are better with this little difference

[wide beacon]

literally just a different color palette lmao

[floral mulch]

yeah

[wide beacon]

anyway final model will be called 225

[floral mulch]

that reminds me I shouldve used a fixed seed during testing

maybe I'll try 1 last time tmrw

but tbh it's not the same

the effect of the layer adjusting is not as good as it was on animagine and other older models

increasing noise actually meant more detail here it just means more noise

maybe vpred related

I'll do the old method of doing like 15 xyzs and using the model that I like the most images from

[heady echo]

i like this one

[jaunty vortex]

same

[gritty lynx]

2nd (counting from index 0) has more contrast I think, personally I prefer it

[floral mulch]

yeah same

but sadly its not perfect in all scenarios I like

and I think I lost detail from all the self mixing

I'm trying ties sum extended now with a bunch of other models and loras

but with base being illustrious the effects are minimal

but with sdxl base it will be unstable imo

oh wait

I'm using vpred 1.0 as base

[wide beacon]

noob incest

[floral mulch]

no wonder the results are very similar, that models is basically where I'm at

[floral mulch]

with the version you liked

[gritty lynx]

I see

makes sense

[floral mulch]

I like that with noob Ican just gen datasets instead of having to fuck around withe xisting images

[jaunty vortex]

https://www.reddit.com/r/StableDiffusion/comments/1i08k3d/the_difference_from_adding_image_space_noise/ this reminds me of how your model with added noise is good at img2img details

[floral mulch]

yeah pretty much the same thing

[floral mulch]

2m, euler a, res

[feral elbow]

current best model? ive been using NTRmix

[floral mulch]

dunno, prob obsession

or cattower

[wide beacon]

https://moescape.ai/models/39711aee-50a2-48cb-bfae-5125d7a76ac7

early access on yodayo

[floral mulch]

3.5 huh

[full lynx]

interesting

[floral mulch]

vpred animagine 4

[wide beacon]

there's #000000, #fefefe and #ffffff in this gen

impressive

[floral mulch]

I think umbra mecha used to look better

nvm thats not umbra

close enough

[wide beacon]

try this https://huggingface.co/AshtakaOOf/lora-extract/blob/main/sdxl/UmbraMecha-Ani30.safetensors

[floral mulch]

is this a lora

I might try to do the same recipe on my old models

umbra mecha and elysia, altho never released the 2nd one on civit or maybe hf either

[wide beacon]

lora extract of umbra mecha from Animagine 3

[floral mulch]

but is it a lora

not lycoris right

[wide beacon]

lora

[floral mulch]

alright then I can use it for merges in mecha methods

[wide beacon]

if it doesn't work well we could try add diff animagine 4 with umbra mecha and then lora extract that

[floral mulch]

I think any model should work with what I did with animagine v4, the question is if it's worth it considering text encoders hold most of the quirks of models

like trying to merge awa/aid or pony is pretty much a no no cuz of te

but from what I heard animagine v4 is also hard baked with the score tags

so idk anymore

[wide beacon]

[heady echo]

success

[floral mulch]

hm

trying to decide if the merge with animagine in it is better

but it's an intermediate model

hm

feel like left might be more clean

but dunno about composition/etc

[floral mulch]

Thoughts

[heady echo]

try with this style
i think think it works w noob

[wide beacon]

it has more textures but I wouldn't say more clean

[wide beacon]

https://civitai.green/models/1021726

might be interesting to merge

[floral mulch]

Hmm, I think the fingers and small details are sometimes more cleaner on the left one, and the textures like you said, look on the image with guns

But also from limited testing I've done some of the things I like are less coherent in compositions

I'll continue later

[floral mulch]

having problems merging the animagine noob into my other merge

the animagine one has troubles with concepts and compositions even though it's a bit cleaner

and looks better with some anime styles

[gritty lynx]

probably a stupid idea, not interpretable
what if you take the low froquency of one and high frequency of the other

[floral mulch]

what would u use for that

[floral mulch]

https://civitai.com/models/1217645 pog you

[plucky dove]

Gonna try when im home

[floral mulch]

It's the same model I put up on hf a week or so ago

Let me know if it does what you want

[wide beacon]

oh well time to post gens

[gritty lynx]

crossover

approximating fisher on low passed/high passed input would give better results but

[floral mulch]

Hmm any tilt suggestions

[gritty lynx]

what's tilt already

1 sec

[floral mulch]

[gritty lynx]

no tilt, tilt=0

not tilted

:>

[floral mulch]

I need to stop my shitty habit of throwing every problem at optimization

[plucky dove]

The cat obs one?

[floral mulch]

Yep

[floral mulch]

can u guys put some pics on the page whenever @glass tulip @heady echo or anyone else

[plucky dove]

nvm then i did try it, it was good but i felt like it still had ai generated feeling that is hard to remove

[floral mulch]

well, it is ai generated

I like this style atm

random find

[glass tulip]

Didn't have time to make any yesterday, I'll fill it up tonight

[floral mulch]

thanks

[heady echo]

sure i can after work

[floral mulch]

these are in que for merges next

[heady echo]

okei done

[wide beacon]

https://huggingface.co/Eugeoter/sdxl-vae-anime-alpha-67500

[floral mulch]

Yes saw that, haven't tried it yet tho

[wide beacon]

I know just noting here

since cake and others are in here

[nocturne cosmos]

its like 1%-0% differences.

[wide beacon]

Lol

[nocturne cosmos]

and sometime its even worse

with, without

[wide beacon]

go to Euge's server

and post that

or I can forward for you

[nocturne cosmos]

sure.

I don't think spending any time on vae is worth it tho.

[wide beacon]

yep

[wide beacon]

a

[nocturne cosmos]

920x1240 + Hires upscale: 2

[heady echo]

i honestly couldn't tell the difference in his civi preview

[jaunty vortex]

A vae that doesn't output jpeg artifacts would be nice but otherwise meh

I just filter pics myself anyway

[floral mulch]

000102

unusable

[jaunty vortex]

wow she stood up

[floral mulch]

feels like a downgrade. sucks triyng to find a prompt thats good for comparisons

[wide beacon]

use one of my prompts or something

[floral mulch]

which one

[wide beacon]

do multiple xyz on various art styles

these for examples #1331696007494369331 message #1331696007494369331 message

[floral mulch]

the styles are the same probably, I'm looking to minimize errors and stuff which is hard to look for

[wide beacon]

this one is very gacha #ai-art message

hands can easily get mangled

can also try this mumei sunset/sunrise gen #ai-art message

[floral mulch]

complicated ass hands for n oreason

cool pics

[wide beacon]

wait that's hannibal lecter

@frail zealot

[frail zealot]

I literally dmed you the link to the profile ya dumb

zoomer memory retention....

[wide beacon]

no that's because you sent it to me that I pinged you

he's using the imi model

[frail zealot]

He uses a bunch of models tbh

regularly shifts between WAI, Noob and NTR

the best question is

if you upload his pics to aibooru

do you still have him labeled as the same artist tag

or nah

[wide beacon]

those are popular

[wide beacon]

add a note on the wiki page

[frail zealot]

I choose to believe umbra mecha is more popular than the download page indicates

[wide beacon]

lol

making a Pony merge during the Pony boom gave you a noob boost

[wide beacon]

do I sound like a zoomer like this ?

[floral mulch]

who

[wide beacon]

Habbibal :3

the top non slop image gen guy on civitai who deleted his account

[floral mulch]

why did he delete it

[wide beacon]

¯_(ツ)_/¯

[floral mulch]

wtf is this boob size

[wide beacon]

my negative is better

[floral mulch]

[wide beacon]

oh yeah btw this is a modified Euge prompt

[floral mulch]

thats where I remembered it from

aid days

[wide beacon]

trimmed down to less than 75 tokens

because i be like that

[floral mulch]

and i always have to reorder other ppls prompt to fit my ocd

1girl char artist solo char traits outfits position background quality

thanks tipo, good stuff

[wide beacon]

oh neat

[floral mulch]

same

[wide beacon]

actually he might've found the model because of me

[floral mulch]

big

[wide beacon]

I commented on this gen https://civitai.green/images/55648565

and I had a bunch of gens on your merge

[floral mulch]

https://tenor.com/view/math-madness-gif-25268021

[floral mulch]

ai comment

bruh how is this r

[wide beacon]

aight lets change the rating

voted for pg

[heady echo]

R

[wide beacon]

it's pg rated now

[floral mulch]

Yeh, pog

[nocturne cosmos]

i think imi's merge is better than v1

wait, is this the final version?
https://civitai.com/models/1217645/sih

[plucky dove]

https://civitai.com/models/1246248/hesperides-illustrious-v10

People merging v1 already

[floral mulch]

it's the same as the bbwm merge-layers 4

I'll try to merge it and see how it goes

but I'd like 1.0, want to try reconstructing the model with that as base

and then add noob on top

[plucky dove]

Good luck!

Looking forward to any good outcomes

[floral mulch]

I wonder if qlip is worth messing around with

[wide beacon]

Ask @jaunty vortex

[jaunty vortex]

no don't ask me idk either

I just tried it and it does stuff but idk if that stuff is better

on furrycord their backgrounds were better but the rest was more glitchy

[wide beacon]

Merge clip and qlip or something

[floral mulch]

i'm trying to use illust 1.0

[floral mulch]

welp

[wide beacon]

cool sample gens https://civitai.com/models/1248959/ntdmix

they mostly all use the same LoRA even though they have a different style

[floral mulch]

I'll take a look

[wide beacon]

merge of ΣΙΗ https://civitai.com/models/1237873/noobmod

[floral mulch]

Ohh

He best gemnight to the chase

Beat

[wide beacon]

@nocturne cosmos @quaint grove background artist https://safebooru.donmai.us/posts?page=1&tags=alu.m_(alpcmas)

[quaint grove]

Error: The database is unavailable. Try again later.

gone very fast

lol

[wide beacon]

yeah so uuuuuh

I sent that link

and danbooru crashed

...

anyway I still had the page open so here what it looks like

167 artworks

but not sure how many are background

it's back lol https://safebooru.donmai.us/posts?page=1&tags=alu.m_(alpcmas)

[quaint grove]

btw anyone know how to counter this data type, when both image tags are "sunset"

i mean

i could "sunset1, sunset2" but hash prompt lora proven that's not how it works

or give up and let them mix together

unstable diffusion

[wide beacon]

artist for vehicles https://safebooru.donmai.us/posts?page=1&tags=momo_hiki

[wide beacon]

mix together

you can use img2img 1.0 denoise with a gradient image to steer toward a color

[jaunty vortex]

could be dawn or sunset what is the difference

[floral mulch]

love that

wish i could find that eastern eu vibe artist

almost like iwzry just like architecture eastern eu

[jaunty vortex]

find photos of eastern europe and train on them 🤔

[floral mulch]

thats work tho

[jaunty vortex]

real..

[plucky dove]

https://civitai.com/models/1254035?modelVersionId=1190596 jemnite version

[floral mulch]

preorder

[plucky dove]

HF link is there

dont have to spend buzz

https://huggingface.co/LMFResearchSociety/CheckpointArchive/blob/main/Jemnite/ŞİH.safetensors

but think u need to be part of org

[quaint grove]

too much quasarcake...

[plucky dove]

[plucky dove]

would reflection as tag work?

[quaint grove]

wdym?

[plucky dove]

what do u mean by counter?

do you mean the sun being too high?

[floral mulch]

no one liked umbra

[floral mulch]

last attempt for trying to get bigger part of illustrious 1.0 to vpred

[floral mulch]

@full lynx how to get black background ur a scientiest

[full lynx]

scheduler

[floral mulch]

which part of the unet is scheduler

[full lynx]

oh merging

final 5 ish blocks help

[floral mulch]

what type of layers tho

already did skip connection

[full lynx]

uh, prob the biases more than weights

[floral mulch]

maybe

they seem to have an effect in output blocks compared to others

maybe I should just merge them at alpha 2.0

didnt go as planned

they did help with 000000 tho

good to know

I need to figure out what does this frequency noise bs

[full lynx]

random guess, look at all out weights std and find the one with largest, merge that

[floral mulch]

Decent idea but didn't pan out

Biggest diff is usually out layers 0 or in layers 0

But as expected the last 2 blocks (together) have a massive influence

Alone they don't help but together it fixes the noise and greyness

Now I just need to eliminate the layers to see which one is fucky

[floral mulch]

[floral mulch]

https://www.pixiv.net/en/artworks/121925391

[floral mulch]

I want a dataset of this

[wide beacon]

might be a decent merge https://civitai.com/models/1271610/orangeflavorednai

[floral mulch]

isnt it a finetune

[wide beacon]

i'm gonna finetune you

[heady echo]

those example gens

[wide beacon]

yeah lmao the description is yikes

I did not look at what else there is since I'm staying on safe

[floral mulch]

@full lynx

[full lynx]

i dont use 4channel

[floral mulch]

based

[heady echo]

omg you found ur model on the frog catching forum

https://tenor.com/view/achievement-unlocked-gif-27073940

[wide beacon]

That’s definitely Sab

[floral mulch]

can someone replace clip-l on my model with this https://huggingface.co/Clybius/XLOmni-Models/blob/main/NoobAIQLIPSuperMerge_clip_l_00001_.safetensors

[wide beacon]

can't you do that in comfy?

[floral mulch]

eah but dunno what nodes etc

[wide beacon]

load checkpoint
load clip
save checkpoint

that's about it

[floral mulch]

I tried yday and the model is really inconsistent with 3 characters

[wide beacon]

like this I'm pretty sure

[jaunty vortex]

It's gonna mess with character details

[floral mulch]

I just need it on a model so I can merge it

[olive onyx]

@floral mulch do you have a huggingface link for the ΣΙΗ model

[floral mulch]

nope

its on civit

[olive onyx]

[floral mulch]

made it so u dont need an acc to dl it

[olive onyx]

thnk you ❤️

[wide beacon]

there is...

https://huggingface.co/Enferlain/juice/blob/main/noob/1bbwm-21821-catobs05-merge_layers-it_4_best-bf16.safetensors

[floral mulch]

has a different name and I didnt add the vpred ztsnr keys so I don't consider it the exact same

[wide beacon]

[floral mulch]

• confusing if there are 2 names for the same model

[wide beacon]

i still use this version on swarm

all my gens have this name in the metadata lol

[floral mulch]

ur like 1 of 2-3 people

[heady echo]

I'm the second person

[floral mulch]

https://civitai.com/articles/12128 comfy workflow

[floral mulch]

still getting downloads surprisingly

and I saw at least 2 merges

[floral mulch]

oof, looking tough

might have to scrap this

[wide beacon]

[olive onyx]

I miss the days when Supermerger was viable and people didn't have to resort to methamphetamine level of cooking to merge

[floral mulch]

u can still use it

[olive onyx]

is it done using comfy workflows now?

[floral mulch]

supermerger works I think

on a1111 or reforge too maybe

[olive onyx]

Does it work when merging the newer models like Illustrious or Noob?

[floral mulch]

sure

[olive onyx]

[wide beacon]

not memory efficient tho

[floral mulch]

yeah

[wide frost]

welp time to install comfY

and get uncomfy

using the windows gui

[full lynx]

swarm

[wide frost]

No

[full lynx]

why

[wide frost]

already tried it

[full lynx]

then just use it for comfy then

[wide beacon]

And Swarm has some of the best custom nodes

In quality

[wide frost]

How do I make an XYZ in swarm?

@full lynx

[full lynx]

tools > grid generator

"infinite" "axis" built in

[feral elbow]

I tried out ΣΙΗ but im getting better results with NTRmix or WAIillustriousXL

[wide beacon]

skill issue

[floral mulch]

lule

[jaunty vortex]

How

[plucky dove]

i can never get over other shitmixes' (ai-generated:1.2)

i like eih and sih since theyre a bit less

[floral mulch]

he listened

just wh at i needed

[plucky dove]

Eih part 2 please

[floral mulch]

doing stuff still

[heady echo]

it's that w ur model ?

[floral mulch]

ye

free ad

[heady echo]

[plucky dove]

ofc take ur time, looking forward to ur cooking

[floral mulch]

figuring out how not to downgrade or sidegrade

[wide frost]

now thats foreshortening

[wide beacon]

https://civitai.green/images/61648308

thick sampling thigh

[floral mulch]

seems familiar

[wide beacon]

https://civitai.green/images/59506505

merge lhc

[floral mulch]

Alright I'll check, but if it needs te I might not

[floral mulch]

got so much experimental dogshit I need to cut down

need to gen a shitton of xyzs andonly keep the ones that are useful

[floral mulch]

furry attack monk ass

[wide beacon]

Actual gaming

[jaunty vortex]

cute

[olive onyx]

how are the alphabets going ?

[floral mulch]

dunno

need a sharp vpred model or lora, tired of getting scuffed results from all the merges

[floral mulch]

wasting time fucking around with shit that gives me shit like

meanwhile my last model

not even same ballpark

time to do xyz to see which models to keep

[floral mulch]

vruh

the fuck is lhc

[heady echo]

[floral mulch]

I can narrow this XYZ down to 4 useful models

[wide beacon]

samples don't look interesting but the description lmao
https://civitai.com/models/1296947/nlxlmix-illustrious-10

[wide beacon]

aesthetic and vtubers

[jaunty vortex]

https://danbooru.donmai.us/posts?tags=chihuri this is not that bad ???

https://danbooru.donmai.us/posts?tags=bb_(baalbuddy)+ ok this one is mostly comics

https://danbooru.donmai.us/posts?tags=iesupa&z=5 this isn't crazy bad either

there's much worse out there, like doesnotexist on e621..

still surprised

[jaunty vortex]

you can't post that unless you actually see his work

gyrotech is even worse

[wide beacon]

oh ye this model got an update https://civitai.com/models/1130584/wais0ftmix

[floral mulch]

Hm

[floral mulch]

I was liking one fo the models I had and turns out it's bad at one of the things I like o7

[gritty lynx]

large hadron collider

thank me later

[floral mulch]

thanks

so when are we doing some giga collider merge

can u rate my new merge methods sometime

[gritty lynx]

sure where code

[floral mulch]

actually here's everything

some have multiple versions just for reproduction

[wide beacon]

how many imports I lost count

[floral mulch]

[gritty lynx]

ill take a look later today and tell you what I think
you can ping me if I forget

[floral mulch]

You can do it whenever

[wide beacon]

found ΣΙΗ in a merge https://civitai.com/models/1347947/r3mix-illu01nai-v-pred

[heady echo]

revive

[floral mulch]

still busy with resolution stuff

https://cdn.discordapp.com/attachments/1230387093793013883/1352836923185430568/image.png?ex=67e01fd0&is=67dece50&hm=d29008edb69d5c98a5a24d1605b6d60a003d8f4a27252d7bbe9ebb1f7b02533f&

but got a few experimental merges I guess, for fun

[floral mulch]

progress

[floral mulch]

which one looks better?

fp64 fp32 bf16 fp16 inference

first 4 are merged to model, 5th is just normal lora use

[heady echo]

they look similar

[floral mulch]

forehead and small details in color and stuff

if u have any prompts to test I can

ngl inference looks somehow better

but

preference

maybe fp64 more correct idk

[heady echo]

maybe this one

pikaole, tareme, 1girl, solo, long hair, looking at viewer, blush, smile, open mouth, bangs, skirt, blonde hair, simple background, shirt, red eyes, long sleeves, white background, bow, ribbon, holding, :d, cowboy shot, hairband, fang, cape, neck ribbon, stuffed toy, stuffed animal, teddy bear, green shirt, red hairband, white cape, orange skirt, holding stuffed toy

[plucky dove]

i like 3

[feral elbow]

waixl still the best model for me

[floral mulch]

u already said that, anyways cattower clears wai

[feral elbow]

theres just no way, cat tower gens look so slop tier looking at the civtai page

https://civitai.com/models/920709/cat-tower-noobai-xl-checkpoint

like these look like they're still from the pony generation

[floral mulch]

yeah cuz ur average civit cuck only looks at the previews, not every model maker is good at genning images

cattower is more detailed than wai

[feral elbow]

i mean im happy to be proven wrong

do you have any examples

[floral mulch]

eye detail test on standing 1grill, wai v20 is one of the models that does alright (including rouwei and obsession 11) but when u do more tests cattower gets the eye symbol detail right more often of course this is jsut one test

if you find that it's more detailed for you in the prompts you do then that's what matters

[feral elbow]

ill admit the sample-bot for cat tower looks good

[floral mulch]

also I found matincore is shockingly a decent model, it's supposed to be some furry shit though?

doN't remember wherte I got it from

[floral mulch]

is that guy actually a bot

[feral elbow]

i assume so

no one is just gonna sit there all day refreshing civtai to try every new model

[floral mulch]

[feral elbow]

https://civitai.com/user/search_bot/images

this one is defeintyl a bot

the one that does the purple hair girl

very useful

more so for finding new models and loras

cuz i go through their accounts and see what looks good

[floral mulch]

hm, yeah I don't remember encountering tghis one

runescape

[feral elbow]

like when i see something like this i check it out

but most of them are the most generic slop looking stuff

I can kind of tell now what model merges are based off now based on these

[floral mulch]

I usually check the top right image cuz that looks like ai on most models

also funn how link becomes a girl on a lot of them lmao

[jaunty vortex]

Too 1girl pilled

[wide beacon]

Bruh

No Wai is so crap

[plucky dove]

wai is okay if you like the ai-generated look

[wide beacon]

https://tenor.com/view/scorsese-absolute-cinema-meme-gif-9523375912274476782

[floral mulch]

https://cdn.7tv.app/emote/01JDS9CMA19X9AZSR8TM89H9DP/3x.avif

[wide beacon]

Animagine at least does something unique

Wai ? is just shit

[floral mulch]

well animagine is not a merge

[wide beacon]

it could almost be considered that with how much it leans into that specific art style even when trying to avoid it

at least Zero alleviates that

[floral mulch]

I havent used it so idk

[wide beacon]

~~ @shut hamlet when vpred Animgine ~~

[feral elbow]

idk what to say but i have had the opposite experience

WAI is the only thing that I can consistently get non slop looking stuff on

everything else has the extreme AI look

[wide beacon]

Maybe learn to prompt properly ?

[floral mulch]

huh, some of these might be serviceable

need to test omre prompts

shit takes forever at 2048x1536 grids

[floral mulch]

[heady echo]

the right one is the best one

[floral mulch]

I'M merging the last and 2nd to last

last looks like shit stylistically, default style is garbo and quality is meh

but it feels the most complete for 2048 gens from what I achieved

and I don't want to spend more time on it

I liked the model I had with that nhl lora merged to it so I want to mess around with that

this is just a sidequest of "can it be done"

(the answer is probably not, since angel said 2048 is pure luck, even if it doesn't gen siamese twins like model strained at lower res)

[heady echo]

🙏

[heady echo]

thread revive

[floral mulch]

[wide beacon]

https://tenor.com/view/mimukauwa-39neko-ぬぬぬぬぬぬぬぬ-nunununununununu-uno-gif-14413518820995350443

[full lynx]

https://tenor.com/view/nunununununununununununununununu-mimukauwa-nice-try-hatsune-miku-uno-dancing-gif-13057743886552416930

[wide beacon]

stuff
https://civitai.com/models/846917/raehoshi-illust-xl
https://civitai.com/models/1400174/oeai-illustrious-xl-more-stunning-illustration-style-or-cute-style

[floral mulch]

yeah heard of the 1st one

but not vpred

[candid sonnet]

dunno if this has been posted or not but
https://civitai.com/models/1255461?modelVersionId=1418511

supposed to be a “newer” version of base noob vpred

[wide beacon]

that's yoinked

[candid sonnet]

yeah ik, might be useful in future merges

for model b specifically, dunno about model a (when I asked yoinked he told me not to use model a in merges)

[floral mulch]

it's just using the experimental vpreds

not sure which one

@full lynx

[candid sonnet]

think it's 24r2

[candid sonnet]

I think the exp vpreds are trained further than vpred 1.0 and have some of the color issues fixed

[floral mulch]

didN't notice it on the one I used

[candid sonnet]

Use model b

Model a is some weird merge

[plucky dove]

24r2+28m

[full lynx]

yea that one is model B's base

[floral mulch]

the one I have is this 41abed00

[candid sonnet]

is that the other noob vpred experiment leak?

[floral mulch]

no clue

its just the hash

[candid sonnet]

I’d say ask @full lynx what the original hash was as model b had model a merged in at 0.05

[full lynx]

euge isnt here, this is the .95 of model B

[candid sonnet]

[full lynx]

shhhshshhhhshhhhhhhh

[floral mulch]

https://civitai.com/models/1386662/shiun-or-shiiros-styles it just doiesnt hit as hard as aid version

too clean

[wide beacon]

they are using whatever this ass model is https://civitai.com/models/1135145 for the samples

[floral mulch]

so many of these dogshit

gummy bear face looking models

with a bazillion downloads

actual civit virus

[reef laurel]

[wide beacon]

not an interesting model but good prompts in the samples
https://civitai.com/models/1296870/hadrian-delice-sdxl-ill

[plucky dove]

https://civitai.com/models/956018?modelVersionId=1633911

[floral mulch]

Oh it's the furry guy from the merge server that thought noob voted couldn't do 1536 res

[plucky dove]

😂

[reef laurel]

It says trained tho

[floral mulch]

I'll just upload the merge that can do 2048 gens

and someone maybe will fix it, and if not someone else will do it sooner or later anyways

[wide beacon]

https://civitai.com/models/1317905/ripplemix-noob-vpred-10-illustrious-01

[floral mulch]

💀

what the fuck

that guy is a schizo 4chin andy

[olive onyx]

"model has good backgrounds" Almost every preview image is porn

[floral mulch]

that guy is shizo

c

havent seen a single sfw image from them

if my current merge goes well I'll put a lot of nice bg images for previews instead of just 1 grill

[olive onyx]

hopefully it does

[floral mulch]

https://tenor.com/view/girls-frontline-2-leva-ump45-gfl-gfl2-gif-15994377690873670730

[olive onyx]

this one iron man meme is just funny to me for no reason

[floral mulch]

the jarvis one?

[olive onyx]

yes

[floral mulch]

its funny

cant think of any examples now toh

[olive onyx]

like you could put the worst caption on it and it would still somehow get a chuckle out of me

[floral mulch]

going well

[wide beacon]

Jarvis call the Chinese I need a schizo negative

[olive onyx]

[floral mulch]

https://cdn.discordapp.com/emojis/951343861903532148.webp?size=44

chicken jockey

time to train an anatomy scorer

[full lynx]

please dont be "danbooru score" [breast size] bias

[floral mulch]

can embeddings have breast si ze

it will be 0 bad anatomy 1 correct anatomy

example images

well, not 100% correct for 1

the goal is to get 0 to be recognizably trash across the board

like even if 1 is not perfect compared to 0 it will be perceivable what the difference is

I'll try the siglip based one

[floral mulch]

It's not working very well, I got one setup that doesn't make sense to rate examples from the bad dataset correctly but it says some good images are bad as well, but the setup that dies make sense is performing like shit

[jaunty vortex]

😦

[languid cliff]

Deres smol one!

[floral mulch]

got a decent result seems like, we'll see in action

bit too strong, got an earlier checkpoint that's less extreme but still gets everything right

except 1 img but whatever

[olive onyx]

babe wake up new Illustrious model just dropped

[plucky dove]

is this one better than 24+28 ? or just different

[full lynx]

different, both good

[plucky dove]

I see, ty. Wheres the documentation tho.... 😔

Euge hasnt updated the doc since january

[plucky dove]

https://tenor.com/view/wake-up-menacing-furious-undertaker-wwe-gif-17536585

[floral mulch]

busy on a side quest

training a scorer for another side quest

[languid cliff]

[wide beacon]

https://civitai.com/models/1571426/urangdiffusion-eqvae-alpha

[floral mulch]

orangutan

[wide beacon]

bruh

apparently dataset cutoff is end of april

[reef laurel]

https://tenor.com/view/always-has-been-among-us-astronaut-space-betrayal-gif-23836476

[wide beacon]

and they are training a vpred version so at least there's that

but since it's rouwei i don't have high expectations

[reef laurel]

why are these stupid people still dumping compute into sdxl

i dont get it

[wide beacon]

because too lazy to train a new arch since there's a decent base here

[reef laurel]

sdxl has been squeezed to hell and back

theres nothing more to gain out of training it

[wide beacon]

we should cosXL

[reef laurel]

[wide beacon]

I don't understand why these trainers use epsilon

like come on get with the times and fix your settings

[reef laurel]

tell that to the thounsands of retards on civit that are still training on illust 0.1

[floral mulch]

waiting for good lumina checkpoint

[heady echo]

omg revive

[feral elbow]

anything good recently? are we back to NAI being better than anything local again?

[floral mulch]

don't think there's anything new, only thing in the horizon is lumina 2 and illust 3.5

but who knows when

[olive onyx]

I cum here once so often but no jew marge updates

[floral mulch]

u r jew

[olive onyx]

but I am old how can I be jew?

[untold oriole]

Just be new

[wide beacon]

https://civitai.com/models/846917/raehoshi-illust-xl

[floral mulch]

there's also

one that is written really fucxking cringe

but says can do amlost up to 4k n ative gen

I forgot what it was

fuck

wanted to try it

but busy making a scuffed ui for the optim lib atm

[wide beacon]

holy mother of blur

[floral mulch]

frosted glass for eeveryxxthing

[floral mulch]

ok almost 000000 but still consistent 2048 capability

[floral mulch]

any cool loras worth to have

[wide beacon]

this one trained on froyo's art
https://civitai.com/models/1645130

it's neat

[floral mulch]

does it work when u dont have fr0y0 in the prompt

[wide beacon]

i'm not sure but it seems to work just fine without

it can also mix into other styles

[wide beacon]

https://civitai.com/models/1615794/shadowforge-naixl

eps unfortunately

[jaunty vortex]

looks cool

[floral mulch]

yeah saw

guess I'll see if it can be used

[eager cargo]

https://huggingface.co/collections/lol20250304/rouwei-karmix-686350268908adfda360455f

[wide beacon]

sounds absolutely terrible

[eager cargo]

https://huggingface.co/lol20250304/rouwei_plus_karmix_is_unknown-readme/resolve/main/Samples/xyz_grid-0001-1987759011.png (Generated by me)
https://cdn.discordapp.com/attachments/1387245081861554186/1389640404051296408/xyz_grid-0176-2727978150-9408-1097-4-64-20250702001041.png (Generated by 6DammK9)

[eager cargo]

To be honest, expect rw08_lol_0.5-yandexAl_0.3-Karmixv0_0.1-PCA-TV-Karmix_0.1, others are okay

[wide beacon]

Also I don’t like 6Damm’s stuff because his settings are always cursed anyway

Imo your grid is enough

[heady echo]

need to check

[floral mulch]

don't think I'll ever use rouwei for anything "stable"

it's good for schizo models tho

[eager cargo]

If wanna "stable" why not using Noob instead, I am just shocked after Rouwei merged Karmix, although base style is being like slop feel but Rouwei's special knowledges works still very nice

[floral mulch]

lmao, can barely tell from this img which ones are shit at 2048 gens

gonna upload all 3. 2nd has animagine 4.0 and illust 1.0 and 1.1, 3rd same + 2.0 and 4th same as 3

I used 2 mostly since april, and 4 in the past few days. 3 passes 2048 gen tests 11 out of 11 times in a row but it looks like that so it's just an ingredient at best. and I'd say 4 is a bit saturated but at least usable

2 maybe best if u don't give a shit about high res

[wide beacon]

the usnr dreambox guy updated one of his merges
https://civitai.com/models/1248959/ntdmix

oh wtf civitai.com not blocked anymore

[jaunty vortex]

https://huggingface.co/rim0/GENESIS/blob/main/Art StyleV3.txt list of the artists 🤔

not bad

[floral mulch]

gonna yoink them

list of artists for what tho

maybe not

has shit thats already in noob

[jaunty vortex]

most are already in noob but there's a few artstation landscape

[floral mulch]

seems tough to pick out

[jaunty vortex]

just look at the names that are actual names

[floral mulch]

also manual download lot of work

[jaunty vortex]

yeah

maybe when I get around to training my paintings

[eager cargo]

After few days, a simple civitai article for these merges got released
https://civitai.com/articles/16659/hugging-face-exclusive-merges-rouwei-karmix

[floral mulch]

oh yeah, here too https://civitai.com/models/1217645

[heady echo]

yay

new merge

[lilac shard]

Heck yeah

[floral mulch]

what new characters should I add to the dataset?

or old

@wide beacon @lime sandal

[lime sandal]

fufu and chinatsu also the new genshin loli

also the tower of fantasy chars

not a lot of models have those chars available

[floral mulch]

which ones, and wha genshin loli, they all look the same

[lime sandal]

from Tower

[floral mulch]

need the names

[lime sandal]

this one called ruby

the genshin one is called aino

and this one is carrot

this one liu huo

they also announced this one but havent released idk why prob lost the license

https://www.youtube.com/watch?v=fe52YLfPAKU heres gameplay idk what happened

oh right looks like they did released her,they just changed name to huang

[wide beacon]

the Reverse:1999 cast

[floral mulch]

got colphne florence leva fufu qingyi encore verina youhou roccia

I'll see if there are enough art of some new zzz chars

oh and cartethiyxa

[lime sandal]

pog

[floral mulch]

huh

there are dozens of great artists in wuwa tags

concentration way bigger than some other games that get a lot of slop artists

[lime sandal]

what about tower of fantasy?

[floral mulch]

dunno the names of the chars lul

anyways just wuwa and some artist pics from wuwa artists = 800 + images

[lilac shard]

You mean fartists?

[jaunty vortex]

1st genuinely good

2nd is kinda questionable if you take away the colors

[floral mulch]

Colors are part of why it's good

[floral mulch]

fuck you

[wide beacon]

it's fixed

[lilac shard]

It's fine guys

[floral mulch]

it's real

gonna merge it now

don't have to rely on illust 2.0 anymore

it's better than it

[wide beacon]

Great

[wide beacon]

hannibal artist https://safebooru.donmai.us/posts?tags=kyezzzz

[floral mulch]

is it

pretty good

gonna add it

added a bit too much slop from blue_archive tag

but wanted the new characters

[jaunty vortex]

wow seele previews look pretty good

[floral mulch]

It looks pretty ass in xyzs comparisons I did

Like, it's very similar to 1.0

But, it can do 1536x2048 no problem

https://cdn.discordapp.com/attachments/1379228092408528936/1410674815093706892/xyz_grid-0007-2113177592.png?ex=68b52c8c&is=68b3db0c&hm=341ce05f9a9e48bd38f537f805bfefc27959905629bd927ce18432d71fb76b4b&

https://cdn.discordapp.com/attachments/1379228092408528936/1411030853269852170/xyz_grid-0025-4190344054.png?ex=68b526a3&is=68b3d523&hm=110c2984e76de85796cf3d211c334c765017527bc199b2aa1d4a000259709d04&

[jaunty vortex]

seele dude put multires noise 😔

pop3 looks nice

[wide beacon]

Either ws or pop3 yeah

[wide beacon]

So basically higher res noob vpred 1.0

[wide beacon]

Yeah imo either ws or pop3

[lime sandal]

still waitin for pop5 release

[floral mulch]

why5

[lime sandal]

cuz i like 5 and 3 gib both

[jaunty vortex]

post 2048b as comparison too

[wide beacon]

and compare only ws and pop3 instead of all of them

[jaunty vortex]

pop5 might be ok

[wide beacon]

it messed up fingers

[floral mulch]

ws is not good sadly, can't post nsfw pics tho

[jaunty vortex]

3 default winner bc all the competition is worse 😔

[lilac shard]

Damn

[floral mulch]

these aren't the end results, just 0.5 runs, I haven't tested the optimized versions

[feral elbow]

where to get seele pop 3

[floral mulch]

nowhere

[feral elbow]

thanks for the help 👍

[lilac shard]

Very helpful

[floral mulch]

uploading latest tests cuz need space

pop3 - 42 + 6 solo
pop4 - 24
pop5 - 34

pop5v2 -
popv3it1 - 5
popv3it15 - 2
bmit32 -

[olive onyx]

why is there no aesthetic lora's for nboob

[wide beacon]

Why do you need one

[olive onyx]

because I have dataset but no lora

why no one else make lora

[wide beacon]

ask someone that trains lora here

e.g. coco, hugs, etc

[jaunty vortex]

I will train it too if you give me it

[olive onyx]

i think i got the trainer working

[floral mulch]

far from 10k images still

idk if I'll even get that high

[olive onyx]

image dataset ?

[floral mulch]

yep

[olive onyx]

very cool

[olive onyx]

Do you plan to run a fine-tune on it ?

[floral mulch]

glora or lokr

full finetune doesnt sound safe

and it would be slower

[jaunty vortex]

what about BOFT or DoRA

[floral mulch]

Takes longer and uses more resources

We'll see

Everyone uses glora now

[jaunty vortex]

what's the difference

there's too many of these variations

[gritty lynx]

glora trains two loras A and B for each layer and inference looks like y = (W + W A + B)x

[floral mulch]

made some lists for what I have so far

[lime sandal]

👀

[olive onyx]

did you finish gathering dataset?

[floral mulch]

No, there's always more stuff and I still need to gather and make some ai gens

[olive onyx]

Why AI gens?

[floral mulch]

Because I want to

There's also some custom things I want

[olive onyx]

[wide beacon]

https://civitai.com/models/820208

obsession vpred 2.0

[floral mulch]

Whaaa

db September

[olive onyx]

doesn't know new zzz characters

[wide beacon]

yeah not the goal of his finetunes

[floral mulch]

seems like it knows wuwa

[wide beacon]

i was hoping to convince Raelina to train on noob vpred instead of illustrious

[floral mulch]

ew

[floral mulch]

tbh not sure it's worth polluting my model at this point before finetune

but I'll see for a bit

[olive onyx]

this is a problem all mergers deal with. You never know which incest session will lead to the genetic anomaly and give you a healthy model

That's why you must keep plapping out models till it works

[floral mulch]

honestly just not really feeling it at this point, like

I got what I wanted which is high res stability, to the most I could out of merging anyways

I like the look of the merges I have

maybe I'm hoping for some slight knowledge update before finetune

idk

[floral mulch]

got a decent amount of the new character from just clip g l and attn2

of course it's not very learned even in obsession

but I want to milk it for what it is

gonna try to get more of the outfit

[olive onyx]

how is the milking going ?

[lilac shard]

The what

[floral mulch]

Idk, not really going for it anymore, don't see tangible benefits, just gathering my images

[lilac shard]

Ok

[olive onyx]

Seele 2.1 out btw

[floral mulch]

yep

[floral mulch]

it has better high res stability but it's a downgrade otherwise

shame

[floral mulch]

there are some weird dupes

300kb vs 3 mb

whateva, taking the 300kb one here

no watermark ez

[lilac shard]

[floral mulch]

weird variations in some

@jaunty vortex do I keep both

[jaunty vortex]

why not ig

[lilac shard]

Yippee

[floral mulch]

this is too similar

[lilac shard]

Aside from lighting that looks just about the same

[floral mulch]

good thing I checked

have way too many duplicates, at least 100 deleted so far

[lilac shard]

Wtf

[lilac shard]

Hell yeah

[jaunty vortex]

same

[floral mulch]

eq training not going well, idk what I'm doing wrong but all configs do the same shit, either stay noisy or become clear with the images becoming broken gonna try on vast ai 1 more time then just stick to normal finetune for now

[jaunty vortex]

ask astro if you can train on his

[lilac shard]

Oooh

[wide beacon]

new ICBINF version https://civitai.com/models/1905816

[olive onyx]

training multiple loras with similar style seperately or training loras with different styles altogether as one?

The goal is aesthetic lora as usual

[floral mulch]

One Lora sounds like a better goal

[lilac shard]

Yea

[floral mulch]

https://civitai.com/models/808215/illustriousnoob-vpred-xl-styles-collection

[wide beacon]

https://civitai.com/models/2093136/rouwei-16channel

[floral mulch]

too jank now but it could be good if it gets more training

[feral elbow]

what models are you guys using at the moment?
I've been using https://civitai.com/models/1217645/sih for the past few months, wondering if there's anything better

[lime sandal]

i use sih pop3

[feral elbow]

I know of sih

[feral elbow]

what does pop 3 mean in this case?

[lime sandal]

the name of the merge that imi was testin

[feral elbow]

he doesnt seem to have posted it though

[lime sandal]

its on hugginface

[lime sandal]

https://huggingface.co/Enferlain/juice/blob/main/noob/2048/v2/21862-seele_pop3.safetensors

[feral elbow]

thank you🙏

[lime sandal]

[feral elbow]

alright interesting I haven't got results like this in quite a while

normal sih and other models work for me

[lime sandal]

its a vpred so u have to enable that on forge idk what version of forge u have tho

[lime sandal]

then idk it works for me