#218 xl
1 messages · Page 5 of 1
same as here
but if i fix it it will be nice, I like the images as they are
BRUH
vlad merger also broken
FUCK
spent the entire day debugging and I'm still rtarded
ask someone that knows more

@floral mulch version of pony that doesn't need quality tags
https://civitai.com/models/280344
I mixed these two models listed below 50:50 In the was shown in the attached Image and the result has been the best model I've ever seen in my two ...
I'll look at it TMRW but holy fuck I'll take anything at this point
i heard that today u can find a vewy hapi neko here that wants to greet everyone!
I dont want that confetti shit, but the other merge it uses might be worth a try
counterfeit guy also came out with an animagine thing
Have you read the Prompting Guide before use? Introduction Hugging Face : https://huggingface.co/gsdf/SwampMachine/tree/main This model is finetune...
I'll use this, confetti, aid
heart of apple, bulldozer
and 2 other scuffed merges I have saved up
and that should be it
I just hope that lite pony mix is not shit
pretty much limited to using non weighted sum with pony related merging
top_k_tensor_sum distribution_crossover and rotate
can do more of last 1 now that it only takes 2~ hours instead of 12
i'll leave one on for the night
i heard that illy nee nee is vewy hapi and has a vewy comfy day today!
Might be worth looking into this finetune of CC mix from Advokat
https://civitai.com/models/291401
VivaldiXL Open DallE 3 Anime Model This model was fine-tuned with DallE 3 output. The base training model is the amazing Confetti Comrade Mix . Tha...
Not a fan of the confetti aesthetic, but it is the best quality one from the models that used pony
Yeah not a huge fan of that style
Apparently the way the clip tokenized works means that prompting will always be kinda broken on current sd models so you can't reach dalle prompt understanding either
looks pretty original
post more pics 😄
Smol floofy mofus are gibbing huggies to everyone to ensure everyone hab good day!
you say that ever time 

i did try w that v6 neg and thought it was good
but i didn't test too much
like i think need to change the negs
it has been downhill since 5.2
52
71 vs 52-62
compeltely lost the style
if u replace clip with the one from previous versions
its almost the same model
thats crazy
Merge clip at weight sum 0.5 0.5 with Animagine
ist alright but not perfect
I want a few other things from animagine
but should be fine manually
right now I'm trying to see which blocks help with char knowledge
other than base
cuz I'm pretty sure a few of them helped
3 has been the best in this test
4 looks more interesting but the hand position kinda wrong
I'm not going for absolute perfection cuz there's too many variables for manually testing, especially without having any guidelines like for 1.5 models, and I'll merge further into this as well anyways but it's always nice to have a decent checkpoint. Best part about this method is you always have something to fall back to when you note down working recipes
Yeah can't tell just from a few cherry picks
That too
Unless one recipe is literally 2x better
Once it gets into "uhh is this one slightly better" territory it's tough
Feels like anatomy fluctuates, but what's more static is style and concepts once you find the right blocks, just what sucks especially is it seems to be also a combination, not just x does y
Yeah
I noticed stable merges tend to end up looking more boring
Well I think it's harder for the model to make nonsense if it makes mostly the same types of poses
Not sure if it worked like that on 1.5 but here if you change only the base the entire image will look different
Hmm which one is the base
Like for aid71, changing it to the ones from previous versions, changes the entire look of the model to the same or similar as those precious versions, even though it has aninagine in it
Previous
I saw you said that earlier
I think aid got the styles by taking the clip from 1.5
But also makes sense since it should kinda be the same model
Dunno what they use, they probably train the te to get the styles for the artists?
I think Euge said they took the 1.5 aid TE and replaced the compatible one into the SDXL model
But we'll yeah they kinda failed with the animagine merge, lost one of my fav artstyles, who knows how many else
This was during 0.5 I think
I c
Interesting
Well it's a decent version
Also does alright natural language
I'll prob keep it at like .7 since I'll do more aninagine side merges down the line anyways
Ill need to try that faster webui version, maybe helps speed up model loading etc
Animagine has great poses
You can buy ram for CyberWaifu
he'd do it

ribbons look cool
So much new text mofu overload but mofu thinks everyone should be hapi and glad!
sadly tons of errors but fun
almost pog

love it

heard smug mofus are here to smug on everyone into hapi!
AuroraXL is out 
https://civitai.com/models/301957/aurora-xl-basepony-xl
The artistic style has been ft base the Pony XL model, following the guidelines of the Animagine XL V3 caption rules, by training with the collecti...
still uses score tags

spent the entire day trying to make a gradio theme
dont give a fuck anymore since it's trash
I couldnt even save it
:0
it's a joke btw
joke model ?
or someone made a model called aurora but it's unrelated to aurora 1.5 ? 
it's a pony finetune trained the same way as Animagine
but of course a finetune will not remove the need for the score tags

no Rella as artist 
Long time no see imi
Last ones kinda bland
it's not aid related
heart of apple 2 and swampmachine 2
both are like 90% animagine
yep
swampmachine is counterfeit guys finetune
tried forge for merging
with bayesian
worse performance and crashes
but at least it starts
The code is kinda different so I’d be weird to just work ootb
(at this point it’s time to do it on comfy)
The memory leak is scaring me away from trying
It's pretty bad, had to do a lot to make sdxl work and even now it might be broken, and that's not even related to the memory leak. I'd use sdnext but I'm not even sure it would work there since the api might be different in addition to meh
hope everyone hab best day!
I fixed glitches (or made them less visible) by merging 2 broken models together
vs
ok I removed
[monochrome environment] since mangas have the type of brushes that cause stuff like the artifacts I've been seeing
conclusion, animebulldozer is broken
most of the journey is done
ngl I think aid v6.1- has away better artist style support than nai3
But ssamabatea
That js2 pony model actually looks a lot better than normal pony from what I tried
Dunno about capability but seems close
js2 ?
Yes
What’s js2
Model name
Uh
will need to try after work
chunk
doing a few more xyzs, need to decide on which ones to use from these checkpoints

@heady echo got any interesting párompts to try? 

let me look
1girl, solo, long hair, looking at viewer, blush, bangs, brown hair, long sleeves, dress, bow, ribbon, holding, bare shoulders, brown eyes, jewelry, very long hair, earrings, frills, parted lips, detached sleeves, wings, pointy ears, blunt bangs, black bow, detached collar, wavy hair, red dress, frilled dress, crown, skull, cage, birdcage
1girl, solo, long hair, looking at viewer, blush, bangs, blue eyes, skirt, shirt, long sleeves, hat, bow, school uniform, pink hair, pleated skirt, serafuku, sailor collar, blue shirt, pink bow, white sailor collar, outdoors, sky, cloud, water, blue sky, sparkle, night, ocean, star (sky), night sky, starry sky, horizon, shooting star, photo background,

@heady echo @wide beacon @jaunty vortex if u guys have time can u check the xyzs and tell me which ones you like? ignoring the glitch part, I think I can remove it but I need to decide on one
surez
maybe 4
think I fixed it
also made it darker
and maybe more scuffed cuz of over details
but its just one ingredient of the 3 so I'll leave it at that
all I wanted is the cringe artifact to be gone
the images darker cuz of stuff, but I think its fine
@heady echo your city photo prompt for example
previously

wtf 

suddently thinking
"did imi merge w pony"

"was the color lora that strong"
wait 
@floral mulch

struggles with denoising
so you might get extra detail like those add detail loras
which includes fingers

but its usable I think
its just an ingredient anyways
the final merge will be aid + this + pony

this was pure animgaine
hm
its cute
I guess the merge is "harsh"
heavier on style maybe
all 3 models used are further finetunes of animagine

flandre pic
awesome 
but yeah its a bit tough to rein in I think
had some issues during inpaint or some inconsistencies
gonna look into it tmrw
i havent tried inpainting with it yet, i really like it
6
but I guess its kinda aschizo prompt
hand up, covering face + fingerless gloves, hiding behind carrot
I honestly love this model, the prompt i grab left and right from other images 

this one is yours 
waw
by potg is my new liked artist tag
kjut
I've been happy wth the merge
animagine is probably capable of all of this but I didn't try as much
or maybe the finetunes add a lot
imi merge vs animagine
I wonder how I'Ll merge this with pony and aid

hmmm, for some reason this model really don't like any sampler with a M.
yes
I tried 2m karras and it was relaly harsh
probably because there is a lot more noise
I've never seen a model that breaks when using 2m karras or other samplers so at least I made something new I guess 
yeah looks like only the ones that usually have added noise work
Umbra mecha works very well with DPM++ 2M SDE Karras from what I’ve tested
Terrible with Euler tho
ngl ive been using euler
model is horny wtf
Chu mean
Completely sfw prompts become quite a bit horny
This was the least weird I got with this prompt
#ai-art message
As I said this was the least weird
Other gens were kinda horny
3rd model is the one that was broken
middle one used it as model b, transferred brokenness
1st one also uses it in a capacity
not broken
at least not visibly
umbra step 1 vs normal step 1
you can kinda see its the same start almost
but it ends wildly different

i heard some extremely hapi mofus are lurking here tonight!
if u lean back they will gib u huggies!

cute cute
new AIDXL to try https://civitai.com/models/124189?modelVersionId=372261
Model Introduction (英文部分) I Contents In this introduction, you'll learn about: Model information (see Section II ); Instructions for use (see Secti...
It's very wonky or idk how to use this version
vewy vewy floofy and hapi nekos are here requesting kindly pats
AIDv8 is good again with the artist style
and I think its even better than the V6
I tried Mecha umbra. It's pretty good but the old SDXL neg is fucked
and euler a 28
just stick with worst quality, low quality, lowres, bad anatomy, artist name, watermark
have to actually use my words now
can't wait until the not scuffed mind-machine interafce is here, can't believe i actually have to type out words smh

real talk
btw we can merge pony
just 0.5 somehow works and doesnt look bad

but you lose stuff from model a and maybe model b too
isn't htat normal with merging? you're just averaging out weights innit?
yeah
well
I tried just individual keys from the first te
BASE:position_embedding:1,BASE:layer_norm1:1,BASE:layer_norm2:1,BASE:self_attn.k_proj:1,BASE:self_attn.out_proj:1,BASE:self_attn.q_proj:1,BASE:self_attn.v_proj:1,BASE:mlp_fc1:1,BASE:mlp_fc2:1,BASE:position_ids:1
token_embeddings ruins char knowledge, but even without ayanami still has animal ears most of the time
token_embedding does the worst
are you just systematically merging every block?
no, just swapping keys from pony
these are just 0.5 weight sum tho
@opaque mantle you're the expert in x/y chart merging, what do you think

hm, hard to tell since I hadnt merged sdxl yet
its a double barrier, sdxl is 1 category and then pony another 
yee 
did u see what yuno wrote
模型融合死路一条,画风多样性和部分提示词的准确性一塌糊涂。
model merging is a dead end. styles are shit and accuracy is all over the place
另外有些人差不多得了,太阳不是鸡叫出来的,没有好的训练模型融个P。
anyway there are some people who almost got it, but the rising of the sun is not because of the rooster's crow, there are not good trained models to integrate
Yuno779
AnythingV3 merger
ah
well dunno how good he is at merging
anyv3 was 1+ year ago
and his anythingxl merge is just 0.25 and 0.5 between models, nothing specific 
SDXL weights are kind of fucky to merge though
you could basically merge any two models on sd1.5 and it wouldn't shit the bed too hard
SDXL sometimes becomes a real shit garbled mess
Cuz 1.5 was all nai based
yeah
Well I mean even for the weirdo finetunes that andite used to make
Lora merging in the huggingface thing https://github.com/huggingface/peft/releases/tag/v0.9.0
I sitll dont know what do shits do
post pic
mofu heard hapi people were here!
Next time
I'm back to merging aids now
Gonna do it auto since the bad ones are diluted enough now
V8 seems fine anyways
Just wanna include my previous good attempts
v8 seems good it's just missing that shiny polished look
ooo 
this was my 218 vs v8
like 218 isn't as washed up
v8 did get the red turtleneck tho 
by ajimita, 1girl, black hair, black eyes, red eyeshadow, red turtleneck, red midi skirt, cowbot shot, street, japanese street fashion

looks weird
not really satisfied with this merge so far, some looks great but others not so much
goat dog 
cow
cow
hmm
replace it with kohaku delta 
I like it but I think aid has more and better styles
I tried kohaku delta it's a little incoherent
very ncie
the imi cooking den
I wonder if I can do 3 chars with a decent winrate
lmao, cant even generate the angle halo anymore, I blame blue archive
yuor angle?
try angel instead of angle

ohshit, im dumb
but wouldnt be surprised
blue archive was probably like 10% of the dataset
or more realistically
Is there a list of artists that umbra mecha can do?
not particularly, just try whatever is on animagine 3.0 or 3.1, there are too many finetunes in it to keep track and I dont even know if they published their artist lists
oke , what finetunes did you include?
heart of apple v1 v2 swampmachine 2 animebulldozer 2
oke
aid8 finetune when
No sure how much AIDXL retained, but here
https://civitai.com/posts/1663973
A post by Hannibal_Lector. Tagged with . made if for myself but maybe it would be useful for someone favorite model tbh
theres no aidxl in umbra, although its planned but not sure when 
it's what Animagine 3.0, Bulldozer and Heart of Apple can do iirc
so all you need is the list in #1197939484638064750
yeah basically
Where swampmachine
ok
actually got to merging pony and animagine now
and it works
but it's still not gonna be like what I wanted, so I just want a small % that can be finetuned on top of
it works as in some of the concepts work in umbra without destroying how it looks
woah
soon(tm)

I still gonna try some more stuff tmrw to make it stronger maybe, now that I know what methods work
its just about finding out the most optimal parts
this model is the goat...how can it rival some of my NAI gens I don't know
are pony loras compatible?
no
probably not
some of them might be at weight 1.5 or 2, or if u prompt all characteristics of a character
if u mean 218xl it's pretty old now in terms of ai shit, but yeah it can still make decent images
animagine might be
I see, it still holds well enough I might say
Holup isn't animagine trained on raw sdxl? 🫠
yeah in the same sense novelaiv1 is trained on raw sd1.5
for now I'll try to get a model thats between animagine and pony in terms of concept knowledge
but tbh umbra is already goat so I won't really take anything if it means ruining the artstyle in it
what do people use for dataset building atm? just grabber or something?
that's what i was told if you're grabbing from dan
floofy furballs are here hugging
@floral mulch https://civitai.com/posts/2122190
A post by Euge_. Tagged with illustration, character, and anime. 🎉 Exciting News, Everyone! 🎉 🚀 Our upcoming model is set to revolutionize your creative experienc...

bro CivitAI notifications panel is so fucking useless
it pinged me about everything except the post
anyway if this is based on Pony I'll off myself
seems to be just AidXL but idk if it's a restart or a continuation
@nocturne cosmos new AidXL soon and seems decent
This is the single weirdest romance gen ever https://civitai.com/images/9777796
doubt
plot twist it's the secret ingredient

Nai3 leak is real

oh wth i didn't even see the 5k
yeah and it's the same for AidXL
there's no way Euge trained 5k artists that all work properly
sure
altho I Feel like theres still more in aid than animagine
at least ones that are stronger
also hard to find the stronger ones in the comparison thread
cuz theres too many
I'd only have noted the ones that do something
imo, art styles in aidv8 is the best in all aid series
ill be happy even if the new model has 50% effectiveness art styles of the aidv8
I'd say 50% is lower than Animagine
Animagine might have 50 artist styles are >= 50% of effectiveness of aid
nanashi \(nlo\)
https://civitai.com/models/410737/neta-art-xl #ai-art message
⚠️ Attention: Please use CFG = 11 as default for testing. For full release note: https://nieta-art.feishu.cn/wiki/PpwqwVDzjiNE5kkUhRtcEsn6nmh I. Ove...
has cringe 3.1/aid6+ like artifacting
new AidXL when
yeah not a fan of neta look so far
kosa is better looking from immediate tests
forgot what it was made of tho
kosa vs neta
Could still be good tho, AID tend to generate low contract and orange-ish img, netaArt could kind of fix that.
I like that look
Yeah neta is less aesthetic there
Damn every time I see ppl draw violin playing it's so inaccurate
What's the kosa model btw
Thanks
Handy
are there any else model merge AID?
(by rella:1.2)

check pinned
New aidxl when
i dont like how neta looks
AWA Diffusion We have released the A rti Wa ifu Diffusion model, designed to generate aesthetically pleasing and faithfully restored anime-style il...
new AID
woah
fresh out of the oven

@wide beacon @floral mulch
yeah getting it
new aid?
yesh

Hmmm
One of the comments asked why they couldn’t reproduce the hitori slime and Euge said it’s Euler A + Hires fix with Dpm++ 2m
It’s even worse than you thought b74
Also the guy saying that no futa = drawback
- What is the relationship between AWA and AIDXL: AWAv1.0 is AIDXLv1.0
Honestly not as bas as I first thought
@floral mulch you seen this ? https://civitai.com/models/402492?modelVersionId=489098
this guy tamed the pony 
All sample images generate directly with hiresfix and afterdetailer, without loras controlnets and inpaint Delice models pointing to the slim teena...
Can it do complex anatomy
from model page it can do lifted leg and feets
it did anything i through at it atm
it nails the hand and feet pretty much 90% of the time
95% even
does it still need score tags?
from model page it seems so
oh god it can do centaur
⁉️
the backgrounds are amazing
currently checking newly released SDXL models
there's one called MooMoo
I don't want to click on the censor
Says it's a merge ??
it says " DeliceXL: SDXL version of HadrianDelice based on ponyXL"
his Hadrian Delice is a sd1.5 checkpoint

I'll tkae a look
need to find situations in which umbra loses, cuz other than missing knowledge I haven't had a comparison of the same shit where I preferred the other models over it, otherwise I'll just have to start finetuning
oh shit I havent tried extract loras
probably decent idea for adding new knowledge
seems like neta doesnt know daizu (melon-lemon)
gonna run through the names from the b74 post in arti
there are still 2000 more but I don't have time to look at those today
too much to test

also wtf
cant even see the face

but seems like all of them can do it
I'm removing the left merge cuz its completely fried

1girl, ayanami \(azur lane\), azur lane, by coldcat, solo, white hoodie, rim lighting, best quality, masterpiece, absurdres
pretty clean
Second and fourth one are neat
typoing daipunch to daipunh
guess its like a base style? dunno
arti winner finally
dankmill instead of dangmill
btw "rim lighting" works well on aid models mostly, not neta tho
neta pretty clean on this
ok with some styles I can celarly see where they start appearing
dino

dismassd
do m kaeru
this ones def a w dr mice
Style transfer failed
havent tried style transfer yet, just random merge
tough without overshooting
and ending up with
although some kinda looks cool
but too far from origins
Should still keep it on hf
Use your test folder for once
i feel like at this point the cycle of imi is
"we fixed prompt space"
"nvm"
Private repo
its "fixed" on a bunch of models now
my main problem is still merging
I dont want to use pony
no im lazy
ok gonna find which tag causes it to fucking shit itself
every pony model
going through my prompt 1 by 1
ces is actually shit
as in
the ponyguy trained scat into his model (least derranged brony)
ok the models die
when I go above 75 tokens
or rather when I Write anything after BREAK
bronies be like:
can u confirm
Broken clip?
like write 71 tokens worth of prompt
BREAK
then rest but not higher than 150 combined
gimme ur prompt
You know Euge is using the clip transfer thing
1girl, 1boy, ayanami \(azur lane)\, azur lane, by quasarcake, standing on a dock at sunset while seagulls are flying, best quality, masterpiece, white trench coat, blush, closed mouth, outdoors, scarf, looking at viewer, perspective, (pov), simple background, hat BREAK (seaside resort:1.11), [beach], leaning forward, humming, half-closed eyes, best quality, masterpiece, absurdres
tried replacing almost ever tag
and still happens
if I remove the part after break it goes away
no shit after break vs full
@frail zealot
nah
its xyzt
being a broken
piece of garbage
poisons the loaded webui instance until restart
none of the models are 1:1 as they should be after the poisoning
cant be evaluated
so far this hadrian pony model seems pretty meh
pdfv2 and tponynaiv41 seem better
Kitarz just dopped a Kohaku Epsilon finetune
https://civitai.com/models/448552
Starry XL enhances the Kohaku Epsilon model by specifically targeting the styles of Pixiv's top artists and expanding the character dataset to gene...
I always thought Kohaku Epsilon produced kind of worse results than animagine 3.1...
Well who knows results might be cherry-picked
Can’t tell until we try
nice
umbra was the best sdxl model for coloring in my lineart in an aesthetic way
worst one I tried was cosxl (undertrained model) vs umbra
the starry night epsilon model does ok at this too
cosxl reminds me of trash merges between animagine and pony
umbra noise ratio will add a lot of details
sometimes unneeded unasked for
I did like 4x noise on some layers
tbh I still enjoy that the most out of all the models i tried recently
I'm trying to merge some awa/pony merges into it now but so far it's not going that well
the model I last got is kionda bland
can look good but
maybe it's just the artist tags I spammed
It's a tradeoff, for a coloring model the noise is good bc it's easier to remove detail than add
"Bland" looks more like actual art ppl draw
eh yeah my style is more contrasted
yours looks nice
I like stuff that's not as smooth, it's just whenever I tried coloring it it was shit
especially witzh fill tools
but then I saw hacks like just rough layer and then you modify the color
and then u fix the out of bounds stuff easier later
I always just used to manually do the colors
I was thinking that it wouldn't be that hard to make a program that can take a psd file with lineart and color layers, then crop around objects with auto tagged prompting at high res + manual fixing prompts, call comfyui api with a lineart controlnet and then put that all back together
mistoline
better than the other ones?
tbh I didn't really try other ones that much
I just heard sdxl controlnets are bad
I would send you examples but it's nsfw 😂

cuz my friend asked me to draw her OC fucking kanae
thought those were both nijisanji vtubers on that pic
cute
I probably have it somewhere
brushy brushy
looks liek this
might be something else
I haven't tried the mistoline lineart preprocessor (anyline) might be pretty good
I downloaded it but
havnet tried yet
maybe Icn get back into drawing if it works well
since I'm tired of gennig with just prompts anyways
This controlnet is a scribble model in the SDXL series trained with an anime style. It specializes in completing sketches and regenerating drawings...
try this one out, its working well for me
Can I use this one in auto1111 webui? nvm it doesn't support webui
i dont want to use comfy
this is not comfy
i couldnt get the preprocess to work in comfyui either, i couldnt find it
it's really good
so clean
This man is into oversized hoodies

💀
I wrote safe in umbra
to get non nsfw pics in dtg
and its generating safes
i ll try it tomorrow it's getting quite late here, thanks
new imi merge? 
just messing around
I cant find any models that would be nice to merge
kinda bored atm
what about a lora with that model to fine tune it to a style instead of having to do the style weight juggling ?
ideally I want all of my favorite styles from aid
which none of them are replicable atm i animagin e models
in
almost none
and not to the same extente as aid
I'll try merging with my aid merges soon ish but not expecting great results
did you see https://civitai.com/models/448552
Starry XL enhances the Kohaku Epsilon model by specifically targeting the styles of Pixiv's top artists and expanding the character dataset to gene...
maybe some way to extract the style finetuning idk
I still think aid is the best
at least for styles
especially the ones before the noise issues
yeah, all other models are not even close
maybe with merging and finetuning on top ?
every time I look back at my aid pics I'm like damn
I think finetuning is probably the way to go
I can probably use the same method kohaku did
but on a smaller scale
just dont like not being able to use my pc for other stuff
although I think it probably won't take much more than 14-15gb vram
can't watch yt
ccan still play stuff like hsr with 14gb vram use
I only have a 3060 so it really takes my whole pc
saj
I still remember when I had an amd card
could afford to use this one for more
AID has better colors than 3.0 but 3.1 is closer
0.51 AID
yeah that was the first good one
raw jumbled mess but the rendering is good
AID 5.1 + animagine 3.0 hires
AID 0.8 + animagine 3.1
another 5.1 + 3.0
there's good colors
Hmm
I don't want to go back to aid without the knowledge of animagine and pony
But I want the artstyles and prompting kinda
i used a mix of euler A and dpmpp_2m_sde_gpu
inpainted only the eyes, rest is as is from #1179290821409910844 message
it's so clean 
idk how you do it imi









