#218 xl
1 messages · Page 4 of 1
too bad we can't make accounts on tusi
i disabled my chinese phone number when i went back to the states 
yeah lol it would just get reuploaded somewhere else
well
i'm just going to take my sim card with me next time when i leave
and drop another like 480 yuan in the account so that it stays active
it's such a pain in the ass trying to access chinese services when the phone number is disabled
It’s on tusiart.com iirc
what model? looks like a nice artist list
AidXL so also 218XL
about v61
Stronger stylization.
Additionally, I added additional noise in the training. Some samplers do not reach zero timestep at the final step, thus potentially causing the generated image to be noisy. So, Euler A or Euler sampler may be more suitable for you to use

wtf, every chinese thing needs phone number
u goofed
That’s what fadingz wrote?
you can be a vtuber and still be braindead
actually the correlation might be higher
Indeed it sure might be
I would say one encourages the other
Probably Why futakuchi mana hates when someone calls her vtuber 
I reserved the number with China telecom so I can just reactive it but yeah
yeh
oh wait
i actually have the account with china unicom
Damn that logo goes hard
but i was using Google Fi for roaming and that is on china telecom's network
got it mixed up LMAO
ah
Kinda a shame the Chinese covers were in audacity
What a beautiful song this is! I ha d a lot of fun working on this project.
Source: https://www.youtube.com/watch?v=W8x4m-qpmJ8&list=PL6x8eYAHnR-0Lxmt99U1QD6AgLuafnLDF
Model made with help of RVC.
Here are links to my socials:
Twitter: https://twitter.com/Rakosz1
Pixiv: https://www.pixiv.net/en/users/88050219
Patreon: https://www.patreon.com/...
i think hololive in general is probably kind of anti-ai
i know ina hates it b/c she draws

vtuber takes on things that actuall matter = 
they're entertainers
not figureheads
I'm pretty sure these companies use ai one way or the other
talk the talk but not walk the walk
Yeah but they won’t admit opently
ame has an ai tag so
True
i guess it depends on who you talk to lmao
Yeah I guess so

I couldn’t care less about vtubers xD it’s funny to watch them sometimes but that’s about it
Or Listen to some quality covers
Btw Imi are you genning rn?
Ah I see
There is something i need to try
Might need to get out of bed rip
This added noise is doing some weird stuff with denoising on hiresfix
Maybe if you change cfg for hires it won’t distort the images
But what is even weirder is that this isn’t issue shared with every artist , just occurs on some
Yeah if you can do that sure
see it on the merge too
It’s just like it’s kinda baked
But I didn’t know that the scope of this finetune is so big , if that’s the case it’s likely it will stabilize
Hes also doing nai approach
Chinese and filipino saving open source
I'd like to do something but dont have 1k to blow on it
Me too bro
Mhm Im kinda leading into conclusion that I was upscaling too much
Seems like 1.5x upscale is more stable
But on your mix I can upscale 2x and it doesnt break nearly as much
What about aidxl0.61?
havent tried it
most right one is nice
am i only one who likes 4th to right? 
i think it has cleaner lineart and has a little more fashion detail
It's not bad but the goal was to make the second model look more like the first
But I don't think I can do that with the methods I use unless I merge 2 into one instead of one into 2
So that's what I'll do next
I hope Lyn releases his new model soon, I think it will be a nice model and help level up sdxl open space
same
.61 seems more unstable
I'll try fixing it with 218 recipe and then merge it with the latest good result if current result will be shit
after that I'm done updating it for a while until animagine 3 or I learn to finetune
if aid is around 25k images then the costs would be doable for even just a single person
and I'd prob start less, for starters if I can run on my own gpu (doubt I can with text encoder enabled) then I'll do that if not I'll runpod/lambda etc it

when the memory leaks are fixed I'll run mbw
ayo
still not a good model, couldn't kill the shitty noise of 61 and the composition is too different from 218xl
yeah I don't like the noise
its even worse
fucking terrible tbh
1 last thing
if it doesnt work I'm skipping this version
negv6 makes the noise better
v5 vs v6

2.0 makes less sense but the style looks better
not much imo, or at least not enough to justify the error
aid already has the style anyways
Yeah might as well just use aid
@floral mulch You should probably tell Euge in a comment that the noise is very heavy in some cases
i heard that if someone pyons in this thread it will gib good new year luck!
pyon
pyon

maybe is the turtleneck ?
tetev2 nice

quality tags at front vs back
1st row: prompt, best quality, amazing quality, very aesthetic, absurdres
2nd row: best quality, aesthetic, detailed, prompt
3rd row: best quality, amazing quality, very aesthetic, absurdres, prompt
any opinion about these? @heady echo 
nice xyz
dunno why auto font resizing isnt at hing 
StableSwarm did or used to have the text at the minimum size
those look super nice 
trying to think which one prefer most
Keep it cooking

@floral mulch you should correct this
i heard dis place with everywan that says anything will be blessed with a extremely hapi day!

he uploaded it after I already asked so nothing to correct
yoo
i heard that today is a day when any floofy pyons will be awarded a big carrot!
🥕
I fixed (I think) the merger extension thingy I use, time to test 

i heard that today everyone who snuggles into there blankets really hard and thinks "hope tomorrow will be extra mofu" will have best day tomorrow 

today everyone is hapi!
hapi hapi
hapi day

@floral mulch did you not bake the sdxl fp16 fixed vae into tete? I keep black images if I use the bundled vae at fp16
i heard that alot of mofus are in this thread 
p-p-pyon!!!
Rink is trying his hardest @floral mulch
that's good
I doN't have much to talk about since I haven't done anything 
I'm testing block by block to merge with hassaku andy model
Lewd model
needs to be better at nsfw



it's for anatomy reasons 

I believe you 
did u know that if ur a good mofu today u get a extra carrot!!!! 
so @floral mulch animagine3 merge when
BEGONE COOMERS, IT'S CAG HOURS NOW
foks
@floral mulch might be worth merging with the pony model. looks like half of the dataset is on anime and there's likelyt to be way less niji bias
shiun (siun 5513) = ziyun
holy shit thanks
the when pog

havent tried the base yet
the normal one is good af so far in pretty much everything I Tried 
I'll try it tmrw
If I didn't stay up until dumb shit hours
i heard that if u open ur cupboards u can find a smol mofu nibbleing on cookies!
need some new payload prompts for merge between the models
illy nee,is,a,mofu,and,vewy,floofy,and,(cute:1.7)
base and finetuned animagine are pretty similar
bruh this model is 
definitely gonna merge this, without the pony parts
otherwise it has chars
artists
concepts
out the wazoo
it's actually artist stripped
but yeah
its understanding of posing and lighting work is ridiculous
yeah the posing understanding is abbsolutely wild
i think if you're going to merge it in though you might wannaa either train a lyco to fix its borky tag shit or do what @wide beacon was suggesting and untrain and then retrain its understanding of quality tags
the hll guy said he was going to train a lyco on this on vtai, might take a couple days. could be worth waiting
no just make the quality tags activate by default with a LECO enhancer
It's actually really cool that we have good ingredients now
cant fucking belive I overstayed because of fucking PONY diffusion
I shit you not
It can do upside down
Without looking like down syndrome
And obscure concepts like disembodied bodypart, I guess that's the furry dataset or idk
i heard theres some smol mofus lurking around here 
woah pony model works well 
i is awaiting the animagine XL 3 and pony merge
@floral mulch check out the ReprogineXL model. It's a merge with Animagine 3 XL. Shit looks cash
Reproduction_LCMv2とAnimagine V3とSD XL - DPO Finetuneのマージモデルです。 レシピはnote参照 Reproduction_LCMv2 and Animagine V3 and SD XL - DPO Finetune merged model...
the merge is more contrasted
Thought it looked like shit 
It’s so extremely noisy and the examples are dpm++ 2m karras at 20 steps
looks like when I tried merging turbo and using 2 steps
I just don't like how the hair is 10x more strands
Agreed
I'm not genning any more
just going thru the stuff I already made
and picking out the good ones
then I'll start the merge shit


TMRW for real
Though maybe bit busy cuz renovating house
But at least I can fix my sleep schedule
i heard that if ur a good mofu today u get three carrots! THREE carrots 
mofu made mofus first xl img 
oh yeah
some stuff just breaks the models
not sure what it depends on
happened in animagine xl too
actually I havent broken 218xl yet even tho I did some schizo prompts
I see ppl having trouble getting LECO working on sdxl
Ok I see the guide on civitai
I think I'm good
my only issue now is vram
Apparently works on 24gb
Oh I'll try this
so what were u planning to do with leco and pony?
cuz I'm trying to merge now and it only barely looks better than when I tried with the nekoray model
Remove the need for activation tokens
cant u combine them into one?
Wdym
I think based on the examples you need to tie those tokens with at least one other token right?
You do know how many clip tokens that is
score up 9
would it make sense to turn those into just "best quality" instead
3 tokens plus the comma token
I have no idea how to do that with LECO
There’s just no reasons to have to waste tokens on so many activation tokens
thats how it was on the civit page so I just copied it
what was that
china webui
yeah
Screenshot from that guy guide
yeah
Yeah so
that would be nice
On webui put just the activation tokens and look at number above the prompt box
Should be number/75
I will next time
positive: "masterpiece, best quality, ultra high res, 8k, clearly fine detailed" # concept to erase
unconditional: "worst quality, low quality, ugly" # word to take the difference from the positive concept
neutral: "1girl" # starting point for conditioning the target
action: "enhance" # erase or enhance
guidance_scale: 12```

this is kinda bad though bc then you only get best quality on 1girl no?
that seems pretty useful
All I’ve done is erase and enhance some concepts not merge
and what do u use to mereg into model?
supermerger should work just fine

From what I’ve seen p1at and srdb guy LECO are pretty light on SDXL
My suspicion is that they don’t train them enough
500 steps is enough for SDv1
He did not give instructions on how to 
maybe if I read the paper I'll understand this better
Anyway if you wanna try I’ve already prepared the config and prompts file
You could tweak them a bit and launch the training
I'd like to try converting the stuff to best quality and equivalents
Basically what happens for me is that after loading the model into vram I don’t have enough vram to run the training
so its compatible with other models
I think best quality is more neutral than masterpiece
if you wanted it to be like animagine you would have to run it like 5 times

Quality Tags: masterpiece, best quality, high quality, normal quality, worst quality, low quality
right now I'm just doing a test merge of an aid merge and pony
only weighted sum, gonna try others later
this was pony for me, but then it might just be cuz I used too many quality tags
Huh
I just don't like zooming in and seeing noise
I dont like it for drawings either

but like its noticable even if I dont zoom in too much
I dont mind that as much
I was trying img2img yesterday and getting frustrated
I dont like the fuzzy look tho
I literally tried to improve the quality by changing settings for a few hours but not sure how
going above the sdxl res causes this
i think it's the cross attention or something
idk why it looks fine when I do txt2img though
All have their own quirks and then there’s Naid 3
I've felt like hires fix wasnt strong suit of animagine, at least not with latent upscaler
Latent upscalers are terrible on SDXL I feel
I'm using it here
it could be better but it's still much nicer than my img2img example and I don't know why
maybe not enough denoising idk
it works fine in aid for me
use it for landscape a lot
I onlyx did i2i with 0.2
there I didnt notice any issues
but I did notice that even 0.2 - 0.3 was able to change stuff decently, which Ithought was weird
cuz on 1.5 I remember using 0.2 and it didnt do shit
it depends on your scheduler
0.2 does a lot if you use normal instead of karras or exponential
actually I don't think so
so idk what sampler they use for euler a
might be normal
In the code
then you cant

I use dpm++ 2m sde karras
idk what you think of it
I use sde for hires fix in bg related images cuz it adds more detail than what I usually use 
dpm++ 2m doesn't like to edit much
"2m sde"
sde deez nuts
restart goes even more ham
restart slow
imi has a 3090
💢
me on my 3070 mobile
I just want my pic to look really clean so I gen at hires
but actually I don't want the extreme detail you get from hires fix
cuz it's so AI looking now
restart lanterns 
my favorite schizo thing is
having low preview ms
and seeing what things could be
when it tries to denoise
I turned off preview bc I'm too impatient 😂

I tried testing perf impact on each setting
it started tanking performance towards real time pretty hard

having higher vram doesn't allow you to make higher quality pics after like 8 gb anyway
coffee inbetween each pic
I get that without undervolting but I think it's bc I have 100 tabs open
got sold shit gpu time to refund on paypal 
batch size 2
20 steps
dpm++
768x1344
40 seconds
yeah thast pretty good tbh
oh lemme try
for me 1 hires fixed 544x960 image took 1min on 6800xt
That was SDv1
yes
That was around 30 seconds for me
25 sec for me euler a
Small d (me) vs Big d
Euler is a lot faster
arent they all slower than 2 m
?
in term of speed Euler samplers have always been on top
first is euler a, second dpm++ 2m, third dpm++ 2m sde

isnt that because we were noobs
I"m doing 60 steps for sdxl
it takes 2 min per gen at 2k res
gn
😮
how much vram is this
oh ok doable
i heard that today mofus are trying to hide inside the creak on the sofa 

https://civitai.com/models/212253/bartstyledb-or-xl-or-blue-archive-art-style
https://civitai.com/models/263833/anime-antifreezing-solution-xl
https://civitai.com/models/264323/sdxlanimebulldozerbeta
SDXL VERSION HAS OUT! Don't worry about the deleted post, I just merge the same version back here, my bad. TL;DR: It is based on Animagine XL V3 , ...
This model has undergone nearly 1 million image training using an A100 server, followed by special training for artists. The total number of traini...
ANIMAGINE XL 3.0 Fair AI Public License 1.0-SD Permitting the fusion of models. Not for commercial use. The total number of training images is 430,...

I heard that today if i call out for illy nee nee u will see her tail from the horizon has it stands up to inform everyone that she has heard u 
my weighted sum experiment failed, gonna check block by block which ones fuck up
What happened here
testing blocks between an aid merge and pony
some of them are completely incompatible
on their own anyways
doubt I'll test all of the options considering there are like 1 million
glad I have 6 less blocks to go through
20 already a lot
there a floofy mofu asking for carrot 
@wide beacon can yo usend over the leco thing
i heard super mario likes mama mia!
progress 
froze shit layers, still broken image
doint get it
should I try add difference or something
Maybe can merge by hand like with 1.5 recipes?
it's fine manually
maybe changes I made to bayesian stuff broke it
not sure
I'll check tmrw
I figured out what blocks are no no
so thats good
and I have a vague idea of what each does from a few tests
probably only applies to these 2 models I'm trying to merge tho
tensor sum kindna works but I crashed it during warmup
top k tensor sum works faster but I've yet to see if I can retain characteristics
i heard mofu forgor to tell mofu wisdom of the day yday!
so mofu double whammy!
today mofus will experiance a dream of carrots raining!
and mofus awake will get there fill of carrots!

floofy furballs have invaided this place and now they are gonna cuddle everyone silly!

next few days gonna try animagine angle instead of pony
pony worst fucking model I've ever dealt with
mergewise
🥲
doing the 2nd hours long leco train
just to see if it helps
spent like 40 hours of gpu on this shit or more
they made a model unusable by model mixers
you just have a skill issue 
leco day again
need to riase learning rate imo
did 2e-4 on 1000 its
not enough
or rather higher guidance scale maybe
2 low

dont get it, another leco fail

7 versions
all failed
actual aids
I'll do 1 last
with shit turned up to 9000
and at worst I'll use it at 0.1 weight
anima pencil x aid?
whast that
The blue pencil animagine merge
I'Ll just merge animagine
dont like blue pencil
but I'll do 1 last try at this leco
without a target
redirecting the score tags into the model if that works
gonna be 4k steps
and alpha 1 didnt do shit for me
16 rank 1 alpha
even 2000k steps no effect
What about that finetune with the feets?
why the feets one in particular 
this one is a good finetune btw imi
https://civitai.com/models/272440
Introduction to This is a model based on Animagine XL V3, which has relatively good performance capabilities. It is trained using roughly the same ...
that's the feets finetune
oh
i heard that if u hab bad day today u can come to me and i will pat you up and show u bad day does not mean next day bad!
F


next attempt
- target: ""
positive: "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up"
unconditional: ""
neutral: "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up"
action: "enhance"
guidance_scale: 11.0
resolution: 1024
dynamic_resolution: true
batch_size: 1```
4k steps
wait what
starting point is what it already knows based on score tags
positive is what to enhance based on
no target as in just the model
no idea if that actually works tho 
shouldn't you also add source_anime or something?
k
i heard that everytime someone mentions score they get a headpat!
rawr!

it's the worst experirence I had
error
just do lora:pony-enhance-v10_last:2 because there's no point to anything else there
apparently its te and unet
Last leco for real, time sink without much progress and not much to learn from it
Was doing block weight to see if it helped
smol hapi mofus are here today to surprise everyone who wants a hapi day!
New AidXL version just dropped
https://civitai.com/models/124189
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...
no changelog but new artists and tags
and this is their compute provider https://nieta.art
看见概念(上海)智能科技有限公司是一家生成式人工智能技术公司。致力于打造生成式人工智能内容引擎 ——用AI释放人们的创造力,并解决其在数据、效果、版权、商业利用上的问题,重塑AI世代的内容生产关系。
wow 0.7
I posted this imi so if you can check up on it while I sleep
#1109884866964754644 message
woah
quick download before it's gone
oh
wth
this introduces A lot
what artist is this btw I recognize but not sure
imlllsn?
bruh the aidxl guy should get their art from danbooru
they have that artist as imlllsn, but on danbooru their name is ohisashiburi with 1.4k pics
I just don't like the twitter jpg noise
dunno, it's not from aidxl
I'm merging heart of apple and bulldozer atm
ooh
honestly it also kinda look like AI with the light back hair
popular artists do that tho
artist metas
the animagine xl defender 
does ur toaster use anixl? if so remember to qoute da mofus on dat toast
AIDV7 is bad 
I heard theres smol floofy mofus currently sleeping by the fireplace
i don't love it either 
the artist style got nerfed a lot and generalized style trigger word doesn't really work.
see smol blobs?
they not sad they hapi!
spent the entire day trying to add a function to the merger that saves every model iteration if I have autoscoring so I dont have to manually merge it later but I'm shit

0.51 / animagine hires has better colors than 0.7 / 0.7 hires
could be skill issue still but the 0.7 previews are not that good either
like this pic
ugly
maybe 0.51 is overfitting or something but even from the previews on civitai it looks better
well he merge animagine so
it's not just the colors all the compositions I'm getting are boring
idk how to explain it
could also be bc I'm trying to prompt for kanae with a lora
0.51
it just looks more interesting with bad prompting
0.51
there's just something cool about these
same everything but with 0.7
actually I fucked up the negative on that wait
0.7
bruh
0.51 just looks cooler even when it makes no sense
0.7 with quality tags
yeah i can drop it in webui
thank you for sharing
do you have more of epic stuff like this? 
it's literally just barebones 0.51 random gens I did like 20 of them
yeah I like 0.51 a lot bc the pics actually look artistic
ill check what i can get on 0.7
if I put some more tags 0.7 is better
I'm going to use 2 different merges that use .52 and .61 after I figured out the animagine side
this is prompt for it? a masterpiece with best quality, detailed 4k illustration, by wanke, 1gir
u fucking wanke
I put 2 different prompts in the text encoders
top one I put by wanke, 1girl
other one I added the other 2 parts
0.7
ah i see
somewhat
this is with the same artist tags I put earlier you can see the influence
he shouldve just add differenced .61 into it or something
well here you can see the yogisya and it's 0.7
style + artist tag
I did
oh nyo spaghetti
I don't really like the hair that much but I like how it doesn't look digital
what prompt?
by wanke by yogisya, 1girl, blonde hair, twintails, looking from side, green eyes, red dress, green leaves with dew drops, looking at viewer, best quality, high quality, amazing quality
you can switch out the yogisya

what about negs?
by wanke by ccroquette
I just realized I didn't have the aidneg on earlier
idk if this really helps
aidxlv05_neg, jpg artifacts, jpeg, paper texture, low quality, worst quality
was this with controlnet?
I guess not
no this is their actual art
ah that explains a lot
i should have realized kek
I just wanted to show that it is somewhat close
I wasn't clear
cuz merge
the 0.7 samples don't use the neg
maybe its niji dataset from the grave
hands pretty bad

mustard gas is highly possible
if only there was an unfucked pony merge or something
do the hands improve when you remove the embed?
dew
with..
damn
maybe its something with clips separated?
my hands looks fine most of tjhe time\
maybe my issues are bc I try to add 2-3 artist names
it was never trained for that
just wanke
meh
hm idk we could use multiple artists i believe
im not getting such hand deformations as you
nah you're supposed to use alternating steps if you do multiple artists
I'm not doing anything special
maybe my neg is bad
what is alternating steps?
bad anatomy, deformed, lowres, messy, abstract, bad, worst quality, low quality, bad quality, ugly, disfigured, deformed legs, fused fingers, extra digits, fewer digits, error, fewer, extra, missing, jpeg artifacts, watermark, unfinished, displeasing, signature, artistic error, username, scan
you can't do it on comfy easily so I never tried
I'll try

is this .7?
yes
hm ill inspect the workflow
yeah hands worked nice in previous version
maybe different sampler
ah
2 left hands
I'm using 2m karras
euler a

quiet
backwards hand here too lol
usually I do a hires fix but it takes 2 min
wtf
it's ok I will just keep my dogshit hands I can just redraw them
merging hard work, feel free to buy me a coffe
!!!
true you're an artist 
go hom lyc
I used 95% finetuning
hahahah
got the grant from stability
but guys merging is hard work
imi does it for free

yea based
whats leco?
enhance or erase concepts
ah that
leco deez nut
not having to use score9 8 7 6 54 to not get shit images
this is my hires workflow
honestly I might change it a bit to make it faster
there's new stuff that came out
could be 50% faster
good luck with that
it would probably need to be finetuned and trained until it forgets
so probably a very long time
the leco thing kinda works
that's cool
gimme workflow when done pls 
covariance shift at its finest
I managed to make best quality work but it made the model look like shit when it wasnt used
its probably possible but its annoying to try over and over
2x workflow
beeg
its fun when I dont get stunlocked trying to add random features to the scripts 
oh shit im missing nodes
you could just use kohya deepshrink it does something similar
I keep getting interesting chair designs
whats that?
I haven't tried it tho
thoughti t was called something else
i see
cute

smol floofy mofus are here now! they are here to pat everyone!
welp this dropped again https://civitai.com/models/124189/anime-illust-diffusion-xl
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...

Merging..
hopefully its more stable. V7 is pretty bad.
looks like it'll be less stylized but more stable (didn't dl yet)
Bruh why did he bake Genshin lora wtf
Why please normies
I also wonder What’s that special algorithm
because he got spammed with a hundred genshin loras from crosspost
You’re better than this euge 😭
the genshin lora is what made it stable /s 
hapi bunnies are mofuing here today!
i wonder what causes this bg
is it my sampler or my comfy setup dying
still think 218 is the best XL except nai3
looking nice
this one tho 💀
was trying to fix my embeds and stuff
i think it's prob the format of the prompt
cuz the tagger made it mostly
hopefully done after this, then I can move onto new aid
🤞












