#Amedira
1732 messages · Page 2 of 2 (latest)
wdym
the .inpainting model
what does your inpainting look like
yeah
if I just inpainted with the regular model
without merging with the runwayml inpainting model you'll have some patterns and artifacts
at least that's what should happen
well that didn't work out as planned

The inpainting model didn't have a VAE baked in
but the other one did
me derp
None of my models should have a VAE baked in afaik
i baked it in
without for this one too
tbh
i either can't tell a difference
or prefer the non-inpainting one
i mean
in my entire time using auto
I've never had an issue with inpaint
Works as intended for me
and always has as far as I remember
Dunno
This is for @neon cosmos because uuuuh
🤝
what is this
wat happened to tusi.art
Who am I to know what happened to a Chinese model hosting website
the name is too close to pusi

It was relatively sfw
On this server level
i mean idk
CCP generally doesn't care about AI stuff
Im guessing there was excessive winnie the pooh content
? I never noticed an issue
or atleast that what it used to be written on paper
The only issue I had was because of vae problems
Good job btw
This is like I'm the top 3 things I've seen in the server
In
For overall aesthetics
I did almost nothing 

Also if you want the LoRA
It has no activation token
Hi sorry if the question is not suitable for this channel but I have used Amedira and it's good but for some reason any Lora that I use make my image noisy
These two pictures are the same prompt and seed except on the second one I used youko lora (https://civitai.com/models/36140) with "strength_model" of 1 and "strength_clip" of 0.5
Is this what supposed to happen or am I doing something wrong and it can be fixed
Don’t use the Youko lora at 1.0 weight
And a lot of these LoRA won’t work well at high weight
which weight is appropriate
And which of "strenghts" is weight, I just installed ComfyUI and it's confusing, even the documentation (https://blenderneko.github.io/ComfyUI-docs/Core Nodes/Loaders/LoadLoRA/?h=lora) doesn't explain what strength_model and strength_clip means
I think I have tried it but I'll try again and report if something goes wrong
Thank yu ❤️
Something is very wrong
Where could I ask what to do
Good model though! I haven't prompted for half a year and the change is real
still need to try bv3 and cv
don't forget mV it's pretty good too
try upscaling the image then passing it trough img2img
it should help a lot on the Youko LoRA
didnt see that
whats mv
cv is the counterfeit one right
bV is the original more Based66
mV is the second which is more MeinaMix 10
cV is the third and more Counterfeit 3

But the first 2 aren't really any close to both anyway
bV has the composition of Based66 with the quality of AuroraOne
mV has nothing to do with MeinaMix 10 honestly
i'll do more xyz with diff stuff after I got some other models, I want to see how they work
so why more meinamix 10

I realised that images turn out noisy because the node with youka lora should go both to the positive and nevative prompt
Thank you nonetheless 🤍
Ok noice
Anyway when bigger xyz
Or less bloated xyz
ah
I was initially checking if any of the other merges do something different
but most of them are worse than what i already had
gonna have to think of something else as far as merging goes
wanna try bayesian cuz I wonder what it does
I feel so happy that b74 is using cV lately
They look good
huh, didn't realize style loras were so strong, they almost completely overtake the base model
and this is only .4
huh
still very similar even withoutstyle loras
yeah it was cuz of addnetlora
maybe it kept weights from the initial model
or something
or it cant flush between each model swap or some shit
dunno
Depends on how they are trained and what you are using them on
But yeah that was odd
Amediera 2 when
(joke)
I have mV bV bV2 bV2.5 bV3 cV you’ll never have 2
was addnet somehow, I think it kept the merged weights or some shit of the first model and merged those into the next ones
no
YES
Some model actually just ignored your LoRA lmao
tfw u have so many colors limited looks the same as a normal image 
I did another one on ai art channel
added munimunimuni
shit looked cursed
wish trhere was a way to stack the xyz in grids instead of infinite
I mean its easier to read
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
Insteado f 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1
its just more clarity
Reminder to add more examples images on readme
And add proper recommended workflow (if I want)
Breasts 
Jk 
It was too tempting
1girl, solo, standing or sitting, full body, [up to you to input hair color/eye color/hair length], absurdres, smile
For negative, exclude neg embedding. Not everyone will have your negative embeddings
Keep it the basic low quality, worst quality, bad anatomy, jpeg artifacts, bad hands
I’ve been adding in contorted fingers and lowres as of late too
And if you want to show off detail or lighting, “subsurface scattering, specular”
And this is optional , “fill light, rim light”

?
what are they doing wrong

i mean its ur model
i figured u knew more about what they're doing wrong
i presumed it was just the schizo quality tags
it's completely overwriting the prompt
@vocal shuttle
4 breaks 
civit ai gonna have a field day
it's a prompt from civitai
ik
was talking bout this
i used the prompt the yotsuba lora dude used in his example
ye
bad prompting
did u make a new lora @steel lantern
can u help me tomorrow or after? I'll test on amedira what kind of quality I get on images and I'd send the image and you could check what result you get on your nvidia card

I forgot to say gex
whats gex
i mean yesh
so what kind of help can i provide
it's late today so next time, but I just want to check if I'm getting bad quality because I'm on amd
like I'm just inpainting hands yday and today and everything looks way worse than I'm used to
Just send your settings I'll take a look
what negative did you use again?
#ai-art message
I use this
But I modified etone weight to 0.5 and steps from 6 to 12
alright settings
does inspector work here
no lora or addnet or seg or adetailer just normal gen
oh wait
let mechange sampler
can u gen this? or u can give me a gen and I'll make it
yesh hold on
I'm talking to neggles about some stuff
heres hires fix meanwhile, no adetailer
can u also try hires fix plz
this isnt novelai vae

Anyway use safetensors and blessed
this?
It's a bit smoother here
here

Don't worry I'm blind too
can u send me an image with decent hands when you get one? I wanna see if it looks the same when I try to make it here
seems like the nv rng does work at least
nah
its def amd
thats makikng my pics look worse
look
left one is what I get
you can see that its shit
so I Wasnt just hallucinating
Probably VAE
Very minor
its huge
especially for eyes
or other finer detail
like fingers
they look seriously bad
when they never did on colab
alright
You could probably use TAESD to get more performance and sacrifice a bit of quality
quality is my issue so I'd rather not make it worse even if its just a little 
Yeah but it's very minor 
Add
"sd_vae_encode_method",
"sd_vae_decode_method",
to quicksettings
You can switch on the fly
Lemme compare vram
Full VAE
TAESD VAE
Lemme try again to see if it's a fluke
Lemme send a xyz
yep not a fluke
dont see it
Did you restart after adding the line to quicksettings?
this is trash
altiht now
taesd, time taken 0:44
A: 6.70 GB, R: 8.63 GB, Sys: 9.1/15.9844 GB (56.7%)
It shouldn't spike like this afaik
TAESD is so light tiny and fast that it should use almost no vram
this is amd
It's so fast in fact you can use it to preview generations before they finish
Are u sure that you are on TAESD?
yes
I'm checking normal now
full 0:40
A: 10.96 GB, R: 13.76 GB, Sys: 14.3/15.9844 GB (89.2%)
Huh it does use less but not to the sam degree
So do you now understand why you can oom even if it doesn't reach peak vram during gen?
The VAE will decode the latent and that uses a lot of vram
Compare quality in image?
Send both gen you did
yea
I'll put them together
also I dont get a spike now when doing hires
?????
wtf is this application
so SCUFFED
this is on taest
d

hooly shit that spike is like 15999
I1ll send you the other ones to compare
gonna try taest more cuz less vram means less heat as well
full
yea this 2x spike not existing is pretty big
maybe could even use sdp
not that I'd need to
actuall the quality seems a lot worse for me
You could use Image Filters extension to fix saturation caused by TAESD since it doesn't use your VAE
when does that happen
But use
I'm not tlaking about the colors
the image looks significantly worse
like the anyv3 days
already sent
full only
Contains 66.69 MB of junk data! Wastes 159.56 MB on precision.
any idea what this means?
what precision
the precision
yeah
fp16 fp32 bf16 etc
what did you use to make this
Could use
You won't like hearing this name
It's Gnu Image Manipulation Program (Gimp)
Yeah u could use Photopea like a sane human being
U could probably find the ideal Image Filters settings
Of course quality is worse
But it's a lot lighter
Full vs me tryna correct colors
Helps a lot if you compare to just this
its lighter but ram usage is not an issue on 16gb
your decoder spike wants to differ
it works though
and this happened even on the t4
not worth that quality hit
and having to use filters
like this just looks like a pic from last october
gonna turn off k quantization since I dont think it does anything visible
maybe it fixes this rng bug
this would be useful for limited vram users though
just not me cuz my main issue is shit quality and this does the opposite
actually can I do batch size 2 with this
naaah
really
did auto get smarter and decode one by one?
wdym while ago
like months ago
I always did thgat cuz I didnt know u could use batch count
so I jkust did batch size 8
This is new to me lmao
yea dunno
I think I have dementia
I was trying to compare current gens to shit I made on colab few weeks ago and I was getting way different shit, turns out the the bigger part of it was cuz of the pad prompt/neg optimization, but there is still an inherent difference even if I use gpu or nv rng, also k quant setting doesn't seem to change much
anyways donezo for today, thanks for helping me with the schizo shit, hopefully I figure it out somehow
left is nv?
yea
right is original
t4 gen
but nv and gpu seems largely the same except for some minor things
something like that
but yeah the resolution so far is looking like amd still makes worse looking things than nvidia
It's like using comfyui with auto seed
but I'm surprised it got that close to the images you sent
like it was almost the same other than woirse looking hair and eyes
yea I Just need to turn off pad prompt and hope for the best when trying stuff I made on t4
u should make a new lora or something, the ssambatea one is still one of the best
yeah
improve it by using Dall-E 3 datasets
maybe we create a project where we gather
a lot of Dall-e 3 images
collectively

Welp I won’t be doing much today
sup
Ahoy
oh ok
like 
Me when
Set a reminder in 5 days from now (<t:1696948145:f>)
View reminders with the reminders command

@rancid plover Can I get Project0.ckpt?
🌙
I wanna try to enhance eyes trough mbw merging of that LoRA
Time to use OUT5
I also wanna see what happens when I negative merge this
Will see if I still have the CKPT once I get back home today.
okidoki thanks
🫡
Hello are you ok Luna?
ghosted
Yesh, fell asleep yesterday tired from the whole trip. Hopefully I still have the Lora - did clean install of SD a while back hope I still have all those files.
Yes, my cats slept with me today since they been away from me for 30 hours (still kittens and I'm concern for their health)
Hope you doing fine as well 😋
noice noice
This is my shack
Reminder for @ivory vine
@steel lantern do it

do it
But I’m busy :(
I couldn't figure out when you want to be reminded. Please try again.
-remindme 4d weekend ig
Set a reminder in 4 days from now (<t:1697294259:f>)
View reminders with the reminders command

what did u do again?
Add samples
ohh
I stole the thing from Lynn kek
nyo
They just are
Set a reminder in 1 day from now (<t:1697175212:f>)
View reminders with the reminders command
<lora:OpenNiji-V25-AnyLoRA:0.6> <lora:letitsinkin:0.2:lbw=MIDD> <lora:evleStyle:0.3:lbw=OUTALL> <lora:Loraeyes_V1:0.3> <lora:beautifuldetailedeyes:0.2>
He's cooking
Someone help me reduce the blush
it even happens in flat color style
Please use LECO
train lora > merge it around 0.5
- I have no idea how to use it
- Would erasing the blush concept not just make it impossible to blush?
- is it mofu certified?
boom less blush
maybe look some of this
https://github.com/ostris/ai-toolkit
I mean it is a lora, so you can adjust the weight ig

I'm installing LECO @ivory vine
good
I'll post my config file for the training
lmk if it went well
I hope it works
Loss*1k: 0.6324: 29%|██████████████████████████ | 147/500 [15:16<48:05, 8.18s/it]
@ivory vine how do I use the LECO?

checkout the second link maybe
ok
Just like a LyCORIS apparently
flip the weight and you get yourself a "Blush Slider Lora"

idk how well it works on other model
look ma, no negative
No one cares in #ai-art lmao
I'll put 3 LECO in queue
One to maybe improve blush erase
Second to enhance eyes without LoCon
Third to make SLE work how it should
Reminder for @fair notch
catch the mofu

Thanks mom
Set a reminder in 2 days from now (<t:1697353089:f>)
View reminders with the reminders command
They finished training but I’m away from home for today
Also I’ll train an SLE with potentially no token activation
tldr I’m cooking
did you get the sauce?
Which one?
for sle
Sauce?
Sure man, I got you covered:
Edit: I sent the sauce copypasta

Sweet Luna lost the LoRA but still had the dataset
so is it just pics with good eyes or what
Ye
wow
<lora:noBlush2_last:0.3>
Set a reminder in 1 day from now (<t:1697354203:f>)
View reminders with the reminders command
Reminder for @ivory vine
weekend ig

<lora:mochaStyle:0.2:lbw=OUTS> <lora:movieFilterStyle:0.2> <lora:siliconLandscapeIsolationStyle:0.2> <lora:jellyfishForestStyle:0.3:lbw=TEF> <lora:littleGrayStyle:0.4> <lora:ssambateaLoconStyleV2:0.2:lbw=INALL>
Me saving all my stuff here so I don't forget
idk if this should be called bV4 or a completely new V
Reminder for @fair notch
use rin nee nee for rituals
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 1 day from now (<t:1697445976:f>)
View reminders with the reminders command
2nd 

what
I chose to not make this into an update but a new variant
This model looks pretty neat imo
yeah looks pretty good
OpenNiji-V25-AnyLoRA:0.8,drawingStyle:0.3:TET,seasonsStyle:0.4:TET,detailSliderStyle:0.2,ssambateaLoconStyleV2:0.5:TEF
just saving stuff
Some loras are so strong even at 0.1
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 3 days from now (<t:1697705267:f>)
View reminders with the reminders command

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Yay
Set a reminder in 6 days from now (<t:1698226534:f>)
View reminders with the reminders command
So
so
so
yes
torturing myself
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 3 days from now (<t:1698486719:f>)
View reminders with the reminders command
I'll put the recipe on the readme
There's also mbw but I have that in txt files
did u test lora blocks 1 by 1?
wut
That LoRA Block Weight
Just some presets I have
TET is Text Encoder and no Unet
TEF is all Unet but no Text Encoder
The others are standard presets

how did u find all of these? did u experiment?
like wtf is seasonstyle 0.4 TET
isnt that just a prompt then
this part tho
I merged the text encoder of that LoRA at 0.4
idk
I just did that and continued with my day
wtf
Sometimes you just don't question
gem knight
Lmao
Lmao
It's just my magic wdym
I believe
Did you try my new model?
wild he doesnt even know wtf it did

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 days from now (<t:1698832467:f>)
View reminders with the reminders command
Amedira
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 days from now (<t:1699184716:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 3 days from now (<t:1699449380:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 days from now (<t:1699795046:f>)
View reminders with the reminders command

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 1 week from now (<t:1700400032:f>)
View reminders with the reminders command
Rip this post
@steel lantern
?
what did he fix, just size of model?
The dude is weird
He’s a professional overfitter because he gets good quality afterwards

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 2 weeks from now (<t:1701611404:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 weeks from now (<t:1704031497:f>)
View reminders with the reminders command
Rink really killed this post

But you killed it a while ago


Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 weeks from now (<t:1706450894:f>)
View reminders with the reminders command
Dead
yes
I'm SDXL now
and I'm not gonna mix a model because uuuh ram
I think the current Amedira bV3 cV and mV are all good enough as is
lemme know if you need an instance with more ram
I have no idea what I could make since I haven't experimented
what model u use?
AidXL 218XL Animagine Pony etc
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 4 weeks from now (<t:1708870213:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1713018642:f>)
View reminders with the reminders command

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1717167781:f>)
View reminders with the reminders command

Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1721315975:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1725463461:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1729610734:f>)
View reminders with the reminders command
rink stop 
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1733758041:f>)
View reminders with the reminders command
When amediraxl
give me a 4tb ssd
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1737905906:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1742053687:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den
Set a reminder in 6 weeks and 6 days from now (<t:1746202225:f>)
View reminders with the reminders command
Reminder for @fair notch
catch and cuddle the adorable mofu in his den









