#🍥|anime
1 messages · Page 27 of 1
My xformers is deterministic and there are still differences in the forms of the image. (2 runs to show determinism)
doesn't seem worth it to bother with installing that
I think I'm more interested in SAG
I feel like I'm looking at the same image 4x in @supple raptor's pics
I tried V2
Same
Details are different
2M Karras vs 2M Karras V2
Are they meaningfully different tho?
When you're looking for crisp results and less noise, yes.
No
Cause to me atm it just looks like non deterministic xformers shennanigans
Have you shared this lora yet?
Oh >.>
Then take off transformers and do a comparison, lol.
No, since its kinda ass I haven't bothered xD
I could publish it if you want it tho
Lazy 
I'll just stick to normal 2M Karras
Fair, I'll go publish it XD
yeah what's this SAG
what is this? its in automatic?
Not in auto
How does one do it then 

I think details look noticable and much better
In comparsion
The image shown also use my model
But idk side effects
Imo, when you have multiple ways with SD to get blurred results:
- LoRA weights too high
- VAE is blurring slightly
- Too high CFG with model
- Regular Karras slightly noising/blurring lines
You can do a lot to increase fidelity.

I don't want to hear anyone here complaining about blurred results AGAIN! 😂
***AAAAAAA ***
a million knobs and no guarantee the tweaking works
not everyone is up to date with everything tho
imagine if you could tweak shit in realtime, how nice would that be
ᴛʀᴜᴇ

v2 better eyes and tip of her right ear has a slice through it on v1 that isn't there on v2. These are deterministic results, not xformers shenanigans.
Wait, how are you using it 
This is like when I almost wrote a paper on how xformers does degrade quality slightly, and everybody called me crazy because they couldn't see what I could.
I agree this does look more interesting.
Its ashen using my model
Before someone complains that their image is blurry they should save it and open it from folder since Webui likes to blur images
when I first started using xformers I felt like quality sufffered
Oooo, that make sense then 
I'm more sad about non-determinism
Indeed. Also, if you're viewing in Discord, then you're getting a severely compressed version. One always needs to open it in a browser to see things properly.
JPEG artifacts are the bane of my existence.
Yeah, it stretches them past the actual resolution.
really?
Yes
the current xformers build is deterministic. I'm using it.

Is it possible to learn such power?
interesting
Same
You'll need to be running pytorch 2.0+cu118 then just activate your venv and run pip install -U --pre xformers
Ah, gotcha.
here's the full discussion on it https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8691
File copying hack is pain in the fucking fuck
eh, my webui is rendering my img in modal at perfect dimensions
Glad its gone
wym?
Back then when you have to update to torch 2.0 its basically hacker stuff
oh
anything with those libs or cuda
I hate doing
python seems like a real mess to me
I don't really know anything about the ecosystem except from using SD
but it seems ugly
I'm horrible at technical stuff, always forget things 
illuminyaati evolution
haha
I mean... I just figured her and Lotta could be friends since they both have green hair. 
I think my upload is stuck 
elf is actually typically my #1 negative
you wouldn't believe how many times I think to myself, "shit I forgot about elf"
Clyde 🙏
O, clyde aprove
short circuited
Thats one amazing elf u have created
I remember this prompt xD
nice hands, also Legend of Zelda vibes
The green elf where I use on model experimenting
Very good prompt, fire flies etc make me notice make me notice everything easily
It's my Saria prompt
gotta be representative of like 99% of AI generations
I'll be honest, when I made the version we both use for testing models, I intentionally added wariza because of the trauma of making 20 different Trinart Characters + Redacted merges to finally teach Trinart how to do wariza 
You experiment on elves??? 
Yes 
Its such good prompt too
what's wariza?
@stable crane this pose
wariza?
Yes
there's a kneeling lora on civitai
XD
that can get really close to this
specifically the leg positions, hands don't have to be on the ground.
wawawa classy woman
That merge was way before LoRAs were a thing. Now it's just a thing I test on basically every model because of how much work I had to do to get it to work on that one. Most of them can do it no problem now.
the kind that will step on me 
looks a bit innocent to me 
Damn I remember we hunting models instead of merging merge of merge of merge of merge of merge models

Yeah, back then nobody was uploading merged models. We just merged them ourselves and shared recipes if they were good. Usually it was just 2 models weighted sum merged together.
it'd be nice if someone could invent a latent sampler that didn't go so nuts with the image
latent always has the crispiest results
but drastically changes it during the preview stage
back in my day...
I'm glad that people are mostly doing LORA and other small size things now
Clyde really does have a mind of his own doesn't he.
This. Single concept dreambooth models are dumb.
I wanna thank kohya for lora extract

I have yoinked dpep4
ᴅɪsᴄʟᴀɪᴍᴇʀ: ᴛʜɪs ᴍᴇssᴀɢᴇ sᴇʀᴠᴇs ᴀs ᴀ ɴᴏᴛɪғɪᴄᴀᴛɪᴏɴ ᴛʜᴀᴛ ɪ ʜᴀᴠᴇ ʀᴇᴄᴇɴᴛʟʏ ᴜᴘɢʀᴀᴅᴇᴅ ᴛᴏ ᴅɪsᴄᴏʀᴅ ɴɪᴛʀᴏ, ᴡʜɪᴄʜ ᴍᴀʏ ʀᴇsᴜʟᴛ ɪɴ ᴛʜᴇ ᴜsᴀɢᴇ ᴏғ ᴀɴɪᴍᴇ ʀᴇᴀᴄᴛɪᴏɴs ɪɴ ᴍʏ ʀᴇsᴘᴏɴsᴇs ᴡʜᴇɴᴇᴠᴇʀ ᴘᴏssɪʙʟᴇ. ɪ ᴀᴘᴏʟᴏɢɪᴢᴇ ɪɴ ᴀᴅᴠᴀɴᴄᴇ ɪғ ᴛʜɪs ᴛʏᴘᴇ ᴏғ ᴄᴏɴᴛᴇɴᴛ ɪs ɴᴏᴛ ᴏғ ɪɴᴛᴇʀᴇsᴛ ᴛᴏ ʏᴏᴜ ᴏʀ ɪғ ɪᴛ ᴅɪsʀᴜᴘᴛs ᴛʜᴇ ғʟᴏᴡ ᴏғ ᴏᴜʀ ᴄᴏɴᴠᴇʀsᴀᴛɪᴏɴs. ᴘʟᴇᴀsᴇ ᴋɴᴏᴡ ᴛʜᴀᴛ ᴍʏ ɪɴᴛᴇɴᴛɪᴏɴ ɪs ɴᴏᴛ ᴛᴏ ᴄᴀᴜsᴇ ᴀɴʏ ɪɴᴄᴏɴᴠᴇɴɪᴇɴᴄᴇ ᴏʀ ᴏғғᴇɴsᴇ. ɪ ᴠᴀʟᴜᴇ ᴏᴜʀ ɪɴᴛᴇʀᴀᴄᴛɪᴏɴs ᴀɴᴅ ᴡɪʟʟ sᴛʀɪᴠᴇ ᴛᴏ ᴍᴀɪɴᴛᴀɪɴ ᴀ ᴘʀᴏғᴇssɪᴏɴᴀʟ ᴀɴᴅ ʀᴇsᴘᴇᴄᴛғᴜʟ ᴅᴇᴍᴇᴀɴᴏʀ ɪɴ ᴀʟʟ ᴏᴜʀ ᴄᴏᴍᴍᴜɴɪᴄᴀᴛɪᴏɴs.
How are you getting such a good looking elf ?
Just prompt:
1girl, (green hair:1.1), short hair, pointy ears, (naughty face:1.4), (seductive smile:1.2), wearing a green (tunic:1.2) and shorts, (illustration:1.1), highres, (extremely detailed CG unity 8k wallpaper:1.1), (mid shot:1.1), (full body:1.25), (solo:1.2), plant, tree, (beautiful eyes:1.15), green boots, leaves swirling around the girl, wariza, leaning against tree at night, blue_eyes, (beautiful face:1.15), (fireflies:1.1), parted lips, highly detailed face, ultra realistic, masterpiece, best quality, the legend of zelda, bokeh, extremely detailed, intricate
Negative prompt: AuroraNegative, KHFB, bad anatomy, disfigured, deformed, malformed, mutant, gross, disgusting, out of frame, nsfw, poorly drawn, extra limbs, extra fingers, missing limbs, blurry, out of focus, lowres, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 818821015, Size: 512x768, Model hash: 1b5f8211ec, Model: AuroraONE_F16, Denoising strength: 0.51, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x-AnimeSharp, Discard penultimate sigma: True
Ah I see thankyou 🫶🏻
@stable crane had a chance to test the epocs?
https://civitai.com/models/47164?modelVersionId=51741 @gloomy scroll It up 
this will be cool to use using swordtember challenges & if i get some cool designs might make them into 3D models 😄
Its very random, but it makes some cool stuff 
Mommy Noel 
Damn, whats that model/lora? 

that chin tho
I'd share the rest of this pic, but I cannot, I like the face tho 
I like big lips and I cannot lie
All my prompts can't deny
That when thick lips come in
With An itti bitty embed
And a round LoRA in your face
You get 
(inspired by guizmoose @wind girder
)
holy
I love the huge sleeves, nice clothes.
Nearing the end of this piece
@stable crane Tried your frog girl prompt... kinda XD
This is pristine. Well done.
I tried inpainting the face and the eyes on the hoodie, even hotter 


The duality of man™️
Cute anime girl VS Sexy frog gf
nice 😂

not yet
I like the green lipstick
Its pretty nice 
Surprised it even worked though xD
TRUE
it almost feels like Blender
lmao
3D prompting, and 3D LoRA 
And I mean, bimbostyle embedding and bimbo style LoRA are both IRL based I think 
only 70% ? lol
I had the same thought
I like the bimbo aesthetic tho, so I just like using it as like a style LoRA 
where is the db with every like cool lora's
posted my review @teal pivot https://civitai.com/models/47067
True 
What you think of my version of Lotta? @stable crane 

there is someone here doing insane thing like neon colors
can't remember his name
川上#2989
his creations are insane
i would love to have the model
nice, maybe a bit uncanny valley
Fair, that'd make sense XD
This is another attempt to get good eyes after merging several times on SD 2.1, they are far from perfect but getting there
Oooo, I think I got a few LoRAs of that 
O I got just one 
I wish I could find the gen I made that looks like a real photograph
Imma try it with my current prompt 
this one is close
its pretty cool, but you just need to add something like bioluminescence & Neon Colours to the prompt and it makes it all pretty and neon
1st 2 images was made with
a devious samurai, cyberpunk, neon colours, orange neon jacket, ((neon tech gadgets)), ((neon glasses)), cybernetic implant, tattoos, sci-fi, fantasy, (intricate details), hdr, (intricate details, hyperdetailed:1.2), cinematic shot, vignette, centered painting art by (greg rutkowski:1.1) (Horror Colors:1.5), bioluminescence
3rd and 4th images using
bioluminescence, neon colours, blue_hair anime girl with a green plant behind her, Tatsuro Kiuchi, raining, wet_hair, sunbeam, sunlight, water_bubbles, water, Grzegorz Rutkowski
neg for both was
bad anatomy, disfigured, deformed, malformed, mutant, gross, disgusting, out of frame, poorly drawn, extra limbs, extra fingers, missing limbs, ((blurred, blurry, blur))
damn
hard to control the realistic models a bit
It do be, still cute tho :o
Also, made disgusted Lotta 
💀
That is pretty dope :O
Very nice. I'd read that light novel.
nice
😭😭😭

what the fuuuuck
really?
Yeah

shirt is thin D:
It's uh, a bit cold in there.
I should delete 
Probably, lol.
ehhh
I can make more, don't worry 

we know you know 🙂



can i load some loras from civitai with google colab?
Uhhh, maybe, you could get the extension for civit stuff
no 🙂
Yesterday for free accounts, afaik yes.
so it realy only effects the free accounts
Yup
dont think i shared these but made them during the Alice in DiffusionLand pow
any link for x4sharp high res ?
lol is this a damaged generation or a sucessful image
Seems successful to me :O
yeah defiantly sucessful lol
That is kinda cool :o
Bimbo time

something wholesome
?
I'm not sure I'd enjoy a realistic FLCL. Not that I enjoyed the original either...
No wholesome, all bimbo 
haven't watch that, but only heard good things about that
am i missing out?
No
downloaded loras, created a hugging account to reupload it because not in civitai
just for my colab

Thank you! You got some great pictures!
@stable crane Do you think this model turned out well?
I had to remove all the """quality""" tags because they made it too horny 
Really dope eyes tho, but really ugly freckles
try removing the freckles?

The quality tags I had to remove btW
(curvy, voluptuous:1.4), (thick thighs:1.2), (plump:1.0)

Those eyelashes are really cool 

(((((thicc lips)))))

O, holographic clothes :o
found a small model
Prompt works fine without all my 3D and bimbo shennanigans tho 
Eyes are very pretty as usual 

Added LuisaP Niji LoRA for even more detailed eyes
LET'S GO i love it 🙂
Those are amazing @lyric zealot
Now @sharp bluff's Niji LoRA 
Also I got a surprising amount more of NSFW LoRAs than I thought
I just wish I could gen proper hands 
Aurora hands update when ? :pog
Lol. I wonder if there's a holding hands inpaint
Good grief... the types of stuff that come up even in the search of Civitai for "hands" -_-
Very much approve 
two different ambiances with ^
THE GALAXY CLOTHES I CANTTTTT
Ayyy 
Even censored I know I have it 


it's excellent
wow, Glad you liked it
Cursed
yeah it'll probably be one of my favorites for some time
hah lol look at this background
The oldest anarchy server...
holo wings?
those look sick
Quite the difficult expressive combo. Sadness, strength, and a hint of catatonic. A mild righteous anger perhaps. Couldn't not comment on this one. I shall now disappear once more.
Noooo, don't go Trexdelussy 
me trying to figure out how to use the LLM models to improve my art
which is actually an LLM prompt
i NEVER should have gotten this 3090
this will only feed my AI addiction
leave it to LLM's to hallucinate the most fantastical shit
hey @wind girder still awake?
i'm here trying to figure out how to get a LLM model (well... vicuna 13b) to what i want do in a single step
this one is based on @gloomy scroll submission
https://cdn.discordapp.com/attachments/1023999442338201721/1098384429388415076/tmpsils1v51.png
feelsbad i coulnd't do the pow due to time constraints
aight. ultralazy mode activated
LLM, analyze pow text, and make me some keyword descriptions of this i need to win this
!
"my eyes are up here"
I really love this


Seamless pattern of cats
Heya Hayat! You can go to #1072220168534642768 or #1072229020520947753 to find information on this server--there are no bots to generate images here! Those links will help you get the information you need.
Thanks!
Lazy COW submission this time, already had made a bunch of monster pics before, what you think tho? @lyric zealot 
i love it
Relatively, yup. 4.5
XD

I’m gonna be honest, I can’t see shit


Is it that dark? 
aight, imma go to sleep tho. didn't get a lot of sleep last night, almost fell asleep driving getting that 3090
On both my screens it looks pretty clear 
O, fair, good nighty Eface 
Goodnightt 
gnight you lovelies ❤️
G'night!
G'night Smol! 💤
Goodnight 
O, good nite! 


Goodnight everyone! 
Now you tell me, is it that dark? 
It’s like 8am so time to sleep
What am i looking at
Well, it’s kinda dark, but it’s more of, I’m on mobile so can’t really zoom in enough to see shit
Vampire
To hide from sun
Fair, was pointin at Smoile tho

We currently living in dark mode
Is this more clear? 
Yes
No sleep for you 


XD
Looks great 
Movie poster!
How to type
Its the same as what I posted, but it darker, it isn't that dark tho
Are you using the new easyui or something
Touch everything 

EasyUI?
A tough style to achieve
I just raised the brightness on the windows photos app 

Its pretty neat
BUTTT
I'm lazy to reinstall extensions

It does look very nice tho 
Can’t symlink that shit?
Extension probably can’t right
But Nubby said they had issues symlinking
Yehh, they said they just had to install manually, so I just didn't attempt to link it
I read him saying that
Yeee 
One day if I get un-lazy tho I may get on it and install all the stuffs I'm missing
I mean
My current auto1111 is also super outdated
Same buddy same
You can edit the UI a lot @lyric zealot
If I ever try to update now, it’s gonna break
Well, the palette, and tis what I settled on
Though also quick settings are pretty easy to set up too

Or is that auto with theme?
Is the UX thingy ye, the fork 
Ahhhh
Tis one
It is really nice to look at, takes a bit to get used to it tho
How much you got?
Really time to sleep
Tomorrow you get it, I'll be nibbling your hairs one by one in your sleep if you don't set it up

8gb
I’m too lazy

That’s still early my man
Too bad, you try it, or you get nibbled hairs 
I'd think that wouldn't matter with the UI. I think you would be fine there.
The sun isn’t out yet

I have my country's equivalent of SAT in a specific topic on sunday
Run with HW acceleration turned off and it won't be a concern at all, but I don't think it should take up noticeably more than Auto does. It's still Gradio, just with different CSS.
I didn't practice for it at all
You know how bad this 3070 is? I have to turn off hardware acceleration on discord because discord is literally lagging my computer when I’m genning 
Not sure if my browser has hardware acceleration turned on or not, need to check tomorrow
Good point tho 
Oh shit, then better sleep soon! Good luck btw! 
I have too many tests
Yup, especially when I’m watching like 3 streams, genning, and chatting on discord

GOODNIGHT SMOL 

True, good night 
Goodnight everyone
Time to really sleep


Frfr
NO PING PLEASE
👍
You got crippling discord pinged cant sleep-itis
I think?
XD

But ye, goodnite 
guys, would you?
I always kiss my homies goodnite 
If I try and inpaint this hand, am I going to hate myself by the end of it? 
yeah but would you do it
I mean with the eye
O
I mean
I feel
Like that was a trend at some point
I wouldn't do it
But I feel like it has been done
bruh
hypothetically, if you could do the act with a sd model, what model would you do it with?
ok nvm
as they say
create a anime
@timid crypt
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
@teal pivot I think I'm going to have some fun with your new model
@teal pivot is sleeping
Ah, I won't ping any more then but at least they can find it when they get back on.
Maybe I'll just ping @acoustic agate instead 
Somehow this is gonna be my LoRA with the most pics in the dataset

Sis, generate sexy anime girls while I train LoRa 
Im outside for days
Oooof 
Hope it cures burnour
Hope so 
Oh damn, I got scared for a sec
I saw 3.94s/it and I almost had a aneurysm
But its cause I had SD open
Now its going down
Landing on the moon pls
The bg really good for aesthethicc
AHHH
Im outside and i wanna try out this style model

Yep I think I stole parts from that prompt when I wrote this version. At the very least I had was trying to remember that one when I wrote it.
So, my idea did work, but it is kind of bold XD
Its here 



ALSO
Sis can you guess what the LoRA I trained is about? 
Nope but it looks so good
I was thinking "I bet I can train a better bimbo LoRA"
My dataset? @mallyboouwu/@uwumallyboo on IG, it did work
But it is very bold
XD
Also do I have to do git commit 22bcc7be to be able to use it? 
@quick hill
Pray
I forgor how to update 
I think I'll just have to do a new install 
@supple raptor You know anything?
Lemm try the first one
what does that do?
no without the 22b
git checkout master then git pull then git checkout 22b if whats what you want
O
Well
At that point I'll just install it again
Cause I know if I git pull from this update
Acording to nubby
It just breaks
I don't know if its been fixed
BUT, I'll just do a new one
lol webui latest commit is 22bcc7b, its the latest master version
git pull
So yeah I have to install again
Wellll, did Auto fix the update breaking thing?
I thought auto havwnt update for month
So things are working fine prob
When was lastest git?
3 weeks ago
this is way better info what this actually dose https://ku-cvlab.github.io/Self-Attention-Guidance/
Thanks!
O nvm
BUT
From what I heard of nubby, doing git pull broke your install for a while
I dunno if thats been fixed yet or not
I wander if this actually works as advertised

When did Nubby said that, 3 weeks is very long to fix bug if this still occuring. Check Github issues etc to find related results.
Try to check issues and such first
If you are concern
I got
2 options
Reinstall
Or
Break it
Then Reinstall
So either way
I am reinstalling
Just check issues
oh Thank you! Did you try it?
Nope I'm outside 
Imma have a look at the issues
i see
Hmmmmmmmmmm

I think Imma just reinstall it
Because it if it breaks in such a way that I cannot use it
It'll be fine
OK so from what I read and someone translate again to GPT
SAG is a regularization technique that can be used to improve the performance and stability of diffusion models. It involves introducing an auxiliary task to the diffusion model that requires the model to attend to certain regions of the input image. By encouraging the model to focus on these regions, SAG helps to reduce overfitting and improve generalization performance.
SAG works by adding an additional loss term to the training objective of the diffusion model. This loss term encourages the model to attend to specific regions of the input image that are important for the auxiliary task. During training, the auxiliary task is jointly optimized with the primary task of generating high-quality images.
It seems to help with overfitting issue and generalization.
Ashen also said it is adding blurs to undercooked areas to tell the model: cook it more.
Although it's more noticable in Low-CFG - From what I seen
If update breaks you can just git reset and revert changes
how to add SAG to automatic?
Follow instruction, if update breaks you could revert same way you update
git stash
git checkout origin/master
git stash pop
^ for updating auto1111
from neggles
tyy
Alrite, I got it set up in a new install
Time to try it on 1 weight
No SAG vs with SAG
It is crisper
Lemme try on a prompt without any LoRAs
Seems to generalize and react to prompt more too from what I seen on projectai
does control net work for u guys?
No SAG vs with SAG
How does it works with pose and such? Bg etc
If you have one
I wanna try w/ low CFG on my oc painting
No SAG vs with SAG again
I'm gonna try that now 
Oh hey nubby is back
There were a couple more commits since I had that problem. Not sure if whatever caused that problem for me was actually fixed or not though.
On this pic it is a pretty massive difference tho
Ooof, fair then
I have run the latest commit since like the day after it came out without any major problems though. Nothing that I wouldn't expect from any random Auto1111 commit anyways.
No SAG vs with SAG
Something went wrong on that one for sure, I feel like it was on my end though
whats this SAG people are talking about
No SAG vs with SAG
I think I have to try LoRA-less, I think LoRAs make it more chaotic
Yeee. I fell asleep after dinner for a bit. Didn't sleep much last night so I probably needed it.
Oof, make sure to take care of yourself
SAG-less vs SAG-ful
Ooof, I feel ya x.x
Probably?
@supple sluice
Check this from ChewToy's link.
interesting, thank you!
Very beautiful lights on 2nd one
sag
True xD
Wake up I have moomer @quick hill 
(sagless vs sagful)
I mean, you woke now, but still, get bonked
Wow
I'm gonna try 2 sag
it made hands worse
xyz for what? 
Hands are RNG, it made the hands better in other ones xD
Ah
I see
I tried SAG 2
Background is more detailed, but it is getting a bit crispy
2 looks cooked
Yes
TIME TO GO FOR SAG 10
O, it is
Frying moomer
But, it was kinda interesting to watch the live preview 
fried bird
More like... horror moomer? Its kinda funky 
its like you start from noise and go to noise again xd
It was kinda good in the middle
But then it started getting very noisy in the upscale, funky stuffs 
That's awsome how it seems to fix hands/broke hands from what everyone else said in project ai and sd server
Oh god
Sagless vs Sagful
And there is a random choker, BUT, I do have "(choker:1.2)" that the sagless didn't do
this is awesome
xD
It helps generalization, that sounds right
SAG almost seems like the karras v1 vs v2 argument lmao
Yeah unless its small details

















