#🍥|anime
1 messages · Page 26 of 1
ohhh
have fun
sweet, trying it out right now 🙂
theres 4 versions but
there shouldn't be any big difference
I recommend the waifu diffusion 1.4 kl-anime2 vae
and bad_hands embed'
aw i broke it lol
Rip
ok generating for gunblade maybe
@proper trellis the images you see as previews are multidiffused in upscale

uh which version did you download @proper trellis ?
cause if you downloaded FP16 while your settings don't want half
then you've just blocked your model
not sure tbh i just clicked the big blue download button xD, but i deleted the file and redownloaded it and it seems to be loading now 😂
gannuburadu'
the menu on the right has 4 variants
WEBUI dose not need yaml files, delete all of them, next to models @proper trellis
fp16, fp16 pruned, fp32, fp32 pruned
oi
I made gunblades @magic wave
you want an upscale
i put this image of a 3D sword i made into img2img and it made the 2nd 1 lol
aha
I just prompted for a gunblade and it gave me one
@quick hill @magic wave @vagrant falcon @gloomy scroll gunbladers
the real test from your model is how well it can do slime 😂
Daaaamn
was that the slime result or?
Oh damn that is pretty dope :o
just a pure prompt
nah that was the pre-test, running slime prompt atm
looks so cool
so this 1 is probably the safest i can share 😅 but its pretty cool, reminds me of nai
this was my other slime prompt,
lmao well a slime lantern
cute lantern tho
https://github.com/suno-ai/bark who's gonna get this running for waifu voices
some times it makes slime, other times it makes glow in the dark cats 😂
blue_hair anime girl with a green plant behind her, Tatsuro Kiuchi, raining, wet_hair, sunbeam, sunlight, water_bubbles, water, Grzegorz Rutkowski
bad anatomy, disfigured, deformed, malformed, mutant, gross, disgusting, out of frame, poorly drawn, extra limbs, extra fingers, missing limbs, ((blurred, blurry, blur))
(version 4)
try adding line art, sketch art, acrylic art, pastel art, oil art, water paint, cg, 3d to negative
it usually makes things a little less flat
@proper trellis
I would really love it if you all can leave a rating or review once you feel comfortable using it
it does have a nice range to it, ill share some results and rate it tomorrow but for now its bed time 🙂
yeah I'm happy about the versatility both in prompting and styles
by that I mean, you can prompt it very simple
she exploded out punis
oh dear
Tried to train a LoRA for cool swords, turned out kinda shite tho
my model just kinda knows it, sort of
try it out and try and use my model as a base for training gunblades
for when you have to pick a lora base
you know what I mean?
unfortunately the body and face are fucked
re.attaching the blade
here though
wonder what that logo is for
elite sniper? elite wargon?
when your gunblade fails you
@quick hill @magic wave well uh....as you can see
@proper trellis
and @lyric zealot you too
look at the adorable puppy
jesus christ
so how do you like my models gunblades then @quick hill @magic wave @proper trellis @vagrant falcon
They're kinda funky but cool :o
Sinlaire?
I did this some time ago but total failure
Fair, I mean, its a really rough dataset xD
Its a kinda thing that only humans can make for now xD
Though the AI does try its best
How to get this art style?
By prompting really really hard

I’m just kidding, find a model that might mimic this style
Seems a little like ligne clarie styling
So might wanna look towards that direction
guys, is it possible to do this position only with prompts?
probably
let me try
you can make a concept lora, and get this look all the time
I haven't learned how to make LoRA yet
you can do image 2 image with controlnet, same thing pretty much
yeah control net would work well
best controlnet to color mangapanels?
I would guess lineart
hey anyone have any ideas
i keep using this as my control net openpose but keep getting these rings
any ideas
I'll try that, thank you.
Remove/lower the emphasis on masterpiece and/or portrait as well as any references to the words painting, frame, or framing. It's trying to draw a picture frame.
gives a new meaning to "in frame" 😛 wait you got framed
Or try 'framed, borders' in neg prompt
I thought this too but even when I do 1 girl 1 boy I still get it
its a trap
seems no new waifus coming, did everyone get banned from colab
do you know who is this ? 😄
looks fantastic @south sable
All these SD skills and months of experience and I still can't figure out how to make inpainting work properly.

lmao I feel that
I don't dare touch inpainting, lest my greatest fears come to life
turn up noise and number of steps
Done that, no help.
Changed prompt, erased negative prompt, changed inpainting model, changed from fill to original to back again, changed from whole picture inpaint area to only masked, changed sampling methods, changed mask blur, ensured resolution was set correctly, cut out a smaller section that I wanted inpainted to ensure the original size wasn't too high, changed CFG scale from high to low to medium. Changed area that was being inpainted. Tried negative embeds. I've tried just about everything one can.
What exactly are you trying to inpaint?
Left hand
ah yea idk if inpainting is gonna b able to fix hands lol
might just hav actually fix it manually
but u can probably paint in an ornament on the empty part of hair for example
or paint out that fish
Ah, yeah. Inpainting hands is just playing a gacha. Around 0.52 denoising would be your best bet with that one, inpaint mask halfway up the forearm and use inpaint area only masked. After that, it's basically just generate until it gets it right.
The better solution is to fix it manually as well as you can then inpaint over the fixed version at lower denoising to make the style more coherent without changing the form much.
Thanks. Yeah, I've cut out a hand from a similar gen and will be trying to meld it in with PS then a local img2img, then I'll ease that section back into the original image. SD is probably taking the wristband and seeing it as a line break of sorts and treating the hand as a separate entity from the arm, which is partly what is giving me such headaches when trying to inpaint. I appreciate the tips, especially your area only and denoising suggestions. That'll help me going forward. 👍🏻
not the best i can make but it looks cool
tensura movie out
#typography
Oh even without the stealth PNG info extension?
That's cool, but why the hell do they keep going back and forth on it?
I guess we'll start working back on the bot for reading those :p
I don't get it either. it's a security problem or it's not, it doesn't change with the weather
Exactly, I really think it probably should be stripped
I don't really know their global userbase on the whole discords so it's hard to say for us, how much of a security danger it is, but I think it should be a setting for admins
by default, all channels would "strip metadatas", and some channels made for image sharing would have the option turned off
that would fit all types of server and solve the security issue
Hii 
Oh damn
The gothening 
What character is htis
She's not very well known, maybe you've heard of her though. Her name is 1girl 
It's just a prompt I stole from someone a while ago.

So sayeth we all
Luna dark mode version
Yes 

Cresentmoon oranement hair+tuxedo
I should try this
XD
Wha? 
Just the hair remind me of my pfp
And I should mske my dark version
Ohhh, thats fair then 
Oh fuck
@magic wave can u try for me xD

Sure 
Tis what I did :O
I did tis :O
I did realistic 
Yeee 
Lemme do 

Fair fair :o
Putting other priorities before waifu? Unacceptable 
First generation
XD
2nd one, cooler angle 
Wider shot 
I mean, 3 generations, and they're going kinda in order XD
Cleanest one so far I think :o
@quick hill
The Vampire char of Paragon 😮
Kinda yeah :O
Where can I find a list of good anime prompts and negatives ?
That's a great question. Sadly I haven't found a centralized source for such things, but if you click on people's images in this channel and click open in browser, then load that image into the PNG Info tab, there's a good chance you'll see what prompts & negative prompts they're using. Usually, people who have black text in their username have been recognized as accomplished prompters, but many of us who post here often are as well.
Otherwise, you can look and find images on Civitai.com and see how people have prompted for certain models and otherwise. I'd recommend starting with the Aurora anime model, as it's arguably the best and most flexible at the moment. https://civitai.com/models/40199/aurora



Is PNG info a auto1111 tab ?
Yep!
The image's info (if it has any) will be on the right of the inserted image.
Created an interesting artist combo for ALunarDream.
yukoring quasarcake torino aqua mika pikazo
G'night folks.
Hi, This is the best face quality I can get, what is the correct way to have good faces ?
What model are you using? What VAE are you using? And at what resolution are you generating?
This was done with WaifuDiffusion, 512x512, what is VAE ?
(masterpiece),(best quality),(ultra-detailed), isekai masterpiece anime boy sitting on a rock in the middle of a forest with trees in the backgrounds, high noon, cinematic, very warm colors, intense shadows, ominous clouds, anime illustration, anime screenshot composite background by mandy jurgens, by irina french, by rachel walpole, by alyn spiller
(low quality:1.3), (worst quality:1.3)
I also use openpose and depth
https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt
This is a VAE, download it and place it in your models folder, then rename the .ckpt part to .vae.pt
Then you can go to your settings, select it and apply it, should be in the Stable Diffusion settings tab
(it's img2img)
And you should try out this model
https://civitai.com/models/40199/aurora
I'm actually using rundiffusion, I cannot install custom models/Loras
Well kinda but couldn't do it at all with my gpu
Whats your gpu ?
Also I assume RunDiffusion is pretty fast, so you should also try using Hires fix, try it with x2, 0.5 denoise, and the R-ESRGAN 4x+ Anime6B upscaler
Well actually that wouldn't work for img2img, but for txt2img, you should try that
What is Hires fix ?
No cleavage, wow 
This lil thing here
Its pretty essential to getting consistent high quality pics
where do you get fatal anime ?
bruh
intel integrated x)
i also have gtx 830m
Yeee, in img2img you can use the SD Upscale script, but its a bit funky to use
tysm
And usually not as convenient as hires fix in txt2img
You're welcome! 
Okay yes thats not enough :/
Its doable with the sd upscale script
I mean, if you upgrade a bit, you would be able to run it fine, I was able to run it with a GTX 960 which I think is like the bare minimum
😅 I'm still lost on how to have accurate faces
Yeee, is what I said, it is a bit wonky tho 
it's a laptop
You can also try to inpaint faces
Oh damn that sucks then x.x
Depends on the settings but surely not as easy as hires fix
Yeee, I usually do SD upscale as a like a final pass for pics I really really like
I think I did a double pass for my last POW submission, which made it end up being like 4K res xD
you generated low qual and then take the seed to highres later?
What you mean? 
What should I do ?
i want to generate a few "batch" but not with highres fix
like you know to create poses and stuff
but if one is good i want to highres it
Well, then you can disable hires fix, and you could re-enable hires fix using the same seed of the one you like the most
oh okay 🙂
Instead of using Whole Picture use Only Masked instead
Awesome !
Yeee, works better that way
After an inpaint how can I make the image more coherent, I mean so the inpaint doesn't seem oddly placed over
Well, you could try doing more inpaint, or use less denoise, for faces I usually go for 0.5-0.6 denoise
You're a god thx
Glad I can help 
I was 0.94 denoise tho
learned so much with @magic wave its insane
Thats too much denoise xD
Ayyyy 
You inpainted more? :o
if you cannot use custom models with your current setup, it is a bad setup
you might wanna try this one https://colab.research.google.com/drive/1wEa-tS10h4LlDykd87TF5zzpXIIQoCmq
Lmao what is that thumbnail
I use an extension that shows me the middle of the video instead of the thumbnail because I despise clickbait
That is a really good idea
yo this config is insane
i want the same in my PC an ultra optimized one
PLEAS SEND ME AKMFNDKDND
Omfg
I need that
https://chrome.google.com/webstore/detail/clickbait-remover-for-you/omoinegiohhgbikclijaniebjpkeopip
Replaces thumbnails and modifies titles of videos on youtube.com
Wow can have prooooompt?
that's gonna be gone soon, stable diffusion with colab gonna be for paid subscription users only
Oh no, based on what you saying that though?
This from the wd server but I mean it's also been everywhere
also reddit post is where I found out https://www.reddit.com/r/StableDiffusion/comments/12tj22r/official_statement_from_the_person_in_charge_of/
183 votes and 103 comments so far on Reddit
Google-kun pls forgive me I won't botch my lora training for the nth time ever again
This is besides the point but Google might (it definetly is) taking a huge L in the AI space
Is there a way to cancel a generation from the command prompt?
upscaled with R-ESRGAN 4x+ Anime6B like you said 🙂
I have no idea how that works
learn
also it's pretty simple
just read what it says, and if you don't understand just press it until it works
then if you get an error, google it
search on youtube to get you started or go over to rentry
800$/month apartment in New York
yes
this room actually quite spacious in new york standards
I love and hate my city.
ah toying around with sound bites I made for my VN, placeholders, kind of, still though. Maybe I can get people to record their voice saying certain things for free in certain tones etc. Then I can use that data to train a voice model and have them sign away the rights to their AI voice
hawhaw
what program are you going to use? I hear elevenlabs is pretty expensive
I managed to scrounge together 3 hours of time through referrals at Replica Studios
which would otherwise cost quite a bit
so I got that
who what?
you
😂
concerning
That would be the dialogue I wrote for the chancellor in one of the story arcs in the visual novel I work on
it's text2speeched, so not me or anyone actually speaking it
I will finish a prologue after my surgery if all goes well
and here is the protagonist!
Ayy das good :o
1girl, bag, bow, cape, dynamic composition, motion, ultra - detailed, incredibly detailed, a lot of details, (amazing fine details and brush strokes), frills, ((texture hair)), green_hair, hat, hat_bow, holding, gradient, (Superdeformation art:1.35), (Chibi Art Infusion:1.25), (Digital Illustration:1.20), (soft_hair), (fluffy hair), disheveled hair, floofy hair, (poofy hair), long_hair, long_sleeves, gloves, looking_at_viewer, ribbon, ((photo realistic)), (1girl:1.2), cute, (full body:1.2), white_eyelashes, ((chibi)), (cute), simple_background, thighhighs, wand, witch, witch_hat, (masterpiece:1.2), (best quality:1.2), (high quality:1.2), detailed, 8k, 4k, (extremely detailed), (highres:1.2), (detailed face:1.2), (sharp detail:1.2), (volumetric lighting:1.2), ((soft lighting)) <lora:minimalistAnimeStyle_v10:0.8> <lora:playfulWhimsyLineless_v2:0.8>
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++ SDE, CFG scale: 7.5, Seed: 1947764678, 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
Omg
i recommend elevenlabs for tts that sounds natural!
not that much you pay for what you use
yeah and 5 dollars i genuinely plenty
30,000 characters isn't a lot
https://www.youtube.com/watch?v=TUSqEQUgNlM this whole video is my vocie cloned and ai generated using eleven labs
Join our discord community at : https://discord.gg/y2zjSsNNtg
if you have a suggestion on what framework we should cover next let us know!
when u go over it u pay a for what u use
check the video above
yeah it sounds natural
but also they have characters
I will likely use up like 50,000 words or more for dialogue
the characters in that may be in the millions
^
I don't have that money
or if u really dont have money use castingcall.club
I mean there is more expensive plans, next one is like 99$ I think for 100k characters
im making a horror game rn you can hire actual vocie actors that want portfolio exprience
and its for free u just agree to pay after profit is made or be payed with credit at the end of the game
so like
crediting them
I'm about to go into surgery for cancer and then I need money for recovery, medicine and maybe other treatment
i used to do voice acting there myself its pretty dope
I can't afford it
500k*
and that's that
Fair :o
Or I go to VAA
so many people make gacha projects there and other things
been around for ages
i prefer castingcall personally i got some amazing VA's for my horror game
or use open source, or use the 3 hours of referals I've used to receive that credited time
for replica studios
I can use adobe audition and whatnot to modify inflection
and formant and such
I also don't have the energy to hold casting calls and direct voice actors from long distance. Most will probably also not have high quality sound recorded in a voice box or audio treated room
there is open source vocal ai
and I can edit things in adobe audition to move keys and formants
so they don't fall flat
O das pretty nice then :o
like this, gonna alter it a bit
Fair fair :o
Where do you live? (Cuz of healthcare)
Sweden. Most of the charges are covered but not hospital stay, medicine up to a certain cost isn't covered, neither is healthcare (but its still relatively cheap.) After-care costs more, like at home help, rehab material
last time I had cancer surgery it ended up costing around 2000 euro and that's with Swedish healthcare
Sheeesh
that included medicine, doctors, uuh "sick travels" (cabs etc that drive sick people), per night stay at hospital, some rehab stuff etc
sure but I live on disability and 2000....I couldn't save up to that in a year
Here thats also covered by public healthcare insurance
or even 2
last time I received help through paypal donations
this time nobody is helping 😛
well one friend is
Damn that sucks x.x
Well, I mean, at least a few people are :o
yeah but I don't mean to sound ungrateful to them
but it's only 50 euro
total
it will help a little but not gonna do much overall
Fair fair, I mean, hopefully you can still get there
I mean, true :o
Though I mean, you can take out loans ye? Worst case you go into debt, and you pay stuff back when you're healthy again :o
or well I dont refuse to pay
I just can't
and no I can't take loans
how do you get a loan when you're on disability?
who's gonna give it?
except for high risk loans that have 40-50% interest
O, thats true, that sucks then x.x
but I don't care as long as I live
Thats fair :o
@magic wave 100 units GPU is good for google colab (pro) ?
isnt it using a lot to do ai?
i would love too
Don't pay for colab for image generation. Use free tier. If you need more than free tier gives you then Runpod is probably cheaper and better.
True ☝️
its better to have it locally then
Ye
I thought they were gonna remove the use for SD on free?
remove?
IDK, I've used it for generation only a handful of times. If they do, then it won't be worth using. Their credits system is too expensive.
yeah but this is new
229 votes and 156 comments so far on Reddit
Maybe, I'm just saying that it isn't worth paying for pro for image gen.
locon and light and shadow made some difference
attached a hypernetwork and her boobs grew bigger
gm
Daemon girl, crimson currant eyes with merigold limbal ring and slit pupil, long slate dove colored hair with charcoal bangs, (graphite colored markhor horns:1.1), in the snowcapped mountains with her furry dragon pets, beauty filter, soft skin, glossy skin, (dark fantasy art:1.175), (digital art:1.15), (digital illustration:1.15), (western fantasy art style:1.1), DND, (Splash Art Style:1.15), (2.5d:1.175), (soft lighting:1.10), ultra wide angle, high contrast, (masterpiece:1.15) (best quality:1.15), 8k <lora:locon-noise-offset:0.76> <lora:lisaLoraCollectionOf_singleLoraLisa:0.25> <lora:lucyCyberpunk_35Epochs:0.36> <lora:crazyExpressions_crazyExpressions:0.55> <lora:daemonicus-000002:0.77>
I think this might be a disaster
For the price of 100 Colab compute units you would get 34.4 hours with a community 3090 on Runpod. Not sure exactly how much time 100 compute units would give you because google obscures that, but it's definitely less than 34 hours.
rare highly clothed gen
yeah it is
CLYDE 
KURAIDU
gm 
Also gmornin @stable crane 
gm 👋
she be daemon
@stable crane I released the model, try it and do not scold too much :D, https://civitai.com/models/47067?modelVersionId=51650
downloading now
daemonus, daemonas, demons, daemon, 1girl, solo, dress, grey background, claws, sharp fingernails, fingernails, breasts, cleavage, simple background, black dress, monster girl, long hair, veil, wide sleeves, full body, long fingernails, long dress, chain, long sleeves, colored skin, covered eyes, grey skin, large breasts, (graphite colored markhor horns:1.1), in the ash covered black mountains and dark clouds with red lightning, (dark fantasy art:1.175), (digital art:1.15), (digital illustration:1.15), (Splash Art Style:1.15), (2.5d:1.175), (soft lighting:1.10), ultra wide angle, high contrast, (masterpiece:1.15) (best quality:1.15), 8k <lora:locon-noise-offset:0.65> <lora:Daemonicus.E3:0.70>
Good luck me
ahhh, wholesome
What did you train? :o
Even if you get bad results, leave feedback so I can improve the model!
O, fair :o
Looks nice from previews
All Exif in previews saved can be used
Yea
no no it was that one
daemonus, daemonas, demons, daemon, 1girl, solo, dress, grey background, claws, sharp fingernails, fingernails, breasts, cleavage, simple background, black dress, monster girl, long hair, veil, wide sleeves, full body, long fingernails, long dress, chain, long sleeves, colored skin, covered eyes, grey skin, large breasts is the caption it received
daemonus, daemonas, demons, daemon, 1girl, solo, monster girl, arthropod girl, black hair, spider girl, long hair, silk, taur, arachne, colored skin, spider web, breasts, full body, nude, arthropod limbs
wonder what this would do
This neko-chan with the nyancast of today...the rain is wet.
@teal pivot first gen, it's quite nice
Woah great
I'll post a review on civit when I get a collection
Thanks!
it's charming
weirdness
Should I keep working on it or find out which epoc isthe best and release it?
i like it
Clyde is bullying 
I've nade 4 epocs at this point
dunno which is best
please do try it though, @stable crane
the daemonicus?
MHM
a bit blurry but a really cute Lotta
will do
THAT CHARACTER'S NAME IS LOTTA? 
not bad i think?
I like the style a lot, lacking bit of crispness in these ones
it's got a neat character to the mech stuff
I think Clyde really dislikes this prompt 
some of my NAI trained LORAs are a bit blurry in general on this model
but I like the model a lot, maybe will try retraining using it as a base at some point
i see
tbf my loras are blurry on a lot of models too lol
In any case, leave feedback with good or bad points, I will improve the model, thank you
Finally one of em sent 
Wow.... vladmandic fork of Auto1111 is insanely slow to setup
What that for do? 
A1111 was insanely slow for me on bad internet
Same
15 minutes and counting with gigabit fiber. Normal Auto1111 is usually done in less than 4 minutes for me.
O that is pretty slow
😮
for me it was hours lmao
For me A1111 is like 2-3 hours xD
Giv prompt 
straight-on, leaning forward, standing, 1girl, (dark green hair:1.1), frog hair, full_body, large breasts, (green frog hoodie:1.2), (bodysuit:1.2), green leggings, (thick thighs:1.25), narrow_waist, thin, lily_pad, swamp, slight smile, arms_behind_back
Imma try it otu 
It's the most up to date fork that I can find and seems to be the most popular. I just wanted to try it out https://github.com/vladmandic/automatic
I dont like the vlad one, it's a lot slower than auto
Tis what I got :o
any idea what is this shadow under the hair is called?
nice
finally done after 20 minutes.
shaded face
the one that gives disappointed vipes
You have a LoRA for that right? :o
if i type that i would get similar results?
I don't think so
sometimes, depends on your model
Also tried adding (ghibli style:1.4)
Ohh, thought you did 
1girl, rosaria (genshin impact), solo, black background, multicolored hair, chibi, simple background, red hair, purple hair, purple eyes, nun, shaded face, streaked hair, bags under eyes, veil, hair between eyes, short hair, closed mouth, looking at viewer, habit, bangs
you might be thinking of bags_under_eyes
found it
ohhh nice
Ohhh, you're right
Yeeee, knew there had to be one :o
damn where did you get that
Was vlad worth ?
Tried anime style :o
Not sure, still have to do more manual setup and figure out what facexlib is and try to get it installed properly because the setup.py failed on that one.
damn I keep losing track of what channel I am lol
Tried (3d, real_Life:1.4)
sounds like pain, xformers support?
real life < what is it ?
grass
xd
That probably explains the one review you had where the image was in .jpg and they were asking why it was worse quality than yours.

Real life? :O
real life exist?
Perhaps 
Also worse quality though. I only accept the highest quality bespoke artisanal waifus 
I haven't found it to be that bad
Real life is a lie
usually blurriness is the gen, not the jpg
this can do avif and such too though right?
gonna do a double check on a blurry gen to be sure lol
hes making extension
it would be hilarious if my blurry complaints were actually jpg
Denoising diffusion models (DDMs) have attracted attention for their
exceptional generation quality and diversity. This success is largely
attributed to the use of class- or text-conditional diffusion guidance methods,
such as classifier and classifier-free guidance. In this paper, we present a
more comprehensive perspective that goes beyond the...
yesh
Self Attention Guidance
Whats the TLDR? 
I'm mostly kidding. I would just rather have PNG. I do save both .jpg and .png of large gens and can definitely see the quality difference.
fml the png was slightly crisper, more tests needed
png always crisper
Basically, more = better?

@supple raptor how do you save both? can I just enter 2 file extensions in the box?
prob ended up like another hires stuff

Gunna spam ashen github in the mean time
Iiii seeee? 
There's a setting to save a compressed .jpg version of images over a certain dimension or filesize. I think the .jpg quality is set to like 80 or 85 by default, but you can change that too. Maybe in image options? I would check but I'm still working on setting up vlad.
This on Aurora btw
gonna take a while, hope it's worth for the efforts.
Probably wont be XD

have you tried comfy ui?
Yeah, it's neat, but it's a pain in the ass to use.
so cute
hq stuff
This is nice, you can specify the default multiplier for LoRAs so that they don't always start at 1.
vlad?
shaded face
prettyy good stuff
I thought that could be done here :O
You know.... I suppose I haven't looked through all of the settings in Auto1111 in a while 😅
Fair XDDD
Lol, I thought the first part of your user name was just short for "a lot of".
Nani dafuq again Vlad. These are the default hidden samplers
Funny. "Comfy UI" is a PITA.
it does 😄 originally, but then I made the character, and named her Lotta because it also doubles as a name
what model
@teal pivot 's new model
It's just a lot more work to setup to make an image. It isn't bad if you don't need to change things, but as soon as you want to do something that would be as simple as ticking a checkbox in Auto you have to rework a bunch of nodes.
try to use grapefruit vae For a lower contrast
Yeah, I don't doubt you, I just found the play on words humorous. 👍🏻
oh yeah I'm on blessed2 vae
blessed gives a very high contrast, I tried haha
? Do you mean high brightness? I've found that it generally is of much lower contrast than anime2
guys
do you know of a good free alternative to colab with extensions and LORAs?
Blessed2 still gives me that awful hexagonal pattern that all of the redacted based vaes give. I have no idea why some people don't seem to have the same problem with it.
sorry if I had to ask, I heard of what happened.
if you have a good GPU, yes
gtx 1650 8gb
very strong dark areas i mean
I wish I could just use the same stuff as the WD 1.4 VAE, but with less saturation, the saturation can be too much sometimes on it 
idk if that's usable
I like her spider butt thing, it's a carrying cart
It is usable 
not too good for models with high amounts of memory.
You should be able to run SD just fine
You can make it https://github.com/sALTaccount/VAE-BlessUp
Like when I tried using Kenshi, or Dreamshaper.
Models go into the RAM, so if you have like 8 GB RAM, you can run fp16 just fine, 7~ GB models may be a lil tight tho
7gb, and 5 gb respectively.
Its better than both of em, and the GOAT™️
My local RAM is 16 GB, VRAM is 8GB
Yeah
8gb vram is fine
ram u dont need much aside from loading weights and merging
xD
oh there's 2 Smiles
XDDD
and 12 lunas
Correct 
Does it take a lot to download? Like how big is the download in GBs? 
There's a 3rd Smile 
Like extensions and LORA good?
Doesn't seem like its worth it 
I mean, they're built in, but they're available in the extensions tab anyway
So its like
It saves you like
2 minutes of time
Of clicking a few buttons to download them
Ooooh, i joined the 3090 gang 😄
So, stick with vanilla Automatic1111 for the time being.
LLM's are pretty much damn faster on that
A1111 has lora built in too
True 
Same
same
Gib me one of the A100s, they wont notice its missing

it's smol this is with the venv and 1 vae included
Boy I wish we had a headpat reaction on this server.
O, I'll have to try it out then 
Little give me two 4090s
Only problem regarding my local implementation is the speed of generations
Main reason why I went for colab is mainly because of that.
...thanks... 🥹
yeah I'd expect you won't get anything nearly as good without local
;)
convert $ into generation speed
XD
O M G
How do I speed up gens in my situation?
Evangelion is mad cool
Yes
Guys, is there any way to speed up generations apart from --medvram, and the --xformers?
Other than smaller gens? Not to my knowledge, no.
Oh, and using a faster sampler, less steps, etc.
I mean, I use 2M Karras at 40 steps
What GPU do you have and which version of pytorch+cudnn are you running?
Try 2M Karras Test. It's fantastic. So much crisper.
Karras test?
Yep
that would require pulling
GTX 1650
let me check on the versions...
New as in back in March, lol. I don't know if you've heard of it before.
Add it as an additional sampler instead of changing your existing one https://www.reddit.com/r/StableDiffusion/comments/11mzsdj/adding_modified_sampler_from_hallatore_quality/
The code listed on my link doesn't change the existing though? It does add it.
I'll try that 
Oh, he edited it. Originally his code replaced it. I should have looked.
O, I broke it
Did you, say, happen to back up the .py file before making changes? Do you still have notepad or whatever you used open to revert the changes?
I can send you my modified files you want
I don't know if that'd be good since I haven't updated 
how do I check for pytorch?
I don't think there's been any changes to those files. But I can send the ones from my UX fork which are otherwise definitely the same as the main repo prior to the gradio changes.
I think I fixed it 
Seems like I did, there was a weird indentation from copying and pasting the first part of code 
interesting
Normal 2M Karras vs 2M Karras V2 
welp
I kinda found an alternative. Someone made this:
https://console.paperspace.com/github/TheLastBen/PPS?machine=Free-GPU&file=%2FPPS-A1111.ipynb
2m v2?
lol
What? There's a clear difference.
lmao
Blind. Lol
ty
hard to say if one is better, just different
what if you use blessed2.vae with that
reintroduce the contrast that got lost
It looks meaninglessly different to me xD
Thats big diff
I prefer left
I prefer left, coloring on left is much better imo
If you can't see it, then I can't help you. Others do and have. There's a reason people use it, lol.
But right have more details
Euler A is a classic. Still great.
XD
Lemme open up original
try grab it from civitai
I feel like I am being trolled 
Is the outputs still good if I'm using the FP16, or would it just be degraded?
open the first pic on discord from this message, and press arrow key left and right
Read the responses of the people in the link that both I and nubby posted, lol.
to see the difference 
Outputs are pretty much the same with no degraded.
in the github PR @untold glacier ?
color grading differences
Yeah. The discussion has people talking about it.
I've tried 2M V2 before and differences is HUGE.
Try a prompt that have very vibrant colors
I cannot see any difference that xformers wouldn't do on its own
Lemme try one of my sword ones
Try it on things like background
Or scenery
Since it focus on details and color
(detailed background), cloud, fruits, colorful
the right lost the little lines on cheek
Yea its noticable
color mostly

Need to try scneery
Or such















