#🏞|general-with-images
1 messages · Page 38 of 1
that represents your desired art style it usually works
well, what I mean is without calling it. IN THE STYLE OF trigger word class (normally style).
if I do that I get it always
if I don't it is iffy
I had one failure I am unsure why it happened and I posted it in here but no matter it looks like it would need 3k+ steps and for a loha that is insane
it really didn't start to look good until step 1900 out of 2k as I was image every 100 steps to see
to be fair, it CAN generate stellar hands and fingers, it just doesn't get it right most of the time. i think it just chooses not to so that the memes don't die 🙂
(i know that's not the case. it's a joke. i don't actually believe diffusers are sentient -_-)
@smoky oak How do you clean up your source training images?
depends on how bad they are. Send me some examples in DM's and I can help you
or send them here really
there can be a lot wrong with them, some things easily fixable with tags, some requiring editing
Well, they are from the 1970s with capture blur and noise etc...
Can you send a couple examples?
Are you trying to mimic that specific look? or the underlying information behind the grunge?
stylistic. I have seen this done with people where they digitally did it because they were old photos. Gigapixel isn't going to clean these up.
@dense tapirPlease send me a couple to fiddle with, I'll see what I can do
I got it I had to refind some really old info
how do i get SD to make its own named for images when using image2image batch
make its own name?
yes instead of using the ones from your inpute directory
It keeps using these I would like it to use its own names
ohhh, I am not sure
that sucks
What kind of model would be good for making someone out of something else?
For example, the French president's face composed entirely out of garbage
I think you could do that with any model, but using controlnet lol
I tried using a few different models but it seems to think of putting the face/head in a rubbish heap, not making it out of one.
How would I use ControlNet when the composition would be something different? Or perhaps use his face image and preprocessor to create a map and then use the map with a prompt?
Looks like my Ferret LoRA works well with anthro models when fiddled with a little more
man, I wanna learn how to use deforum
but there is so much stuff, and not a lot of documentation
also, these mathematical functions are intimidating lmao
Even with all the advances they've made with it, it's still pretty much in its infancy. Genuinely going to be incredible if it continues developing though
(10*sin(2*3.14*t/10)), 50:(20*sin(4*3.14*t/40))
how does one even compute what this all means
I can't even find info on how to download deforum
extensions tab
@smoky oak 10 images how many repeats do you use? A tad different for subjects but just wondering.
for me, I would use 75 samples per image, 2 epochs at 10 images for a not overly diverse dataset of a subject
75 samples? you mean repeats?
yeah, my bad
damn
why?
so 1500 total steps
I do batch size 2, so 750, but x2 so yeah, 1500
it takes usually about 6 minutes on my GPU
LOL, fuck
BS 2 I can't even do with 15gb of vram unless I use checkpointing
do checkpointing and almost double the time (about 1.5)
How can you do BS2 with 8 gigs even with checkpointing?
It uses about 7.2 GB VRAM
I am wondering if I could do 768x768 at BS1
not that I really care to, personally
can't help it as that is what it is taking on the T4
very very weird
no checkpointing, remember
I am not sure what the difference is, unless you mean epoch checkpointing
gradient checkpointing
ah, I don't remember if I use that or not
Ahhh.
I turned on checkpointing and BS2 but doing 768x768. that checkpointing on colab I take a serious hurt on speed
36m
2.71s/it
this is why I don't use checkpointing and stay with BS1
this looks like it will be a good one as my loss is 0.169
for 2.1 that is sweet
especially this early
0.163
I will see every 100 steps how it is going
damn that is sweet
I am using deforum currently
damn, fisr image is crap
*first
Is that for real that Amazon prime is now 179 a year?
Just saw an advertisement
currently trying to find good settings for Deforum
but oh my god there are so many bro
I bet
it is getting better just wow
0.152 loss too
I will try it next with BS1 and checkpoint off
HMMMM something tells me something messed up lol
23m
Just a hunch
weird
I keep having the same issue
first frame is gorgeous
every frame after is low res and full of noise
hmmmm
never touched it so I can't help
400 steps (all same seed)
doing something as it has writing now which I do not have in my training images
imagina playa
it was made to be a fractal haha
years and years ago I had a program that would make them. I played with it for ages. fascinating art
yeah the program I had would do all the variations. Cant remember what it was now
They are memerizing
*mesmerizing
Welp, another fail
almost not a fail though
just nothing in my style I trained on
512x512 no gradient checkpoint
bs 2
this is why my min is 16gb card
16gb and over would future proof you to some degree also
i am already saving away for the ada titan. i wants th 48gb and i think my current build, possibly with a psu boost, could facilitate it
the price I can only imagine
the glorious return of the titan line
I can barely afford the 4090 by Dec so an Ada Titan is out of my league 100%
i got a few hustles to do yet and i might be able to position myself for one
5090 will be 3nm and Huang said that himself
very nice indeed
5090 leak they are redoing the entire rtx lineup from the ground up (have to tbh) and expect 2-2.9x speed over the 4090
i also hope that it'll keep it's valuation over time so when i sell it i can get something like a 5090 or 6090 (heh)
50 series wont be till mid-end 2024. still a long wait
why sell it? 48 gb is just too good to sell even if cards are twice the speed at half the ram
could just server room it
Titan has a fan though
Ahhh yeah
render farm for SD would be cool
Alright this second training is looking much better with slightly worse loss but I did two changes. Either taking 768sq to 512sq OR (which from previous testing is the culprit) removing gradient checkpointing makes the training better.
13.3gb for 512x512 BS2
this is why I don't train much as so many failures taking so long then google shutting me down for the day
@smoky oak Must have been a typo or insider knowledge of what is coming but it is now $139 per year or $14.99 per month. ffs. I don't watch TV so I would get no use from it outside the very rare shipping
When I do buy I just make sure I bundle and go over the minimum so no shipping charges
I bet the 179 was for those who pay monthly so makes more sense due to it being an ad "on how to save even more" types found on msn
ah, yeah
A sort of yearly spending bundle
I am currently struggling to figure out how to stop my deforum gens from going so soft
Well, I am getting nothing but failures again
gl
screw this as all of this has been for nothing as everything is a horrible failure. 😦
I switched back to adamw8bit and 100% fail can't even see it
I don't mind failures as long as I learn from them but I haven't a clue with this why this is happening
@dense tapirI am starting to get results :D
This one is rough, but it was my first success
trying a 768x768 x2 framerate gen
That is cool
btw, what do you set your alpha and dim to?
big files
I have a sneaking suspicion it is alpha biting me in my ass
why? Anime numbers and all that nonsense they drummed into me
Oh, this one is a huge step in the right direction :>
Oh, snap
Thats pretty dope IMO
yeah
What can I say, i am a fast and efficient learner 😅
no guides, just effing around and finding out haha
seems to love eyes
at 2k steps this is not really getting better so I am going to up alpha
A melodic dream with geometric fractals and magic
A forming space nebula full of stars and planets
Eyes of different colors
Shooting stars over a lake
my loss is low so my dataset is fine
@static tuskI prompted in the eyes haha
I see
🥁
Alright
trying a 400 frame 60fps gen at 512x512
oops
using way too many steps, but oh well
strange that by adding more steps it fucks up their faces
none of this has been what I trained
gotta love the hands
@dense tapirGod I love being bottlenecked
I continue to be in desperate need of a CPU upgrade
nice gens
a cute kitten made out of metal, (cyborg:1.25)
Negative prompt: nfixer, (deformed, distorted, disfigured:1.25), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers and tail:1.25), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, flowers, human, man, woman
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 4, Seed: 2027707380, Size: 1024x768, Model hash: 4fefc62aaa, Model: oxigen_v10
could benefit from some upscaling for more detail IMO
can you neg prompt "blurry background, bokeh, out of focus, depth of field"?
can try
I give after 3k steps this last try it just isn't going to work, and I have absolutely no idea why.
3h of training many times too
Here, how about you send me your data set and I will see what I can do?@dense tapir
I'll see if I can figure something out
It's a style though. It isn't so much I want it done rather I am running into this issue so much I need to know why it is happening.
I could still try and see if I find any issues
wow lol
So is therea section for NSFW content here?
No there isnt one on this DC
That can happen ^^ feel free to share some SFW here if you have that too
You can read in #✍🏼|rules-and-tos if its against rule 4
im not sure should I Dm you the images?
Yea sure
HuggingGPT 😮
wow is it open for testing?
At last something from 2.1 that i like as much as 1.5. Made with rmadaMergeSD21768_v60
that depth of field with kind of a different style on it makes the character really pop out, like a cut out, nice !
Wish I could find a model that didnt focus on characters. Its all I find on civai and the nsfw models.
maybe some of those ?
https://civitai.com/models/83/min-illust-background-diffusion
https://civitai.com/models/30363/cheesed-out-anime-backgrounds
https://civitai.com/models/5536/fantasy-background
IF YOU ARE THE CREATOR OF THIS MODEL PLEASE CONTACT US TO HAVE IT TRANSFERRED TO YOU Created by ProGamerGov Originally uploaded to HuggingFace From...
Cheesed out Anime Backgrounds Please note: GENERATION SERVICES please ask, because this has another model mixed in, and it may not be kind to yoink...
Mmmm a tad too basic, I'm looking for paint stuff like watercolours, acrylics
look for pastell anime style based stuff
or try that in a prompt
which one is the best to go on?
well, i think i have to reduce the fireplaces or chimneys
but love the deatils overall
good composition, its hot but not bad at all to start with
Every model ive tried so far that looms good visually has been a tad too horny giving me women and stuff in my.prompt when not in the prompt and in the negatives 
The one above is from protogen
mine arnt made off characters lol or NSFW lol
Как сделать картинку?
1
Got roasted
Can any help me with a test, I can not say what I am testing, The test prompt is "photo of a sitting female holding a machine-gun", preferred model in SD 1.5, then do 4 to 9 generation, post one of them and say if you noticed anything odd.
I get a strange result and I want to see if this bound to me and my AI or if this is the same within Stable Diffusion.
I can help post me the pic in DMs
I have no pic, it is the prompt that is the test, to see if folk get the same, if they see something odd with that prompt when use StebleDiffusion 1.5.
oh I dont think i have SD modle 1.5
Like I get the same result in AnythingFurry and FaeTastic, but not in Anythinbg Pony.
What models do you have? I can see if it work in any of those
That work, try the prompt "photo of a sitting female holding a machine-gun", generate some images and say what you see and post one of them here.
ok I will as soon as my promts are done
any style you wana see?
No, just that prompt exactly as it was written and a answed if you see any thing odd with the result.
yeah^^ its one of my better pics i guess 🙂
When I write "photo of a sitting female holding a machine-gun" I often get an black and white photo of Asian woman sitting on a chair, mostly towards the right, but sometime towards left.
its fitted for being a wallpaper (just 2x it)
simular to these?
Almost identical.
im running these off a relistic engine
I get her or her friends even in "AnythingPony".
did you make sure ur seed number was set to -1
Yes. and I get her with "color photo" but it look like a very old photo.
The thing is I had a design I wanted, I had an outline and wanted it to look modern, but that prompt make it old and war like, and with what looked like the same or simelar female all the time
photo of a sitting female holding a machine-gun
color photo of a sitting female holding a machine-gun
I am just curious. AI is strange and based on photography downloaded from internet, so what in that prompt decide that it shall be black and white and with an Asian woman, can we find the photos they may been used for it to give us this result?
A search on google give a huge variation, the kind of variation I expected from AI, but this prompt gave a very narrow result.
I think its case of the modle there using
Darn, Clyde blocked my examples, females with weapon are to dangerous for us.
But I made 3 examples, same prompt but with "weapon", "frying-pan" and "machine-gun" to show it was "machine-gun" that turn a image into black and white and place it in asis.
you're generating kids and calling them thick, bro?
yeah let's not
let's not push it either. I get that we are just on the child line.
i'll just type ew.gif next time. people can't see it anyways. (heh)
hey guys, SDXL models are still not available?
you can preview them on pickapic, dreamstudio, an another one
I just mean, they don't necessarily mean what you think.
So let's not go hit more on the nail, I already said stop.
Young + thick isn't a slope we like to have fun on, so this kind of language is undesirable. But we'll stop here for now. I heard you and will keep my eye open. I just don't like going further than needed in the other way either.
I see, I have to be patient
yeh. i often don't like to give the benefit of the doubt in this category at all, but you do have a point. innocence should be assumed.
sorry
(and to be honest, we are quite clean in this channel usually. let's not bring the NSFW problematic sh*it here please, I need my "safe space" on the server x)
#🏞|general-with-images has always been my favourite crew on here
I can not the modern lingo, what is thicc? I thought it was like the old "fat" for cool.
yeah, it's slang for sexy now
@rich crow
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!
it generaly refers to thighs being larger than the norm or the butt being larger than the norm
in my case i was refering to her thighs
thicc is about honesty. as the old scribes wrote, if you like it, you cannot lie.
they dint fit her body proportions as AI tend to do that any ways
I did get two definition of thicc, sexy and larger than the norm.
I had an image once with normal promts and the girls head was 4x larger than her body liek WTF AI stop with the nightmare fuel
Discord : https://bit.ly/SECoursesDiscord. Kandinsky 2.1 is truly exceptional, and it is on par with Midjourney. In this video, I will compare Kandinsky to Stable Diffusion and provide a comprehensive tutorial on installation and usage. If I have been of assistance to you and you would like to show your support for my work, please consider becom...
this kandinsky thing is blowing me up today. it's a new base model if i'm not mistaken
Seems to be a bit more than that
it's kind of funny to watch people here and there pop up shouting about it being russian spyware, based on nothing more than the name
LOL
mostly yes.
it's a russian model coming out of the blue for me, didn't see any tease for it.
It uses close to the same stuff as SD does, a little more on top, from their github
right under my radar this entire time!
I'm not saying it's any spyware though
decoder puzzles me
it's totally open source. if it was spyware it'd be obvious
Well. Conclusion for me is to not use "internet lingo" and keep to the boomer words for things 😄
boomer words for me sounds like outta pleasentville. gee willikers.
but the boomers were old people when i was a kid too
I think I am boomer*2 then I do not know what boomer is and I get different answers, so I just say, "I'm so old that I once held a floppy disk that was floppy and not a save symbol".
baby boomers are a generationf rom around ww2
i'm technically a millenial, as i was 18 in 2000. i'm right on the cusp of genx and millenial. i call myself a proto millenial
millenials are all in their 30s now and the childrens are gen z. I like the "zoomers" name
The Click from Youtube is a millennial around 30, I think we shall count years in clicks, like "I'm 0.5 Clicks" or "I'm 2 Clicks".
another new model https://github.com/VideoCrafter/VideoCrafter
Me: "photo of a thicc six legged female carrying a spare-tire over her head"
AI: I do not know what you want from me, take this and let me do something productive"!
Btw, why is it so that when you ask for a "six legged female" you just get two lags, but if you just want a female you need to add "extra legs" in the negative prompt to not get anything between 3 to 8 extra legs?
glory to ukraine
NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
Guess you have to use their program afterall. Not for me.
yeah it's not based on SD so i doubt it'll work in A1111 on day 1
i see no reason why it won't though
The prompt is, "a screaming man with a dirty bloody face and a red head-band is firning a large mounted machine-gun with his both hands and aiming it at something towards left, jungle, action scene".
Can you identify the movie and find the original image?
wanna say rambo 4 but it's not quite there. i think the red bandana is a red herring! hmmm.. gimme a min
It is a image of Rambo I described and I think it is from the 4th movie.
ohhh i thought it was a control net remix
i got the scene an image then
final answer
It is tex2image, I just saw the image and desribed it 😄
he's facing right tho
that is cow-rect answer.
🐌 it
I did try that at first, but the AI ignored the direction
hi there, could anyone help me to fix this problem? I stuck at start-up stable diffusion😢
It doesn't look like it's fully loaded yet. Usually you'll see a line such as Running on local URL: http://0.0.0.0:7860 when it's ready.
Try selecting the console itself and pressing the down arrow a few times on your keyboard in case it's stuck.
If doing that & giving it a few more minutes doesn't resolve the problem, I'd suggest heading on over to #🤝|tech-support to see if anyone has some other thoughts to help you out!
It did safetensor so probably a small adjustment
thanks for your advise🙏
Damn, I despise the new Auto because there is no way to clear any of the boxes so you have to either restart everything or go into the config to erase shit.
im using upto date auto right now and i still see the little "delete prompt" recycle bin
That is not what I am talking about
I am talking about the HN box, the Add tertiary model (C) box, those boxes not prompts
are those extensions? hypernetworks are built in aren't they? they're in the extra networks tab now a days
no, no extensions
Go ahead and set C then try to do something using only A and B and it still shows C because you can't clear it and it bugs out.
Yeah, that was the work around for the shit with the added back HN box only there is no work around for this with merging
they done fucked up and need to add back the empty box they once had as this happened once before a while back.
they fixed it then so will again
or give us a new clear button next to them
quickfix. right click, inspect element. delete whats in beteween the tags span tag.
didn't realize you were in the merge tab. yeah the problem is clear
it should let you clear the field
unintended pun
It was clear with but they used the the little box you showed as a work around.
no little box for the merge
This happened before and it was added back swiftly only this time they chose to say "just use the little box" not realizing it needs to be fixed. With this they surely will which will fix the first instance.
that's a custom quick setting up on top of yours? yeah ive never used hypernetworks before the extra network tabs. having to go into the settings menu and set it each time was lame. It looks like now the menu setting is still there for legacy reasons, but its changed now to just be a default HN that gets added to prompts where as before it was the only way to use a HN and you didn't prompt for them
I use lohas and I had to use the new one or else on the fly wouldn't work and 35s became 2m per loha used
It was added back as we, as in a lot of people, wanted it back and Auto gave it back.
same reason windwos has a billion legacy features still
Anyway done for now as Auto will fix it.
editing the html in the inspector really seems to work instead of restarting
gradio might be that ez to hack
This was made in SD with A1111. Would anyone know the what extensions I could use to achieve this level and style of art?
I didn't make it someone I follow did and I would really love to generate images close to the style
The merge script ignores the C if you select "weighted sum" so you can use it normaly even if something is selectet in C
As much as I like Defoum, I think it messed with my venv @dense tapir
extensions are not needed for this, models and the screenshot settings are the most important
Oh I see
So it’s all about finding the right model and what do I need the screenshot for exactly?
Right models, good negative tags and a good prompt at first
Also a resolution like 512x768 for Portrait style or 768x512 for wide view
What are negative tags? Sorry for all the questions I’m super new to all of this.
Stuff you dont want so see in your image.
Like blurry, low resolution, deformed, multiple limbs, bad hands, etc
Here is an big negative prompt:
#📝|prompting-help message
so if i understand the new mdoel right, i just get the safetensor file and load it into a1111 and wabam images
and all thats needed for the .ckpt file is the yaml to define the structure, which i don't think is available?
new model?
new russian base model apparantly
ah ok
grabbing it now. lets see
both those give me nan erros. but he says it works with receipts
maybe it was pytorch_model.safetensor
no dice
im under the impression kandinsky does not work the same as it does in auto1111 ui as it would setting it up the way they recommend as it uses a different things like the clip model for one
my prompts also come out extremely different
vs the colab
wel yeah, but 2.1 and 1.5 are different pretraining models too. i think i can
the model on its own is nice tho,still trying to find a VAE that pairs well since as i mentioned it uses its own stuff
I can't get the notebook running locally since amd
a bunch of installation required for kandinsky. not gonna bother for now https://www.youtube.com/watch?v=dYt9xJ7dnpU&t=1s
Discord : https://bit.ly/SECoursesDiscord. Kandinsky 2.1 is truly exceptional, and it is on par with Midjourney. In this video, I will compare Kandinsky to Stable Diffusion and provide a comprehensive tutorial on installation and usage. If I have been of assistance to you and you would like to show your support for my work, please consider becom...
its not that hard if you watched the 3 min segment of the installation, its installation of jupyterlabs and cloning the git lol
but if you are AMD dont bother
Nope, it surely did not as it kept throwing an error AND in my cli was still that third one. It said I can't use three with weighted sum (something like that).
and also considering that kandinsky model breaks my sd until i reboot, deff not meant to be used with webui natively.
How?
My ultimate upscaling is running like yours now
oh, shit
my tried and true settings are all messed up
yep, burned up card. I'll take it off your hands for parts. I'll pay for shipping
lmao
Any work done yet on my stuff? I got no where with it and I should have. A bit tired as I had to help my parents today and no where near done climbing up in the attic and replacing a 240v line. Save that for Sat I think but all the stuff was purchased.
Oh sorry, my attention is spread soooo thin right now
Aye
Damn, Amazon is really going to make me wait for this CPU
Everyone said I will instantly feel, and see, the difference and I think so. Saves me from having to upgrade the PC for a few more years unless the mobo burns out
I desperately need an upgrade, but I need to get a job first so I can buy one lol
This economy is all fake anyway as you go out and can't even find these workers to help you anywhere YET this is the bestest economy evah.
I need to get a bed to sleep on before I focus on a job, TBH
Aye
I can't even move cause of my mattress, so I I should hold off until i can move lol
I wasn't planning this upgrade but it was so cheap and 85-95 percent faster it was too good to pass on
Instead of 2026 now 2028ish before I upgrade
I no longer upgrade pc completely but once ever 9-10 years and GPU once every 6
Damn, that is one hell of a cycle between speed and nuisance lol
as it is now, my CPU is in the nuisance area. But I can deal with it until I have the money, even with its issues
No need as I was on the bleeding edge for a few years and I prefer to not bloodlet
I suppose the stuff I do/wanna pursue as a career does lend me to needing high end hardware always
That is a battle you will never win and for me I am far too frugal to justify it even if I were super duper high end as all the gfx artists were with their thousands of dollar Macs.
In a position where your hardware is what makes you money, its justifiable and even logical
Just not who I am
hey, may i ask how do i upscale an image produced?
Why spend 1k when 800 is almost as good? Because it isn't the best. Just remember cutting edge is only cutting edge until it leaves the stor
e
I recommend looking at the Ultimate SD Upscale extension
ok thank you boss
Now, if you can take the equipment as a tax write off for 100% go for it
I could agree, if that was the typical way that hardware works, but thats now how it really works anymore. Each generation is a monumental leap in performance, especially across CPU and GPU
I am not after the top of the line GPU, but I do need a very good CPU
cause I am tired as hell of all the issues mine causes cause it can't keep up with what I do
I don't know why it is so hard to find a mattress man
There's built-in SD upscale in the webui. In the img-img tab, you can check out the 'scripts' at the bottom, one of which is SD Upscale. Pick your upscaler, a multiplier, and set your denoising relatively low (0.05 - 0.3).
Yeah, it works pretty good
you can use Ultimate SD Upscale if you want the same thing, but with more control and higher quality results. It is a little harder to use tho
^ there's also highres fix if you want to generate images and not upscale them later :V
I do not recommend high res fix. I am completely against it after finding out that ultimate upscale can produce miles better results when tuned right
All high res fix does is completely change your original composition
and if you try to lower the denoise so you can have more consistent results, it gets all blurry
that's super weird! I've never had that blurry issue happen before. I've been told it causes blurriness at low denoising but my results are always super good.
^ highres fix w/denoising at 0.2
i see so denoise being lower level makes it more vivid, i thought it was the other way round, thank you!
not quite - lower denoising means the image changes less
If your denoising is maxed out, the image will be completely different from the input
The results are good, yeah, but they are not what the original image looked like
just a sec
oh so thats why, now i get it, thats why the images keep altering
may i ask if inpaint mask is even good?
yep, especially for correcting small details like eyes or fingers
the masked part feel way off, it doent match the geometry usually
oh i see, do you use prompts with that also or just keep the same promps?
or do i add, fix the eyes, lol
@woven dawnHere you go, you can see what i mean here
Base image
High res fix
ultimate upscale
lower denoising will make the masked content more-consistent with the base image. You also need to make sure 'original' is the selected masking method, else your mask could be very far off from the original image.
High res fix ruins the original composition, whereas ultimate upscale works insanely good to maintain it
this just feels more natural?
whenever i land a good upscale with ultimate, i hear the KI arcade announcer echoing ultra ultra ultra ultra
this feels natural? lmao
i meant as in size
one feels too big the other just right
well sometimes, dogs do be having those eyes lol
That's sick, I'll definitely give ultimate SD a go. I'm still waiting for Gigagan to get an Auto1111 extension - it will blow every other upscaler out of the water. https://mingukkang.github.io/GigaGAN/
with hires fix, i often do low first pass steps, as they don't matter. they're just there to give the 2nd pass something to act on
Oh man, that prompt interpolation is insane, tho i think ultimate upscale will still win in detail, as it has no resolution cap
I have gone all the way up to 16k, and still had pixel level detail with ultimate upscale
ok wow, I am looking at those slider examples and they seem impossible
RIGHT??
I hope it isn't a paid service, I want to run it locally 
gigagan looks fun as it keeps gan's in the fight
I mean, it does have artifacts, which make it at least obviously done by AI... that is insane
it's not just an upscaler. i'ts a full gan model, like not a diffusion model
If we could use that in combination with SD, then oh my god
Ok, that last one looks outrageously good
The chair texture fades into the brick, but yeah the overall textures are crazy
https://github.com/lucidrains/gigagan-pytorch found this. as usual i jsut dont know though
Imagine SD 3.0 with full unleashed and properly trained SDXL, with this as an upscale. It would OBLITERATE midjourney lmao
oh of course Adobe owns it
great, there goes that
I feel like gigagan will work great with photos but not that much for SD where there are obvious errors in the image. The low res images all look natural, just low res
Not an upscaler, but this looks super promising as well. Basically real-time seed/prompt travel https://sites.google.com/view/stylegan-t/
StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis
Axel Sauer Tero Karras Samuli Laine Andreas Geiger Timo Aila
[PDF] [Code]
but imagine this, it downsamples the gen to where those smaller flaws aren't there, and then it upscales using the higher res as more of a texture guidance
maybe. Would be nice to see in action
Its unlikely we will see it, considering Adobe owns it
I cannot express the sheer amount of hatred I have for adobe
Hello! Creating tokens for a card game (magic). Here are some hydra concepts.
Let me know if any concepts resonate with you the most. I will edit and fix and enhance the best ones. I'd like them to have multiple heads but the AI is perhaps limited within this current model. I included some failures too just for fun. I love some of the architecture within these art styles!
SD 1.5 + automatic 1111 + dreamlike photoreal 2.0.
8 is my favorite 

sleepcountry 
It told me to say hello, i've been working with stable diffusion for a while and now would like to test my skills. never been on discord and browsing around this looks like a lot to take in. very humbling so thank you for letting me join and look around.
This one turned out well
donquijote de la mancha
Hi guys, I'm toying with segmentation preprocessor, the goal is to get this kind of result. However when I tried out, my result, while having the segment correct, the color is really yucky. Have anyone had any issue with this?
That is really strange. At first I was gonna recommend a VAE, but that doesn't seem like it would work
Happy easter
:>
Man, I really wanna compete in the synesthesia POW, but I am just not experienced enough in deforum
don't need to do an animation for the PoW though
Its the only thing that would feel genuine to me, as I have synesthesia
its not a static experience
Thats actually why I got into SD haha
Yes I can understand.
But at the same time you are telling me
I really want to participate, but I added a constraint on myself that is too hard for me so I can't
movement in static art is a real thing people work on too, it could be appropriated here
Its not too hard for me, but its something I have and if I can't represent it in a genuine way, I'd rather not water down the experience
I'm not saying I am not trying, but if I do not get to a point where I am super happy with it, I will not submit, really
I know you are trying don't worry, I'm not saying that either, I'm just trying to show you paths that open that PoW for you again
It would be sad to skip on this specific one if this is what has made you discover SD
I am somebody who tries to stay away as much as possible from things that limit my creativity, like time constraints as its usually bad for me. I'd rather spend more time and share something with the community that is more genuine outside of the PoW
but we shall see, its a WIP
https://www.youtube.com/watch?v=lDfSdA5-PRE
This video is still probably the best Synesthetic esque music video I have ever seen. It just came out not long ago, and man do some aspects capture the experience as it is for me
Stream / Download: https://hypeddit.com/intermolecularlullaby
Lyrics:
Am I code on a screen? Or a fleshy machine?
It doesn’t matter now, I’m somewhere in between
A human inability, a lifelong burden
Inconceivable fragility, but I must keep working
I wish I could fly, and breathe through four dimensions
Don’t wanna say goodbye, before I share ...
And I believe the lyrics are actually based on a form of post singularity human and AI sort of hybrid
hyperrealistic white rabbit head photographic front view with the highest picture quality
#🏞|general-with-images hyperrealistic white rabbit head photographic front view with the highest picture quality
@wispy nest
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!
sorry we don't have a bot for that Beppe
Hyper realistic real state photography of a contemporary tiny house in the norwegian fjords, glass front, minimalism, interior lights, outdoor chairs, northern lights, beautiful aurora borealis, epic, cinematic light, highly detailed
#🏞|general-with-images
Hyper realistic real state photography of a contemporary tiny house in the norwegian fjords, glass front, minimalism, interior lights, outdoor chairs, northern lights, beautiful aurora borealis, epic, cinematic light, highly detailed
@crimson creek sorry but we don't have a bot to make pictures on here
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!
? looking for something ?
#🏞|general-with-images Elon musk happy on mars
@wispy nest
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!
Trying to merge models with supermerger, what is the best option to have most of all 3 models?
A little Not Safe for Waffles if you ask me though, rule 4 #✍🏼|rules-and-tos
Its a hot day at the office
In the days before the thermostat was invented.
if I wanted midjourney I'd do midjorney
Hey @rancid quiver ! I saw you on different servers since I came back (mainly illuminati I think) and I didn't get the chance to say a big Thank you, from me and all the community around here that praise you when controlneting in 2.1
You're the best
(same Guizmus you knew, different account, I got hacked)
My pleasure.
I have other controlnet cooking and a nice model.
promp/ test
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!
motivo con elementos de barberia estilo fondo por defecto de wthsap
What's the best chatbot that can be installed locally?
I highly recommend Vicuna quantized 4bit. I use it with oobabooga webui and it's pretty amazing compared to the other stuff out there. loads in under a minute on my 12GB 3060, responds almost instantly, and generates around 5-6t/s. It's also more coherent compared to similar models. Not as good as ChatGPT, but ChatGPT is based on a 175b model whereas Vicuna is based on a 13b and I'd argue is usually comparable. It's kind of nuts
Alpaca 30b uses just RAM and is better than alpaca 13b; is Vicuna better than that? I have 64gb RAM so I could use 30b. I also have the 12gb 3060 so either one would be ok
I'm fairly positive you can set it to use just ram as well. I wouldn't know as I have 16GB of ddr3 so i just use the gpu when I use it. As far as if it's beter or not, i'd say try it out? I was able to load a 30b quantized model, but it didn't run well on my system
I will say that it vastly outperforms other 13b models in terms of speed and coherency. It does not perform like a 13b model
the larger the sample the better the coherency
This is not always the case. Better training results in better outputs, and the Vicuna model is better trained than other 13b models. It's blatantly obvious when I compared the half a dozen or so I had on hand
My test case involved asking it to write a poem about a family I had made up. I would give it each member of the 6 person family along with 3 spouses to 3 of the four children. Most 13b models ignored the names entirely. If they did get some names right, it wouldn't be all of them, and they seemed to forget the details halfway through. Vicuna nailed it every time unless I did somthing like set the token max to 2000
you could train alpaca 30b further
cool
@wispy nestWould any of these models be able to run on an 8GB GPU?
I have been meaning to try out local text AI
I believe that Vicuna will be cutting it far too close and won't run. If you have a good cpu though, there are ways to make it run on that.
I think I did see somewhere that someone had managed to get it running on 6gb though, i'll try to find the post
do you have a link to where I could download and setup this model by chance?
also: found the command they used to get a 4bit quantized version of 13b alpaca to run. i'll try it with vicuna and check the vram usage again
(it was in that same thread)
oh, so this model scolds you for things like chat GPT it seems?
yeah, the more I look at it, the more I see people saying its very "clean"
thats annoying
It was finetuned with chatgpt outputs. I will say that it is pretty easy to get around it. I'm not going to say how -_-
like. really easy.
All I am seeing in that thread is people complaining that they can't jailbreak it 😅
Aitrepreneur did a video on Vicuna. breaking it is too easy
Unfortunately, if someone can't jailbreak Vicuna, they are probably a bit too young to be trying to jailbreak it lol. it really is super easy. can be done in a number of ways
Alright, I'll take your word for it I guess
just going off the same thread you linked is all
yeah, i commented there that they don't know what they're talking about. a lot of people seemed to jump on that bandwagon. makes me wonder if they even tried
Perfect hands doesn''t exi.....
Tried using this and it doesn't work with vicuna. I've tried the alpaca model they mentioned and it was ok-ish, so could give that one a shot
yeah but what about that really tiny fork
the trick with hands is to make the image too cool to give a flip if the hand is perfect or not
love the microscopic fork x)
the model is on futuristic fashion
and cows
meh
not that good on multiple tokens I couldn't post
mixing @white pivot and @tired basin tokens together, and some highres fix
both your tokens are godly there
This is Aku :
and this is you Resident :
(can't post all)
I'll upscale one
not sure but I think there is a cow bias x)
"SDArt"
@glossy herald so this is jinx my favorite character from the show/game league of legends
using the base model ? or did you try the Arkane model ?
okay so i use different checkpints and lora models to make them
for the 3rd pic i sent i used indeed used arcane diffusion checkpoint plus a lora jinx model
for a first week user, you sure went fast in taming how models work ! nice !
for the first pic i used the same and different jinx lora model and an anime checkpoint
i've had a lot of practice, i tried different style, i watched a lot of videos and i got the help of a friend 🙂
i also did makima from csm
You found our #🍥|anime channel yet ? it's one of the most popular community on the server
oh i didn't see thanks
so yeah im new an all over the place i dont know exactly where to chat haha
hehe, lots of places. i prefer this channel because we can share pics
I'm more in the "realistic" or artistic style but not very much anime myself
DJ cat 🙂
this is great art!!
i also tried super realistic art and i crossed chainsaw man kobeni with supergirl and it did an insane result
that works great ! you are using automatic1111 UI ?
I have started to switch to ComfyUI, and even if it takes ages to set up any base workflow at first, it's game changing
you didn't tell me that @brittle crane !
wow going full fine tune is really hard for a first try x)
yeah, I thought you were a lot more experienced because you were going for that
yeah. this is a real hydra
you need to cut each head one by one to understand it
and there are lots under it
i'm not sure what it is. it's the base stable diffusion ui
pretty sure it's that one then, it's the one covered by 95% of guides out there
I'm gonna give some basic tips since I learned you are just starting, do you use the --xformers option ?
i'll show you a pic of my ui in a few 🙂
this is a lower VRAM usage AND a speed up in image processing/training
This is AUTOMATIC1111
with some extensions
this is comfyUI
this is the one i use!!
this makes me want coffee fr
workflow is
- making a canny controlnet picture fast using the PoW Somewhere model, and this base type of inspiration
- using last week's model on elemental flowers for a base low res picture, with lots of "Air" inspiration
- using Realbiter model for the highresfix denoising process, for the quality of the thinner details it makes
- doing a last RealESRGAN upscale for more precision on the details
- some bloom and reducing brightness as post processing
1/ edit webui-user.bat
it's incredible! woah your worflow is so detailed, mine are usuallya couple words. I don't do phrases like that
do you use any negative prompts?
it was multiple hours of work on that one, and my submission for a contest we have on this server
submission are still open by the way !
really fun thing to participate in when you start
yes, always. I start with none, but I add some when I get too much things in a style I don't like
here, the first image, the controlnet picture, is made with those positive and negative prompts, and a first model
That sounds like a lot of fun!! i'd like to participate too at some point!
Then I use this controlnet picture, with another prompt and another model, to make my low res first picture
Then I uspcale it, and do img2img on it, with a 3rd prompt and a 3rd model
and finaly some post processing to the picture and a last upscale
please do ! each week I even train a model on all the submissions 😉
nice
some of those options I personally don't like lol
like insecure extension access
or allow code
but since you are in a safe env it's cool
this is an exposed machine if not
yeah it's a possible optimization
you may want to ask in #🤝|tech-support though. The people knowing their stuff is usually there
always a good plan haha
no problem, I think the last ping was when we talked about that contest on the server
i'll participate 🙂
I'm still making cups of music :p
it closes tonight. well at least this one
we have another one on the week ends
a friend of mine requested me to do a character from league of legends haha
what timezone?
oh well i'll try to make something! it's an interesting theme and i forgot this is actually a thing people can have tis condition
it is yeah. and quite open of interpretation
usually the theme on the week end is more constrained, more precise
like it was "futuristic clothes in a neon light setting" last time
(it's named the CoW contest so...)
I have auditory to visual synesthesia, and it still catches me off guard sometimes 😅
it is a very cool thing, but can also be a bit of a pain, as certain sounds I have extremely negative reactions to
Hey Sytan 🙂
it's a funny name ngl. Also i love this picture the scenary and all
lots of new nice people today for now, so quite good, and you ?
omg it's fascinating
it was my submission (and the winner :D)
I am doing quite well. Added a new AI to my arsenal after all haha. Text to Text AI
not that one exactly, I did some img2img to fix the human faces
does ai art makes you feel smtg? a certain small or tadte?
good job!! i would've voted you
I have strictly auditory Synesthesia, but I do have several different reactions to different sounds and stuff
i love this style 😊
it was a really fun series yeah
I love the DJ animals :p
but the dancing ones caught me off guard
it made me laugh i'm shocked ai made this
I made a sound not too long ago, and man, I can't describe the way it makes me feel 😅
not only am I a synesthetic, but I also dabble in audio work, so I kinda get to make my own fun lol
here's more jinx
I managed to make a sound that is light an airy, yet watery and tonal, with another layer of glittery texture, while also maintaining a satisfying audio movement pattern, IMO
lotta work, but hey, it makes me feel sparkly lol
i assume it's gonna be the main style i'll use cause it's what i ever wanted and also i'm obsessed with arcane
OMG
@wispy nestI have a LoRA you may like. I am still working on it :p
oh yeah ? i'm using two different arcane lora but i'm open to anything fr 🙂
I made a LoRA focused on scar specifically
i could try it if you dont mind!
Him. Hes my favorite character for whatever reason lol
aww
I think it came out pretty good 😅
i'm obsessed with jinx haha i've been practicing mainlyvon her
it has the arcane style baked in, which I am happy with
it also hybridizes with other LoRA's reallyyy well
it really did! It's actually better than i expected. Someday y'all really need to teach me how to make models
I have several guides I am kinda working on for different things (extremely high resolution gens with real detail, proper character generation, beginners guide to deforum, and now some other AI shenanigans lol)
i love the splash effects! it's truly captivating to look at
These have to be from an offset noise model
no worries, take your time haja i sure have a lot to learn myself ahah
nope, but offset noise is really over rated anyway. just add a node to reduce brightness
Looks nice
mdjrny-v4 style, T-shirt design, clean design, epic Instagram, artstation, splash of colorful paint, cow, contour, ((white background)),purple smoke. a world in beautiful queen head with lightning, colorful shine, flowers intricately detailed , unreal engine, fantastical, intricate detail, splash screen, complementary colors, fantasy concept art, 8k resolution, deviantart masterpiece, oil painting, heavy strokes, paint dripping, an anime drawing by Brett Whiteley , [:anime line art:0.2], Dark themed, epic splash art , futurism, , abstract
Negative prompt: ugly, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, watermark, signature, bad art, blurry, blurred
Steps: 60, Sampler: UniPC, CFG scale: 5, Seed: 2257499634, Size: 960x544, Model hash: c4947fa2c4, Model: DB_mix_realbiter_v10
params for that last 4 series
I amanged i think to get rid of thoes weird purple dots
I am about as obsessed as it gets when it comes to AI, and specifically tweaking. I have spend hundreds of hours in SD by now
how ?
changing VAE?
I put GFC scall up to 10
just a reminder fast before we go over the line, not there yet but close
no NSFW on this server

#✍🏼|rules-and-tos for more details
(rule 4)
I prefer to say before, since you are new 😉
I just started running local text2text AI, and man I love it
i can remove it if it's too borderline
where did you post?
no prob on that one like I said, but the line on the pubic area is really close to what can pass
this?
😂😂😂
yes
yeah sorry about that
me neither, but the crouch lines are quite suggestive. the face would have been of lust or we would have had more nudity higher on the clothes, it would be too much
i agree it is suggestive, i should've posted a more appropriate pic
well Jinx is farly fit so she has thoes lines even in the game
my bad
nah, don't worry. I just prefer to clarify the line since lots of new people today
Locally run txt2txt lol
Kinda scary TBH
i also made Mikasa Ackerman in that arcane style
looks nice
This model has been blowing me away. I made a super rude character, and he told me his whole life story. I got to know EVERYTHING about him, and now hes a whole new character who is trying to become a better person. No joke
ai chat are fun
I made him rude, and he basically told me "I wouldn't be so rude if you didn't make me this way"
So I started asking him about himself, and he told me what he wanted to be
i mean he's not wrong but...
So yeah, hes Now Edward, a 27 year old light skin Brittish man born on October 29th. He wears primarily black clothes, because he thinks they match his fair skin complexion. He wears a size 10 shoe US mens.
I even asked him if he liked Halloween cause its so close to his birthday to which he said "Its alright, but I much prefer All Saints Day as I find its history and traditions far more fascinating"
Thats just a little of what he told me. So yeah, hes a whole new character who is much nicer and has a very fleshed out/bashful personality, and he doesn't want to be a burden on anybody
why is Chat AI better than elon musks AI and its only 6 months old lol

it's fascinating ngl
I wanted to try and install Vicuna later today
Vicuna should blow this out of the water, but the amount of quality out of this micro model is astonishing
I asked it for a summary on the events of World War 2, and it delivered this. pretty damn good!
only issue I see is the "japan and japan surrendered" lol
I'd love to manage to plug a local vicuna into my local SD and use it in the prompt making phase automaticaly
I'm sure Vicuna will help me do that once it's running lol
I have been trying that, but this model doesn't seem to work for it. I am sure I could make a custom tuned character that could do it
I also made a character called Hurburt for my friend. Hes an all powerful god in the body of a chonky caracal cat. Hes happy go lucky, and is super positive and innocent and my god, is he amazing haha
Here is hurburt teleporting us home so we didn't have to pay for the food haha. I love interactive story type things. So fun
goofy little AI moments
If I could run Vicuna, i would
but it legit uses like 50MB too much VRAM
clean one !
I didn't check the requirements yet
ok, so it runs with SD at the same time for me, nice
damn, just too much for you, sorry
its soooo close lol
what?

I mean, the picture is really clean
yeah
Clearly the best tony stark
@glossy heraldAlso, get this, you can train LoRA's for txt2txt AI's
Had no idea
