🏞｜general-with-images

[split valve]

True.

[rigid heron]

this shit come out good

[split valve]

The analog filter creates a disconcerting effect, its like uncanny valley.

[sterile temple]

the imperfection helps make it more convincing

[split valve]

Disconcerting in a good way i like the weird effect the analog filter has.

[sterile temple]

photographic style vs analog film style

[split valve]

good filmgrain

[rigid heron]

waltuh

[sterile temple]

pretty decent guitar

[split valve]

yea its pretty good! only about 1 generation behind midjourney i would say.

For comparison here is midjouney

[sterile temple]

SDXL can draw clean lines when it wants to

[split valve]

wow looks like an autocad archviz drawing

[sterile temple]

Victorian house, garden, brick path, lush vegetation, detailed sketch, architectural drawing, reminiscent of Frank Lloyd Wright's style

I actually felt something when I saw them 😄 like genuinely impressed

[dense tapir]

2.1 base model, lol

[woven iron]

Deep-fried is a new negative prompt for me. What is it weighting out of the images?

[woven iron]

This is amazing!

[sterile temple]

the overcooked look you sometimes get. it's a style of meme, so a good prompt to add to the negative since there should be lots of examples in the training data (in theory)

[sterile temple]

not bad, but SDXL is nailing that style

still get some meh stuff from SDXL but occasionally you find something it does really really well

this is from a 1.5 model. not exactly a drawing, more of a render

[sterile temple]

pretty house

[dense tapir]

Yeah, but notice the chimney?

[sterile temple]

good for leaning on

[dense tapir]

LOL

Always something

[woven iron]

Thank you! I always enjoy learning from you all! So much talent in here 💜🙏

[sterile temple]

I got another 20 similar prompts from chatgpt using Frank Lloyd Wright's name. So I will try them all out

that's the house we all know from him

[dense tapir]

2.1 refuses

Best I can get

[sterile temple]

kind of reminds me of a landscape artist style

[dense tapir]

Yep, 2.1 says no

This is cool

[sterile temple]

like a silent film

[dense tapir]

Almost

[sterile temple]

that's better! 🙂

[dense tapir]

Well, this was a surprise

all of these are the same seed

not sure why it makes the paper funky

that was 25 steps vs 20 and this is 50 steps

[sterile temple]

interesting

[dense tapir]

even has their signature and the dude is a lefty.

[sterile temple]

well... annoyingly it's late just as I'm getting into this. have to get some sleep now

[sterile temple]

a bit blurry but good

ok bed time 🥲, gnite

[oak osprey]

sdxl looks derpy now

[rigid heron]

jesus what is with your hate boner for sdxl

theyre testing things out with the bot its obviously not representative of the final thing

[oak osprey]

which part is confusing to you lol

[rigid heron]

why do you hate it so much i dont get whats wrong with it

[oak osprey]

then it's the perfect model and i'm just a fool, and you can ignore me

[rigid heron]

of course the currently being tested base model isnt as perfect as the current super-finetuned previous model but you can clearly see that when released and finetuned in the same way it will be much better

[oak osprey]

you don't see any issues in that picture?

[rigid heron]

i do

[oak osprey]

that picture was shared by the project lead to hype people up

[rigid heron]

so?

ive seen plenty better made images with this

a less than ideal advertisement choice doesnt mean much

[oak osprey]

🙂 i guess that means it has no issues, see earlier comment about me being a fool

[rigid heron]

thats not what im saying at all can you read

[oak osprey]

you are so pissed at me for having an opinion, i get it

[rigid heron]

what?

[oak osprey]

what can i do for you?

[rigid heron]

i asked a question and explained my opinion and you started being a snarky prick about it

[rigid heron]

you started the aggression

[oak osprey]

aggression? no way, i'm confused by how someone couldn't see the issues lmao

but then you said you do see the issues

so... i'm still confused

[rigid heron]

its just that ive seen you talk about sdxl only negatively

as if youre expecting it to be perfect

[oak osprey]

i don't think you've been here long

[rigid heron]

ive been here long enough to see you complain about it several times

[oak osprey]

cool

[rigid heron]

of course its still going to have some of the problems of a base sd model

[oak osprey]

oh, it has pretty much all of them

[rigid heron]

yes

because its a base sd model

but when you compare to previous base sd models its much better

[oak osprey]

yep

let's just leave it at "i don't wish to get into it" and "they've asked for this feedback"

some of us remember 2.x 🙂

the model architecture of SDXL is disappointingly banal

[oak osprey]

lmfao whyyy

[rigid heron]

why what

[oak osprey]

that morgan freeman is cursed

[rigid heron]

a little but he looks more like him than other peoples names ive tried

[oak osprey]

its the eyes going opposite ways for me

[rigid heron]

well yeah theres that

[oak osprey]

and the mouth

but otherwise good

[rigid heron]

though i imagine when inpaint becomes available thatll be easy to deal with

[oak osprey]

teeth are hard for diffusion

[rigid heron]

also ryan reynolds works pretty well

[oak osprey]

chances are GA isnt using sdxl anyway

thats probably 2.x

[rigid heron]

ill try morgan freeman with sdxl then

[oak osprey]

they dislike 1.5

[rigid heron]

i havent heard particularly good things about 2.1

[oak osprey]

and btw i love sdxl, i hate how easily solved issues are ignored because other less important stuff now passes tests

[rigid heron]

sdxl seems to do faces pretty easily, there arent usually glaring issues with things like eyes and mouths

[rigid heron]

oh i agree with this

i think i was misunderstanding how much you were actually complaining about it

[oak osprey]

sdxl has the same thing about faces as 2.x and 1.5 actually

[rigid heron]

oh

[oak osprey]

portraits are amaaaazing

[rigid heron]

i havent had it show up nearly as frequently

[oak osprey]

but faces at a distance = gore

[rigid heron]

it was its own decision to give him the hat

it knows the man needs it

[oak osprey]

lmao

[rigid heron]

of course

[oak osprey]

interesting

[rigid heron]

a human artist drawing a tiny face in the background would struggle too

not as obviously as ai but still

[oak osprey]

its because they used the old vae architecture from sd 2 and 1

[rigid heron]

man i wish i knew what the fuck a vae did

[oak osprey]

its like compression

[oak osprey]

like jpeg

[rigid heron]

i thought vaes were like post-processing

[oak osprey]

they are

[rigid heron]

honestly i dont know much about what goes into making ai im really just a consumer, my only knowledge of vaes is that i put one on and my images stopped looking washed

[oak osprey]

they handle luminance and chroma and other graphics crap and fine details

the latent space goes through the vae to be decoded into rgb space

[kind quartz]

sorry just about backgrounds issues

[oak osprey]

vae is variational autoencoder and.it is trained by giving it an image and seeing how closely it can recreate it after encoding and decoding

the vae is lossy meaning it doesn't faithfully reproduce inputs

there are visible artifact patterns in the sdxl bot outputs because of the low precision of the vae. it is arguably the heaviest component of the model with respect to vram use, so keeping it small and shitty works well for most use cases while still running on consumer kit

all of the people wanting this to run on 8gb vram or less is why it is not better

[rigid heron]

im not even gonna try it on what i have now im just upgrading

ive been running sd on 6gb for ages

[oak osprey]

vae tiling can help but with dynamic thresholding it can.cause artifacts like brightness going all over the place for each tile.

so you see squares of brighter and darker sections of the image

[rigid heron]

so thats why sometimes two things in an image just seem like theyve been photoshopped in

they have weird mismatched lighting

[oak osprey]

thats possibly a text encoder issue

the text encoder captures relational semantics and the unet is trained on that embedding which the text encoder generates from the prompt. some overfitted concepts in the text encoder can make it difficult to have them appear in other contexts

[oak osprey]

and when the embeddings dont mesh well, you get that pasted on look. tuning the encoder before tuning the unet is essential

if the text encoder has a crappy representation of a prompt it is an uphill battle for the unet to learn correctly

[oak osprey]

openclip for sd2.x has human deformities baked into it. the hands it produces and limbs in general are poor quality. so the unet struggles to learn a coherent image of them

tuning openclip on its own to produce hands and.faces, and then popping that into stable diffusion 2.0 without further training, ends up fixing most of its issues

[grim hound]

Guys, I need some help. What am I doing wrong? I've followed so many guides, but I can't get this working with the tile ControlNet model. It works fine with a QR code model for ControlNet downloaded from civitai, but it doesn't work at all with tile. I've tried reinstalling all my ControlNet models, but it didn't fix it. I am literally using the same settings as the guy sharing his workflow. And this is the fourth different guid I try to reproduce:

[oak osprey]

use one of those free websites that lets you make the QR codes, problem solved! i don't think anyone scans them anyway

[rigid heron]

i dont think the images even work as qr codes

[oak osprey]

btw @rigid heron 2.0 can be fine-tuned pretty easily and i don't know why that guy has a baby or who is standing behind him - some kind of AI cryptid, possibly - but it's not as bad as people made it out to be, even for faces-at-a-distance

pop that into hires fix and watch the magic unfold

[fossil spire]

Can someone explain why when I put a prompt from Lexica into my local SD, I get this creepy ass little demon with like 4 arms that was obviously drawn by a 12 year old. However if I put the same prompt into Lexica website, I get a really well done image. Am I missing some setting?

[oak osprey]

different model or embeddings

are you running the default model sd-1-5?

if so, thats the issue =p

[fossil spire]

Yea 1.4 I don't recall seeing a 1.5

Another example lol... Any recommended model/embeddings?

[oak osprey]

try Artius v2.1, Illuminati Diffusion, epiCRealism

depending on what you're looking for it's those three ranked from adventure to photorealism

oh right, Vodka v4 is really good and on pretty even footing with Illuminati, and could be equal to Artius but i'm so in love with that mix it's hard for me to put anything in its league

epicrealism doesn't really need negative prompts (but it benefits from them)

[weak sage]

i like to do additional post processing in gimp like unsharp mask, vignette, sometimes oil painting and canvas noise,

with unsharp mask i do it in several passes, i do like .5 pixels and then 1.5, 4.5, 15, 45, 150

[oak osprey]

nice

i enjoy the gimp, too

[weak sage]

and with every effect i reduce the opacity and merge down onto the original when im sure the effect isnt too much

i dont actually paint in gimp, it doesn't feel good to me

[oak osprey]

throwback to early april when i was working on a project whose code was entirely written by GPT4 to see if it can do it. it used DALL-E2 for the images, Elevenlabs for the audio, GPT3.5 for the script generation, and then the whole project ties it all together via a MELT XML file it generates, which is rendered into the mp4 file. all in PHP. god i hate that language. i had to keep correcting GPT4 "we're not coding in python, we're doing this in PHP" because it's not its default

they've definitely updated DALLE-2 since then

[weak sage]

yeah neural net software has kind of reached a critical level of acceptability with image editing

not there yet with music

[oak osprey]

OpenAI jukebox was really cool tho

https://www.youtube.com/watch?v=Ca-Q6ElMetY

i unironically enjoy all of these attempts

i just wish the codec were something better than 22kHz

i think it uses EnCodec but i've not looked

sometimes it's even lower fidelity than that, like an 8kHz AM radio station

[weak sage]

anything that doesn't even run FFT on the signal is going to be garbage

[oak osprey]

fast fourier transform?

[weak sage]

yea

[oak osprey]

not really "in the know" on the lingo, sorry fam

[weak sage]

you knew it tho

[oak osprey]

have you seen RVC? retrieval based voice ?

[weak sage]

i havent

[oak osprey]

my colleagues are the math wizards, they ramble on, and on, and on, and i guess something got stuck in there, but i wouldn't say i know what it means lmaoooo

[weak sage]

it means they take the wiggles of the soundwave and turn it into what your ear sends to your brain

all the frequencies or pitches are separated

[oak osprey]

well, my basic (flawed, likely) understanding of FTs is that they can find a function to fit a waveform?

[weak sage]

yeah the waveform is a big mess, so your ear, or the FFT function, looks for any cyclical patterns it can find, at every possible frequency, or pitch

[oak osprey]

that's like me saying "the parameters in the unet are updated", i know that concept but not the exact math of how it works. someone, at some point, tried to teach me. and it's like talking to a brick wall, let me tell you

[weak sage]

i dont know anything about the math either

[oak osprey]

well i know "looks for" likely means "a derivative/integral"

[weak sage]

if you look at a piano keyboard, every key is a different discrete pitch, it's like an integer version of it

[oak osprey]

nope that doesn't work for me lmao

i had trouble understanding the whats & whys of calculus until my professor explained the process of finding the area under a curve using smaller and smaller boxes until we have an infinite number of them and i was like hoooooooo okay i like this

[weak sage]

https://en.wikipedia.org/wiki/Spectrogram

[oak osprey]

well i think that's how AudioLM is working, right?

or stable-audio-diffusion

can't remember which it's called but they trained it on text to image pairs that described audio waveforms

[weak sage]

https://www.youtube.com/watch?v=37zCgCdV468

[oak osprey]

it's a really cool technique and i'm curious what the impact would be of training a model from scratch on audio waveforms vs fine-tuning stable-diffusion checkpoints

i am now curious if i can improve the audio generation capabilities by applying zero terminal SNR noise schedule to their models

the inference process of SD leaves behind some trash that we can ensure is gone about 90% of the time, which would mean the audio we hear is more precise

SDXL doesn't quite have this but instead would result in overly smooth outputs due to the "ideal denoiser" from k-diffusion

but the increased parameter count of SDXL...

VAE replacement whennn

[kind quartz]

funny sometimes look back, probably 1.4?

[oak osprey]

that could honestly just be them testing SDXL at smaller resolutions

it's interestingly working VERY well if that's the case, kudos. not easy to do

you have to evenly display a number of image sizes to the model during training if you want to do that. you can't just give it different aspect ratios at one base resolution. you have to do that, with multiple base resolutions. i assume they used 256, 512, and 1024.

the 768 model for 2.0/2.1 end up heavily artifacted at 512 because it was trained exclusively on 768 for about 800k or more steps and the transformers inside the unet begin to expect to express details they commonly see at 768

i've trained the 768 model on 1024 images for 100k steps at a high batch size and it similarly loses the quality at 768x768

i'm tempted to do that style of multiple resolution bucket training, but i'm not ready to put that work into the training code

[cyan snow]

Holy shit, remember when it couldn't reach the same quality as an extremely fine-tuned 1.5 model? This is insane

Man, this is an insanely good model

I have no idea how they managed to implement 2 text encoders in a single safetensor, but this is awesome. I can't wait until it's released

[oak osprey]

they increased the dimensionality of the text embedding layer

i haven't looked to see whether it's concatenated or if it's stacked

[cyan snow]

I don't know what it is, but I gotta tell you, I was never able to make this level of detail with ANY AI model

[oak osprey]

probably easier to concatenate but that's speaking as a noob

[cyan snow]

It's just so seemles

[oak osprey]

the text embedding size is more about coherence and overall "knowledge" but you have to train the unet on that stuff or else it just shows up as confusion

one thing i'm curious about is how much the parameter size of the model increases the requirement for training data or if it's learning rates that get adjusted, or what

like their dataset emad was mentioning has 12 billion image in it and they've captioned it using, i believe BLIP2/a finetune of it, or something similar that leverages the dual embeddings

[cyan snow]

I have no idea, but I must say, SDXL is the greatest AI diffusion model I saw yet, and it's not even finished OR finetuned

[oak osprey]

i hope they replace VAE with something else before 1.0 but there's a fat chance of a major change like that

they might fine-tune multiple VAEs, which would be awesome

[cyan snow]

They released a VAE for research access

On HF

[oak osprey]

there's ways to fix its remaining issues but not much time

that said they plan on regular updates

so not just YOLO'ing a release like 1.5 and 2.1

can you elaborate for me what kind of details you're seeing in SDXL that you can't make in legacy models?

[oak osprey]

we're using it already on the bot and clipdrop though, that VAE isn't useful on 1.x or 2.x models

the size/architecture or advanced capabilities of the VAE itself hasn't changed, only its training schedule - it could be better simultaneously at photoreal and anime now, for instance. but there's issues with how it expresses patterns of noise still

[cyan snow]

It's not compatible, different architecture

[oak osprey]

it's literally the same architecture, comfy has said it, mcmonkey has said it, TwoDukes has said it

[oak osprey]

it's such a heavy (VRAM-wise) component that they're really not interested in increasing the size of the VAE or the tensor space it operates in

[cyan snow]

Really? I tried to use the SDXL VAE on A1111, it fucked up my UI

[oak osprey]

yeah it doesn't work and i don't know enough to explain why but i know it's the same VAE architecture

[cyan snow]

It can't be the same architecture, if it was, it would work=\

[oak osprey]

i'm telling you lol it is the same >:| unless SAI devs lie to me

[cyan snow]

But regardless, Automatic 1111 will be forced to update the webUI to support all SDXL stuff like ordinary models, so I don't give a fuck rn. I can wait, it will be so worth it for this model, this model is by far the best diffusion model

[oak osprey]

apparently if you retrain a vae from scratch on a different data distribution, it will be incompatible with the learned data in the older model

[cyan snow]

I have access to it, it doesn't work. It literally can't be the same

[oak osprey]

could have a different loss function it used during training, but i'm not sure how that would interact with the unet/text encoders.

[cyan snow]

Also, how the heck did they manage to implement 2 text encoders in the model?

[oak osprey]

the VAE is just a encoder/decoder @cyan snow it converts into a lower dimensional latent space and then back to a higher dimensional image representation

[cyan snow]

VAE is Variable Auto Encoder

[oak osprey]

i don't know where to begin explaining this one because i don't know how much of it you understand already

if you know how a single one is used to guide diffusion it's not that big of a leap to guess how two of them work, especially if they're concatenated. but they could be stacked and averaged etc too

it probably helps that they were both trained on 77 token captions and they can either be padded from left or right or both

[cyan snow]

Maybe, whatever vodoo magic SDXL is it's fucking amazing

[oak osprey]

text encoders are hard to train because you can't directly do inference from them, you have to build a unet as a backbone on top of the embeddings produced by the text encoder, and that doesn't / can't happen concurrently, you first train text encoder and then the unet

you can fine-tune the text encoder and try again but it's a great expense and by the time you discover you've been fighting a cursed text encoder quanta, it's a few hundred thousand dollars in

i've messed with reverting a unet to an earlier checkpoint, and using a more-tuned text encoder, which works very well. i assume at Stability's scale, this would work better, and be even easier

[wispy nest]

the "showdown" seems like such an apples vs. oranges comparison

[oak osprey]

yeah it's like the highest level of a weird tournament

it makes sense though once you see the charts they've produced from the data

[oak osprey]

[WARNING] Image path does not exist: /datasets/midjourney-c/fullbody_pose__dynamic_pose__perfect_blend_of_beefy_teyana_taylor_and_beefy_janet_jackson_as_perfect_goddess_sekhmet_the_goddess.png

midjourney users

god bless that image for being too small and getting deleted.

[cyan snow]

Dude, I can't wait to run this on my PC

[oak osprey]

i know what's likely to replace those donuts once you do

a jar!

a jar full of shiny apples with multi-dimensional fractals embedded within their essence

[cyan snow]

Yeah, I will make some crazy shit with that model

[oak osprey]

"sick burn bro" prompt

[oak osprey]

training SD 2.0 is so fuckin easy

2.1 is like getting punched in the face and asking for more

stellar at composition, really bad at denoising

[cyan snow]

And SDXL is a fucking legendary model

By FAR the best, it has so much potential, and SAI will make specialized fine-tuning tools

The unfinished and not finetuned version on clipdrop already CRUSHES everything else

[oak osprey]

yeah the base SDXL feels like Bing image gen and the refiner just blows away expectations

and i don't mean Bing in a bad way, that thing is really, really impressive

[cyan snow]

Bing's image generator isn't that close to XL

It is as good as 1.5 finetuned models

[oak osprey]

what are you talking about when you say "XL". it is two models, and you can run the base on its own, which is what we've seen tested in the bot channel vs the refiner before

i've been seeing mcmonkey's pictures he shows which he repeatedly says are just the base and it's like Bing image gen

[cyan snow]

The refiner should already be in one of the variations on release, the version with refiner should be a single safetensor

[oak osprey]

it's not a single safetensor file, at least i don't think it is, that doesn't matter much

it's an img2img model

[cyan snow]

They confirmed it to be 1 or 2 safetensors

[cyan snow]

What?

[oak osprey]

the refiner.

[cyan snow]

Isn't it like just more parameters in addition to the base?

[oak osprey]

you mean like a LoRA? nope

it's more like Controlnet Tile

[cyan snow]

It's something new, isn't it? I bet it's a new component, just like versions of 1.5 have VAE baked into them and some don't

[oak osprey]

it is its own model with just one text encoder and a vae, but you can reuse the vae among two models in like, ComfyUI

[cyan snow]

ComfyUI?

[oak osprey]

yeah a node-based thingy like A1111

you should try it, it's the new hotness

[cyan snow]

Huh, does it have all the extensions?

[cyan snow]

Like making game shaders and materials?

[oak osprey]

it has nodes and community members do write stuff for it but i'm not sure if they're called extensions or what

for example it has ultimate upscaler as of a couple days ago

[cyan snow]

If it has hires and CNET and Adetailer, I'm sold

[oak osprey]

bingo like the blueprints system in Unreal Engine

[cyan snow]

That's awesome, I'll try that when I'll return to my house in 2 weeks after SDXL is finally released

Also, there is a new A1111 extension that allows LLM integrated prompting

It can in fact get instructions

Oh shit, it does have hires fix =]

And Adetailer

And even prompt gen

Oh, it has the ability to fuse 2 images together, this is insane

[oak osprey]

it's got everything but i honestly can't answer generally whether "does it have extensions" covers your needs 😂

[cyan snow]

I might switch to it when I get back

[oak osprey]

well i hope you like it, it's probably the best tool for doing local inference at the moment

[cyan snow]

And I'm familiar with that kind of node interface, this seems to be just right for me

[oak osprey]

i don't mean in a "it's the best, period" way, i mean in a technical sense

so an interesting feature is that you can drag images from the examples page on the Github into the ComfyUI and it loads the workflow up

[cyan snow]

Idk, next time I will use the AI will be after SDXL releases. We'll see

So far I'm impressed with ComfyUI, I really like the node design, it will be much easier for me to work with

Thank you for introducing me to it =]

[oak osprey]

i can't wait to see what weird shit you do with it

you can load Kandinsky alongside your model and alternate steps of denoising between them

@sterile temple did something like that, i believe

[oak osprey]

@cyan snow trying kandinsky

might need to check on something in a sec to fix that

oh it's pretty good

it has an interesting take on the girl with the pearl earring

can't say i've seen that one before

@wispy nest kandinsky does your oil digital real thing

for such a simple model it's very impressive

a little toasty on certain stuff

[cosmic palm]

I doubt anyone will make a fanart of Hu Tao

[stoic lotus]

some more RDFX-PRO images

[oak osprey]

surprisingly coherent on human photoreal subjects in Kandinsky lmao what the hell

[smoky oak]

You know, I see why AI is taking over news reports when humans are still using shitty grammar like this:

"Much faster than by using"

What the don't mention is it has less params than 1.5, it can only do 256x256, and it's from Novel AI lol

And it's results also look pretty terrible lol

It is a cool little piece of tech, but once again, just click bait to try and make people click to see if there is something new better than SD

[oak osprey]

maybe it's trained on 10% of the hw

emad said something like that was coming and impressed him

[oak osprey]

🥴

[smoky oak]

https://tenor.com/view/whimsy-wednesday-deathxmon-gif-27105040

[oak osprey]

i love the kodak commercial about saving moments in life where the little girl's flying toy like that goes straight into the fireplace

[smoky oak]

yesss lmfao

https://tenor.com/view/barbie-fire-fail-gif-4451510

Also just wanted to share this snippet of a song, cause this has to be the single biggest and punchiest kick I have ever heard, and I am not sure how you get something to sound so massive while not being any louder than the rest of the song

(Loud and intense warning)

Like, I checked, and the kick is no louder than the rest of the song, its just that fucking saturated 😩

And now I am working on the new and extremely over engineered mounting system for my AC unit lol

Looks like I am taking a break from AI, which I can't believe I am saying lol

[oak osprey]

she just lets it happen

[smoky oak]

Gady laga

[oak osprey]

Gravy Lager

[oak osprey]

closest i can get to Ligma Gravy

[wispy nest]

one of these uses that artist's name, the rest don't. Which is it? ;P

[shrewd trench]

Who else hyped for Biortsmer coming out later this year???🤩🤩🤩

[oak osprey]

@shrewd trench

i meant to send the picture of this supermodel first responder rescuing someone from a mud pit

[shrewd trench]

lmao thats epic

looks like the netflix adaptation of biortsmer 2

[oak osprey]

she's gotta go in or they're gonna die maybe

[shrewd trench]

oh no

im getting titan vibes

[oak osprey]

supermodels just love getting stuck in pipes 😦

[shrewd trench]

oh not only pipes...

[oak osprey]

the clothes dryer?

[shrewd trench]

YES

ah damn

i had found a funny image

of an ice bear in a drier

dryer?

[oak osprey]

driar

[shrewd trench]

droia

im not on my pc rn

[oak osprey]

get back to your desk, champ

[shrewd trench]

do you know regional prompting

[oak osprey]

i had dinner with them last week

[shrewd trench]

ill go to sleep in a bit. its 2.30 here

[shrewd trench]

that is good to know

can you try making 4 pipes?

like with regional prompts

with many stars

[shrewd trench]

like this but 4x

[dense tapir]

@smoky oak A new sampler just dropped and from what I am seeing half the steps of SDE.

https://old.reddit.com/r/StableDiffusion/comments/14n6vt7/new_sampler_restart/

[shrewd trench]

wai wut

how is it?

[smoky oak]

oof, old reddit

[dense tapir]

I have no idea we must now wait for Auto1111 to get it

[shrewd trench]

ok!

[wispy nest]

noice.

but hows the quality? 🤔

[shrewd trench]

in the meantime i will go to sleep

[smoky oak]

Their example images don't seem to show literally any benefits IMO

who generates at 100 steps

[shrewd trench]

gn guys

[dense tapir]

night

[smoky oak]

I mean, if they are gonna make claims about being faster and better, show me some 20 sample comparisons, with times listed lol

[shrewd trench]

dont get lost downloading loads and loads and loads of loras bc of the previews : D

[dense tapir]

I haven't seen them over hype before.

[smoky oak]

I guess we will see when it gets implemented

[dense tapir]

All academic and just like opinions, and assholes, we all have one so I just wait to see. Might be like UniPC, lol.

[smoky oak]

Yeah, I remember how much people thought UniPC was gonna be good lol

Its just slower and less efficient DDIM lmao

[dense tapir]

ffs, no tensor core cards can't even use it, like my 1060

[smoky oak]

Restart?

[dense tapir]

I am unsure what restart is as I have never seen it before.

[smoky oak]

oh, so what was that statement referring to then?

[dense tapir]

I am still wondering if the right course of action when training is to use the base mode from SAI?

for dreambooth

[wispy nest]

Help

[fleet goblet]

Gold.

[rigid heron]

is this the house md guy

[oak osprey]

hugh laurie

@fleet goblet

[fleet goblet]

Cybernetics have come a long way.

[fleet goblet]

This ghost says hello.

[dense tapir]

Nope. Kinda looks like Hugh Laurie though.

[rigid heron]

i drive

(literally me)

[dense tapir]

My power supply arrive tomorrow so I can finally spend all day on Sunday moving my entire system over to a new case. Next up will be a new video card and I can hold up to 400-430mm long cards.

[oak osprey]

the hat looks so impractical

[rigid heron]

it does but also there are like 4 hats I could think of that look like that

so I guess it's not unrealistic

[unborn light]

Action Figure

[dense tapir]

Well, damn. Sexy beast.

[dense tapir]

I think this training was a success.

[rigid heron]

for some reason I always default with sdxl to just casting actors as characters they didn't play

[unborn light]

Another action Hero DT

[rigid heron]

I tested what would happen if I did a character as another character instead of a person as a character

It almost worked

Prompt was ("iron man", "as Doctor octopus").and(0.9,1.0)

Im really running out of ideas I can do with this clip drop I need this thing to release

[rigid heron]

Yknow in this drought of creativity in my brain I also may resort to funny ai boob

Also every time I try and get a skull wearing a helmet it just makes a skull helmet

[feral pebble]

"A young lady looking at her cellphone in her lap. The glow of the cellphone is becoming a small pegasus, with its wings spread out flying through the magical glow of the cellphone. Realistic, but magical."

I don't get it. Whenever I use SD (this is Invoke), it just mixes up my terms, like it doesn't know grammar at all.

any of you have any insight (or especially solutions) for this?

[green socket]

i gave it a shot in the bot channels and got this

https://cdn.discordapp.com/attachments/1080261341362786384/1124580680219754516/Realistic_portrait_of_a_woman_looking_at_cellphone_magic_pegasus_standing_on_cellphone_realism_volumetric_lighting_octane_rendering_hyper_realistic_seed-0ts-1688191323_idx-0.png

ill fiddle some more and try to get the pegasus on the phone

getting closer, maybe?

https://cdn.discordapp.com/attachments/1100170514670039070/1124582838004633761/Realistic_portrait_of_a_woman_looking_at_cellphone_face_lit_by_cellphone_screen_tiny_pegasus_standing_on_phone_screen_realism_volumetric_lighting_octane_rendering_h_seed-0ts-1688191839_idx-0.png

[green socket]

which one of these is closer to what you're imagining?

[feral pebble]

@green socket hey, thanks for helping :)) .. this was because my wife was hospitalized and stuck awake for 5 days straight.. she saw a friend of ours sitting there (she was actually there)

but she saw the glow of the friend's cellphone which turned into a pegasus flying above it (kind of small.. like, dunno, 1 foot tall or less.. )

after that it turned into a narwhale

[green socket]

https://cdn.discordapp.com/attachments/1100484581037195384/1124588925504204900/Realistic_portrait_of_a_smiling_woman_looking_at_cellphonesmall_pegasus-1_3_flying_out_of_cellphone_screen_realism_volumetric_lighting_octane_rendering_hy_steps-40_seed-0ts-1688192640_idx-0.png

it might be something where you gen the lady and then inpaint the pegasus

i like this one even though there's no pegasus lol

https://cdn.discordapp.com/attachments/1101178530865352815/1124590048571052125/Realistic_portrait_of_a_smiling_woman_looking_at_cellphone_small_pegasus-1_3_crawling_out_of_cellphone_screen_realism_volumetric_lighting_octane_r_style-Fantasy_Art_seed-0ts-1688193558_idx-0.png

[green socket]

Prompt to try: Realistic portrait of a smiling woman looking at cellphone, (small pegasus:1.3) crawling out of cellphone screen, realism, volumetric lighting, octane rendering, hyper realistic, intricate,

Negative: angel wings, blurry

[feral pebble]

I wonder if it's me or if invoke ai just sucks

@green socket

[green socket]

oh stable diffusion 2.1 really sucks IMO.

use a better model. I like deliberate but there might be another one

[green socket]

rev-animated or dreamshaper are the more popular fantasy models. I would link to them but Civitai is down... again.

[feral pebble]

@green socket i've not yet tried the alternate models .. just getting back into trying SD after a lot of issues in life.. civit's neat

yikes. why's my stable-diffusion-webui come up all in chinese??

[green socket]

i tried revnimated and dreamshaper and they just wanted to make a random lady with no phone. :\ might need to play around with weights.

[feral pebble]

I think the language embedding we use is garbage

it really doesn't have enough capability to understand what the sentence(s) are saying

[wispy nest]

GUYS GUYS

Who knows how to make stable diffusion concept files?

With .bin extension

[smoky oak]

Comfy UI is genuinely pissing me the fuck off

it works, then I try to run it and it tells me I need an A100 in the error message? I assume it installed wrong, reinstall

now xformers is broke, so I update it

it can't update cause now torch is broke

I follow their doccumentation on the page to fix that error

It can't cause it says it doesn't have torch to rebuild torch

What in the actual hell happened to comfy UI, why is it so convoluted and such a pain in the ass to use now?

I have used it several times in the past with 0 issues

[wispy nest]

@smoky oak

[smoky oak]

What?

[wispy nest]

If you have perma trouble with a100 ypu better use...(i ll send in you dms if you dont mind)

[smoky oak]

What are you on about?

I am so confused lmao

[wispy nest]

Oh, my apologies, i thought you re not satisfied with sd webui

But this is different stuff you re talking about

@smoky oak by the way, do ypu know how to make stable diffusion concept files?

(With .bin extension)

[smoky oak]

no idea

[wispy nest]

Crap

[smoky oak]

Just trying to find my way around the dubpster fire of comfy UI right now

It has you install torch vision and torch audio, but they require torch 1.12.1, yet it has you install 2.0.1, which breaks everything

IDEK what to say

[wispy nest]

Im trying to become a god of nsfw arts(i cant tell any further since rules are against it)

[smoky oak]

m'kay, you do that I guess lmao

[wispy nest]

I just said im nsfw creator, i dont discuss it any further so im safe

[smoky oak]

great, now stop talking about it lmao

Plenty of us here do tons of NSFW, no need to bring that into the server

[wispy nest]

Also i cant say anything bad about xi qingping since i dont want my xiaomi explode when i get a call

[smoky oak]

https://tenor.com/view/boooooo-this-guy-stinks-serious-spongebob-squarepants-gif-11981386

[sterile temple]

I just use the portable version of comfy with no issues

[smoky oak]

I have used comfy plenty of times, but my god this is pissing me off

the xformers it auto installs needs torch 2.0.1, but the torchvision and torchaudio it installs need 1.12.1

update the vision and audio, break xformers

Update xformers, break audio and vision

so now I am having to back one by one through the X formers versions until it auto installs a torch version that works

oh yay! Now it broke torch again! Woohoo!

alright, uninstalling all of the messed up shit they installed

[sterile temple]

did you use this one?

[smoky oak]

I used that one, and direct install

2 times each

same issue every time

[sterile temple]

huh weird

[smoky oak]

installs torch 2.0.1, needs 1.12.1

I intall 1.12.1, not xformers is broken

Downgrade Xformers to 0.13, now torch is missing

Install torch, its already installed

try again, no torch found

[smoky oak]

oh wait a sec, I did NOT try that link

trying now

I tried the direct zip download, not the portable one

time for my CPU to off itself uncompressing all of this haha

[sterile temple]

hope it works 🤞

[smoky oak]

ok, it at least LAUNCHED this time

time to see if what I am trying to do will work

@sterile templeIT WORKS YOU FUCKING MAD MAN

WOOOOOOOOOO

[sterile temple]

https://tenor.com/view/y2kpod-millennium-welcome-to-the-year2000-y2k-audiodrama-gif-15970306

[rigid heron]

did clip drop get more servers or is there just nobody using it

Every time I press generate the queue is like 10 images

[wispy nest]

Guys please, can someone please help me with textual inversion training?(i only have 6 gb vram) I just need one trained .pt file

[sterile temple]

API is alot faster today too, nice

[split valve]

SD XL0/9

Analog filter

[sterile temple]

getting the squid robot vibe from the matrix movies

[split valve]

clockpunk

[split valve]

slightly

[kind quartz]

dancing house, and gaudí, maybe Hundertwasser. If not, try him @sterile temple

[sterile temple]

yes gaudí 🙂

[kind quartz]

fift remind me dancing house.
Gaudí was giant 🙂

[sterile temple]

was blown away by these frank lloyd wright images last night

[kind quartz]

yes i like architecture as well. I think i love some sort of architecture most, because kindergarden was in that style.

[sterile temple]

sdxl does architecture sketches so well

[stoic pivot]

I found a odd thing, larger images follow the prompt better, I wanted to test models and made a prompt that was not very specific other than it should be small creatures floating in space. at 512x512 the character just stood on the floor, but at 1920x800 I got images of floating animal characters that I later could crop out.
AI is strange.

[dense tapir]

LOL

[stoic pivot]

Now I also want to make scary images

Sorry, here is a nice image that is not soo scary.

[stoic pivot]

A monster under the bed.

[kindred willow]

Dr. Brenden Williams, CTO of Three Corporation, 2047
Born Jan 3rd, 2013, Dr. Williams was a top student throughout his education, getting his Ph.D in computer science while minoring in Business, he became known as the "Ruthless Programmer" due to his no-nonsense attitude towards handling situations and his education.
Sometime in 2040, he befriended the CEO of Three Corporation, a Technology Megaconglomarate, by pure chance and due to his charismatic charm, he was appointed CTO of Three Corporation.
He was eventually fired in 2051 as a result of his behavior and involvement in the cover-up of the Baton Rouge Teleportation Experiment Incident.
Despite the fact that during these events, he terrorized the rural town of Huntington, shot a Public Law Enforcement Officer in the knee to prove a point on who was in charge of the situation, and his Machiavellian handling of the situation, he was not charged of any crime by either Public Law Enforcement or Pyramid Security Solutions.

Thoughts on this Character?

[wispy nest]

He looks like mr blue eyes from cyberpunk 2077

[kindred willow]

Interesting

[west tree]

It's a butthole.

[split valve]

SDXL seems to have a very long token input limit...

[oak osprey]

you know, the unfortunate thing about the SDXL fine-tunes that are coming out is that they're all going to be based on SDXL 0.9

[split valve]

works really well at incorporating every word you give it.

[oak osprey]

the 1.0 release isn't available for fine-tuning yet, and so the base model at release time will possibly be better than the fine-tunes

[split valve]

im going to post a bunch of images... same prompt but different filter styles in clipdrop.

[oak osprey]

i don't get why they did it this way because there's zero motivation to actually fine-tune it

[split valve]

and of course arigami

[oak osprey]

this one looks weird to me, am i the only one?

like some of the detail to me looks like artifacting until i look closer and then it... doesn't?

something uncanny and extreme is happening but i don't know what

[split valve]

i think thats the photographic filter. The detail will tend to artifact like when you run an AI upscale and it comes out like reptile skin

[oak osprey]

the styles are prompt keywords that get added to the end fwiw

[split valve]

its amazing that they are just keywords, because they feel like fine tunes or something

[oak osprey]

prompts are very powerful in SDXL but as a result, you can go really far in various directions eg. too smooth, too crusty, etc

you have to somehow be very precise in what you mean

know what i mean?

[split valve]

i can tell the prompts are reacting different to the old versions. i like it

it gives you what you ask for.

[oak osprey]

it does and it doesn't. you can ask for watermarks and you don't get any, which is why the negative prompt for watermark won't remove them

you can't mess with negative prompt or seed gen in clipdrop

i have local copy of the SDXL weights and run it in Python

[split valve]

nice... i would run it on my machine but its just a GTX 1070

[oak osprey]

you can probably still do that, albeit slowly

there's many layers of optimizations you can attempt, as this is a modular architecture, not all of the model has to be loaded at one time

[split valve]

yea.. its pointless running up an electricity bill tho rather than paying for clipdrop

[oak osprey]

oh true, true. you should try StableStudio or DreamStudio. they can plug into the SAPI and give you more control

[split valve]

i have dream studio as well.

[oak osprey]

does it let you set negatives and seeds etc?

[split valve]

dream studio does.

i got this prompt from ChatGPT Psychedelic alchemical garden: A realm where psychedelic plants and magical elements intertwine, creating an alchemical garden of vibrant colors and surreal shapes that invite viewers to explore the mystical depths of the subconscious.

i added Alex Grey style to the end and it worked... cant believe the input token limit must be high

[oak osprey]

they've been messing with that stuff in secret

idk if they're going to keep the current prompt parser in the end

@prisma iron any details on that prompt parsing implementation yet?

[sterile kiln]

Nice prompt

[split valve]

Thats pretty good.. usually AI will totally screw up windows and such.

[oak osprey]

idk lol the reflections are weird

[dense tapir]

wth is this? LOL

[oak osprey]

a chorse

[wet cypress]

can someone tell me how can i generate more images like this? What's the style

How can i find the seed

[dense tapir]

Unless it is in embedded info or a .txt with the info the seed is lost forever.

[wet cypress]

do you know what style this is

[dense tapir]

I don't as I don't know Anime.

[wet cypress]

if i find the seed, how should i format the prompt to use the seed

[dense tapir]

Looks a bit like 2000-2010 style anime

[wet cypress]

i generate images on discord

/ dream

[dense tapir]

I don't think you can as most of us do this locally.

maybe someone who uses those discord bots can help more

[wet cypress]

thanks

[dense tapir]

you are welcome

[oak osprey]

that is SDXL?

it looks a little more cooked than i'd expect but some of it has "the look"

tbh it's more like a very good 1.5

[split valve]

yes SDXL from clipdrop

[oak osprey]

damn, wouldn't have expected that kind of halo'ing from clipdrop

[split valve]

whad do you mean?

[oak osprey]

if you don't see it, i don't want to point it out lmao i love the images

[split valve]

i can see things i want improvement on such as details being more realistic and not looking blatantly upscaled but its still good, i can live with it.

[oak osprey]

i've been playing with it in some toy python scripts but it's not very easy to tweak and re-generate that way, so i'm working on getting it going in my Discord bot code

[split valve]

if people can evolve on it and improve the tools thats great.

[oak osprey]

fucking laughed so hard

[earnest solar]

before I actually started reading the text I thought it was some sort of real chart for a moment and wondered why you were posting it here.

[proud dagger]

I agree.
Mlued tlhe biee lardre cont?
21.4%
1.4 on scare
velb a3%
10 caa
veri .S3%
7.5.3%
The bard yflne
war Yre o werßt
hale gh.?'
38.%
Hae srades fof eprile preotk

[oak osprey]

it is better than 90% of infographics i find

it doesn't have an axis that doesn't start at zero, has no unlabeled axis

[proud dagger]

(shoutout Windows Powertoys for being able to transcribe XLesian text)

[oak osprey]

things are, if anything, over-labeled

[oak osprey]

one neural net giving another a helping hand

feed that thang back through BLIP2 and get the text perfectly too i bet

[earnest solar]

Asbo Zaprudah, magic baby seal

[west tree]

Is it a butthole?

Really, though, loving the use of Giger.

[oak osprey]

still clipdrop?

[split valve]

yup its all clipdrop because ive subscribed now,

[split valve]

how did you know i put giger in the prompt?

[west tree]

Because I art.

[green socket]

Very awesome! Did you happen to save the prompt for this one?

[split valve]

Decaying Interdimensional Rift: A grungy, crackling portal reveals terrifying glimpses of alien realms. The stark contrast between corroded metal and cosmic swirls creates a blend of ancient and futuristic aesthetics, ominously suggesting the risks and unknowns lurking beyond the boundary. By H.R. Giger.

[west tree]

Also, dood, seriously? Maybe the f* xenomorph.

[oak osprey]

just kinda hoping it's not lol

they're not great looking to me

[split valve]

yea true. its a dead giveaway.

[oak osprey]

i don't know if it's the prompt or their upscaler on clipdrop

[earnest solar]

Masterpiece, best quality, Apollo Berrun, handsome young man, detailed short golden hair, detailed blue eyes, slender athletic build, DND, monk, digital painting, by Akira Toriyama, lora:akiraToriyamaStyle_toriyamaAkira3:0.7
Negative prompt: Sexy, Big Breasts, NSFW, lewd, risque, nude, hat, laser, lightsaber, showing skin, lowres, bad anatomy, bad hands, bad faces, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blind, bad eyes, ugly eyes, dead eyes, blur, helmet
Steps: 20, Sampler: Euler a, CFG scale: 12, Seed: 390771138, Size: 768x768, Model hash: 22c84a4a0a, Model: mangledMerge_v3, Version: v1.3.1

it's a 2.1 diffusion.

[west tree]

Who TF uses Euler a?

[green socket]

I do still

[earnest solar]

me, What would you recommend switching to?

[oak osprey]

DPM++ 2M

[west tree]

Why TF 2.1? I just loaded it up and I can't get a good prompt for shit.

[oak osprey]

try my 2.1 model 🙂

[west tree]

I use DPM2, and avoid Karras unless I'm stumped.

[earnest solar]

mangled merge is a good 2.1 model

[west tree]

or UniPC when I'm broke.

[west tree]

Where do I get these 2.1 models? Does civit have them, or do I have to learn how to use huggingface properly?

[oak osprey]

https://civitai.com/models/73341/pseudos-journey?modelVersionId=78058 < this one is midjourney-trained

https://civitai.com/models/97808/pseudo-flex-base < this one is more photoreal and does portrait and landscape and a base res of 1024x1024

apparently they're both 5 stars now

[split valve]

Is SDXL going to gradually improve or do we need to wait for an entirely new model to be trained?

[oak osprey]

people are using my model as a base for merging now with fantastic results

i'm pretty happy about that because i have no energy for taking my model further at this moment

[split valve]

Thats above my head i dont know how you guys manage to combine models and stuff.

[oak osprey]

i honestly am not a model merger, i am a model trainer. there is a major difference, both are their own form of art

[west tree]

It's a four-click process. I'm trying to learn the dreambooth extension if anyone has advice

[earnest solar]

https://civitai.com/models/5395/mangled-merge

[oak osprey]

^ that has pseudo-journey-v2 in a high percentage as a merge too

mangled merge is possibly better than journey v2 as a result

i think it 'knows' more stuff but has a lot of 'magic' from my model weights, no idea though really

[split valve]

whoa!!

[west tree]

Nope. Need photorealism.

[oak osprey]

try a stunning portrait of jesters at the twisted carnival

[split valve]

Clipdrop you cant use any deep controls i dont even think negative prompting works.

[oak osprey]

is this photorealistic though to you, i don't know. to me it looks like studio portraits but not what 1.5 models refer to as photorealism, which itself is like an AI art style.

[spring sail]

sorry I've been inactive for a minute, got covid. how y'all

[west tree]

https://tenor.com/view/icp-insane-clown-posse-insane-clown-posse-gif-20063408

[oak osprey]

here's one with better contrast

yeah someone else said it looks like the prompt involves the terms 'juggalo convention', i don't know what that means but now i do. thanks

anyway these are the results of pseudo-flex-base

[oak osprey]

works best at 1024x1024 or like 1280x1024, 1366x720 is really cool and the widescreen vibe feels very different

portrait mode at 960x1280 is great too

nowhere near the number of deformations but honestly i have to admit, some subject matter that isn't in my 480k training images doesn't work well in portrait. stormtroopers mostly

none of that uses hires fix or img2img, LoRAs, textual inversions..

[earnest solar]

I like this mangledmerge

I'm pretty shit at prompting and it still spits out awesome stuff.

[oak osprey]

yes! it works without negatives because of the pseudo-journey-v2 text encoder being slowly/carefully trained on midjourney image-caption pairs successively freezing the layers of the text encoder until just the last two were being tuned for a few weeks straight

[earnest solar]

I'm using negs anyway as I'm using it.

[oak osprey]

the freezing happened pretty early on, like, you want to do it within a few hours of starting training or you'll be having problems later

[earnest solar]

used to it*

[oak osprey]

negatives still work very well yeah

[dense tapir]

Yes, he put a lot of effort into that

[earnest solar]

i can tell

[split valve]

its more coherent.

Hands look correct and armor details

[oak osprey]

fine-tuned text encoder ftw

out of the box the thing is so damn cursed

[split valve]

im hoping clipdrop will improve.

[oak osprey]

you can do that for clipdrop, too

well

i mean sdxl's text encoders

i'm nervous to try tbh but tuning the text encoder is something i can begin now and still have that work apply for 1.0 (i think)

[earnest solar]

Yo I love it when the botto spits out non-Euclidean geometry

[split valve]

it cant get the details.

[dense tapir]

Mangled merg ftw

[smoky oak]

So I have been messing with SDXL's base weights after I finally was able to figure out how to get them working, and I amr eally hoping that SAI is purposely giving us bad models just to try and still leave some improvement for release day, cause these models are both pretttyyy... blegh

The refiner is the star of the show, and its still pretty disappointing

It cooks anything ever close to realistic, and the base model is basically 2.1 but a little less bad

VRAM requirements are a bonus tho, I will say that

[smoky oak]

8GB will definitely be enough for 1024x1024 with SDXL base

and if you use it strategically, like with comfy UI, you should be able to run base and refiner just fine in 8GB

also, whichever mod said that SDXL uses barely any more VRAM when going to higher resolutions, that does seem to check out!

1024x1024 uses 6.4GB VRAM on average for me, while 4096x4096 only used 7.1GB, though the VAE decode did push it over 10GB VRAM, but when it switched back to tiled decode, it was able to complete in just over 8 GB at 8.2GB VRAM

Very very interesting how they did that

[split valve]

stupid nvidia locking most cards to 8GB

[smoky oak]

I would have been able to share this all sooner, if my dumb ass wasn't having problems installing comfy UI

[oak osprey]

whut

[sterile temple]

Looks like a big difference between the API model and the weights you have just based on running the same prompts

[oak osprey]

they're allowing just anyone in to the beta now

[smoky oak]

looks like lmao

[oak osprey]

did you roll a D20?

it's a D&D reference, young'un

[smoky oak]

Yeah, that was very interesting to me. It does seem to run artistic stuff about as good as the API, but its shit at realism

[smoky oak]

I actually play D&D, thank you very much :<

I have for 7 years >:C

[oak osprey]

thats the joke

such a nerd

[smoky oak]

[oak osprey]

i didn't even know you did. but i knew

i knew, the lady at the grocery store knew, the librarian knew

[smoky oak]

base SDXL 0.9

after upscaling and adding the refiner

I am quite happy with this models artisti stuff, but flat out pissed at its realism. It does about as bit, if not maybe even worse than base 2.1 with photo realism

[oak osprey]

base 2.1?

[smoky oak]

yeah, that bad lmfao

[oak osprey]

what

no way

2.1-base or 2.1-v?

2.1-v is overtrained gore

i have not tbh tried 2.1-base a whole lot

[smoky oak]

like in some ways, it can do good, but man, tyou really need to beat the shit out of it with CFG finetuning and some extremely specific settings

Honestly, I am not too sure which ones I have used

just know this base is not good for realism, like at all lmao

Its all smeary, and blurry, and full of what looks like MJ V4's artifacts

[oak osprey]

2.1-base might be the best of that series considering it's "middle of the road" in terms of training step count just like 2.0-v

but the added benefits of the reduced LAION filtering

ohhh

are you able to put a custom text encoder now?

[sterile temple]

How do they run it with the two clip models?

[oak osprey]

you can use the file from my pseudo-journey-v2 branch and the one from the pseudo-real branch, with any 2.x model, and improve coherence

[smoky oak]

like here for example: 1.5 finetune wolf

[oak osprey]

@sterile temple they're combined in a single huge layer on the input

[smoky oak]

1.5 finetune high res fix wolf

SDXL refiner wolf ._.

[sterile temple]

Can you do the same in comfy?

[smoky oak]

you talking to me?

[oak osprey]

tis such a weird combo of realistic and digital look

[smoky oak]

if so, I am using comfy UI exlusively for this test lol

[smoky oak]

right, the refiner cannot do realism at all

[oak osprey]

@smoky oak pfft give it a human from EpicREAlsim

or whatever that model is called

i have faith in this refiner

otherwise what's it all been for

more bouncy lighting, i guess

[green socket]

Like that game from stranger things? 😛

[oak osprey]

E.T. the board game?

[smoky oak]

ask the refiner for an image that is realistic, with an extensive negative that tries to discourage being artsy

get this:

the refiner cannot do realism at all lmao

[oak osprey]

don't give any negative at all, what happens?

negatives i find, tend to hurt SDXL bot outputs

[smoky oak]

it just leans more into artsy stuff

[oak osprey]

ugh

wait

"ask for realistic" how

[smoky oak]

this was no neg

[oak osprey]

with SDXL some of the realism terms from 1.5 days are counterproductive

[smoky oak]

I tried several different ways of prompting it, all of which 1.5 understood, and base SDXL understood, but not the refiner

[oak osprey]

gotchya

i hate how slow the sd_xl branch of diffusers is to start up, it makes it a real pain to test all of this

[smoky oak]

same prompt in SDXL base got me this

far more realistic, but it really just looks like MJ V4 up in here

[oak osprey]

that looks like what i get from 2.1 with controlnet tile on a 1.5 model

i was waiting 15 minutes for torch to start up

[oak osprey]

turns out the GPU crashed

[oak osprey]

the patterning on its forehead wasn't that bad before on clipdrop

[split valve]

well theres lots of options i could chose. like the various upscalers

and filters

[smoky oak]

yeah, those images lead me to believe that the 0.9 SDXL beta weights are just a bad version to try and validate training, but not make great images

[oak osprey]

there's uh, different upscalers on clipdrop paid?

[split valve]

yep

[oak osprey]

sigh

[split valve]

just different models, like sharpen, denoise, remove compression, then the smooth or detailed option

2x clipdrop sharpen upscale

on scrutiny it looks bad

[smoky oak]

that does not look very good at all, so thats reassuring lol

[split valve]

patchy and what not

[smoky oak]

As a paid service, I would mega pass lol

[oak osprey]

thanks for burning your credits for us, for science

[smoky oak]

clipdrop is unlimited IIRC

[oak osprey]

not when you pay

it's the ultimate Thank You

[oak osprey]

hmm i guess it's just skipping queue for 1500

i see

[smoky oak]

per day, yeah

doesn't cost them anything more than other users using for free haha

[oak osprey]

it's confusing keeping track of DS, Clipdrop and SAPI

[fallen dock]

imagine having to pay to render images lol, stable diffusion my dude is the way to go

[oak osprey]

this is stable diffusion

[fallen dock]

nah

[oak osprey]

that's stability AI's paid service for SDXL use

currently the only way to use the model for most

the bot is here, but... uh... YMMV

[fallen dock]

bot?

did you just call me a bot?

[oak osprey]

#bot 1 through #bot 10

[smoky oak]

yeah, used to be good, recently got neutered lol

[oak osprey]

#1101178483608137918

[fallen dock]

oh nvm, that bot, lol

so I assume yall haven't heard of......AUTOMATIC1111

[oak osprey]

that thing is junk 😄

[fallen dock]

excuse me?

[smoky oak]

yeah, Auto is rough, its what I use day to day, but its not very good lmao

[fallen dock]

[oak osprey]

it's not a reflection on you

[smoky oak]

comfy UI is luring me in, I really do hate to say it

[oak osprey]

you are fine, Poly

[fallen dock]

ya but yall are paying for something, ya really don't need

[smoky oak]

Yeah, Auto is fine, its just not made well for what it does

Its inefficient

[oak osprey]

oh, i don't pay for clipdrop

[smoky oak]

no? lmao

[oak osprey]

sono does and that's their choice to do so because they saw value in SDXL

[smoky oak]

I have like 300k local image gens at this point, and I train my own LoRA's

[oak osprey]

mysticago made 14,000 image gens using the SDXL bot here so far

for free

no need to heat their own computer/house up

it's okay if others are fine with using even Craiyon to make their AI images if they're really happy with it

[smoky oak]

I have never paid a penny for SD

[oak osprey]

in fact i'm going to test Craiyon again lmfao

i haven't used that little boi in like, 3-5 months

oh it has negative prompts now

and art styles 😮

and it uses some language model to predict your next prompt

annnd it has an upscaler

yeah this is totally different now

[smoky oak]

oh wow, thats insane haha

I wonder how wombo dream is doing now

[oak osprey]

i guess i'm not a craiyon prompting expert 😦

you have to squint pretty hard

[smoky oak]

wow, that looks WAY better than when it was just dalle mini back in the olden days lmao

[oak osprey]

i know

i feel like there's more you can get out of it but i don't know* how and it takes a while to gen lol

ahhhh the prompt is too extreme

simplify it, and results are way better

[smoky oak]

Interesting, I am trying wombo dream directly against base and refiner SDXL

[oak osprey]

dude why lmao

[smoky oak]

it actually did pretty damn good lmao

[oak osprey]

shouldn't have said that lmao