🏞｜general-with-images

[oak osprey]

i had GPT come up with 50 random prompts about candygore and it did some really interesting stuff, some of which i'm going to visit more deeply at some point

https://images-ext-1.discordapp.net/external/a4C15h7XoVJ9eN3QdjECwl2d0tFR5IqCOCB5sBV3mKY/https/tripleback.net/public/discord/1687933437.13697084956391d9265ceb742f38271d3075084.png?width=1756&height=1003

twisted carnival seems to be a particularly good theme

[dense tapir]

Nice. I never understood why a face 5 feet from the virtual camera is janky in 2.1

[oak osprey]

because they are so far away from the 256x256 training epochs is my guess

or maybe too close to... i train on base res of 1024x1024 on these models

i had the assumption that the transformer/attention layers are learning how to create small faces in a larger context by looking at a high res image, where the face is a very small percentage of the image

the other thing there is the training schedule changes really impact faces on their own. doing inference from the trailing timestep really does produce a more congruent result to the loss witnessed during training on the backward pass.

[dense tapir]

Something I am flabbergasted over is how Nvidia is still benchmarking a 2k series card when showing a 4k series card. That should not even be mentioned.

[smoky oak]

@dense tapir pseudo is saying they are janky cause most of 2.0 is 256x256, then 512x512, with 768x768 slapped on top

And he is trying to fix the aspect buckets and resolution discrepancies by fine-tuning only on 1024x1024, which has enough detail for faces to look good far away still, and all I can say is his results are really speaking for themselves

[oak osprey]

well i'm sure it'll piss him off that you're relaying my findings but thank you for your sacrifice 🫡 lol

i wish he and i could work together again, and i'm willing to accept my part in our escalation and apologise, even if he won't accept it

[dense tapir]

FT scares me away due to PTSD and the people with their 350k datasets and captions all cranking around daily for 3-4 months.

4k series sure did keep the 3k series' values

[oak osprey]

crank dat hawg

[dense tapir]

that is the cheap one too

[oak osprey]

mine fine-tuning took like 2 weeks and he can try tuning on top of my checkpoint, i guess. there's no need to throw 3-4 months at it

[dense tapir]

FTW model is 850 and FE is 899

[oak osprey]

twisted carnival sister

fuckin' love that

twisted carnival is mine now, no one take that >:|

[smoky oak]

Me constantly

[oak osprey]

god that's not the original text

it's i smell the smelly smell of something that smells smelly

[smoky oak]

Nope

that smell, a kind of smelly smell, a smelly smell that smells... Smelly

Great, now smelly sounds stupid AF and fake lmao

[oak osprey]

[smoky oak]

God, the first 4 seasons of SpongeBob were so good man, and then it went to shit lmao

[oak osprey]

candygore attends twisted carnival

[oak osprey]

the contrast is because the prompt is whack but i like getting weird stuff from this model since it's so hard to

@smoky oak me taking 2.1 to the sacrifice grounds

[smoky oak]

So many iconic moments and creative quips from early SpongeBob

[oak osprey]

you're gonna love sandeez NUTS

yeah early spungbung was not for kids

[wispy ether]

what

[wispy ether]

https://tenor.com/view/misfortune-tragedy-sponge-bob-why-gif-18089335

i got a towel of that lol

[smoky oak]

Call that PH 14, cause that shits BASIC

[severe swallow]

I know what it isn't tech support channel but I need help

[smoky oak]

What GPU? What model?

Oh, I see the model, but I don't recognize it

[severe swallow]

Gpu-gtx 1660ti
Model-humaruyahidekazu (named after hetalia author)

[smoky oak]

Interesting, so you have to use no half and full precision

Is that a 6GB card or 8? I always forget

[severe swallow]

6gb

[smoky oak]

Ah, yeah, so med VRAM is recommended

Lemme look more

[smoky oak]

Oh, did you get a NaN error down below? @severe swallow

[dense tapir]

1650 and 60 have no fp16 support either

[smoky oak]

Yeah, that is brutal

[dense tapir]

yep

[wispy ether]

What the hell does that even mean

[severe swallow]

Yes and I have medvram

[smoky oak]

The devs for SD were talking to me about trying to get SDXL down to 8bit

And they said they also currently have decent results from 2 bit quantized LLM models

[wispy ether]

4000 series

Useless for anyone below

[severe swallow]

So what about my problem

[smoky oak]

Not true, it uses less VRAM regardless

[smoky oak]

If it's a NaN issue, it's likely you need a VAE

[oak osprey]

we made a discovery

[wispy ether]

Ok

[severe swallow]

Wdym

[oak osprey]

2.0 > 2.1 @wispy ether

[wispy ether]

ok

[severe swallow]

And write a command prompt to copy-paste

[wispy ether]

I use 1.4

[smoky oak]

I am eating with my mom right now, and I am away from my PC, so I can't walk through the steps of adding VAE's, but you should be able to find somebody or a video to help

[wispy ether]

Me reading i was eating my mom

[oak osprey]

if you love 1.4 you'd love 2.0 they're both very broken out of the box, but 2.0 has no crispy radiation burn noise

[dense tapir]

1650/60 has to use the -full flags like an amd does

[wispy ether]

1650 is worst

[oak osprey]

on my 4090 i have to use --no-half-vae to make 2.1 models work

[severe swallow]

I am Russian I don't have money to rtx

[wispy ether]

Same on 3060

12gb

[oak osprey]

with ALL 2.1 models?

that's so weird man

[wispy ether]

Idk bro

I dont have ALL

[oak osprey]

me neither

i use realism-engine to compare against. or Artius v2.1

no, nep

[wispy ether]

Ok

[wispy spindle]

???

[oak osprey]

[wispy ether]

Bye

[oak osprey]

fruit holy shit

[wispy ether]

BRO COMES HERE AND SAYS ???

[oak osprey]

usain bolt

did you get friction burn from moving so quick good catch

[wispy spindle]

u come in here and say 🪡

[wispy ether]

https://tenor.com/view/fusionanimations-godspeed-the-flash-flash-gif-22825410

[smoky oak]

Just wait till nep meets his daily quota of insulting some form of marginalized group lmao

[oak osprey]

honestly a fitting character for you to reference

[wispy spindle]

red bull gives u wings

[oak osprey]

quota. makes me think of the "bot farm in india" video that's actually a coriander testing session

[smoky oak]

Lmao

[oak osprey]

some dude walking along literally flinging pieces of coriander at dudes sitting at computers

[smoky oak]

Nep really gives off the energy of 2 11 year olds in a trenchcoat

[oak osprey]

almost none lands in their bowl

it's wild

just totally bonkers

[wispy ether]

?

[wispy ether]

?

[wispy ether]

Keep your opinion to yourself

[oak osprey]

yes, let's be constructive

was not possible to generate image of 2 kids in trenchcoat tho?

[severe swallow]

How about generate 2 kids in postal dude and hatred guy clothes

[wispy ether]

i only get 2 staying next to each

[wispy ether]

[severe swallow]

Sadly it's nothing like these dudes

[fossil spire]

Can anyone explain why my generation, on the right, is more anime styled? I used the exact same settings as the image on the left. I suspected a different variant of the same image but not the style to change entirely.

[wispy ether]

bruh

[wispy ether]

how are we supposed to know

[wispy spindle]

nep, play nice

[wispy ether]

what

[wispy spindle]

not too sure on this, but sharing your settings might help, fellow miku enjoyer

[wispy ether]

https://civitai.com/models/97799/mike-ehrmantraut-sd21-768

finger!

finga

finguh

[fossil spire]

Just started all of this today. Which settings would you like? And I had to google miku lol. I just picked an image that looked nice 😂

[wispy ether]

em

[oak osprey]

any of you have kids with messy rooms?

[tawny forge]

gtx 1650 😎

thanks open source community 🙏

[dense tapir]

\

[oak osprey]

you gotta crack some omlettes to make some eggs

[oak osprey]

ahhhh she's on her own now

2.0 with the fuckin killer fine-details on point

it's the large details we tryina fix

[finite dagger]

What was your prompt for this?

[dense tapir]

Let me check

Underworld growing plants, upside downgrowing plants

no space it was a typo

[cunning geode]

this is a pretty amazingly unique superhero outfit
4k upscale using controlnet-tile and ultimate SD upscale (here it worked, as opposed to my earlier attempts with it for a piece of digital art)

[weak sage]

SD ultimate upscale script crashes if there is no negative prompt

[oak osprey]

whoa, that's Maya Hawke

Ethan Hawke's daughter, from Stranger Things i think

could she battle this robot?

[weak sage]

just updated controlnet

[cunning geode]

yes i included maya hawke in my dataset, only 7 images but thatw as enough

[weak sage]

yeah its working now, SD is one of those things, if you update one thing but not others it breaks

[oak osprey]

i started some training on my 4090 Just for Fun and think i forgot to look at the training params carefully and what the fuck is happening

[sterile temple]

wassup?

[oak osprey]

sup fam

2.0

google street view 1969

[dense tapir]

10 17 12

[oak osprey]

imagine how this mars rover wouldn't have ever gotten stuck unlike the dumb one those dinguses at NASA built

[oak osprey]

i dont get why 2.0 can do small faces but 2.1 cant

[sterile temple]

when i finally get my new computer setup, I think I might make a lora using images from this prompt grumpy/happy old bones full of <random colour> dust

it has this soft chalky aesthetic i'm really digging. plus the characters are hilarious

[oak osprey]

@wispy nest it's pretty frustrating having the carrot dangled

[sacred fulcrum]

Soon.

[cunning geode]

now this is epic

[smoky oak]

@dense tapir Ok, I knew the 4060ti was bad, but oh my god is the 4060 just a pathetic joke of a GPU lmao

[terse sleet]

Does anyone know what would cause an image to have these sort of horizonal line patterns running across it?

[smoky oak]

lets see it

[terse sleet]

This is the image in question:

[smoky oak]

Like in the background?

[terse sleet]

It sort of doesn't show up sometimes, depends on zoom level

[terse sleet]

Yeah

I forgot what the name was, but I think it's similar to that effect that happens when you take a photo of a screen

[smoky oak]

oh what in the hell, something is very off with that image

its like... pixel filtered with icons or something

[terse sleet]

Exactly

[smoky oak]

That is so weird, I have never seen that

huh...

[cyan snow]

man

[terse sleet]

I generated multiple ones on same model and with same prompt, and it didn't happen again. Just that one image

[smoky oak]

Im sorry, I genuinely have no clue

[terse sleet]

All good, just interesting to see that

[smoky oak]

It looks like maybe some form of corruption

[cyan snow]

man, this kandinky thing can sometimes be so impressive

[smoky oak]

@terse sleetOk yeah, so the image saved at a way higher res than it should have, it like subsampled all of the pixels into icons

that is so weird

[terse sleet]

I see

I didn't have any hi-res keywords or anything like that btw

[smoky oak]

Also, I do see the stripes you were talking about as you zoom in/out, and yeah, its the exact same effect as when you take a picture of a screen

its called Moire

[terse sleet]

Yes that's it

I had to deal with it earlier when I was taking pictures of a screen

[smoky oak]

yeah, you can see it here

[terse sleet]

Ended up running stuff through gaussian blur

[smoky oak]

extremely weird... if you find out what caused it, please let me know cause I have never seen anything like that before

so bizarre

[lime lotus]

Is that a lora?

[terse sleet]

A what sorry?

[lime lotus]

If you use the same seed and same prompt with a similar checkpoint does it do the same thing?

[lime lotus]

A lora is how you get specific characters or styles. Since you didnt know that its not the issue.

What model is it? Just curious

[terse sleet]

Anything Diffusion

Plus codeformers and animesharpx4 post processing

[terse sleet]

Sorry, I don't know much about Stable Diffusion, since Im completely new

So not sure how I would get the seed

[lime lotus]

You hit the recycle icon to see the seed of the last image you generated. You can also drag and drop the file into the PNG info tab and that will give you the seed and generation prompt

I would try using the same seed and the same generation prompt and try a different upscaler

[terse sleet]

@lime lotus, I see, thanks for the info

[lime lotus]

Np bud

[cyan snow]

nah, I can't outdo this with Kandinsky, the SDXL on clipdrop was fixed

[smoky oak]

Yeah, it does seem like they improved the one on there, but also downgraded the ones in this server unfortunately

[cyan snow]

I'll take that anyday, I was making a comparison earlier today with Kandinsky2.1 and the SDXL bot, Kandinsky won, but lost to the one on clipdrop

[smoky oak]

weird, the one on clipdrop shouldn't be any better than the ones on here, worse, actually

[cyan snow]

and when XL will go public it will be even better, so yeah, looks like SAI is going to win this round

[smoky oak]

Not sure what any of the benefits of kandinsky are at all ll

[cyan snow]

you can scroll up to a few hours ago and see, it follows prompts better than the SDXL bot, but the SDXL on clipdrop defiantly beats it

[smoky oak]

oh god kandinsky looks terrible

[cyan snow]

maybe, but it followed the prompt pretty nicely

[smoky oak]

I could care less with how bad it looks-

[cyan snow]

but yeah, this is way better than what i made using kandinsky

[smoky oak]

I don't doubt it, kandinsky looks like base 1.5 lmao

[cyan snow]

nah, it's similar to base 1.5, except it follows prompts in a different way

(this is what i made with kandinsky)

[smoky oak]

Maybe its just trash at making people, IDK

[cyan snow]

well, I think the one on clipdrop did a better job with this prompt, but it is a close comparison.

[sterile temple]

I just think that particular prompt is working differently between the two models. You can prompt dog made of wood or dog made of water fine, it's just dog made of fire is not working. If you try flaming dog or something different it might work better

[cyan snow]

no, no, the one on clipdrop did this perfectly, even better than kandinsky, is it 0.9?

[sterile temple]

it's supposed to be

[cyan snow]

also, the model on clipdrop improved by a lot, I don't know what you guys changed, but the model on clipdrop is as good as SDXL0.9 from the presentation.

[rigid heron]

im new here and only just heard of sdxl what is clipdrop

[cyan snow]

where you can access SDXL0.9 before it's open sourced

[rigid heron]

oh

[rigid heron]

is that not the bot channels

[sterile temple]

that's SDXL 0.9 with randomised config for them to test the outputs

[cyan snow]

also the bot, but the clipdrop is WAY better in my opinion

[cyan snow]

I thought it's also past epochs of SDXL training?

[rigid heron]

oh yeah that makes more sense than just using resources for it

smart way to train it too

[sterile temple]

yeah maybe that too, make sure the new checkpoint is not worse than the old ones

they need a thumbs down button if both images are bad

[cyan snow]

yeah, XL is a masterpiece, even before it's finished

[sterile temple]

do you know the prompt for that? I could test the API output

[cyan snow]

yeah, ''dog made of green fire, concept art,''

[sterile temple]

of course I won't have the extra prompts they add to clipdrop

API XL

[cyan snow]

i didn't use any style, it shouldn't matter

[cyan snow]

huh, is that the latest epoch?

The SDXL on clipdrop is definitely better than Kandinsky

[rigid heron]

So does sdxl just have access to the entire Internet or something

[languid pagoda]

I think I successfully have done my first model training, it ofcourse didn't pan out well lmao but you can see some of the style I think

[rigid heron]

There are so many concepts and characters and things it just knows by name

[languid pagoda]

but right now it's have some real issues figuring out wtf is what

[cyan snow]

nope, AI models are so dense they can make pretty much anything just from a ~12GB file

[languid pagoda]

but it's uh progress

[sterile temple]

supposed to be, but it's just the positive prompt, no negative

[cyan snow]

idk then, whatever they use on clipdrop is definitely the best

[languid pagoda]

that's a neat style btw

really curious about the results i'll get from the new stable diffusion model when it comes for local use

It seems this one may be a game changer

[cyan snow]

same, so far if the SDXL1.0 will be better than the one on clipdrop, it's pretty much game over for other AI art generators

and that's just the base model, and it already beats everything else

[languid pagoda]

nah midjourney will still do well

but I love seeing the jump in quality for stable

[cyan snow]

midjorney isn't even close. even a 1.5 finetune can destroy MJ

[rigid heron]

yeah but midjourney is paid isn't it

this is free

[languid pagoda]

the thing about midjourney is that not everyone wants to run it locally or can

so it'll keep doing well

but im glad stable is doing well

[green socket]

Midjourney is easier to set up than running it locally

[languid pagoda]

I love having options and local use

yee

[rigid heron]

I don't think running midjourney locally is even an option is it

[green socket]

Nope!

[languid pagoda]

definitely not

it's unlikely it'd be even possible on consumer gpu's even if they wanted to rn

[cyan snow]

I never saw MJ get to this level of detail, SD is definately better.

[rigid heron]

also running in discord kinda limits how much you can change and finetune things

[languid pagoda]

also no custom models etc yeah

[rigid heron]

but most people who aren't already locally running it probably want to do that

[oak osprey]

channels with bot in there when there is where the bot with you the thing, the dreamer

[languid pagoda]

yeah p much, midjourney is a easy out of box solution that works everywhere and on a toaster

different crowds

[oak osprey]

idk i maintain a discord bot that runs on many servers and serves custom models

[green socket]

A lot of people only use phones, and SD can't really run on an old iphone or any android

[oak osprey]

SD runs on iphones

[green socket]

That's why I said old

[rigid heron]

yeah at the cost of your fingertips when the phone burns them off

[green socket]

5 generations burns 20% of the battery 😂

[cyan snow]

maybe, but it isn't able to make images in this kind of quality. and the model i used to make this is a 1.5 finetune, SDXL finetunes will be WAY better than 1.5 finetunes.

[rigid heron]

god damn what the fuck

[languid pagoda]

yeah idk we'll see it doesn't really matter which is better to me. The fact that stable will look amazing is all that matters to me

[oak osprey]

https://twitter.com/pcuenq/status/1669413165963300864

Pedro Cuenca (@pcuenq)

Stable Diffusion on iPhone is much faster now!

Same model, same phone (iPhone 13 Pro), same settings 🤯

The trick?

• 6-bit quantization in Core ML. Announced last week in WWDC
• Additional optimizations to the attention blocks

Check our post for details https://t.co/87Pvlrocrb

Likes

386

Twitter · Jun 15, 2023, 6:34 PM

i honestly don't care who "wins", i have all of them available to me to use, but i get triggered by incorrect assertions

[languid pagoda]

I only have and will use stable

I be poor plus customization is good

hoping i can get training to go well

so i can really make some cool shit

[green socket]

Midjourney is pretty useless for me because I can't guide it

[oak osprey]

have you tried BlueWillow? tis free

[languid pagoda]

and help with 3d modelling

stuff like textures etc

currently using dreambooth

and testing shit out

[rigid heron]

I just use sd because it's what I know

[languid pagoda]

I haven't, stable working fine for me

so i have no reason to switch especially with xl coming

[rigid heron]

Also does sdxl have a built in language model or something? Like it has the reading comprehension of a person

[cyan snow]

SDXl can do TEXT

that's how insane it is

[rigid heron]

I'm used to separating all the different aspects I want into individual prompts but with this I can write a descriptive paragraph about something and it churns out exactly that 20 seconds later

[oak osprey]

sdxl can do text because they used ViT-bigG/14 to make embeddings to train a captioner (BLIP2)

the captioner is able to essentially read images

i've made a few fine-tunes to BLIP2 that do this and can zero-shot transcribe like, "HAPPY BIRTHDAY" from a banner, "FOR SALE" signs... for some Kodachrome pictures it got "HAIGHT" for the street sign in San Francisco

i also assumed they didn't flip images during training that had a text/ocr score

[rigid heron]

Oh wait I thought this was just a weird way of saying it can read well, are you saying it can produce an actual image of text

[oak osprey]

yeah

sometimes it works, but not as frequently as deepfloyd does

once you can run it locally it's fine because you can do 100 gens

[cyan snow]

still, it's insane that we will have a model that can do the best generations and text and pretty much anything.

[oak osprey]

it ought to be good, it uses like 21gb of vram to run both models 😛

[cyan snow]

it needs 8gb, emad said it

[oak osprey]

that's just the base model

mcmonkey said it

the refiner is 6 billion parameters

[cyan snow]

yes, but what kind of person loads multiple safetensors at the same time

[oak osprey]

have fun with that lol

still uses 10gb

[cyan snow]

they said in the presentation that pretty much all the 40xx series cards can run it nicely..

[oak osprey]

the base model

have you seen the output from just the base?

[rigid heron]

I'm gonna upgrade soon to a 3060, it has 12gb hopefully that's enough

[oak osprey]

they pitted 1.5 against refiner, 1.5 against base, and also, refiner against base, and 2.1 against base/refiner

the RLHF scores put 2.1 over 1.5 and base over 2.1 and refiner over base, but they don't make the chart available to show how much difference it is

[cyan snow]

idk man. whatever they are using on clipdrop might be the best model yet, and it's not even the finished SDXL1.0

we'll see in half a month

[oak osprey]

i got bored of how smooth everything looks and then i read the research paper they referenced and i get why that's the case now

[rigid heron]

I genuinely can't resist taking whatever prompt I'm testing in the bot channel and just making it a robot

this model makes robots so well

[oak osprey]

2.1-v finetune does as well

[rigid heron]

but look at this shit

[oak osprey]

looks like Halo

[rigid heron]

Its completely symmetrical too i don't even know how it did that

[cyan snow]

SDXL is likely to be the best diffusion model, even before we start finetuning it

we'll see in about 2 weeks

[rigid heron]

it's that soon??

[cyan snow]

like, this is better than what kandinsky makes on max settings

[rigid heron]

I don't think this even needs finetuning it can just do everything

[oak osprey]

you'll see 😛

[cyan snow]

i sometimes can get to it's level of detail, but that's after finetuning the shit out of 1.5, just wait and see, SDXL finetunes are going to make all other AI art look like wish.com

[oak osprey]

[rigid heron]

I don't even know what separates this from normal stuff

[wispy ether]

Kaspersky

[sterile temple]

There is no option for style in the A1111 API extension, so I manually added the 'enhance' style to the style. Now I think that looks a bit closer to what you got from clipdrop

[cyan snow]

I get it, so you aren't running XL locally, it's the API, isn't it

[sterile temple]

yeah

another one. i think adding 'enhance' style to the request is making a big difference

[cyan snow]

Yeah, if what clipdrop made was better than what Kandinsky made, this is like stepping on a dead bug

[sterile temple]

not sure how these styles work with the checkpoint, maybe they're a lora?

hopefully whatever it is, they release it with the checkpoint file

[cyan snow]

I already asked that, we will find out when it releases

Hopefully, if everything goes according to plan, the SDXL fine-tunes will be the best diffusion models yet.

[languid pagoda]

flat color model progress

[smoky oak]

Jesus Kandinsky sucks at humans lmao

uh, weird way to copy that lol

[languid pagoda]

yeah that's a decision alright

lol

[smoky oak]

she looks awful lol

[languid pagoda]

I just realized I included no human references

so animals look way better huh

this is actually.. a pretty good result I think yay

[rigid heron]

You know flat colors is something you can just prompt for right?

[languid pagoda]

Yeah I mean ultimately you can do pretty much most things with prompts

But don't worry this was just a test, I intend to do some very specific things in the future to combine with 3d modelling I think

[oak osprey]

kinnyeu reaves

i tried to make him shaggy from scooby doop

[languid pagoda]

oh no lol

[rigid heron]

That's given me an idea

[oak osprey]

well be sure to give me credit

[rigid heron]

I'm gonna do shaggy from scooby doo as William afton

[rigid heron]

I mean I don't post this stuff anywhere

I just make something, think "that's cool," and then make another thing

[oak osprey]

the prompt segmentation stuff is so weird lmfao

the prompt segmentation tends to make the output contrast kind of weird, so, ignore that, it's mostly just an experiment but that on the left is:

("stunning portrait of keanu reeves", "as scooby doo shaggy part of the scooby doo cast, 8K").and(0.85, 1.0)

and the right:

("stunning portrait of keanu reeves", "as scooby doo shaggy part of the scooby doo cast, 8K").and(0.70, 1.0)

somewhere in between 0.7 and 0.85 is a pretty good middle-ground

dear lord

fuck, it's so bad

willem dafoe supposedly

[rigid heron]

The shaggy as William Afton thing didn't work

Even though they're the same actor

[weak sage]

revitalized an oblivion screenshot

[languid pagoda]

that's pretty neat

[oak osprey]

SDXL's black level maximum sucks vs mine

git gud, noobs

more often you'll get something like this from SDXL or worse, straight-up artifacts and trash remaining in the image

[rigid heron]

why does this man hate sdxl so much

[oak osprey]

just disappointing when they leave a lot of work up to the community

[rigid heron]

what

[oak osprey]

i've been fixing 2.x for months

have you ever used it?

[rigid heron]

sdxl is incredible what

[green socket]

But think about the sense of pride and accomplishment you'll have when you fix it!

[oak osprey]

i will probably not have that still i am broken

fixed 2.1 and then just set out fixing 2.0 and realised 2.1 was cooked to death by stability

[rigid heron]

ok so I asked for spider-noir which it didn't recognise but it did give me the coolest fucking version of sam raimi spiderman?

[oak osprey]

there's now a fixed base to fine-tune from for 2.1-v but it wasn't easy

just curious, did you ask for oil paintings? i find they have very high aesthetic scores in laion datasets

or.. i guess that's watercolor?

[rigid heron]

No I asked for an impressionistic painting, as well as in the style of disco elysium

[oak osprey]

ah

you can use prompt segmentation too

("one thing", "another", "maybe a third").and(weight, weight, weight) where weight is a float value

[rigid heron]

I don't know what that is and it's probably not a feature in discord sdxl bots

[oak osprey]

it changes how the backend interprets different pieces of the prompt, and helps it listen better

it is

they added it with (attention emphasis)1.4 in general

[rigid heron]

I don't really know what you mean by float value so I don't know what this actually does

[oak osprey]

1.0 is a float value and 1 is an int, "1" would be a string

it allows you to change the attention for whole chunks of a prompt, and ties the contextual embeddings to that prompt piece

[rigid heron]

so basically "this is one thing and what it looks like, and this is another thing and what it looks like"

Or no

[oak osprey]

the way openclip works, it basically takes your text, and finds a ... well, a bunch of numbers, that end up somehow describing the relations between different "features" that appeared in images that were seen with tokens/similar tokens to the ones you present.

example: a person has a face, a face has eyes, a mouth, etc. these are all in that embedding space, in a sense, and the unet is trained to guide diffusion based on these embeddings

this prompt segmentation allows you to break that process up into distinct prompt chunks

so the context doesn't get lost or ... what's the word.. uh.. bothered

so for example zuckerberg is really heavily trained in and you can't make him necessarily show up as Belle. but if you do ("mark zuckerberg", "as belle").and(0.9, 1.0), it worked

i'm not terribly certain how it differs from the (bracketing terms) stuff, as i haven't really looked at that closely

[rigid heron]

As belle?

[weak sage]

Yours is.... Darker Than Black

[oak osprey]

Bell Daphine or whatever her name is

[oak osprey]

it turns the pixels off in an oled display

[rigid heron]

Oh so itd be like ("character", "as character").and(0.9, 1.0) and itd be the first character dressed as the second character

[weak sage]

oh yeah i ahve a lora for her

[oak osprey]

hopefully! yeah. and it's surprising how much change it has between 0.9 and 0.8

like, there's an infinite world of space between those two values. 0.91, 0.911, 0.9111, etc

we can only represent 16 bits of it, but good enough

[rigid heron]

I will try it out

(And probably do it wrong)

[oak osprey]

it's actually kind of hard to get it correct, i have a bit of code in my discord bot that retrieves prompts from GPT and then creates them in that format, and sometimes i just run it overnight and get a buuunch of good ones

[rigid heron]

run it overnight?

bruh my house would burn down

[oak osprey]

("cinematic shiny candygore", "a film still of a twisted carnival", "vivid colors", "gory details", "retro aesthetic").and()

loved this one

[rigid heron]

It may have been foolish to test this prompt segmenting thing with a character it probably doesn't know

yeah no it had no idea who gimli was

Shit it worked

("Jack black", "as santa claus").and(0.9, 1.0)

[oak osprey]

told you 😄

[weak sage]

sometimes you just know theyre not going to be nice about it...

[oak osprey]

so my model can generate fantasy stuff.. hmm, interesting

@weak sage can you match that or better? curious if you can come up with a better prompt because i think this one is limited in scope/grandeur

[weak sage]

what prompt, and what are you going for

[oak osprey]

like a dark mardi gras jester

stunning++ photographs of jesters+ at the twisted carnival-

ah holy shit ok lol i'm starting to get very different stuff, i was on sequential seeds and now am on random

i really love the ones on the left side

not unhappy with the right side but i did want jesters

wow i like it, this prompt ❤️‍🔥

@sterile temple ^

[sterile temple]

nice. is that the 2.0 model?

[oak osprey]

2.1, the published flex-base

[weak sage]

oh i should go download that

[oak osprey]

i don't have the 2.0 one packaged up for playing like this just yet, it has too many anomalies

in a couple weeks it might be 🔥

i really need to sit down and refactor my datasets again. i have more options now. i just need to clean up and do it

i can run one of these through the 2.0 though because i'm super curious

[oak osprey]

it is so dark 🕶️

[weak sage]

probably noise offset lora

[sterile temple]

5 steps?

[weak sage]

i start out with 4 or 5 to test before i commit

[sterile temple]

SDXL

[weak sage]

skull on the right supposed to be purple like that or is it an artifact thingy?

[sterile temple]

i guess it's mixing the balloon and skull?

I used the same prompt you had

[weak sage]

12 steps, not much difference

[oak osprey]

for a sec i'd thought you'd loaded up my model

[oak osprey]

oh man 2.0 really is a different model from 2.1

it looks like it's learning...

i'm excited i can train this for thousands of steps though

the works is my oyster or whatever

[sterile temple]

here i am thinking i'm sending the style to the api, but it's style_preset and not style so probably wasn't doing anything this whole time

[oak osprey]

i did that with my guidance scaling in my discord bot

i thought i was using different cfgs and they were all around 10.0++

fucking typos

[weak sage]

ive never used styles, am i missing anything

[sterile temple]

it's a thing with the api and the bot

must be a TI or lora or something

people have mentioned this magical 'enhance'. i noticed it's one of the styles in the api documentation

there we go, left pepe before fixing the api call, and right with 'style_preset': 'enhance'

[sterile temple]

what I thought was enhance, and actual enhance 😄

[oak osprey]

@sterile temple the decider

[weak sage]

how do i add enhance magic to my images then

[sterile temple]

how are you using sdxl?

[weak sage]

im just in automatic11111

[sterile temple]

did you install the stable ai api extension?

[weak sage]

whats that hmm

[oak osprey]

well, im out. lata

[sterile temple]

ciao bella

photographic style

[weak sage]

doesnt mean well stop posting stupid jester pics

[weak sage]

barghest?

[sterile temple]

okay working with styles greatly improves the output of sdxl

[weak sage]

looks more real than reality

[smoky oak]

It's getting annoying how many people keep "correcting me" when I say there is a second version of SDXL that has a second model attached, saying I'm lying, and that there is only one model

[rigid heron]

why are these so many of this green fire dog

[weak sage]

idk someone post it and i thought cool so i imitate

there are barghests in the witcher 1, common enemy at start of the game

[sterile temple]

I'm getting MUCH better outputs using a 'style' whatever they are

[smoky oak]

it's just preset words for the prompt

[cyan snow]

i think it has something to do with the refiner

after it will release i will have a better idea of this godly architecture

[sterile temple]

I tried lots of prompts without a style and the output was meh

[cyan snow]

the SDXL on clipdrop makes possibly the best stuff without a style

[smoky oak]

As far as I was told, there is no refiner in SDXL right now

thats why the quality of the results has dropped so considerably

[cyan snow]

idk, what ever they did on clipdrop, it beats Kandinsky, MJ and most 1.5 finetunes

[sterile temple]

I am thinking there is a default 'enhance' style they use on clipdrop

[smoky oak]

@oak osprey got more info on all this for me?

[cyan snow]

idk, whatever it is, it's magic

[smoky oak]

All I remember was that when we all saw a drop in the quality on this server with SDXL, it was the same time they said they removed the refiner from SDXL

[cyan snow]

idk man, what ever is being used on clipdrop, it's actually insane

it's WAY better than it was 2 weeks ago

[smoky oak]

It looks about as good as the bot in this server before they nuked it

so now you have to go to their crappy site in order to use the betetr version where you are trapped behind huge wait times to try and upsell you. Kinda scummy IMO, but I already hate clipdrop in general for how they handle their sketchy claims

also, the upscaler on clipdrop causes more issues than it fixes, which is unfortunate

[cyan snow]

it does detailing way better, i remember me and another guy made a comparison of Kandinsky and SDXL, kandinsky won against the SDXL bot, but the SDXL on clipdrop destroyed it

[smoky oak]

yeah but like, which of the dozens of SDXL models were you guys testing against? Cause I have gens from the SDXL bot in this server that are still better than clipdrop

[sterile temple]

I just modified the code for the api extension for A1111 slightly so I can prompt for styles and it's made a HUGE difference

[rigid heron]

robobitch

[smoky oak]

so wait, does that API extension use like dreamstudio credits?

[sterile temple]

yes

[smoky oak]

ah, nevermind then. Was gonna use it

[sterile temple]

it's about 1.5 cents per image

[smoky oak]

since when is it WAY cheaper?

it used to be like 20 cents an image

it was absurdly expensive lmao

[weak sage]

insert coin to continue

[sterile temple]

20 dollars I got 2000 tokens, a 1024x1024 image with 30 steps costs 1.6 tokens

[smoky oak]

wow, they massively reduced their prices

when SDXL beta first hit, it was like 8 credits for 30 steps of 512x512

it was less than half the price on the old site, then they removed access to it and hiked prices ._.

[smoky oak]

I wish we had a way to do direct comparisons between clipdrop and the good SDXL bot in this server, but they removed it

[rigid heron]

I'm starting to think I should just use this clip drop thing

[smoky oak]

All I DO have to add is kandinsky is yikes IMO lol

[smoky oak]

they seem to have made it the only way to get better SDXL images now

I did like 100 gens in this server, and the current model on clipdrop just doesn't seem as good IMO

[weak sage]

why use their server instead of your gpu

[sterile temple]

I'm pretty happy with the API now. before I was like 'is this it'?

[rigid heron]

I mean the models just not out yet

[smoky oak]

Those were all from the bot in this server when it was good

[rigid heron]

Why would they make their own product worse

[weak sage]

oh so youre beta testing 👍

[smoky oak]

to try and sell you on their subscription on clip drop

They ARE a company after all lol

[rigid heron]

Oh it costs money

[smoky oak]

it does, if you don't wanna wait forever

[cyan snow]

it doesn't =/

[smoky oak]

its free if you wanna wait for it (it can take up to like 10 minutes an image), which is still way better than MJ to be fair haha

[weak sage]

i prefer not having some corporation decide what i can gen

[cyan snow]

or just wait 2 weeks =]

[smoky oak]

Same here, but unfortunately we don't have access to it yet

[smoky oak]

Assuming they hit that deadline, but I am actualyl very hopeful of that

just curious how long its gonna take after release for us to be able to use it in the UI's we all run haha

[rigid heron]

oh wow yeah this clipdrop is much better

[cyan snow]

shouldn't be that hard, AUTOMATIC1111 will be forced to do it in the same day it releases

[weak sage]

theyll rebuild a proprietary version of our UI's with DRM,i guess

[smoky oak]

ok, you seem a little too worked up over corporation greed lol

[smoky oak]

I just have no idea what that looks like yet, its kinda concerning IMO

oh shit, they increased the speed of their clipdrop servers massively

ok then, my bad

[cyan snow]

the UI? it's just a webUI with a ton of features

[smoky oak]

wow, from 400 queue to generated in like 30 seconds

DAMN IT

I DID NOT MEAN TO ADD THAT LMFAO

kill me lmao

[rigid heron]

Yeah this clipdrop doesn't seem all that slow to me

[smoky oak]

no, I will admit, its WAY faster now

like a week ago, 300 queue took like 10 minutes

[rigid heron]

And I'm getting shit like this

[cyan snow]

wait, you aren't using the A1111 webui?

[split rover]

good quality on the bot will be back and likely is still possible now but ofc there is lots of variation internally, we are actively testing some very weird stuff right now for 1.0. We didnt nerf the bot to get people to go to clipdrop, clipdrop we passed our best found settings and they run it by default, we still run experiments that change daily here on the bot in contribution to 1.0. It can be a mix of model variants, using the piped refiner or not, using really nonsense inference settings, etc. It honestly changes a few times a day typically so no guarantee current variations stick around long. No specific nerfing, just general info gathering on how various things perform at scale to help make better models

[smoky oak]

yeah I am, but I mean like... They will have to implement support for SDXL, and we are not sure what that looks like yet

[smoky oak]

Thats fair in that case, I just know there was a lot of backlash when people found out we would not be getting teh quality of the refiner model from the SDXL bots, and I was assuming it was removed to try and manage expectations

[rigid heron]

I mean fucking around is the best way to find out

[smoky oak]

also, IDK if there is something wrong with my clipdrop, but these results look just as bad as they did 2 weeks ago... hmm...

[cyan snow]

I don't get it, how SDXL will look like? we already have a good idea of that. it will probably just be a massive safetensor of ckpt

[smoky oak]

It could be that way, or it could be very different, no way to know at the moment

[cyan snow]

it was confirmed to be a huge safetensor of something

[smoky oak]

They said that was the end goal

[cyan snow]

if that's what you're asking

[split rover]

Nah definitely not the case, there are active/non-active refiners on the bot right now. We are working to try and match it with a single model but ofc splitting workloads across specialized models typically leads to better results and we also want to make the prettiest stuff possible. Lots of work going into the next version (1.0) right now though to improve it further

[smoky oak]

Hmm... alright then in that case, I feel like so much info about all of this is just kinda he said she said, from the devs to other staff and stuff

regardless, I suppose we will see in the end anyways

[sterile temple]

is 'enhance' the refiner model in the style_presets? or using any of the styles employs a refiner?

[smoky oak]

also, side note, its insane how different the visual styles of Photographic on the SDXL bot and clipdrop are

I am assuming they are prompted drastically different

cause on clipdrop they always look super washed out and flat

[split rover]

There are a lot of staff who work around it on other parts and then there are a few of us directly who work on it daily so I am sure stuff has been telephoned a bit. I am also sure there are some other crazy rumors floating out there too by now haha

[smoky oak]

out of curiosity, one of the devs said that 12GB should be able to theoretically generate SDXL images of any resolution, any info or response to that?

Seems like a huge claim IMO

[split rover]

That I am not sure 🤷 its possible they modified it on their own or just that we do some wacky things on the bot atm

[smoky oak]

they also said that 2048x2048 should work on 8GBVRAM

everything just feels so messy 😅

[cyan snow]

bro, just wait 2 weeks, we will all find out

[smoky oak]

Biggest thing I WILL say right now, is that clipdrop is massively faster than it was about a week ago, so mega props for that!

[split rover]

cant say for certain on that, im not the optimization/inference expert haha. Id prob ask Comfy if you see them around, they are the optimizer wizard. I just hurt gpus to make pretty images

[rigid heron]

12gb will generally be good enough though right?

That's what I'm upgrading to soon

[smoky oak]

Again, that date isn't guaranteed either, its all just kinda telephone

I remember when 99.9% of the SD reddit was SWEARING SDXL was coming out that one Friday lmao

[cyan snow]

Emad said it in the show and tell

[smoky oak]

there have been several statements in documentation and from SAI themseleves that have said that 8GB on an RTX GPU should be good, and 16GB on an AMD GPU should be fine

[smoky oak]

||You say that likes hes reliable with what he says 😅 ||

I have seen several things saying they wanna "try to aim for middle of July"

[rigid heron]

Never tried isometric before this shit good

[smoky oak]

in the end, I would much prefer to have to wait longer to really get something thats properly concieved, personally

Last thing we need is another 2.x fuck up haha

[cyan snow]

but man, this is a good model, Kandinsky, MJ and even 1.5 finetunes can't compete with this model

[smoky oak]

on a single image, sure

I still really think you should have tried that extremely good SDXL bot model that was in circulation for a few days, it was insane

[rigid heron]

Speaking of bots I need to make more robots

[smoky oak]

bobot

[rigid heron]

I swear it's all I do with this thing

[smoky oak]

oh thank you! you gave me a gen idea :>

gonna try something no model I have ever used could do lol

[weak sage]

star ocean 2 fan?

[smoky oak]

nope, just toddler english enjoyer lol

[cyan snow]

the SDXL model that will release will be the best version before it gets finetuned

[rigid heron]

Damn, the model didn't recognise what I meant by freddy fazbears pizzeria

[weak sage]

is it gonna be compatible with all our loras tho

[smoky oak]

not necessarily, again, there was staff in here saying that the "best" version of SDXL has a massive like 6B parameter refiner ontop, and that version was not capable of running on consumer available cards

And then they said their current goal is to try and fintune base SDXL to be as good as it is with the refiner

(the images I sent above were from presumably the full refiner model, cause god they looked so much better than what we even have now)

[smoky oak]

nope

gonna need to make new ones

but, SAI did say that they are releasing finetuning tools

which is MASSIVE news

[cyan snow]

we will find out in 2 weeks, this conversation is useless until the model finally releases

[smoky oak]

I really do feel like I am using a different SDXL bot to you guys, cause man these results I am getting from clipdrop are off

[rigid heron]

beefy ass bobot

[smoky oak]

See, I have been trying to generate muscular and attractive male cyborgs 😅

[sterile temple]

what's the prompt? I'll try it out with the API

[rigid heron]

I think the main part was "huge muscular robot made entirely out of thick black cables and wiring, white metal plating"

[smoky oak]

gonna have to try that in some of my 1.5 models

[sterile temple]

guess which one doesn't have a style 😄

[rigid heron]

I'm curious I'm gonna see if I can make a robot out of gravel

I could not

[smoky oak]

@sterile temple I'm not sure what you are doing to get these results, but I can't seem to get any results like these out of the server, or clip drop

[rigid heron]

What kind of results are you gettinf

Send an image

[smoky oak]

Just a sec

[rigid heron]

also this dude funky as hell

[smoky oak]

results like these

no real musculature in them

[rigid heron]

What's your prompt

[smoky oak]

"huge muscular robot made entirely out of thick black cables and wiring, white metal plating"

the one you suggested haha

I tried several different ones

[rigid heron]

Oh I think I forgot stuff

Add humanoid in there

Should help

[smoky oak]

YEESH

[rigid heron]

Yeah I'm getting that too

It's just luck

It'll still only be a minute or 2 this shit is fast

[smoky oak]

before bro, 300 was like 10+ minute wait lmao

they really upgraded their servers

thats considerably better!

Thanks!

[rigid heron]

Now I'm getting queues of 3 images

[smoky oak]

Training LoRA's genuinely has me so mad, cause I don't know what broke in my installs

Went from being very good at LoRA's to not being able to get a LoRA for anything to work with literally 0 changes to my kohya install

[cyan snow]

this is trippy

if you look closely you can kinda see windows HUD

that's so weird

[smoky oak]

ha! you kinda can lmao

[dapper cape]

my shit gpu using 500% of its power to gen images

[cyan snow]

Bro has mother board in his laptop monitor

[dapper cape]

gaming laptops ads be like

[cyan snow]

it made the alien windows HUD again

wait, is that an apple logo?

[dapper cape]

where

[cyan snow]

in the left of the other obstruction of a computer

that's an apple logo

[dapper cape]

oh

[rigid heron]

man Oscar isaac makes everything cooler

he's just one of those cool looking guys

[cyan snow]

did i have a stroke

[smoky oak]

You genuinely just made me snort

I don't think I have ever, EVER snorted before lmao

[cyan snow]

these are AI generated memes

this is what aliens see when they see memes

BRO

WHY IS THIS HAPPENING

i just prompted ''meme''

AAAAAAAAAAA

I'm going to hell

[cyan snow]

oh, did you get the joke? BECAUSE I DIDN'T

[smoky oak]

I was so dumb founded that you violently attacked my haha receptors

[cyan snow]

man these AI memes are #relatable

why does this look like my mother

IT JUST MADE FOOD WTF

this is what clipdrop makes for memes

this is like the Kandinsky architecture diagram i swear

[sterile temple]

this is what I got from the API for meme

[cyan snow]

i get it HAHAHA

this is comedy

EW WTF

these are better than most memes on tiktok that the fucking children make

[rigid heron]

Cybiden

[severe swallow]

Just for master and margarita enjoyers who also a helltaker fans

[primal lance]

How to open negative promps???

[wispy ether]

Nobody knows what website that is.

[primal lance]

Stable diffusion

[cyan snow]

y'all like pizza?

[sterile temple]

pizza pizza pizza!

https://tenor.com/view/pizza-jerri-gif-13104200

[dense tapir]

@smoky oak I was informed of this on another discord "ROCm 5.6 just release and the docs don’t look like they support the 7000 series tho". WTH?!? They were supposed to and they even did a PR for all the 7k cards to put them in.

[dense tapir]

Told ya. really both are 4050s.

[dense tapir]

@smoky oak What gets me is how Jensen can actually put this stuff out knowing just how bad they are? Does he have no self respect left?

[wispy ether]

4070:

[oak osprey]

in my opinion its already overtrained lol

it was good to go like a month ago

lots of overfitting but wow so smooth

its been training since february

[oak osprey]

dr evil?

[sterile temple]

yessir

wouldn't let me prompt fat bastard

API slowing down, just in time for bed 🥱

[oak osprey]

oh god

i changed the prompt up a bit to ask for mardi gras monkey jesters

im going to like actually die from these being so damn funny

luchadors at the juggalo convention

[fallen dock]

https://tenor.com/view/hmm-suspect-gif-22611582

picture on the lower left bottom be like:

[worldly flume]

https://creator.nightcafe.studio/creation/vz2Sp3iQLPorTKWoQeA5

[oak osprey]

thats my favourite of that batch tho

[worldly flume]

what do you folks think...

[fallen dock]

.....

[stoic pivot]

Pizza?

Asked for 'a painting of a feminine sheep character holding a banana pizza', got a nun playing a pizza lute.

[split valve]

Its now making amazing images.

The styles things on clipdrop is working amazing

same prompt different styles.

Erotic Cakes cinematic artist concept. Style of Salvador Dali and Roberto Bernardi

[oak osprey]

nice, it's like open-source midjourney

v5.1

[wispy nest]

you using the api for xl on auto for these?

[oak osprey]

that's SD 2.0 that i fine-tuned

you wannnt?

[wispy nest]

seems like Clipdrop is giving the best XL stuffs

[oak osprey]

agreed

[wispy nest]

nah Im good but thanks Im onto the future only lol

[oak osprey]

the bot is the horrifying look at what is to come

they're trying to distil the model

[wispy nest]

do they own Clipdrop or something

[oak osprey]

yes lol

[wispy nest]

ah that makes sense then

[oak osprey]

chaz

[wispy nest]

was wondering why the Dreamstudio site is like an after thought to them

[oak osprey]

that honestly seems to be a different team

[wispy nest]

not that they dont obviously own that but they dont seem too concerned

[oak osprey]

it's just all kind of left hand vs right hand, i think it'll all be ironed out soon enough

[split valve]

It cant do hands, but that styles thing is ahead of Midjourney.

[oak osprey]

it CAN do hands and quite reliably but not if you ask it to

[wispy nest]

yeah its just confusing sometimes did a person use clipdrop, the api or whatever, but like you said itll all get ironed out soon... i hope

[oak osprey]

well, yes and no

[split valve]

its improved massively within a few weeks

[oak osprey]

i mean that clipdrop and dreamstudio will equalize

[wispy nest]

but it seems like you shouldnt burn your credits on dreamstudio or the api then unless you want inferior images atm

[split valve]

Analog film style

[oak osprey]

but for general images it's only going to get more confusing once 1.0 drops and we have fine-tunes

[wispy nest]

sure

looking forward to that kind of confusion though lol

[oak osprey]

you're not going to have any idea whether something is base, refiner, fine-tuned base, etc

oh for sure as long as it looks good, it's all good

but like, why didn't you play with DeepFloyd

[wispy nest]

its just more as a humble consumer who bought credits on their dream studio, when can I get images there like clipdrop and they never give a real answer

but Clipdrop works just as well I guess. the results are dope

[split valve]

its coming to dream studio.

[wispy nest]

yeah but when

they said Monday

then kinda Tuesday

[oak osprey]

my model uses the new noise schedule from Bytedance and is fairly groundbreaking in that no other base fine-tune has been done on that model's scale yet 🥹

[split valve]

Probably when its v1.0

[wispy nest]

its coming before 1.0 according to people here but at this point id not bet against it being 1.0 either i guess lol

[wispy nest]

love this one. clipdrop?

I wonder what diffuser clipdrop uses

or sampler i mean

[west tree]

Nope. Just stock A11.

Not my best, but trying to play with new ideas.

This is my best

[weak sage]

10/10 would eat, who the fuck sliced it so thinly tho

[west tree]

Can somebody help with hands? I'm trying to use adetailer without success. My shit is mad deformed.

I'm trying to get good anatomy down pat. It's not something I've worked on, before.

[oak osprey]

apparently i've made 540 images with SDXL

probably not as many as some have

[west tree]

Does SDXL prompt the same as 1.5?

[wispy nest]

Ive probably made 50-100

[wispy nest]

XL should be powerful as hell though. Cant wait to see 1.0 fine tuned

[dense tapir]

You realize SDXL is just to satiate the masses for 3.0 then that is it

3.0 is supposed to bring us back around and be the end all of be all

[oak osprey]

no

it is different from 1.5 and 2.x and maybe is more like DeepFloyd, which is similarly powerful and thus difficult to prompt in a precise way

[west tree]

But is the syntax the same?

[dense tapir]

I sure hope not

[oak osprey]

you'll get a lot of good stuff but until we can really mess with the prompting locally with stuff like Compel and a1111's prompt code, we don't know how to really access its deeper abilities

the prompt syntax is based on your image gen software and you probably don't want to use the same prompts with the same terms and definitely not with the same weights applied

[cyan snow]

huh, 1.5 knows Splatoon

[weak sage]

the kind of work that would have taken two whole months...

[oak osprey]

is that Euler?

[weak sage]

i used the broken sampler 😄

[west tree]

Scrotum monster.

[oak osprey]

i cant stop looking at this, someone accidentally prompted my bot ill

just that

another accidental gen

[lilac pilot]

Not sure what is causing the added white look and what you call it... latent noise in hair and upper right painting?
Image
I was using the inpaint+llama for outfit varients.
Trying to get o- ren ishii from kill bill and Kang Sae-byeok squid game . might have to make a textural inversion for the clothes any suggestions

[oak osprey]

Real-ESRGAN is kind of bad lol

[oak osprey]

premtp?

what the hell

[west tree]

Please do not reuse this, except for personal use. I should've thought before posting.

[oak osprey]

it doesn't work in my model anyway

either the prompt syntax is bad or the prompt is lol

[west tree]

Are you hosted or local/remote?

[oak osprey]

i use Craiyon

it's pretty good

[west tree]

Ah. That's why. This is for A11. It incorporates stuff that doesn't work anywhere else.

[oak osprey]

make it work

[unborn light]

My new pet 😄

[oak osprey]

i made a family

[wispy nest]

is there a very general noob help channel where i can ask noob questions

[unborn light]

Hi @wispy nest You can try here

😄

Shoot

[wispy nest]

such as: I'm in the #bot 1 channel and it's letting me create images - so is this free now? I thought they were charging for this

[green socket]

It's free for now because they're testing a new model

They might pull it once the model is released

[wispy nest]

i see and when i tried to google stable diffusion it seemed very difficult to discern which was the genuine article, there seems to be a lot of piggybacking and imitators would i be correct in that assessment

[oak osprey]

there will likely be a Huggingface Space where you can do inference for free

[green socket]

There are a lot of versions because it's open source so a ton of people make their own versions of it

Like how there's a billion different versions of Linux

[wispy nest]

before arriving here i did check out some web based free ai art generators, why did they all seem weaker than this one? the images generated were not as impressive

[green socket]

This is a newer model with more training. The other web ones are older. Stuff like craiyon is from last year

[wispy nest]

i see. i started on youtube and there is a lot of hysteria about this there. I think people just lack imagination. especially ironic considering all the artists freaking out about it

[wispy nest]

new question: I'm still wrapping my head around all this, but I'm seeing guides for installing stable/unstable diffusion locally (?) So am I to understand that this constitutes a totally different approach than using the bots here, and bypasses the need for paid subscriptions and such?

[smoky oak]

Good job!

Now unmake them 😀

I know I have sent this before, but I wanna send it again lol

[mortal kernel]

That image makes me feel oddly uncomfortable

[kind quartz]

should be cutties 🙂 with big eyes. But it seems they are not happy.

[cyan snow]

IT HAPPENED

I CAN"T BELIVE IT

FINALLY

[cyan snow]

also, i figured out Adetail

that's a useful extension

[severe swallow]

Denji in style of chris chan

[open sun]

guide updated - DATASET SATURATION BALANCE [DSB] - https://civitai.com/articles/397

[kind quartz]

Snowman that winks 🙂
Those builders are supposed to be women...

[rigid heron]

the pixel art is a lot better than i thought it would be

[split valve]

Almost nailed the pixel art but not perfect, and i think it needs to be perfect to make it work.

[rigid heron]

yeah i mean the ai understands what pixel art looks like but not why

it doesnt know its made of equal squares but it know it looks like that