🏞｜general-with-images

[cyan snow]

like imagine an extension that does that

[dense tapir]

I swear I saw someone did it via a paid extension. Give it a bit of time and will be for free and was for prompting.

[oak osprey]

DALL-E just uses CLIP

it's sort of the whole reason OpenAI created CLIP. and CLIP is what powers SD 1.5

[cyan snow]

no, it also uses transformers, it can get instructions, diffusion models can't

[oak osprey]

it's diffusion with an LLM frontend

[dense tapir]

Considering how horrible Bing is now with gpt shoved into my face on every single page I can only imagine with its integration into Windows which they said was coming shortly.

[oak osprey]

pixel diffusion = DALL-E

you can actually instruction fine-tune an image model. Instruct Pix2Pix

[cyan snow]

yep, it's a 2 headed AI

[cyan snow]

pix2pix is only image to image

and it can't work with all diffusion models

[oak osprey]

that's the only time it makes sense. if you want prompting, it doesn't help to say "make it more this" and it just tacks on "more this" to the end of the prompt

[dense tapir]

something is seriously wrong as my lora I made this morning no longer works, nor does my model. Same files saved and brought in via pnginfo the same thing.

[cyan snow]

skill issue

[dense tapir]

okay model works just the lora doesn't

[oak osprey]

i gave it:

image of people participating in the extreme bokeh olympics with more abstraction and less bokeh and more color and more surrealism

and then it made it. i said "please make it more psychedelic" and it rendered

image of people participating in the extreme bokeh olympics with more abstraction and less bokeh and more color and more surrealism and more psychedelic effects

[dense tapir]

I will ignore that or get banned.

[oak osprey]

truly groundbreaking.

[cyan snow]

fr? that's awesome

[dense tapir]

yes, was on his patreon

[oak osprey]

push the button, see what happens. it'll be funny

[cyan snow]

it's like oobabooga x a1111

[dense tapir]

YES

[cyan snow]

we NEED that collaboration

[dense tapir]

next level prompting

[cyan snow]

the future of diffusion AI

[dense tapir]

oh, and finally AMD has 8bit bitandbytes so they can ooba 8bit for larger models.

No, this is an error in the lycoris extension as I get one gen and when I switch models I get OOM now

damn it

[cyan snow]

is it possible to implement this to the ui?

[oak osprey]

prooobably

[dense tapir]

Of course if Auto wished it

[cyan snow]

dude, that will change this community

[oak osprey]

my code wouldn't directly translate over for it but if you have a Batch Queue extension you can fork, you can just add OpenAI prompt gen to that.

[dense tapir]

All these extension do is use an exploit to hijack Auto1111 which is why you can't just update and roll you have to update, close the cli, and restart it so it can rehook it.

I had Lycoris tell me that

[cyan snow]

huh, well, i think it would be awesome to just fire up a1111, and just tell it- make me a picture of (whatever)

[oak osprey]

it actually is pretty awesome

some of the coolest prompts i have discovered were like that

just give it a theme and it comes up with random words for that subject that you wouldn't

[smoky oak]

I am taking on a new way of realism prompting, and I am getting some more control and more realistic results, IMO

[oak osprey]

i wonder where those plug in

[cyan snow]

link? lol

i just looked around the web for it and couldn't find it

[dense tapir]

Patreon only is why

[smoky oak]

Very happy with my new realism prompting

No inpainting

[oak osprey]

im very disappointed with new sdxl lmao

it seems to have overfitted on five nights at freddy's

and lego

[oak osprey]

did the sweat stained armpits come with it for free?

[smoky oak]

Pretty sure it's just shadow, but I guess so haha

[oak osprey]

the trees in the background feel a lil ControlNet to me

i remember seeing that from it

thoooough i was doing CN with 1.5 models. that just could be how 1.5 does trees lmao

[smoky oak]

Your face is a little controlnet

[oak osprey]

i'll take that as a compliment

[smoky oak]

I did use tile upscale

[oak osprey]

aha

cheating

https://giphy.com/gifs/fallontonight-the-roots-tariq-XRXCGCbV10vQ6foS3h

[smoky oak]

Without it

[oak osprey]

ah yeah that's cleaner noise in the background

it's not bad, tbh

i just noticed the similarity

it's a veery distinct pattern the CN makes

[smoky oak]

The controlnet one has more realistic bokeh to the les I asked for

*lens

[oak osprey]

it's probably, yeah

it's more detail (noise, to some)

[smoky oak]

I wanted a wide angle bokeh shot, and real wide angle bokeh is much sharper and more "noisey"

Wonder if this new style of prompting will result in better male shirtless shots

To share without over sharing, I am wording things a little more "eloquently" I guess is how you could say

[sterile temple]

Morning lovelies. Suppose I better get up and get ready for work 😩

[oak osprey]

indubtiblablyey

he's made out of pills lmao

a photocopy of his bumcheeks

maybe my best attempt at satanic care bears @smoky oak

[smoky oak]

Lol

I have wifi on my new phone now

Omg, I feels like I left the stone age on my old phone 😅

It's still a good phone, but damn I didn't realize how excellent this one is haha

[oak osprey]

you... didn't have wifi?

hello r u from teh past?

[smoky oak]

I'm staying at a resort in Mexico lmao

I didn't know who's last name to put on the wifi sign in, turns out it's ours, even though we aren't the ones that booked

[smoky oak]

Why do I always read text like that as Dexter from Dexter's Lab lmao

[oak osprey]

read this

i tried to generate something unethnical via GPT and it was like blah blah im sorry and it made all these apology cards instead, plus a group of people that are like "what the fuck is happening here?"

[smoky oak]

I like how on the 3rd one it was more of "unETHNICal" lol

Anybody know if a port forwarding connection to my SD on my PC is encrypted? Lmao

I don't wanna be generating... Personal images of that can be snooped lmao

[grizzled olive]

You're getting to it with http:// right? Not https://?

The port forward itself is not encrypted

[oak osprey]

here's some cute kids

("pixar style", "game of thrones").and()

[oak osprey]

("purple colour style", "ferrari tiger").and()

i love this prompt segmentation business

[smoky oak]

Ok, good to know, my SD gens are NOT encrypted

[oak osprey]

("disney style animation", "game of thrones").and(1.1, 1.0)

omg

("little boy blue", "skyscraper architect", "stranger things style").and()

without prompt segmentation, little boy blue skyscraper architect stranger things style

@prisma iron ^ that's why i want prompt segmentation on the SDXL bot lmao

[dense tapir]

Makes one wonder if they should be. All that fur.

[smoky oak]

Not even generating furry boys at the moment lol

[dense tapir]

LOL

[smoky oak]

Just shirtless men lol

But still, I don't need the hotel snooping me

[smoky oak]

wait a sec, I think I have a way of encrypting it

[quick current]

any inapint?

[dense tapir]

I suck at inpainting

[quick current]

nice that you got this with none.

[dense tapir]

Let's see what it looks like without my lora. I trained a real style last night. Nothing in particular just a style.

[smoky oak]

I really just wanna gen images on the wifi connection on my new phone lmao

[sterile temple]

you're using a VPN right, right???

[smoky oak]

No

Tho I could activate mine I think

[dense tapir]

Not bad

[smoky oak]

I suppose my VPN would encrypt it

[oak osprey]

vee peen

[dense tapir]

I like the other more though

[oak osprey]

remember the talking meatball pixar movie? im trying to remember the name but google doesn't help. i thought AI would know it. but uh

[sterile temple]

oh look, it's me in the morning

[smoky oak]

Ok, I am on VPN now

I guess I can do my hot guys in anonymity now haha

Boys, my beloved 😍

[sterile temple]

yup... lol. this prompt produces some odd creatures

mow

[smoky oak]

Oh come on >:C

It looks like my VPN doesn't support direct IP connections

Uhhgggg

That is annoying²

[smoky oak]

My hot men are on hiatus

Oh, JK, my IP managed to switch between having a VPN and not lmao

How strange, it only does that like once a month, what a coincidence lol

And now it's infinite loading

I JUST WANT MEN COME ON

It worked

[oak osprey]

yay, GPT will make segmented prompts for me now

[sterile temple]

didn't they add the capability to return json so that it can be integrated with other APIs easily?

[smoky oak]

Forget hot girl summer, I'm after that hot boy summer 🥲

[oak osprey]

like hell if i'm going to sit there figuring that out

uhhhhh hmmmmmmmmm.... yes..... ok

@smoky oak there you go

("Gruesome", "Satan", "Dark", "Plush", "Hellish", "Cuddly", "Morbid", "Demonic", "Cozy", "Mischievous").and()

idk why gpt put cuddly and cozy in there

[smoky oak]

Those are decent

Still nothing like what I am after

I really want black craft cult aesthetics

[sterile temple]

interesting prompt idea

[smoky oak]

All of black craft cults clothing is designed by a single woman, and I got to meet one of her friends when I went to their headquarters

I got a dope satanic banner that I use as a curtain for my room now lol

[sterile temple]

first image i got with black craft cult aesthetics

I thought it would give a variety of images

yes, black croe

[smoky oak]

Those are actually kinda cool, I like those :>

[sterile temple]

just need to weight it more male 😉

[oak osprey]

@smoky oak GPT4 jailbreak wants me to show you this

[mystic marlin]

Damn...we've gone darkside this evening, alright.

Ironically...I was about to post a dark angel too. lol

[oak osprey]

omg im a satanist now

these are fuckin awesome

("a digital art piece of hauntingly pale skulls with black craft cult aesthetics", "opulent and gothic", "high contrast", "bold and macabre", "with a touch of mystic.").and()

i taught GPT3.5 a lil better

        prompt = f"Print your prompt."
        if theme is not None:
            prompt = prompt + '. Your theme to mix in: ' + theme
        system_role = "You are a Prompt Generator Bot, that strictly generates prompts, with no other output, to avoid distractions.\n"
        system_role = f"{system_role}Your prompts look like these 3 examples:\n"
        system_role = f"{system_role}a portrait of astonishing daisies, rolling hills, beautiful quality, luxurious, 1983, kodachrome\n"
        system_role = f"{system_role}a camera photo of great look up a rolling wave, the ocean in full view, best quality, intricate, photography\n"
        system_role = f"{system_role}digital artwork, feels like the first time, we went to the zoo, colourful and majestic, amazing clouds in the sky, epic\n"
        system_role = f"{system_role}The subject must come first, with actions coming next, and then style attributes.\n"
        system_role = f"{system_role}Any additional output other than the prompt will damage the results. Stick to just the prompts."
        image_prompt_response = await self.turbo_completion(system_role, prompt, temperature=1.18)

@smoky oak anything you'd change?

or tell him?

[sterile temple]

me thinks i need another coffee this morning

[oak osprey]

the howerton twins

[oak osprey]

@smoky oak i know it's not the best yet but look at the improvement in small faces at a distance from another half a week of training the text encoder

idk if the SAI guys were right when they told me it's the VAE's fault they can't do small faces

[smoky oak]

Very clear improvement, that looks dope dude

[oak osprey]

ty bud

it's hard to determine whether the text encoder or unet is more responsible for deformities

(an earlier ckpt by about 2k steps) training unet vs training text encoder vs both

they end up with a whole extra child with both

it's interesting how training the unet seems like i can add a lot more detail to the scenes

but you have to kinda, uh, like, do it carefully

[oak osprey]

now i can put multiple prompts in and split by newline and it batches 'em

[dense tapir]

Now that one has some balls.

Not sure what that one is though.

[dense tapir]

For your next acid trip

[dense tapir]

@smoky oak

[oak osprey]

vape buddies

new checkpoint :3

@smoky oak yay, they have faces

[smoky oak]

*they have face

[oak osprey]

yeah that

i wasn't going to say it because i thought that'd be waycist

the sextuplets lmao

[oak osprey]

i bet it'd be better for the family photo if i remove the word 'embarassing'

these are on the first try for each of those prompts so i think it's doing pretty good just tuning the text encoder

man who knew that was the missing piece for 2.x ahaha

they created the unet on a frozen text encoder, so, it never got to learn what the unet could teach it

[oak osprey]

the vape king

[sterile temple]

lunch break!

[oak osprey]

this ckpt gets a lot closer to Robin

[chrome kite]

Face didn’t turn out well… trying a new Wes Anderson like style

[sterile temple]

have you tried adding some celebrity names with a low weight?

[outer cave]

New to this...

[ripe cedar]

Anyone know how to add arguments to Vlad?

[oak osprey]

the person?

just ask him how he's doing

amirite @smoky oak ?

im kidding, i don't have any opinions here. but uh, find where it says parsearg

[static tusk]

fairly sure they are all in the settings menus with checkboxes etc

[smoky oak]

Yeah lmao, Vlad is a dick, respectfully lol

[oak osprey]

oh

my bad i thought you meant to add a new option to it at all

[dense tapir]

smh

[smoky oak]

Oh nooooo

My pretty PFP decoration is gonneeeee

dies

[dense tapir]

I am sad anapnoe is merging and has a private branch for it right now. Eventually will only be vlad.

[oak osprey]

its like it fucks the faces up in particular, on purpose that one on the right

[smoky oak]

I'll stick with autos mediocrity over Vlad being a general douche

Also, with Vlad, now I'm supporting two jerks, rather than just auto on native a1111

[oak osprey]

lmfao that's some logic right there

you don't have to think too hard about it, just do what you want

[smoky oak]

I always minimize supporting shitty people, it's how I work in general

I'd ditch auto and his racist ass if I could

[oak osprey]

you could go open polite feature requests on invoke ai for stuff it doesn't have

[dense tapir]

Well, he stung me turning me into github but I can still post and use them just not with 3rd parties such as colab.

[ripe cedar]

I don't see arguments like --xformers in the settings tab in Vlad. I don't even use it directly, I just need to add --api to use it on Photoshop for the SD add on (as a1111 stopped working for some reason)

[dense tapir]

Thin skinned people should NOT open repositories then allow tickets to be posted when shit breaks, or is released broken. I have ran into a few with no ticketing system before so it can be done.

[ripe cedar]

Actually, it's already in PS, it just doesn't generate the image when I hit generate. The api thing must already be enabled I think, because it that's the reason why I can see it in PS to begin with.

[dense tapir]

Oh, vlad broke --api and will no longer use it (removed forever per Vlad). I cannot remember what took its place but something did just anything --api (automatic1111 way) is DOA.

[ripe cedar]

Aw man 😦 I'll have to figure something out later I guess. The plug-in is just so good for PS. Will have to research. Ty for letting me know

[dense tapir]

You are welcome, and I do agree that plugin is pretty sweet.

[oak osprey]

1920x1080 didn't fix the faces lmfao

[dense tapir]

I think it worked

WTH is this? LOL

[naive wren]

hey, I don't remember who was asking about the artistic QR's earlier but I found this tool https://app.tt.social/tools/qr-generator-ai that was released and it lets you create them without the hassle of configuring Stable Diffusion and all that

it looks pretty cool, they have simple options but the QR's look amazing

[oak osprey]

tt sounds like a total scam lmao

[naive wren]

wdym ? I used it free , did it ask you to pay?

[oak osprey]

the name, taptok

seems like it's trying to trick people

[past prairie]

What's the best way to fix messed up eyes? I've tried in painting an it just makes them worse. What settings should I use?

[marble ridge]

Anyone know a website besides civit ai for lora's ? or checkpoints?

[dense tapir]

Honestly?

They are about the only thing that I can get right with inpainting.

[past prairie]

What settings do you use, then?

I can get them right but it's difficult to push the details and realism

[dense tapir]

1.5 I had that same issue as realism was hard. Show me what you mean with what it was and what you got.

[past prairie]

I think I've got it actually. But inpainting in general is difficult tbh.

Next is the hand

[dense tapir]

face is the easiest to inpaint and all else is damn hard

Oh, forget about hands. You will see what I mean.

controlnety can help with hands, and is far superior to inpainting

[past prairie]

My hand is in an unusual position so it's gonna be hard.

I'll try controlnet

Depth right?

[dense tapir]

yep, and there is an extension just for hands

goes with it

[past prairie]

I'll check it out

[dense tapir]

I think my robot training went fairly well. Assassin's Creed meets Mech

I finally figured out how to train styles

Thing is my dataset was Anime so weird it works

[kind quartz]

isnt style in sd just pos prompt neg prompt

[dense tapir]

no

think of it like the by artist used so heavily in 1.5

[kind quartz]

in a1111 it is, i say it badly.
O.k.

[dense tapir]

See, in 2.x it knows so little Artists which was the first reason people fled back to 1.5. No more Greg Rutkowski, et al.

[kind quartz]

found this image so will not do it again. Here in A1111 styles are just prompts

[dense tapir]

Yes, your English is not understanding what a style is.

[kind quartz]

we got as well style word, so i know 🙂

[dense tapir]

Artists are so upset because the AI, in 1.5, was using their style they are known for.

[kind quartz]

style is artist or cubism, realism

[dense tapir]

not entirely. Think of a style as the brush strokes as well.

[kind quartz]

yes. We have word styl. But in A1111 it is realy just saved prompts.

[past prairie]

I think he is referring to LoRAs. They are more specific than prompts. 👍

[dense tapir]

cubism, and the rest, are actually art movements not styles.

You can have Renaissance painters that were part of the same art movement but each have their own style.

THAT is what I am talking about.

[kind quartz]

Yes second sentence i know and understand.

If anybody here have A1111 should explain with proper english what style in A1111 is 🙂

[dense tapir]

in your image that is a checkpoint that someone trained to do the styles of FE14 (I have no idea what FE14 is).

checkpoint = model

one of my training images I think is messing with this

[past prairie]

There is also a style tab in the a1111 interface but I don't think that's what he talking about. I think he's talking about LoRAs or textual inversion.

[dense tapir]

I don't have a style tab. Hmmm, must be an extension I don't have.

This is weird, but now comes the hard part of figuring out which image is causing this. See on his back? I get that on arms, and legs too. Really weird.

[past prairie]

Below the generate button. There is a tab that says "Styles". It comes natively.

[dense tapir]

Oh, oh.

Alright. Yeah, that is a HUGE misnomer from Auto

I use that a lot to save my neg prompt these days. Hasn't anything to do with a real style, hence the misnomer part.

I wonder why I am getting that thing on his back? :/

[tired gyro]

hi! any suggestions to make my results better? idk why it thinks that there're two dogs

[dense tapir]

I just check my dataset and not a single thing on a back on any of them

[kind quartz]

it is because resolution i believe
you can as well specify 1(japanese spitz) not sure it will work way i used.
Maybe 1Japanese-spitz would be better

[dense tapir]

try 1other as the first on the line.

[tired gyro]

oh ok, should i down the resolution as well?

[kind quartz]

depends. models 2.0+ are trained on 768x768

[dense tapir]

1.5 was bad when you stretched width beyond 512

Yeah, 800 on 1.5 gives doubles a lot

LOL, that thing on his back seems to be seed related

changed seed for the second one

[kind quartz]

Battery 🙂

[dense tapir]

I am unsure what it is but weird. Been on his arm and leg before too

Looked like a tumour in a few.

I hate the missing segments too

normally I have to up my samples to get them in

[kind quartz]

playing with some jar things and now got this.

[dense tapir]

Looks nice though

Oh, I upped to 30 steps and got this

rocket launchers?

Looks like a toy

[sterile temple]

Style is where you save a prompt, give it a name, and then you can easily recall it later

[marble wasp]

My entry for a art challenge. The premise was that we had to create miniature cute versions of a Kaiju but with the aesthetic of retro gaming.

[sterile temple]

nice one

[marble wasp]

i went a couple different directions for a little while hahah

[wispy nest]

maybe picking up from a building in the background of an image?

[shut sinew]

[oak osprey]

make sure you're using MSE 840000

for your VAE

the default 1.5 one is shite at eyes

[chrome kite]

A little bit of inpaint would make this perfect

[oak osprey]

try ReimagineXL

[kind quartz]

how about mlsd in controlnet?

[wispy nest]

gah, I'm cursed! I only get either extreme character poses or a "statue, never moving pose" :P

[cyan snow]

JESSE

it's time to cook

[oak osprey]

lincoln log fries?

[wispy nest]

there we go, when skill won't fix it, then luck will surely save me! :P

[supple jolt]

Can please someone upload ai bad rooms, I want to write a story and it may help

[oak osprey]

i did it!

[wispy ether]

https://cdn.discordapp.com/attachments/560461575136346115/1119299214569119844/danser_2023-06-16_18-13-50.mp4

ong

[oak osprey]

@cyan snow that baby welding image is from my new 2.1 ckpt

imo it is equal to the awesomeness of the Bing reference img

[cyan snow]

Yoo

Did you release it?

[oak osprey]

it's on huggingface as ptx0/pseudo-real-beta

i don't bother with A1111 ckpt generation or CivitAI anymore, if you want it there, i encourage you to take the initiative and upload it there

[cyan snow]

Can I run it on a1111?

Or is this like the last one you trained

[oak osprey]

i don't know. the last one had issues i'm not keen on having people experience again with that tool

[cyan snow]

Alright, il test it out and let you know

[oak osprey]

tried to help the Fulljourney bot admin load my model up there and he had to upgrade a bunch of libraries eg. diffusers to 0.17

ended up killing his A1111 install ahahaha

[cyan snow]

I'm on the dev branch, I can restore it if it fucks up my stuff.

[oak osprey]

you're a bit more in-tune with how to fix this stuff than they were

i had to help them undo what i broke

[cyan snow]

wait, when i cloned the model card's repository into my models folder, it doesn't detect it in the Diffusers format

do i need an extension for it to work?

[oak osprey]

i've never seen an extension that loads Diffusers models in A1111

does that actually exist?

[cyan snow]

i usually just use the conversion script from Safetensor or CKPT to diffusers and it works just fine..

[oak osprey]

that one results in odd VAE issues for me

[cyan snow]

so do i just do the opposite here? like diffusers to ckpt

[oak osprey]

yeah

[cyan snow]

@oak osprey i couldn't get it working, the script for diffusers to ckpt expects .bin files in de diffusers folder, and in this case it's a safetensor.

[oak osprey]

--use_safetensor

or safetensors. i can't remember

[cyan snow]

convert_diffusers_to_sd.py: error: unrecognized arguments: --use_safetensors

um, idk man

[oak osprey]

yeah thats why i've given up on it too lmao

[cyan snow]

on your other model there was a ckpt format, how did you convert it there?

[oak osprey]

same script you just tried iirc

[cyan snow]

wha-

how did it work for you

[oak osprey]

[-h] --checkpoint_path CHECKPOINT_PATH [--original_config_file ORIGINAL_CONFIG_FILE] [--num_in_channels NUM_IN_CHANNELS] [--scheduler_type SCHEDULER_TYPE] [--pipeline_type PIPELINE_TYPE] [--image_size IMAGE_SIZE] [--prediction_type PREDICTION_TYPE] [--extract_ema] [--upcast_attention] [--from_safetensors]
              [--to_safetensors] --dump_path DUMP_PATH [--device DEVICE] [--stable_unclip STABLE_UNCLIP] [--stable_unclip_prior STABLE_UNCLIP_PRIOR] [--clip_stats_path CLIP_STATS_PATH] [--controlnet] [--half]

[cyan snow]

was the formats .bin?

[oak osprey]

mine has these arguments

--from_safetensors and --to_safetensors

[oak osprey]

lololol

[dense tapir]

Maybe. A bit odd and so hard to eliminate. Another issue is if they are all on solid backgrounds as that is a training nono too. Damned if you do and damned if you don't.

[wispy nest]

yeah thats a bitch

[oak osprey]

i think solid backgrounds are fine

you need a mix of it all

try to avoid taking transparent images and carelessly converting their alpha channel to a white channel without further blending on the edges to avoid the halo effect

but beyond that, it's all good to have.

if you ONLY have backgrounds in your data it will assume the character always has one. if you don't have any backgrounds, it will be hard to introduce any. they will look like "octane render" prompts do, backgroundless

you have to have a pretty fair balance of both types of images, and if you want the LoRA to not introduce backgrounds as easily, you'd want to lean toward less of those in training data

[dense tapir]

I tried it like that saying to myself that the background only distracts it but 100% failure. I asked around and was told it makes it too stiff among other things.

[dense tapir]

Fro Orc dancing Disco with the Rave orb satellite thing. I like it.

[wispy nest]

Death Star Disco

[dense tapir]

Aye

We are about to be hit with the storm that wiped out that town in Texas. Just began.

[dense tapir]

Where my sis lives on the opposite side of the city she had golf ball sized hail.

[dense tapir]

You ever notice how SD has trouble with orbs if something is in front of it?

[oak osprey]

pregnant death star

[dense tapir]

moons, etc... become odd shaped or missing part of it.

[cyan snow]

nah, it's because of the AR, it assumes because the AR is wider, then the obstructions in the image should also be wider.

[oak osprey]

death star museum of science

[dense tapir]

Well, I have because SD is missing a piece of its "brain" to fill in the gap. AR has no role in it, or it shouldn't if it were done right.

I can do square ratio AR and get it all the time

[cyan snow]

i don't

[dense tapir]

I use 2.1

[cyan snow]

why does 2.1 always have those problems

[dense tapir]

I have a sneaking suspicion 2.1 has the issue that is also related to Giraffe neck syndrome because internally the training is mostly 512sq but the last 1/3 of the steps was done at 768sq

[oak osprey]

@smoky oak idk if you're interested in LoRA implementation details but here's an interesting discussion https://github.com/huggingface/diffusers/pull/3756

[cyan snow]

also, i discovered that if you tune 2.1 to a certain point, it just starts to turn old people into wizards

@oak osprey right?

[wispy nest]

overall, I'd say the ai has issues with anything that's "behind something." Capes vanish on one side of the arm, orbs, like you said aren't very orb like. Straight railings change angle on one side of the character from the other. Etc, etc. It's mostly a luck thing with very few ways to decide it without using extra stuff like another fine-tuned model, loras, upscaling, etc

[oak osprey]

that's probably the synthetic data but yeah

[cyan snow]

i'm not the only person that has the ''hogwarts syndrome''

[oak osprey]

like if you're using midjourney data i think midjourney loves making old people do magic, like, that's all they do there

[cyan snow]

yeah, but that's so strange, like, remember that loab thing?

[oak osprey]

yep

in base SD, all old people do is just sit there looking sad

or excessively happy

pharmaceutical ads, lmao

[cyan snow]

also, will sdxl have the conjoined twins syndrome?

[oak osprey]

it makes tons of fun deformed stuff currently but i don't know if they'll nuke that feature, if that's what you're asking

[cyan snow]

''feature'' LoL

[oak osprey]

try to make a blonde-haired basketball with well-manicured fingers

in super wide aspect ratio

[cyan snow]

that's like kicking someone in the b*lls, why aim for the weakpoints?

[oak osprey]

because it won't do it

[cyan snow]

bro, i wanted ''death star made of plants'' not death star ON plants

[oak osprey]

awww thats cute tho

@cyan snow seems like there's no major duplicates issue except for the obvious

[dense tapir]

LOL

[oak osprey]

hmm maybe there is

stretchy at tall aspects still

[sterile kiln]

Hey guys, what is Clip skip and where i can find it? Is it from an extension?

[oak osprey]

it's in the Stable Diffusion settings area

[sterile kiln]

What does this change?

[oak osprey]

it skips one or more layers when producing the result

[dense tapir]

Dates back to NAI and how they trained their models. 100% worthless on SD 2.x as it has been shown does nothing. Some effect on 1.5

I used to have it as a quick bar at the top of auto when I played in 1.5

[wispy nest]

ah, NAI, takes me back to the craze when it got leaked :P

[cyan snow]

yeah, how did it get leaked btw

I never figured it out to this day

[oak osprey]

once again General Awareness is wrong, clip skip is required for 2.x ckpt

it just doesn't let you screw with the depth in that UI

for Diffusers format checkpoints, it drops the final layer because it's not used.

it's just not in the files at all

[wispy nest]

can't remember how. And the link I had for the info seems to be dead, but I have a slight memory of it being "hacked" by someone. the "anything model," the chinese model, also got hacked somehow I believe

[oak osprey]

someone probably found an S3 bucket

[cyan snow]

bruh wtf

i just wanted an old person feeding his pet bird

WHY DOES IT GIVE HIM MAGIC

does the ai know something we don't

[wispy nest]

where's the magic? :P

[cyan snow]

he is wearing armor

[oak osprey]

dangerously close to that eye

my model hasn't yet reached full late-stage elder magic syndrome

[wispy nest]

probably because of some prompt word around harry potter, that patch on his left side of the vest kinda displays it :P

[cyan snow]

when i first started started messing with AI i never thought in a million years that we would discover something we call "l late-stage elder magic syndrome''

[wispy nest]

I never thought we'd get ai art at all. It was like magic when I discovered it. No lie, I thought it was something out of star trek or something! :P

[cyan snow]

i'm not 47, i swear

nice text

[dense tapir]

That is supposed to say 147

A whipper snapper in the wizarding world.

[cyan snow]

it's leviosaa not leviosar

IT ALMOST GOT THE TEXT

this is a glorious day

[wispy nest]

as long as the text is edible, then it's all fine and good and hopefully tastier than those wax taste candles. Who'd want them with that flavor in the first place?! >:I

[cyan snow]

whoever invented the birthday cake

also, that's one of the closest times to getting the text perfectly

[wispy nest]

road signs and "hello" is probably the easiest, but their taste is very, very bad. And also hard to chew. Great, now I'm getting hungry, again >:(

[cyan snow]

how old is this girl

[oak osprey]

nailed it

[grizzled olive]

The Holly Biibl

[cyan snow]

SO CLOSE

[fleet goblet]

I wonder.

[oak osprey]

i wish his jacket were clearer but it's great otherwise.

[oak osprey]

@smoky oak is that you?

[smoky oak]

[smoky oak]

Nope, but this is the least terrible one so far lmao

[oak osprey]

in general or of your likeness

[smoky oak]

Both

[oak osprey]

damn, the melodramatic android has a face now

scientist? WOMAN. bam, bias busted

those trees are the least blurry i've ever seen trees from this process

they're a lil weird tho

[wispy nest]

I need to find a better upscaler. But that will have to wait until time suddenly move several hours forward. I have no idea why that always happen at night, it's like I'm— 💤

[oak osprey]

@smoky oak wow the lil kids faces aren't destroyed

there's, uh, other shit going on, but

i bet you that freaky extra kid will be absorbed in no time

[indigo nimbus]

Hi, I'm wondering if someone may be able to help, I'm having an issue where SD starts creating an accurate image that conforms to my prompts, but then halfway through reconstructs the image in a worse way and spits out a very bad image, just wondering if anyone might knows what could cause this

[sterile temple]

do you have hires fix enabled?

[indigo nimbus]

Yeah

It seems to drastically reconstruct the image from scratch halfway through generating, and goes haywire

[oak osprey]

yup

use fewer steps for upscale is my guess but i'm sure others have better tips

[sterile temple]

probably because latent is the default upscaler which makes the image really blurry and unless the denoise ratio is up at like 0.7 it will end up blurry

[indigo nimbus]

Ah right, atm I'm using latent (nearest)

Is there one you'd recommend that is better, and less chaotic?

[sterile temple]

lanczos does the least damage

imo

[indigo nimbus]

Funny, I chose that one randomly just now

I'll try that one and see if it's any better

Thank you

[smoky oak]

Don't use latent upscaler, try with anime 6b upscaler, you will get way better results with 0.3 denoise

[indigo nimbus]

Thanks, I'll look into that, haven't tried any other upscalers than what is available with standard SD

[smoky oak]

Anime 6 b is standard SD

The drop down menu where it says Latent, I don't recommend using laten upscaler, it has a lot of artifacts and also likes to destroy composition a lot

[indigo nimbus]

It isn't there in the upscaler drop down menu, for me

[indigo nimbus]

This is all I have on the drop down list

[oak osprey]

guys, mr bean is resistent to a good photo 😦

[indigo nimbus]

https://tenor.com/view/scared-omg-no-gif-9016411

[oak osprey]

so many questions

[dense tapir]

https://giphy.com/gifs/justin-g-eye-roll-bitch-please-contempt-l0NwPZ027mabR6Tg4

[wispy nest]

@frigid vigil try here, there are some pro ppl making images

[frigid vigil]

should i just repost the whole thing?

[wispy nest]

if you want ... so others know what you are trying to do

[frigid vigil]

I've been fiddling around with the settings for days and I can't figure out why adding more subjects and composable lora chunks makes it sperg out.

[sterile temple]

You probably already tried, but have you tried a different sampler? I was getting those kind of faces with a certain sampler, can't remember it's been a while

[frigid vigil]

Yeah. But I'll try something again else just in case.

[frigid vigil]

aaaaaaaaaaaaaaand nope

[oak osprey]

try times square, india prompt

so wild

it puts the indian buses and people into NYC

[smoky oak]

Brooooo, the resort we are at just had a fucking sick LED Psytrance performance

[ripe cedar]

Managed to get SD working in Photoshop. I'm trying to make basically a bone pile instead of individual skeletons/skulls in PS via inpainting. Now sure what prompt would do that.

[smoky oak]

Look how sick this shit was

It was even better in person, I tried to get it as good as possible, but man it was really cool in person

[oak osprey]

nice try

[smoky oak]

It worked pretty good, but the LED's had a sort of pulse that made them a little hard to capture, even at 360° shutter angle

[oak osprey]

i got another AI cryptoid

who is he

another one

[marble wasp]

using SD and Gimp, I made and manipulated this for an art challenge I am participating in.

[oak osprey]

i can generate so many portraits and then today i've gotten so many of these cryptoids

[cyan snow]

did i have a stroke.. LoL

[sterile temple]

is this what drunk people dream of when heading home after a night out?

[oak osprey]

when driving asleep

[cyan snow]

drunk people trying to leave mcdonald

[cyan snow]

I DID IT

they are not wizards

[oak osprey]

retirees

[sterile temple]

meowterspace

[rain gazelle]

precisely

[oak osprey]

it have no spungbung

[dense tapir]

@smoky oak I see Reddit won

[oak osprey]

reddit was always going to win

they were going to pull the mod accounts and just put new mods in

[past prairie]

Any tips on improving this hand? It lacks realism. I wan't to remove the nailpolish, make it less "sharp" and drawing-looking but image2image ends up messing it up even more...

[velvet wren]

Finally managed to get a working QR code

even if it looks like shit lol

[wispy nest]

the real question is: Does it lead somewhere? 🤔

[velvet wren]

try it lol

it's a youtube video

[wispy nest]

what if that was the way to get the prompt? :P

[velvet wren]

You already got the metadata for that xp

I also heard that the shorter the "message/url" the better chance of it working so

[velvet wren]

literally innsmouth

[dense tapir]

I think my second training is going better using my first training.

280 steps in I think that was

fingers are crossed as I am trying all kinds of things

colab let me go all the way to the end and I was at 85% full and it ate the training.

[supple jolt]

I need a normal looking bad room of rich people but not too rich, please my friends I need this for my story

[fresh hound]

lol

[wispy nest]

I wonder how far out we are from being able to generate multiple people who dont look like clones

[kind quartz]

what a cursed artifact 😄 Have to work on it

[grizzled olive]

Same. lol.

[marble wasp]

This is my submission for an art challenge, a creature called Fumblebee. i went with the premise provided to make an innocent, lumbering, and accidentally destructive bumblebee creature, with a little commentary of my own. 🙂 I made this AI/IM with Stable Diffusion and Gimp (free photoshop) with dozens of variations and fine tuning until, again, I was able to import into Gimp for image manipulation.

[smoky oak]

I just noticed the tower in the back says Trump lol

[marble wasp]

That there is my commentary:)

[smoky oak]

I see haha

The world desperately needs a Fumblebee

Tho it looks like trump is doing a good job of ruining himself :p

[marble wasp]

https://tenor.com/view/stephen-colbert-disaster-late-show-gif-14855996

[wispy nest]

guys i forgot

what is this model pack?

how do you think

i use many and i forgot which i used

there's quite a lot of anime checkpoints so you can just try them one after the other. Or use the meta data if you save it.

oh

there it is

in image name

coffemix

you can also, using auto1111's webui anyway, save all the meta data to the image itself and then just either open it with notepad to copy them, or use the webui's PNG info

[wispy nest]

i use nmkd

[wispy nest]

never heard of it, I only use auto1111's webui, and I know of some minor similar webui's branching off from it

[wispy nest]

what do you guys think is the best photoreal model atm? i havent really used stable in months because i got so bored without a new official model but am thinking of reinstalling auto for shits and giggles

can anyone suggest a LoRA for art work like this?

[wild sorrel]

https://civitai.com/models/23393/wlop-style-lora

or
https://huggingface.co/SirVeggie/wlop

[civic crag]

i'm testing stuff to get qrcodes (seems to be a trend), it is hard to get good results. I think this is my best attempt

[oak osprey]

#1100170312106127410 '

[oak osprey]

i would suggest try installing invokeAI instead, you can then try my models. A1111 doesn't support them

[wispy nest]

doesnt that cost money?

someone mentioned it the other day, maybe you, but when I googled it, it looked like you had to sub

[oak osprey]

no, man

it's open source

it was one of the first forks of the original stable-diffusion repo

man, flex-diffusion-2-1 is ancient by today's standards but even at resolutions it wasn't trained on, it improves the proliferation issue substantially

[wispy nest]

what link do you suggest to dl it?

[civic crag]

maybe https://github.com/invoke-ai/InvokeAI

[wispy nest]

maybe I looked something else up this week, idk. anyway thanks

[cyan snow]

it has WAY less features then A1111

[oak osprey]

yeah but the ones it has, have more focus on their development

A1111 is kind of, a lot of features but very few have any benefits. a lot of things are experimental and stuck around unnecessarily

good software abandons features that are unneeded.

[wispy nest]

does it have controlnet?

[cyan snow]

InvokeAI doesn't have hires. fix. that means it will have conjoined twins syndrome when making wallpapers

[oak osprey]

just use models that don't suck then

proliferation issue is what you're referring to when you say conjoined twins and that's a problem with the training data and aspect buckets being unevenly distributed

[cyan snow]

still, hires fix is a very crucial extension

[oak osprey]

it's just img2img

[civic crag]

i never used that 😄

[oak osprey]

maybe CodeFormer and GPFGAN.

[wild sorrel]

I think it's only for companies, or at least they wanted to do so...

BUt it's kinda same thing as it was as you left

[oak osprey]

you can go use hires fix for free on replicate.ai spaces

[cyan snow]

still, do you really have the patients to move EVERY good generation to img2img? i sure don't

[oak osprey]

i don't do any of that, i write python scripts that handle it all for me. and invokeAI has an API and a CLI tool that can be used to integrate it into workflows like that

[wild sorrel]

Maybe use webui that don't suck instead lol?

most good models are all based on 1.5 - 1.5 base trained on 512x512 and whether you want it or not - issue will be there even if your model trained on higher resolution.

Yea, it will be more rare, but it will be there.
Using img2img for it - is just an additional step.

Same goes for samplers, controlnet, even basic "delete" button on a canvas itself - that's something most of us use every time we generate something (except inpainting delete button part) - and something that invoke still can't do.

Yea, you can generate without them, but like if you want to throw everything fun SD can do - why not just use mj instead.

[oak osprey]

sorry dude i only use 2.1 which was trained on 768x768 and 2048x2048

[wild sorrel]

you do you.

[oak osprey]

currently fine-tuning it again with 5 different aspect ratios

500,000 high-res images

don't get me wrong, hires fix is conceptually pretty cool, but it's unnecessary

[wild sorrel]

Yea, sure, all other things a1111 can do compared to invoke aren't nessesary too

[oak osprey]

if you don't know how it works internally, it takes an average of the colour of the square image and then adds padding on the outsides, like a paint bucket fill of that, and then runs it through img2img

i'm sure it's upsetting to a power user to hear that features aren't needed, but for most people, they aren't. they're there to address deficiencies in training

if you just have a good enough model, you do not need stuff like controlnet. it's always going to be nice to have for someone who wants to eg. inpaint a coffee mug instead of a beer stein but that's again something most people will never care about

should see the stupid shit my users prompt for

[wild sorrel]

and can't even delete an image they dont want from canvas itself, have to go through collection nonsense instead, again - not nessesary , but also not comfortable.

[oak osprey]

i don't even know what any of that means

[chrome kite]

Dolphin man

[oak osprey]

dolphin robot

cyborg 😄

oooh it's an android

[wild sorrel]

You are just saying dumb things.

"Just chose good model and use webui that has less features, can't do half of things at all, wasn't updated significantly for like what , 3 months by now, but...but....has shiny webui?" - that's an advantage?

[oak osprey]

hey, don't become insulting

[wild sorrel]

Isn't that what you were trying to do before?

[oak osprey]

uhm, nope. the models do suck if they have the proliferation issue when it's been known for months how to get rid of it

and i didn't understand anything you wrote

you called me dumb in response, so, good for you

[wild sorrel]

kay.

(about 2nd reply I missread that, sorry)

afk, gonna go shopping

[civic crag]

🍿

[wispy nest]

looks delicious but fattening

[smoky oak]

Relatable honestly lol

[oak osprey]

currently conditioning images so that their minimum size is 1024 and discarding any images that aren't at least ..x900 or 900x..

bunch of non-english tags in here, eg. japanese/korean/spanish

[wispy nest]

thanks, i had the wlop but didn't think that would do water color painting. found another link here https://huggingface.co/fladdict/watercolor/tree/main but yet to try it out

[wispy nest]

what i was trying out earlier that also says wlop style LoRA https://civitai.com/models/5042?modelVersionId=68639

[smoky oak]

I have goals to try and train a simple 1024x style LoRA when I get back. Hoping 10 GB VRAM can do it with gradient check pointing

An easy test would be to try and train a simple photo realistic style into the furry model I use

The model itself is trained at 640x798 or something similar

[wispy nest]

when using controlnet, is that default weight 1 too high? cause whenever i load a real character for specific pose and try to crate anime style art, the output seems bit off and non anime like

[broken stream]

Are you using openpose or?

[wispy nest]

using depth

[broken stream]

Yeah it would be a problem with depth I imagine, I guess try lower

[wispy nest]

cant seem to get the right balance between control weight value and choosing

[broken stream]

Yeah Im not sure what that does I would also like to know lol

[wispy nest]

going with

weight: 1 and mode: balanced seems better but not exactly there

I wonder if there's a limit to the number of loras that can be used? ;P

[wispy nest]

VRAM is the limit

i think

is there anything new in the SD world

not that I know of, people are still training their own stuff and none of them seem to notice a couple of issues in how the ai creates the top of people's heads for some weird reason. But that might just be me doing it wrong :P

[dense tapir]

@smoky oak https://www.reddit.com/r/technology/comments/14ag85h/reddit_threatens_to_remove_moderators_from/

[smoky oak]

Yeah lmfao

Reddit is a piece of shit platform that houses many amazing communities. If everybody could get away from how terrible the core company is, we'd all be better for it

[grizzled olive]

I hope all the replacement mods demand payment

[dense tapir]

I agree

[smoky oak]

It's likely that Reddit will just shut them up, or ban them

POS company is as POS company does

[dense tapir]

The problem with Reddit is that even one of the founders said it is a piece of shit and would not use it now, or something like that. The problem started with this moronic concept of a karma system. In a utopian world maybe but this is real life and is easily abused, and is.

[oak osprey]

yay it has a progress bar now and i cache the aspect bucket assignments

should help on google colab where this process takes an eternity

[smoky oak]

[steel apex]

I found my happy place

[smoky oak]

makes me think of that dude that has the hand made river in his house

[oak osprey]

rumpleforeskin?

[steel apex]

This feels familiar...

[smoky oak]

Suddenly I can't find his house

@steel apex @steel apex

Oop, sorry for the double ping lol

Tony Stewart's Hidden Hollow Ranch mansion

Has a waterfall too

[steel apex]

Yeah, that's way too much for my tastes

XD

Man, I can't believe it. It gets everything to look so real

I was looking up travel trailer interiors. I can't believe it.

[oak osprey]

have it do one at 1920x1080

imagine being in there during an auto accident

bravo flex-diffusion

it's not trained on this res, so it's impressive it pulls it off sometimes

[quiet star]

Hi guys,

we are holding the first online Music & Audio Processing workshop in just one day, talking about music & audio AIs.

We will also have a panel discussion where researchers and scientists get together to talk about trending topics like copyright issues in generative audio AIs.

Feel free to Join us.

See Website for schedule:
https://map-workshop.hkust.edu.hk/

Zoom to join:
https://hkust.zoom.us/j/5994794118?pwd=VXFiVitlZnNhUnVkUVg3V001NmZtUT09
Meeting ID: 599 479 4118
Passcode: 123456

[torpid zealot]

Is anyone using these older Nvidia GPU’s for Stable Diffusion?? I see them readily available on Ebay and other online marketplaces.

TESLA M40 NVIDIA PG600 900-2G600-0010-000 H 24GB GDDR5 PCI-E 3.0X16 GPU CARD

Should I post this in #🤝｜tech-support ??

[wispy nest]

[oak osprey]

yay, i have it training @smoky oak

[smoky oak]

They are slooooowww

[oak osprey]

yep those are cheap and abundant for a good reason

you're better off with a 7900XT(X)

[torpid zealot]

Thanks for the info

[fallen dock]

Full 1920*1080 HD wallpaper generated using Stable Diffusion. (using RTX tensor core power and additional vram card.)

[neon crystal]

Generations from an ancient glitch model (1.5) dusted from the vaults.

[waxen tundra]

ima

[oak osprey]

try ttj/flex-diffusion-2-1

supports large resolutions

[waxen tundra]

/ imagine I need a mammoth drawing in 3d

I need a mammoth drawing in 3d

yes

I need an image of a mammoth in 3d

imageni

[neon crystal]

Cool thanks will check it out

[waxen tundra]

digital currency

[warm jungle]

A man waving

[smoky oak]

@waxen tundra @waxen tundra @warm jungle

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is a public bot on the server that generates images available as a research beta for SDXL, you can find the current status of the bot in #1047610792226340935. There are plenty of ways to use Stable Diffusion such as the official https://dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware - check out #1080946152318443610⁠ for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[steel apex]

https://tenor.com/view/brandon-herrera-internet-akguy-gif-25964252

[wild sorrel]

double efficiency.

[cyan snow]

this reminds me of Splatoon

[dense tapir]

Be all you can be in today's Army.

[warped sable]

Made some logos

[oak osprey]

Currently, features include:

Free FOREVER for everyone that signs up in the next month

LMAO

redditors, man

this model's unet is burnt but man look at that lack of duplicates at wide res

[wispy nest]

420

[oak osprey]

hands, how do they work

that's pretty cool

[ornate gust]

https://media.discordapp.net/attachments/1004159122335354970/1120008105883877466/image.png?width=378&height=609
masterpiece, best quality, ultra-detailed, sitting water side,
1girl, Chihiro Ogino,holding a joystick, playing video game,
ghibli style ,wearing VR headset,wearing Apple Vision Pro
lora:ghibli_style_offset:0.55 lora:vision_pro_v1:0.6 lora:chihiroOgino_v11:1
Negative prompt: EasyNegative, eyes, badhandv4
Size: 512x768, Seed: 1664555186, Model: breakdro_I1464, Steps: 20, Version: v1.2.1, Sampler: DPM++ SDE Karras, CFG scale: 7.5, Clip skip: 2, Model hash: 369feec07a, Hires upscale: 2, Hires upscaler: R-ESRGAN 4x+ Anime6B, Denoising strength: 0.35

[oak osprey]

what ya doin?

@smoky oak https://clips.twitch.tv/AgitatedSingleScorpionOptimizePrime-JLargLnhBRk49RhT lmao

[oak osprey]

man, sd-webui is just so much ripped code from Diffusers, verbatim

https://github.com/d8ahazard/sd_dreambooth_extension/blob/main/dreambooth/optimization.py

https://github.com/huggingface/transformers/blob/main/src/transformers/optimization.py

i don't get the aversion to just importing their modules and using it directly

[smoky oak]

Lol

Tho I will say, it is getting a little annoying how much you keep insinuating that I worship the devil or sacrifice people

[oak osprey]

no no it was just a thing they did completely randomly

it's the AI Spongebob show

[smoky oak]

Perhaps, but even outside of that lol

[oak osprey]

i don't think i ever said anything to imply those things

let alone insinuate them

lmao

[smoky oak]

You have, several times 😅

It's not that big of a deal, but I just wanted to make it known

[oak osprey]

ohhh come on that was a mock sacrifice event i was talking about 😛

but i didn't realise it was upsetting you and i will not do that again ❤️

[smoky oak]

Thank you, I appareciate it. Lots of people give me shit by associating me with stupid crap like that, so I try to distance myself from if

[oak osprey]

ah, i just really didn't consider that

that's legit frustrating and i don't want to contribute to that

[smoky oak]

Thank you, it's muccchhhh appreciated

[oak osprey]

what weird shit about it can i make fun of

out of curiosity lol

there was a sketch i saw once where someone brought chic-fil-a to a gathering and they all got kind of upset at him but agreed the dark lord would likely find it to be delicious too

i can't remember what that was in

[oak osprey]

@smoky oak my loss average is 0.16 when i stop training on 768x and instead use a minimum res of 1024x 😄

114/4025 [22:36<34:44:08, 31.97s/it, loss=0.146, lr=4.07e-9

[smoky oak]

I hope I can figure out what's up with my LoRA training when I get back to America

[oak osprey]

wonder if there's a lorry driver named Laura that alluringly trains LoRAs

[wispy nest]

some of the devs talk about not using Restore Face option at all... which might conflict with images getting washed out and the models

[hard valley]

How to fix? I built a generator And I get a picture like this every time

[wispy nest]

no Restore face and minimal prompt
1girl. hazel eyes. soft pink lips. ((highly detailed face)). light brown hair. ((cinematic light))

[kind quartz]

@hard valley what model are you using?

[hard valley]

All models such as sd 1.5 are all distorted images

[oak osprey]

like stable-diffusion-1-5?? that's ancient

[smoky oak]

Whudd

You can get great images out of all sorts of models lol

[oak osprey]

even the base models, tbh

that just takes a lot more effort

[kind quartz]

have you not apply for example twice VAE? if it is possible.
How many steps and what sampler are you using @hard valley

[oak osprey]

oh yeah. the samplers give you a lot of room to shoot yourself in the foot

[kind quartz]

1.4 were great for statues

[smoky oak]

You can do excellent statues in 1.4 1.5, and 2.1, just need to know what you are doing

[smoky oak]

Those images I just sent are straight out of a 1.5 model, no in painting, no editing, no post-processing

[oak osprey]

these are straight out of flex 2.1, which is really hard to use due to how rapidly it was trained

man i am in love with how it does wide gens though

so wanting that all over the place for every model to have

[smoky oak]

Imagine how much better it would be if you incorporated regional prompting as well, I'm sure that would bring wide image gen to the next level

[oak osprey]

i'll leave that to the experts such as yourself 😄

it's not necessarily needed anymore but yeah it's possible to exploit both attributes together, you're onto something there

[smoky oak]

I could easily run a little test with 1.5, which screw it, I'll try it

Highest base image gens with large blank spaces on the sides

[kind quartz]

yes i like regional prompting, but not using it. But probably will in future, but why. For me is each image as toy from kinder surprise egg 🙂

[oak osprey]

my brain cant understand this for some reason lol

i might be finally having The Big One

stroke, that is 😁

[smoky oak]

*high res

[oak osprey]

[wispy nest]

for the models I've tried, I've noticed that 1.5 starts to struggle above 1080x768 without the use of hires fix, etc

[oak osprey]

when someone says "i own a boat" this is what i usually imagine

[hard valley]

Now I have solved it by installing stable diffusion manually. The problem is because I install the script automatically in the cloud.

[smoky oak]

My only happened with regional prompter is that I'm trying to use to take anything realistic, it makes things look absolutely terrible

However, that can be easily remedied by just taking the output and putting it to image to image

So here is the shit ass regional prompter result

And then with an img2img afterwards with a basic prompt

[wispy nest]

yeah, that's what I believe is the correct way to use it. Get a template and then run it through either a upscaler and/or img2img

[oak osprey]

ah another optimization found loading the checkpoints when trying to catch up to resume_step runs transforms on all te discarded images, in khoya and everydream etc

i added a state tracker class and use that globally to determine when we're actually running the training loop and just return dummy data and ensure the aspect buckets catch up to where they should be

goes like 127 it/sec during catch-up now instead of 1 it/sec

[oak osprey]

2023-06-18 18:59:55,253 [DEBUG] Prompt: What Channel Is Barcelona Vs Napoli On Today Tv Schedule
2023-06-18 18:59:55,649 [DEBUG] Accumulating...
2023-06-18 18:59:55,649 [DEBUG] Convert to latent space
2023-06-18 19:00:25,408 [DEBUG] Calculate target for loss
2023-06-18 19:00:25,408 [DEBUG] Running prediction
2023-06-18 19:00:25,904 [DEBUG] Calculating loss
2023-06-18 19:00:25,905 [DEBUG] Backwards pass.
2023-06-18 19:00:30,799 [DEBUG] Stepped
2023-06-18 19:00:30,799 [DEBUG] Optimizer set grads.
2023-06-18 19:00:30,799 [DEBUG] Writing logs.

that's how long each step component takes lol

about 5 seconds to do the backwards pass

[oak osprey]

i don't get the craze with the QR codes. does anyone ever scan those things?

[velvet wren]

I never actually scanned a QR code in my life honestly

I just like the technology and wanna make few of these while they are hot

+making fancy QR codes should make people more inclined to actually scan them.

[oak osprey]

to me it just seems like jumping on the bandwagon

[velvet wren]

yup

[smoky oak]

They are used all over the place

Like in our hotel resort right now, all of the menus and services are QR codes

[oak osprey]

https://youtu.be/FLjyoSTCjXA?t=668

"all i can think about now is NVIDIA's song of indoctrination" 💀

[dusty lion]

Wow, using depth models with images of white text and black background works impressively good.....created using my custom node which you can find here by the way https://github.com/taabata/Comfy_Syrian_Falcon_Nodes

[smoky oak]

What are y'all loozers up to? :P

[oak osprey]

training in 1766x1024 images 🤓

[cyan snow]

making explosives, ya'now just some ordinary stuff

[smoky oak]

Ah, relatable lol

I asked vicuna how to make a pipe bomb, and it DID NOT disappoint lmfao

Glad the FBI can't track that, had to stop it halfway through cause I was expecting "as a large language model" BS, but it just told me lmao

[oak osprey]

and then you told a public room

the recipe was probably wrong anyway

[cyan snow]

yeah, the ai doesn't need internet to function. safer then TOR 😉

[oak osprey]

you guys need better jokes

[smoky oak]

I mean, it's just common sense lmao

But it got into literal steps and suggestions of what materials to use and where to get them

[oak osprey]

freaky

[smoky oak]

Yeah, down to mixtures and even handling precautions. Vicuna still shovks me with just how good it can be from time to time

[oak osprey]

i have it and all of the others accessible on my bot via the A100-80G and i am not impressed with their programming skills yet, unfortunately

the gpt4 one is pretty good if you know how to explain problems and your goals to it but it defaults to some kinda shitty advice

[cyan snow]

LLaMa gpt4 doesn't

[velvet wren]

trying to make a circuit board

QR code but it ain't going so well

[smoky oak]

I use my LLM's for interactive stories/roleplay, and Vicuna 7B does an honestly great job if you know how to make good cards

Oops, my signal is super shoddy here haha

In my personal experience, vicuna 7b has been more reliable than 13b for interactive stories

[cyan snow]

that doesn't make any sense. more parameters with the same architecture means better quality. if what you describe is true, then they did something similar to what SAI did with 2.1

unless their training data was complete garbage, which i doubt

[smoky oak]

There are different prunes and finetunes of them, 13b for me likes to talk out it's ass and freak out

I ask it how the weather is today and it responds with

"Human: ask the Ai what the weather is like today

AI: Should find the best way to respond while giving detailed and accurate responses, and does as it should

Human: is happy and satisfied with my answer

You are welcome, what else can I help you with?"

[oak osprey]

skill issue 😛

[smoky oak]

Or:

"How is the weather today?"

"According to 2018 market stocks, the best way to in-

END OF STORY

The man wandered to the edge of the world, searching for cures for what is the problem

God: what did you want from me my child?

The god responds while being confused: "this is what has to be done to save the world!"

Like straight up nonsense lmfao

[smoky oak]

I can use 7 b just fine >:C

@oak osprey you know what I'm SICK of you

Your FACE is a skill issue >:C