🏞｜general-with-images

[dense tapir]

1600, heh

[smoky oak]

heck yeah, even works on 512x768

[smoky oak]

2.25s/IMG

[dense tapir]

cyberpunk?

[smoky oak]

I am working on something to make firefly users weep lol

somebody made a comparison between bing AI, firefly, and MJ

and I wanna submit an SD gen that clearly blows them out of the water lmao

[dense tapir]

what are you smoking firefly is the bestest out there and I know cause they told me.

[smoky oak]

loool

[atomic crane]

hello

[smoky oak]

MJ

[atomic crane]

HELLO

[smoky oak]

Dalle2

firefly

[dense tapir]

btw, news alert MJ V5 does guns and damn they do them well

[smoky oak]

one of my submissions lol

it shits on the others sooooo bad lmfaooo

the prompt is supposed to mix art deco with cyberpunk

and I think mine also does that the best

I also have this one

[dense tapir]

from what I am seeing from firefly it is more for kids and a comic style. Sort of akin to "Fisher Price My first AI".

[smoky oak]

the sheer carnage that SD wreaks on these other image generators is insane lol

[smoky oak]

right

[atomic crane]

omfg

[dense tapir]

@smoky oak MJ V5

[smoky oak]

baby's first mid image generator lol

[smoky oak]

oh wow haha

this looks like its an image straight out of modded skyrim lol

[dense tapir]

Unlike SD it does guns

[smoky oak]

true

tho SD can do guns wayyy better with a LoRA

but I guess the point is MJV5 didn't cuck weapons out lol

[dense tapir]

exactly

consider how much was cucked out previously this is a HUGE advancement for them.

Why a third hand 😦

reach arounds 😦

[grizzled sage]

isn't midjourney a refined 1.5? i don't think they're training base models over there

might even be a refined 2.x

[dense tapir]

I have a question about img2img and that is why would I stick an image in there then interrogate it and it comes out with a shit ton of noise no matter what? Sort of the same look as if I applied unsharpen to it several times?

[smoky oak]

It's a lot more than that

[dense tapir]

tbh, that is exactly what it is doing is unsharpen so wow

[smoky oak]

MJ has a significantly better Text Encoder

[dense tapir]

yes, it really does

[grizzled sage]

they certainly have an internal collection of extra networks. you can see a distinct one in action with symmetry

[dense tapir]

I still prefer anything I can run locally but I was floored to see it do guns

[grizzled sage]

do we know that? or are they passing it through a processor to decide what the final prompt should be against their training style, so that every prompt deterministically provides complex 75 token long prompts...

I think they're doin the ol 2step. it works though

[dense tapir]

MJ used to be the worst censorious one out there but that may be Bing AI now

[grizzled sage]

the censoring is really nothing of concern. keeping nudes out. oh no. the internet has plenty

optics are gonna be important and we've seen how bad it can look

[dense tapir]

in other news flow is wrong yet again. This had nothing to do with nudes. Hell same prompt SD posted the pic in here and Bing warned me if I do it again I can get banned. Nope, nudes, no nudes, fuck that.

[grizzled sage]

yeah. not a big deal. it's just a filter gone haywire. nothing new.

once again, mountain out of a molehill

[dense tapir]

Yeah, but I wonder how many will get banned, really?

[wispy nest]

hello

[grizzled sage]

if you're trying to push their stated limits, you're asking to get banned from their system. warning seems apt there. i'm sure it's all flagged for review and if it's all content that was false positives, an appeal could surely be had. if not, then screw it. there's going to be dozens of ai services to use.

[dense tapir]

Shoe leather with salt and pepper or shoe ketchup?

[smoky oak]

When I was originally with midjourney, it was all their own thing, then SD came out and they interfaces their stuff with SD, and the quality went up massively while maintaining their very good color processing and such

So I think it's safe to say it's definitely its own thing

But I do believe it works in symbiosis with stable diffusion

[grizzled sage]

the image tagging style that they've been training to is likely highly proprietary and something they've been developing internally the whole time. now with the new mjv5, you're supposed to talk to it like describing the image in a chat like style, instead of a prompt style. its a big reason why i think they take the input and process it into their proprietary tagging style that is kept completely internal

helps prevent reverse engineering of theri models too as it adds another layer of black box to it

[proud birch]

Realistic #https://cdn.discordapp.com/attachments/1004159122335354970/1089118879332184074/IMG202209071809132.jpg

[ruby thorn]

Does anyone know if you can select multiple cells with the latent couple extension? like using 1:2+3 in the "position" option?

[sterile kiln]

Hey guys, where am i supposed to paste my upscalers files?

[wispy nest]

typically stable-diffusion-webui\models\ESRGAN

[sterile kiln]

👍

[true lance]

wild

[glossy herald]

I'm sure it's not real time, but the result still is top tier

[sterile kiln]

Hey Olivio, just saw your video about Latent Couple.
good presentation and good editing 🙂

[dense tapir]

I had Latent Couple working then it suddenly stopped working and no longer obeys me :/

[sterile kiln]

stupid question but, did you tried to update extensions and SD?

[dense tapir]

Yep, everything was up to date. Not sure what happened or if another extension is whacking it

[sterile kiln]

I often did repports on Github for this kind of problems

[dense tapir]

Yeah, I would but I use it so very little it doesn't bother me that much just thought I would mention it. If I need it again I will try a reinstall if it hasn't updated by then.

[grizzled sage]

lol you can lick boots all you want. it is quite common etiquette that when you go into someone else's house, you follow their rules

while i'm anti corporate, anarchism is dumb and i'm not anti authoritarian

[stoic pivot]

Prompt: "crayon drawing, mountain, sunny, cat girl, character orange cat, solo, standing, hands on hips, looking at viewer, smile, open mouth, by CraiyonUnicorn, trending on artstation"

I think it was "trending on artstation" that tied it all together 🙂

[potent hornet]

Guys, any idea what this error might be? I can't launch stable diffusion because of it

[stoic pivot]

I just thought those images was funny. I just wanted to be silly and a bit trolly and wrote a prompt like " Rick Roll troll face cat girl, 8K, 9000k, UHQ, UHD, super detailed, best Quality, even better quality, super realism, hyper realistic, RAW photo, master work, aesthetically pleasing, by Rick Astley" 🙂

[heady mesa]

A stunning 3 - story apartment building, composed of 3 modules, the total soil area 1000sq nestled on a gentle slope in the picturesque countryside overlooking the charming city of Tavira in the sunny Algarve region. Ultrarealistic, With its cutting - edge eco architecture, this building boasts a unique blend of neo tradicionalism and eco architecture seamlessly integrating with its natural surroundings perfectly balanced by the natural materials that enhance the building's eco credentials. Surrounded by lush greenery and bathed in the warm glow of the sunset, cinematic --ar 16:9 --v 5

[indigo scaffold]

I hit random keys and somehow THIS thing came up...

[floral flicker]

I’ve read a clear explanation of what “mask blur” is in inpainting (applies a gaussian blur to x pixels from edge of mask), but can’t find a clear answer on what padding size affects

[wispy nest]

I think that's info around masked area AI will get.
More padding = more context for AI on what it exactly is masking, visually

[floral flicker]

Ahhh, thanks for the explanation

I need to play around with that

[wispy nest]

well, idk for sure

[floral flicker]

I have a tendency to ask questions, then realize I could probably just do a quick test heh

[grizzled sage]

pickapic has fun zero shot experiments. cookie monster as the t-800. other models struggle with this concept quite a lot. it's one of my common zero shot tests lol

the cookinator

i know now why you dunk milk, but it is something i can never do

[floral flicker]

What happens with like [cookie monster | T-800] ?

Those are pretty great

[grizzled sage]

i dont think picapick has that syntax. [] there is negative prompts

pickapic rather

[floral flicker]

Ah yeah, I overlooked that detail

I watched a YT vid on Dreambooth, which began with how to upgrade to PyTorch 2.0 with compatible x-formers. It increased performance. I was expecting all other popular youtubers to have some tutorial to upgrade, or at least some other mention on Reddit or something

Seems like something significant that’s under the radar

This was the vid https://youtu.be/pom3nQejaTs. Instructions for pytorch 2.0 are at 5:20

I’ve upgraded this in SD over a week ago, haven’t had any adverse effects as far as I can tell. Only better performance

[umbral scaffold]

garota linda

I just thought those images was funny. I just wanted to be silly and a bit trolly and wrote a prompt like " Rick Roll troll face cat girl, 8K, 9000k, UHQ, UHD, super detailed, best Quality, even better quality, super realism, hyper realistic, RAW photo, master work, aesthetically pleasing, by Rick Astley"

[glossy herald]

I got a rock roll model if you want

https://huggingface.co/Guizmus/Experiments#rick-roll-v1

[stoic pivot]

Then all love cats I had to make a cat, I made it smile and gave it big glossy eyes and used the tag "naturally cute", I had to edit out the bottom part even it was not dirty.
Oh, and I wrote a bit of text on the black bar.

[sterile kiln]

after a big discussion with the dev of Multidiffusion, he completely fixed his extension, this guy is awesome !
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues/63

[wispy nest]

Points up. Mods please. @glossy herald

[dense tapir]

@glossy herald

TY

[glossy herald]

yep

fudge

[stoic pivot]

TY.

[glossy herald]

seriously

was is happening for long ?

[stoic pivot]

No.

[dense tapir]

Apparently so

[stoic pivot]

Maybe 5 min, you where here fast.

[dense tapir]

I mean longer than it should, yeah

[glossy herald]

I had not had any report before your ping

I have put notifications on any reports

[dense tapir]

I reported it but you got here fast with the mention

[glossy herald]

you reported ?

I tried too

but it didn't land

I'll ask the bot master

[dense tapir]

I reported both of them then I threw up

[stoic pivot]

I am so new with this I did not know how, before I understood other was sending ping

[glossy herald]

ok

you did well anyway pinging us

5 minutes of this is a lot

I'm really sorry

[dense tapir]

I blocked him after I reported but it appears reporting is a bit broken

[stoic pivot]

The worst is, I feel guilty then it was aimed towards me as an reply. I may made someone angry.

[dense tapir]

Naw, take nothing personal from those types

might just hate cats

[glossy herald]

I'm checking with the team. I did tests following this and reporting works OK for other messages. there may be a fail we need to fix on specific ways to post, not sure, but I can confirm other reports are working

[stoic pivot]

Just for that, support cat want to say thanks to @glossy herald

[dense tapir]

Spam doesn't?

[glossy herald]

I'm gonna throw up too now damn, that was long since it hadn't been that dirty

[glossy herald]

not the same bot, some does, some gets by discord, ... I'm no admin, I just see right clicks and reaction ⚠️

[dense tapir]

The second one that was disgusting and it wanted me to star it. Ahem, no. ugh

[floral flicker]

Whoa, looks like an amazing extension

[crystal delta]

@wispy nest

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[toxic moth]

i'll start here then hahah

for reference, using automatic 1111 webui

reloaded my ui so my prompts are lost

but these are the sorts of things i'm getting out of SD

[floral flicker]

Drag n drop an image you generated into PNG Info tab of Auto1111

[toxic moth]

very simple prompts like 'bear logo, minimalistic, vector"

oh wow

thanks!

[floral flicker]

😉

[toxic moth]

yeah that first one is:

a trendy bear logo, minimalistic, corporate, detailed, HD,
Negative prompt: ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, blurred, text, watermark, grainy
Steps: 50, Sampler: DDIM, CFG scale: 7, Seed: 1336646256, Size: 512x512, Model hash: bfcaf07557, Model: 768-v-ema

[floral flicker]

There’s some good checkpoints for vector looking output

[toxic moth]

alright

[glossy herald]

I had that old prompt I used for different sports logo.
What do you think of :

2d ferocious bear, vector illustration, angry eyes, football team emblem logo, 2d flat, centered, highest quality, very detailed
Negative prompt: ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, blurred, text, watermark, grainy
Steps: 15, Sampler: UniPC, CFG scale: 5, Size: 512x512

I kept your negatives

[toxic moth]

I'm going to check that out

[glossy herald]

changed the sampler

and lowered cfg too

[toxic moth]

any reason behind switching sampler?

[glossy herald]

I find UniPC on CFG5 very effective, gets the same results as DDIM but on just 15 steps there

[toxic moth]

I see

just generated one out of interest

[toxic moth]

can someone explain me what's going on on the background hahah

[glossy herald]

hum that's a strange artefact

go higher in steps a little maybe but yeah

[stoic pivot]

All make so good images today, then I come... Giraffe!

[toxic moth]

yeah I keep getting things like these

[glossy herald]

you had that with your method before my changes too ?

[toxic moth]

yup

turned up the steps:

[glossy herald]

what model is that ?

(testing my last training)

[glossy herald]

ho ok

then don't stay in 512x512

go 768x768

[toxic moth]

OH

that makes

[glossy herald]

you are using a 768 model

[toxic moth]

so much sense hahaha

[glossy herald]

x)

it performs quite badly on lower sizes

[toxic moth]

let's try that again now :p

[glossy herald]

^^

let's try on my side, using 1.5

I added a little style for fun

[toxic moth]

nice

i'm on 2.1 rn

[toxic moth]

aye that's a lot better

damn

[glossy herald]

works correctly

nice 🙂

it works with lots of animals

quite easy to change the prompt

[toxic moth]

tysm for helping!

[glossy herald]

also, try removing "ferocious"

no problem 🙂

[granite trout]

@strange jungle an example of a drywall knife

[strange jungle]

Ok, that's a dry wall taping knife\

Looks like SD 1.5 doesn't have that , but you could train a model with it

For what you're doing, you don't need a fancy professional rig. The 8GB is probably enough

[granite trout]

How are you looking up what SD does/doesn't know about?

[strange jungle]

I am generating images

when I use "dry wall taping knife", it comes up with a plain dry wall knife

[sterile kiln]

it is

[granite trout]

Fair enough

[strange jungle]

Even when I use an image of a taping knife, it tries to make it a different type of knife

[strange jungle]

however, what I have can generate and upscale for what you need. It would just be a matter of training the knife into it

8 or 12 can train

I used 1.5 based models. Not sure what a 2.2 model would do with it

[wispy nest]

I think DDIM should be good on 15-20 steps too, I think DDIM is one of those samplers which doesn't really benefit (or even change) from high steps.
I was doing 15-25 on anime

[granite trout]

Interesting, that probably explains why the model tends not to recognize tools in general

[glossy herald]

well, the training worked well

[granite trout]

Common things like hammers don't give it trouble but a lot of things you'd find on a work site trip it up

[strange jungle]

If it wasn't trained with images of tools, it won't have them

8-12 vram GPU system will handle training

[granite trout]

Is there a way I could contribute images to the training data the mainline models use? Seems silly to just do it for myself without contributing to the community

[strange jungle]

there are groups that do that, but I only know they exist, not how to get involved

as far as I know, SD doesn't take contributions for training their new versions

If you trained it with images of tools, you could then upload the LORA or whatever for people to use = contribute

[wispy nest]

AI decided it's not headphones, it's a diadem halfway through 🐧

[granite trout]

Kinda neat tbh

[stoic pivot]

Last image for today, and I do not know what happen in it, but some try to get her ears.

[wispy nest]

pig ears yay

[dense tapir]

Looks like Chester Cheeto is trying to gank the ears off.

[strange jungle]

that was my first thought.

[tidal kite]

just sharing a couple images from a model i'm trying to train on ghibli mixed with old dreamworks 2D animations. a lot of the current anime models have a very similar look to the face so i'm trying to train a model that has more variations. :)

[glossy herald]

that looks nice !

how many pics in that dataset total ?

[wispy nest]

really like the one on the right. very nice

[strange jungle]

are you making that model available to others?

[wispy nest]

reminds me a LOT of atlantis

[glossy herald]

people like animation 🙂

[floral flicker]

Very cool model

[tidal kite]

thanks guys! there's 52 images total, tried to prioritize characters with unique variations in facial features and expressions. i trained it with 5,200 steps, but i'm still learning the best ways to finetune these things.

[toxic moth]

what do you mean by checkpoints and do you have any specific ones you could send my way?

[floral flicker]

Ok so Multi-diffusion extension definitely puts Ultimate SD Upscaler into the trash can

[tidal kite]

thank you! yeah i'd love to make it available to others if it works well. it still produces jumbled garbage sometimes for some reason but if i can fix that id love to make it available

[tidal kite]

Hahahaha! Yeah! my bf actually said the same thing that it looks like atlantis art 🤣 thats the dreamworks style coming through i think

[glossy herald]

1-2 e-6LR?
that feels like good diversity/training without overtraining, but I haven't seen the dataset so hard to say.
Well done figuring this out !

[strange jungle]

models are saved as checkpoint or safetensors files

[floral flicker]

I haven't actually used any yet, I have so many things bookmarked to get around to.
Here is one I have bookmarked for example
https://civitai.com/models/4618/vector-art

[toxic moth]

ooh so .ckt files

[toxic moth]

thanks!

[wispy nest]

Do we have a way to prompt "portrait" without saying "portrait"? I want to get face+ maybe some shoulders, but not upper body.
"Portrait" might trigger painting style...

[floral flicker]

You could just use openpose with controlnet

[tidal kite]

Hahaha yes! I remember seeing 1-2 e-6LR in the colab somewhere, not even sure what that means or what it does but that was included. 🤣 Hahaha and thank you for the encouragement!

[wispy nest]

nah

[strange jungle]

try controlnet openpose with only the top part.

Portrait style

[wispy nest]

I don't want to have control over how it will look, pose or view 😦

[strange jungle]

Try prompts like "portrait style" or "head and shoulders"

[floral flicker]

Well, if you use a veyr short Guidance End, it should start generating relatively where you want it

then let it get creative

[glossy herald]

if you want some guide/definitions on the params, how to do all that, there are some resources runing around

[strange jungle]

or, generate an image of the upper body, then crop out what you don't want

[granite trout]

Prompt: "lithograph of knives"

💀

[wispy nest]

The problem is I have robe in my prompt and usually AI trying to include it too, so goes for upper body
I want to keep it for furr on shoulders...

[vapid raft]

i am new here

what is exactly you do here ?

[wispy nest]

we just chat there, you generate locally or using cloud services
#1072220168534642768 #1072229020520947753 #1080946152318443610

[strange jungle]

Knives, lithograph style

[glossy herald]

hey new, I'm guizmus (and am sorry)

we do pictures

we share techniques and resources

we have dat fun

woooooo

x)

foreveralone in controlnet

[vapid raft]

i think i am in the wrong place

lol

[glossy herald]

depends what you are looking for

[vapid raft]

i was looking for ai painter

just to kill the time

[glossy herald]

this is Stable diffusion, an AI that does pictures

but we don't have a bot for that

there are websites though

here is the faq

[crystal delta]

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[sterile kiln]

Stable Diffusion and its extensions are the best for that

[vapid raft]

thx guys

have fun

[glossy herald]

you too !

[vapid raft]

one more question

how do i delete my acc here ?

[strange jungle]

This image is from a training on images from the Battle of Cut Knife Creek: http://travelphotobase.com/v/CDNSK/SKRP153.HTM

[glossy herald]

you just "leave the server" ? or do you mean on discord itself ?

[vapid raft]

yes

[strange jungle]

"lithograph of knives" got this

[glossy herald]

click on the little gear here

then delete account if you are looking to delete discord account (first time i'm asked that)

[granite trout]

Yeah I’m working with one of the cloud models and this one’s probably not too well trained

Will have to dig out my tower and get a local install

[strange jungle]

looks like 1.5 has a lot of images from lithographs of the battle in its original training set

[granite trout]

Hmm

[mossy creek]

Latent Couple + some inpaint.

[granite trout]

Maybe I need to start researching tagged images of more common household items

Seems like most of the data out there is aimed at humans/animals and fantasy stuff

[glossy herald]

targeted conditionning like that is so strong

[strange jungle]

Someone around these servers looks at the database for the original training - posted a link to where you can search for what was trained into SD

[strange jungle]

six fingered hands?

[mossy creek]

Negative prompt has parameter for that... but what the hell. Thanks for feedback.

[floral flicker]

Anyone had any success using the Region control with Multi-Diffusion extension? Initial tests, my foreground region is just trash sitting on a beutiful background region

[strange jungle]

that's the worst type, when two fingers branch from one - it kind of sneaks in the extra finger

[earnest solar]

Trigger discipline is important to mandos.

... usually

[topaz heron]

Generar una imagen similar a la enviada

[floral flicker]

Surprised to find that Multidiffusion also works with Latent Couple

[wispy nest]

She almost figured out how to hold that sword

[ripe cedar]

I read that the recent update on auto messed some stuff up and extensions don't work properly. Anyone know what's a good version to use?

[wispy nest]

any release before today should be fine

can do 2 days ago to be safe

[dense tapir]

Underworld biker is resting

[toxic moth]

anyone else got some tips for generating logo's? preferably something very simple, clean and corporate

[glossy herald]

controlnet

[toxic moth]

in what way exactly

[glossy herald]

here is an example

I used controlnet, and img2img

canny model for controlnet

base logo :

or I could go a lot more technical and complicated

[toxic moth]

i see

[glossy herald]

depends the liberty you let it have

[toxic moth]

time to figure out how to get controlnet into webui then hahah

[glossy herald]

but to keep the likeliness of the logo, controlnet is almost essential

https://www.reddit.com/r/StableDiffusion/comments/119o71b/a1111_controlnet_extension_explained_like_youre_5/

that's an easy guide

[toxic moth]

lifesaver

[wispy nest]

hmmm

[smoky oak]

What's up dingle hoppers

I wish I had access to Adobe firefly so I could make a comparison on their subreddit just exposing how bad it is

So many people just think it's so damn good, and keep being rude to people using SD/trying to show off the strengths of SD

"oh, but it can't do graphic art designs"

[wispy nest]

Interesting model...

[wispy nest]

graphic art designs like what?

[floral flicker]

The txt 2 vector generation it can do

[smoky oak]

Really, tell me more

[floral flicker]

that's really the meat and potatoes of Firefly

[smoky oak]

They think for whatever reason SD can only do realism in stable diffusion

[floral flicker]

It'll probably be an extension in A1111

[wispy nest]

idk, based on what I've seen it's not just for gen , it's just tons of different tools and photoshop integration

Like...tools literally took from SD community and integrated into their systems

[smoky oak]

Yup, and they aren't even good implementations of them

I just can't believe they marketed it/released info on it with it being so bad

Mmmm, delicious generations out of firefly

[floral flicker]

lmao

[smoky oak]

And the people praising adobe for not using the work from the people they actively screw over with outrageous pricing is... Hilarious

[wispy nest]

yea their model is kinda...reminds me of 2.1 SD model...

[floral flicker]

I mean, at least the kid does look like their offspring

[wispy nest]

lol

[smoky oak]

True I guess lol

Somebody fixed the Firefly gen with SD

[floral flicker]

He has more of his mother in him, unfortunately

[wispy nest]

yea that's better

[smoky oak]

I contributed to a reddit post comparing all the image generation AI's, and don't worry everybody, I made it clear that SD blows all the others out of the water

[glossy herald]

nice

[smoky oak]

The prompt was to get a art deco inspired futuristic cyber city

[glossy herald]

you closed the debate

I knew I could count on you

[smoky oak]

MJ

Dalle2

Firefly

Bing AI

And SD

[glossy herald]

MJ, more like Mickael Jackson, already dead ! AMARIGHT ????

hum

[smoky oak]

Lmfaooo

[wispy nest]

That's weird

[smoky oak]

This is rancid lmfaooooOoOoOoOoO

[wispy nest]

mj 1024x1024, but details like 512x512

[dense tapir]

ngl, MJ, Firefly for the animated look and SD for the realism

[smoky oak]

Here was my other SD one I submitted

[smoky oak]

I used two different realism models rather than art models, TBF

But also, SD is the only one that's even slightly art deco like the prompt said lmao

[wispy nest]

I think most MJ users just dont do as many gens as we do...since it's not as comfortable to use with discord and shit

[smoky oak]

Oh yeah, and MJ is slow

[dense tapir]

I hated that having to use Discord

[smoky oak]

I remember when I used it, even on the paid plan a full high res image was several minutes

[dense tapir]

I never did figure it out way back when and never returned to it

[smoky oak]

Not even including the wait time, just the gen time

[smoky oak]

In all honesty, they actually were really impressed with SD's result, and several people said they were gonna have to look at SD again, cause they didn't know you could get such detailed results

So many people sleep on SD, and it annoys me

Here I am, trying to sleep with SD

😩

Lol

[fickle gorge]

All I want (at this moment) is a telephone, but it doesn't seem to know very well what one is. I expected at least the 12 keys to be consistent, but not even that I'm getting ☹️

[dense tapir]

I recognize an embedding in there but forgot the clothes torn one's name.

[wispy nest]

It's not embedding

wearing black dirty (ripped:1.2) (torn:1.2) cloak, (robe:1.1), gloves,

[dense tapir]

One was made for 2.1 that did that

lol, cool

I wonder if 2.1 can do that?

[wispy nest]

idk

[dense tapir]

apparently not

[wispy nest]

I think I haven't even tried it on base 1.5, I stopped using base models pretty quick

welp

[dense tapir]

Now I see why an embedding was made

a girl wearing black dirty (ripped:1.2) (torn:1.2) cloak, (robe:1.1), gloves to keep it simple to see

torn looks like it sort of worked

[wispy nest]

maybe increasing weight on ripped will do?

but it didn't work on some 1.5 models too

I think rpgv4 didn't do it too, clothes was always too clean

[dense tapir]

Interesting.

Have you ever noticed certain seeds will force two subjects when you prompt should be only one?

[wispy nest]

I think that only happenned to me lately when I tried to do too high resolution

like 1024x760 gen straight away without highres fix

idk about normal gens

cut out right side manually due to a broken hand 😄

[dense tapir]

I changed the seed and 2 subjects gone. Always distinct but it acts like I prompted for two

[wispy nest]

768x768, yea , still can happen from time to time

[dense tapir]

certain seeds, yeah

like how some seeds are always comic books

[wispy nest]

I think my pb was 3 persons at 1024x768 when prompted one

[dense tapir]

I added full body

[wispy nest]

lol

[dense tapir]

I'd shit my pants if that came walking towards me

I mean you see the first image coming then see the second

[wispy nest]

How do you describe this emotion?
eyebrows

[dense tapir]

frown

[wispy nest]

ty

[dense tapir]

yw

[wispy nest]

I might've gone too far lol

but I need to know how it'll look on realistic models now

[dense tapir]

Which model is that?

[wispy nest]

https://civitai.com/models/7371/rev-animated

Update on "screaming" on realistic models...it looks terrible..

[dense tapir]

😦

You know it probably to prevent bad people doing bad things easily.

[wispy nest]

I don't think it's on purpose , it's just looks bad

mouth opens too wide making it just...ew...

[dense tapir]

never know

[wispy nest]

and on some images it's just an open mouth , while whole face has no emotions lol

[fickle gorge]

One thing to test would be lowering the weight of screaming

[dense tapir]

yeah, try that

[wispy nest]

I think for realistic it's better to do angry or something instead of just screaming

[wispy nest]

it looks like it's trying to do screaming in pain, not anger

[wispy nest]

That's a bit different emotion tho, my image was more like worried, this one is sceptical

[grizzled sage]

well i was describing more the physical brow movement. it's over a range of emotes

[wispy nest]

hm, aight

[grizzled sage]

any wrinkled brow is furrowed. i learned it from old poker playing navy guys

[ripe cedar]

Furrowed brow

[grizzled sage]

mix an match for differennt exciting results lol. furrowed anger, furrowed sadness, furrowed joy

[ripe cedar]

The problem is how to control facial emotion to a fine degree

[wispy nest]

That's probably dumbest thing I ever created 😄

[grizzled sage]

furrowed brow, joy, happy grizzled anime man

[ripe cedar]

If it's sequential art, how would one control it to add a little more or less ie 10% frown vs 13% from panel to panel? I suppose it's still something AI people are working on

[dense tapir]

Frown lines.

I hate them but they come to us all

[ripe cedar]

I tape my forehead with duct tape so it's not possible

[dense tapir]

Super glue works too and can shower

[wispy nest]

okay, yea...
angry scream works on realistic models too, but takes alot more attempts to look somewhat decent...
Add sword problems on top of that , hand problems and eye feckery and here we are....
Really hard to get decent image without inpainting...

[wispy nest]

since we're posting spooky stuff, here's "the nightmare king" i made about a month ago. had to remake it because i had gotten rid of the original but still had the generation info. never got around to doing a proper upscale on it, but with what's available now could probably turn out pretty good

[fickle gorge]

Is the... VIUIE's keypad fixable somehow? It seems there is still trouble with numbers and letters

[wispy nest]

You can get rid of it with inpainting, but numbers...I don't think they make alot of sense for AI...idk if you can do it correctly

[wispy nest]

it can occasionally get them perfect, found that out when i was attempting to make victory stands (the 1,2,3 places people stand on after an event). most of them were pretty much trash but a few got the numbers perfect. but yeah...seriously hit or miss.

[dense tapir]

Yeah, 100%

[wispy nest]

occasionally yea, but with so many things that can go wrong, that'll take alot of time

probably good idea would be to finally do controlnet, but naaaaah

[grizzled sage]

thinking something like a lora could be made for specifically that level of tuning.

[ripe cedar]

Maybe the equivalent of blendshapes/shape keys

[grizzled sage]

there is that new face marker contnrol net. or was that just somebody using sketch mode with markers?

https://www.reddit.com/r/StableDiffusion/comments/11v3dgj/new_controlnet_model_trained_on_face_landmarks/

[smoky oak]

I have been playing with SD and new styles of generations, and I am pretty impressed at how granular and high quality the control over the style and the quality of the final product are

[fickle gorge]

Ok, I'll let two SD instances generate forever. I hope they can get a at least a single good result 🙏

[floral flicker]

Just started an X/Y/Z plot using MultiDiffusion extension enabled - was wondering why the ETA is only counting upwards - it keeps doubling the dimensions of every generation lol

Image # 4 is about to finish with dimensions of like 12k x 12k

actually nvm I think it's just at a standstill

So there's a little checkbox "keep input image size" which I unchecked and now is running the XYZ plot as originally expected

[smoky oak]

I recommend ultimate SD upscale over multidiffusion

I have tested both a ton, and I just get faster, more consistent, and higher gross resolution results

I made an image that is 15630x6528 in less than 20 minutes on an 8GB 3060ti

[smoky oak]

I wish somebody would make a more granular controlnet for poses where we can actually do proportions and such

[floral flicker]

The one thing about SD Ultimate, is I kept having visible seams

None of that is happening with this

[floral flicker]

Seen this? https://zhuyu1997.github.io/open-pose-editor/

[smoky oak]

yeah, this isn't what I am talking about at all

doesn't really help to get unique proprotions

[floral flicker]

Each segment that is clicked, has its own big list of parameters

bone thickness, lenth, etc

Unsure how much more control you could get than that

or do you just mean, controlnets execution of the openpose input

[smoky oak]

yeah, that is cool and all, but still not what I was meaning. I was meaning a controlnet where it adheres to extreme proportions, rather than fixing them

like you can give somebody a huge head and it gens them with a huge head, or a really long torso and stuff

[floral flicker]

ah

[smoky oak]

I would love a way to control their torso proportions as well

like hip width, waist, shoulders, all that

[light gazelle]

my first creation, prompt was about a soviet propaganda poster

dont really know what it means

but its kinda cool

[spring kelp]

any way i can get a decent upscaler? amd 6600

[light gazelle]

a thing ive noticed with all ai image generators is they really shit the bed when it comes to text

even when you tell them exactly what letters to put

[floral flicker]

1340 x 2048 in 2m 20 sec, on RTX 4070ti using Multidiffusion + controlnet

And, Surreality is a really cool checkpoint

[dense tapir]

I wonder how long on a 4080 and 4090?

[floral flicker]

basically same, but with noise offset 🙂

[wispy nest]

still nooby even after doing this for awhile a few months back, doing what I can with a weak gpu lol

[smoky oak]

Daily reminder that some models just aren't good at some things lol

This is a specific prompt using photorealistic dreamshaper

Absolutely horrific results

Versus the result I got out of deliberate

Some models work, some don't lol

Looks like dream shaper doesn't work for this prompt yaha

[ocean citrus]

cat on ice

[mellow dagger]

Red rose

Red Rose#

[glossy herald]

hello 🙂

no red roses nor cat on ice, no bot for that around here

[crystal delta]

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[grizzled sage]

name that control net source photo

click for answer:

[smoky oak]

@dense tapir So apparently SDXL isn't even 3.0

its just another addon to 2.2

[grizzled sage]

my branding theory is that each version going forward will potentially have xl variants

[dense tapir]

Looks like it.

tbh, I saw NOTHING, save for slightly better text, that we can't already do with loras, embeddings.

[smoky oak]

I am messing with it a bit

[grizzled sage]

i was reading that lora creator thread on reddit today and was wondering, what if someone trained a lora on videos, like every frame

[smoky oak]

it does seem to do massively better, at least on dream studio... but thats cause the other base models are trash

[grizzled sage]

he asked that and i wonder if he was hinting at capability hes glinted

[smoky oak]

@dense tapirok, have you seen the 3 new SD 2.2 models?

2.2.2 is very interesting imo

[dense tapir]

Yes

Only the third one intrigues me

[smoky oak]

it prompts completely differently

HmMmMmMm

it seems we can only beta test 2.2.2 now

[glossy herald]

I think comparing it to the base model 1.5 is a way to try to estimate the potential it can have
like, 1.5 on its own wasn't perfect at all, but loras and all make it a beauty for example, same to a lesser measure on 2.1 that took off slower.
So if 2XL seems to perform massively better than 1.5 out of the box, it could mean it could be trained to even higher standards than what we have

It mostly depends on community's engagement in perfecting what will come out (and when it will come out too I guess)

[smoky oak]

That is what I am hoping for big time

just really hope that quality isn't locked behind another terrible text encoder ._.

[glossy herald]

but I think other things could push the community to change

[wispy nest]

Does it understand prompts better \ knows more words?

cause there are lots of issues with that on 1.5...

[smoky oak]

it has wayyyy more parameters

I think its 2.6 billion now? IIRC?

[wispy nest]

that's interesting

[glossy herald]

like inner structure/more params, leading to just more possibilities for the model.
Imagine if the models could store more data for real and be in 2048x2048 base, with better prompt understanding, more token limit, .... Even if the results were initially worse than current 1.X 2.X, people would invest big time in it because of the potential

[wispy nest]

my gpu can't do even 1024x1024

[glossy herald]

(I'm not saying 2XL has any of those to be clear : I have no info on that)

[smoky oak]

unfortunately, that would also be massively restrictive

[glossy herald]

yeah that's an extreme and stupid example

let's say it would run on 128MB VRAM too

[smoky oak]

SD 2.x should have been so good, but to say the ball was dropped is... an understatement

[smoky oak]

so now we are going going Absurd

lol

[glossy herald]

I'm just making the point that massive tech upgrades in the base model would be a real push towards them, even if the results are initially bad

[smoky oak]

Yeah, huge agree

[wispy nest]

yea

[smoky oak]

I would use 2.x and deal with all of its bountiful issues... if it worked as good as 1.5 for training/listening, but its a massive leap back and thats offputting to me

[glossy herald]

2.1 didn't lift off because of this imo. 768 is an upgrade, but it wasn't enough at the time to convince people

same for 2.1 training, never managed to work with it correctly

but even today, the incentive for a trainer is to use 1.5 because people do mix

[smoky oak]

The issue is training LoRA's and other things off of 2.x's horrible text encoder is a nighhtmare

That is what has lead to me not even looking at 2.x

[glossy herald]

I tried, but yeah, I gotta say the repeted fails when it came out, when I was feeling like "i knew what I was doing for training", that participated to my burnout and break for a while

[smoky oak]

I do have to say, the text legibility is a good sign

That is damn good and consistent text

[glossy herald]

yeah, it reduces the need for controlnet for that

[wispy nest]

oh wow that's clean

not perfect,looking at "A" , but still

[smoky oak]

so if thats something we can expect to translate into more fringe things like facial expressions, then I am stoked! assuming its not ruined by another lackluster TE

[glossy herald]

doesn't always work (2XL)

[grizzled sage]

Can always make an extension that detects generated text and then cleans it up a bunch

ie : we'll fix it in post

[wispy nest]

try with other actual words, maybe the reason behind it is cause it has some kind of visual representation for this word

[smoky oak]

to be honest, all of the base models are trash on their own, and they rely on the tuning, because they are not tuned for greatness... but to see that SDXL isn't doing that bad at all out of the box is... hopeful

[smoky oak]

if that is SD 2.2.2, its an aesthetic fork of 2.2 XL, which takes what you say as a description rather than a tag

AItrepreneur shows it off, how SD XL tile, and SD XL beta will do text of what you say, but SD XL 2.2.2 will do a stylized prompt around it

his thought is it works similar to midjourney, which I could see, based off of my very little use with it

my only massive concern for SDXL is the size... and the VRAM contraints for training

[smoky oak]

I must say, these watercolor sunsets look fucking great

[dense tapir]

Yeah, if I can't train it then it is DOA

[glossy herald]

I'm on Dreamstudio, not on next, it's libeled "SDXL beta"

[smoky oak]

ah. We don't know which one that is

[wispy nest]

yea it looks like it understands "written" , probably we'll see someone doing embedding or something on it for better results, interesting

[smoky oak]

I am using the old dream studio, which states its 2.2.2 XL beta

[dense tapir]

that is the third one

[glossy herald]

yep, it wasn't out when AItrepreneur did his video

[smoky oak]

The third one is a style, more than a base model... but not the same

I am trying 2.2.2 beta

[dense tapir]

I dunno but what I saw in his vid I was not all that impressed. Just too many things that must happen that if they don't then most people will stick with what they are already using

[wispy nest]

Does it have same problem as 1.5 has?
Like if you type "white boots" - AI applying white not just to boots, but to everything around, unless you specify other colors?

[glossy herald]

maybe on the public DS they auto select the 2XL model depending on the selected style/prompt?

[glossy herald]

an advertising board on a city bus stop with the words "Soon" written on it in pink

[smoky oak]

ok... I must say

[wispy nest]

yea it looks like it is still applying similar colors to other elements

[smoky oak]

I am seeing some massive improvements in some areas, and I do not wanna get too excited if there is a chance they drop the ball with the TE but...

[wispy nest]

but it\ll be easier to tell with humans and clothes

[smoky oak]

SDXL seems to stop all the others with stylized simple prompts

[dense tapir]

2.x TE sucks

[smoky oak]

the prompt is all

"A pencil sketch of a forest"

1.4

1.5

[dense tapir]

I saw a lot of 1.5 in his vid, btw. Far more 1.5 in it than 2.x

[smoky oak]

2.0

2.1

SDXL...

I think the results kinda speak for themselves in terms of detail pulled from simplicity

will be interesting to see if that means we need short prompts for the same detail, or if we will be able to leverage same length ones for massive detail improvements

[glossy herald]

an advertising board on a city bus stop with the words "Soon" written on it in pink, a blue car drives by, a white bird is on a bench
CFG10

xD

[smoky oak]

ok yeah, SDXL is killing it for simple style prompts-

[smoky oak]

"Water color painting of an autumn forest"

1.4/1.5/2.0/2.1/SDXL

[glossy herald]

changing theme

[smoky oak]

Like I think its safe to say SDXL is winning big wins here

[glossy herald]

a photo of giant hammer falling on a city, with the word "BAN" written on it, RAW photo, 8k

[smoky oak]

Any style prompts to compare?

trying to think of styles

[glossy herald]

I'm using the "photographic" style

[smoky oak]

cause so far SDXL won both of them by a lot IMO

[glossy herald]

let's try cinematic

tiny head

[smoky oak]

I will try one right now

[glossy herald]

wow that one is good

[smoky oak]

"Digital art of a cyberpunk city"

1.4/1.5/2.0/2.1/SDXL

[wispy nest]

To expand on this and why this isn't what you always want:

If we want to apply style of some kind to only one element - it won't work.
Like a simple example, if we'll prompt a car with one deflated wheel - it will apply "deflated to all wheels.
More then that - if there's something else on the image that can be deflated - it will be deflated.

Same goes for clothing, if we want let's say just a sleeve of a person to be ripped in a fight or something - that won't work too, cause it will try to apply "ripped" to everything that can be ripped.
prompts aren't enclosed by "," too, so separating words with it doesn't work.

Only overspecifying all environment \ clothes details fixes it.

---

With coloring it is not bad most of the time, cause it keeps image in same pallet and usually it makes image more visually appealing to an eye, good for design perspectives too, but not great for little detail changes and accurate prompting

[smoky oak]

ok, ok, ok

I am getting really hyped for SDXL...

And I am worried I will be let down

but so far, its impressing me compared to the other bases

[wispy nest]

yea, that's interesting base

[smoky oak]

"A picture of a pretty woman in a pink dress in an autumn forest at sunset"

1.4/1.5/2.0/2.1/SDXL

Like its another clear win IMO

SDXL did by far the best. It listened well, it has the best pose, and is the most aesthetically pleasing...

HmMmMmMmMm

please for the love of god, do not ruin it all with a bad TE 😭

[wispy nest]

TE?

[smoky oak]

Text Encoder

The part that converts what we say into things the AI understands

guys...

[grizzled sage]

thats the part that makes it XL i thought. has way more parameters in that

[smoky oak]

SDXL is... Shocking me

its listening alarmingly well

[wispy nest]

Where can I test it?

[smoky oak]

"A photograph of a husky puppy sitting on the beach at night"

1.4/1.5/2.0/2.1/SDXL

Its the only one sitting, and its the only one visibly at night-

[wispy nest]

How good is it at holding things?

hands, eyes?

[grizzled sage]

The usual test of willem dafo terminators is doing good

the base anime mode looks good, but it doesnt do mecha versions of deloreans

all the novel ai anime derived models can pull off delorean mecha

the results are great and i hope to see it refined. i wonder what it will take

[smoky oak]

"polygon art of a frog"

1.4/1.5/2.0/2.1/SDXL

SDXL is crushing the accuracy on all of my prompts-

Like, it has not missed a single time

[sterile kiln]

Repported Clip_Vision with T2i_Style was not working on MultiDiffusion, he seems to be interested in it, we will probably be able to use Clip_Vision with MultiDiffusion! Top !
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues/66

[ebon gate]

Top is 2.1, bottom is sdxl
"Ice cream candy"

[wispy nest]

What the heck is sdxl

[smoky oak]

its the new stable diffusion model

its being beta tested early

its stable diffusion 2.2.2

[wispy nest]

ah

Is sd2 still way more demanding to run

[smoky oak]

@dense tapir SDXL is showing a massivveee improvement of prompt understanding...

statue of liberty from behind

From below

from above

at night

being demolished

like its listening pretty fucking well

[wispy nest]

make a kawaii baphomet

[smoky oak]

As a satanist, I approve lol

[rocky saffron]

EF85mm f/1.4L IS USM, 5:4, pretty, woman, legging, brunette

[smoky oak]

@dense tapir SDXL can do weapons...

really fucking good ones

VS 1.5

[wispy nest]

looks like a handheld grenade launcher

for toddlers

[smoky oak]

It can even do shit like red white and blue braided hair

"A photograph of a Native American chief smoking a pipe by a campfire in the middle of the desert at sunset"

@dense tapir

SDXL is insanely accurate to what you ask for

"a photograh of a 3 layer purple princes birthday cake with the number 16 at the top"

[sterile kiln]

@glossy herald x)

[smoky oak]

"A photograph of a man with sunglasses, a cowboy hat, blue denim jeans, a red plaid vest, leaning against a beat up pickup truck at sunset with a hand on his hip"

I cannot believe how good SDXL listens....

[glossy herald]

Ho yeah ! Super Jacques !

[smoky oak]

@glossy herald SDXL is insane

[glossy herald]

Super bien fait !

[sterile kiln]

haha, merci

https://www.youtube.com/watch?v=7ejrN0NexaY

[wispy nest]

Lol it learned text

[smoky oak]

it learned a helll of a lot more than that

and yeah, it can do very damn good text now

it blows the other base models out of the water man

[static tusk]

what about a macro photo, super zoomed in

[fervent rune]

"Macro Photo of a ladybug on a half eaten leaf"

[static tusk]

not too bad. missing a lot of detail on the bug

[wispy nest]

damn, that actually looks even worse then 1.5 object styling 😦
colors all over the place, even specifying each object didn't help

[stoic pivot]

Updated.

[sterile kiln]

creepy

[foggy dome]

beauty studio advertisement

"beauty studio advertisement"

how can I get a image here?

[wispy nest]

You don't, you can install SD locally.
We don't have discord bot

[crystal delta]

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[foggy dome]

Ok. Thank you

[glossy herald]

thanks Soul 🙂

[wispy nest]

We really need a way to |enclose| part of prompts with styling, to apply certain words only to a part of prompt

SDXL

ripped, torn clothes doesn't work on it 😦

nvm, kindaaaa works on some images

[sterile kiln]

linda, secretary in the podiatry service, warrior in the Xena series on Sunday April 22

[ruby thorn]

what do you mean?

[blazing crow]

That 4x Ultrasharp Upscaler that is going around recently really is something...

Combined with HiRes passes it does create some crazy stuff.

[wispy nest]

I mean so if I I say man in a |pink hat| - I get only pink hat exclusively, everything else should NOT be pink or with pink shades mixed in.

SD tends to apply colors\ styles and well, words overall to whole prompt \ image.
Not for a specific words.

syntax is just and example, it's not a thing

[ruby thorn]

I agree, that would be super helpful

[ebon gate]

Oil painting of a cyberpunk samurai, magnificent, elegant, beautiful, dynamic lighting, killian eng, ocellus, theme park, fantastical, light dust, elegant, diffuse, grimmer, intricate, light dust, orange and teal contrast volumetric lighting, triadic colors, splash art

Midjourney got some competition

Same prompt but without orange and teal part

[wispy nest]

eh, idk, sdxl isn't doing well on somehwat complex prompts
But if it actually better at understanding words I'm sure we'll get good community models later on

[blazing crow]

It does look stunning that 4x Ultrasharp Upscale in Automatic 1111

[spring kelp]

how do i install this upscaler to webui?

[blazing crow]

One sec let me look up that link again... give me a min

[spring kelp]

youre beautiful

thank you

thats some very goos upscales wow

[wispy nest]

it's not good for photorealistic images, unfortunately

[spring kelp]

thats fine I'm aiming to get good at anime prompts

[blazing crow]

With Realistic Vision it actually does some amazing stuff

[wispy nest]

idk, based on my tests it makes faces look not so realistic

materials are fine, but faces not as good

[blazing crow]

You can Download the pth file from here: https://mega.nz/folder/qZRBmaIY#nIG8KyWFcGNTuMX_XNbJ_g
Just plunk it down into your models/ESRGAN folder and reload the WebUI then you should have the Upscaler ready.

[spring kelp]

e z p z

thank you so much

just plunk the pth file in and thats it?

[wispy nest]

maybe with low denoise...like 0.1, but at this point it's better to use any other upscaler

[blazing crow]

Yup nothing more then restart your webui and it should run it's course, it'll be available as a Upscaler during creation and in Extra's as a regular upscaler as well

I usually do about 20 HiRes passes with it and that more then enough otherwise it can sometime go a little freaky.

[spring kelp]

awesomeee appreciate it, i just got this set up yesterday so still looking/learning about this tool's capabilities

[blazing crow]

But like @wispy nest says it is better on Artsy stuff then Photography, but you can make it work though...

[spring kelp]

mmhmm~ Im not a big fond of super realistic stuff, if its realistic its tasteful paintings, i love that kind of art

[blazing crow]

As with everything AI Diffusion, the only limit is your limited imagination... once you see what it can do and your imagination goes apeshit with it the results can and will be stunning!

[glossy herald]

Here are some little tests I was doing on the mix model of the last 7 PoW ( #⭐｜pow-info ). total dataset is around 330 pictures currently

more there : #1089542080801542207 message

[spring kelp]

very nice upscaler

[wispy nest]

"SDArt someone" prompt fpr 5th image or is it not reading png info properly?

[glossy herald]

it is surely reading it properly there yes

I didn't do all combinations but this model is trained a little differently

[wispy nest]

hm, aight

[glossy herald]

I trained on 7 PoW dataset at once

so I added a token to name the PoW

"cats", "underwater", "loop", "someone", "something", "somewhere", and "friend"

I'm not entirely sure where to go there.
What I can see, is that individualy, each submission is less well trained that it was in each PoW model

but at the same time now, we got some really great tokens working

in particular, the main token is great and very diversified

and the users that did participate in multiple PoW have also very good tokens to use

I got that comment on civitAI on one of the models

to make merging easier with other models u might want to consider adding a color rich tag, and perhaps just art . Would really help those of us merging

[rose pike]

blob:https://discord.com/994f63df-7b08-4e56-97f6-52ad46e654f7

[glossy herald]

for the "just art", we are good since SDArt splits in SD and Art, token wise. but not sure what they mean with "color rich tag"

[glossy herald]

blob to you too and welcome

[wispy nest]

I don't understand that whole thing of what's happening

You made a model based on #1023999442338201721 submissions or what's happening?

[glossy herald]

well yes

I made 7

[wispy nest]

What even PoW stands for 😄

[glossy herald]

Picture of the Week

and for 7 different week, I made a model, that lets people have fun mixing their creations together, doing variations, and each time creates a "style" token that fits the theme

but here, I took those 7 datasets, and trained them all at once

[wispy nest]

interesting interesting

[glossy herald]

it makes quite an artistic model, where you can call on specific themes of the PoW, or on specific users (each user has a unique token too)

[wispy nest]

but I don't get how'd you get anything with it with so little options to prompt

anything specific*

[spring kelp]

pow, right in the kisser

[glossy herald]

well you are right in part, you can have a hard time getting something specific.
there are 174 tokens trained in total in here, and almost none mean a real thing.
Mostly "SDArt" is trained and applies a style to the rest of the prompt
it's an interesting token since most pictures in the submissions are quite high quality
but it's mostly for fun and doing variations on the submissions of each user, or mixing submissions of multiple users at once

you can run lots of small prompts like I did, and get a lot of good results, select some tokens that you like, and mix with a "real" prompt that will use the styles of the token you choose

[lone anvil]

Had some fun generating Chinese princess :

[wispy nest]

hm aight

[tall haven]

A accidental dweller from aladdin

[jovial frigate]

Behold... idk what this is since I made this while I was drunk a few days ago

[wispy nest]

That's more like a princess 😄

[dense tapir]

As long as I can train with it with the ease of 1.5 if not I have no use for it.

@smoky oakAs we know the 2.0/2.1 TE fights you every step of the way in training it. Same data 1.5 loss 0.05 while 2.1 0.4-0.5.

A magnitude larger is nothing to overlook.

[stoic pivot]

I tested Bing, id do not allow me to generate images, so I asked how a tag would look for Stable Diffusion and it suggested.
ADJECTIVE: Cute
NOUN: Hamster
VERB: Holding a wooden rake and standing in a field
STYLE: Cartoon

I think it had been cool if AI could use a structured format like that to separate things, like Character1: Long hair, blue eyes... and so on.

[onyx anvil]

Hahaha love it

[wispy nest]

You can make chatGPT do that

[grizzled sage]

used to look through my phone at pics i took after a night of drinking that i forgot. now we're in the 2023 and i'll look at pics i synthesized while drinking

[stoic pivot]

ChatGPT is online and made by a company I do not trust for now, so I hope the innovation happen within Stability and the opencource.

[wispy nest]

I mean...that's just prompts, nothing bad will happen if they'll see your prompts 😄

technically bing and other companies using gpt too

so pickapic.io is using sdxl now?
https://www.reddit.com/r/StableDiffusion/comments/11vu1n8/pickapic_is_now_using_stable_diffusion_xl_22_beta/
Does an ok job with what little testing I've done. Here's the best "Professional heavily stylized digital illustration of a beautiful ginger freckled female fantasy cosplay mage with freckles, grey eyes, and a closed mouth. She is 20ish years old. She is wearing modest long blue robes with many small details and is standing in a a dark alley in a fantasy city with many details. She is looking at the viewer." imo., though there were lots of different styles shown

[outer spear]

Its not the best for illustration styles, as compared to some of the community models, however if you compare XL to regular 2.2 its such a huge leap

it can do all artstyles pretty well already, fine tunes with XL will be amazing

probably midjourney level

[wispy nest]

Yeah. I was just impressed that it followed a (relatively) natural speech prompt so well. not everything was that good, obviously, but really nice.

not sure if it always uses sdxl though, as i'm under the impression it should output 768x768 if it does and quite a few are 512x512

[outer spear]

SDXL is only 512 right now looks like

i dont know why

and yes

pickapic switches between models

[wispy nest]

another pretty good one, but 512 and didn't follow the prompt as well imo. bet it didn't use sdxl 😛

[stoic pivot]

Just ust the prompts "8K", and "UHD" to overcome the 512x512 limitation 😄

[wispy nest]

definitely didn't work lol. though it did completely change the artstyle EDIT: nvm, was just a fluke. lots of the previous artstyle in the generations as well

[stoic pivot]

He he, and it did not make it better, I like the first one better.

[thin jolt]

image in the Old West, where two men are facing each other, one of them handing a stone amulet to the other. The image has a focus of mystery, with the men's faces detailed and very close to the viewer. The men are seen in profile, one of them is about 30 years old and the other about 70 years old.

[wispy nest]

yeehaw

[light gazelle]

ai animation

[crystal delta]

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[wispy nest]

When 8 fingers actually doesn't look bad 😄

or whatever those things are

[light gazelle]

what prompt did you use to get her to hold a plate of burnt food lmao

[wispy nest]

that wasn't on purpose, that was just burning part, SD often makes human hold something burning instead

[light gazelle]

trange

[wispy nest]

funny accident ¯_(ツ)_/¯

if you are trying to get an image with this - we don't have bot.

[modest geyser]

Ok

[crystal delta]

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

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[ripe cedar]

stderr: ERROR: Error [WinError 2] The system cannot find the file specified while executing command git version

ERROR: Cannot find command 'git' - do you have 'git' installed and in your PATH?

Should I upgrade the pip? It won't make the new version of SD install, will it? I don't want to use the new updates SD yet because it has some problems

[wispy nest]

Do you have git installed?

I think at some point it suggested to add a PATH variable

[ripe cedar]

I only installed python and SD

[wispy nest]

you can type
git version
into cmd to see if it is installed

[stoic pivot]

I sometime write sarcastic prompts to make fun of the promp-myth, but even the words often are pintless they may add something, this is "(Rick Roll troll face) (red nose:0.3) (cat girl), 8K, UHQ, UHD, super detailed, best Quality, even better quality, super realism, hyper realistic, RAW photo, master work, aesthetically pleasing, by Rick Astley trending on artstation", with Stadle Diffision 1.5

[wispy nest]

idk if there are reasons to upgrade pip, in most cases newer versions are better, but it also might lead to compatibility issues with other things

[ripe cedar]

I typed git version in the Command Prompt and it says 'git' is not recognized as an internal or external command

[wispy nest]

yea git probably not installed

[ripe cedar]

I never had to install anything else before to use sd, I'll look up git and see

[wispy nest]

It shouldn't be necessary if you aren't going to update it tho, idk why is it asking

[ripe cedar]

2.40.0 is the most recent apparently, should I download?

I see

I'm ok with manually updating SD every few weeks, I wouldn't need git pulls

I'll look up the errors and try to find out

My python is 3.10.6, should I use an updated 3.10.X?

[wispy nest]

Made an idea based on this

young girl feeding big massive demonic creature with glowing eyes, 

I'd want to know where did she get that thing she's trying to feed creature on img2

[wispy nest]

demon girls > cat girls

[wispy nest]

yea...okaaay...almoooost what I wanted...lol

[ripe cedar]

Wouldn't learning a bit of blender be really helpful? Getting a few 3d objects right where you want them and then using depth or img2img would go far imo

[floral flicker]

3d modelling is a bit tricky to just get into

[floral flicker]

But yeah, it’s very good for depth + controlnet

[grizzled sage]

I think another clever part of the midjourney situation is that they're the central host and control all the content. So Their model may very well likely 6 every image that's generated their aesthetic filter doesn't rate high enough. When they generate 4 varieties, the person feels invested in picking the good one. Only shares the good ones and the rest are all tossed.

This affects perception quite a lot. I say this because people always say a very high quality peice is "MJ level" but.. i see those and they just look like any other post coming out of a place like /highendai or other high quality generations from any model.

It's important to distinguish because people seem to think that "Midjourney level" is impossible on stable diffusion, but again, highendai exists and is FULL of very quality renders from people using a variety of tools. Many are SD. Why are we insisting on using a trademarked brandname SAAS to describe "good art" ? Brand names don't deserve such pedostles

Jimmy. She's an Ad Jimmy

[ripe cedar]

Getting the error

No module named 'fastapi'

I looked it up and it says the fix is some !pip install stuff. Where would I paste the code to install?

[dense tapir]

in the root folder of webui but make sure you activate the venv

[grizzled sage]

command line into the root webui folder. .\venv\Scripts\activate will activate it on windows. then paste your command

[ripe cedar]

How do I know which is the root webui folder? There are a few webui things on there.

[wispy nest]

/stable-diffusion-webui/ - that's root

[ripe cedar]

I'll ask in tech support, don't want to clog this chat. Sorry, I didn't notice tech support chat before. Ty though

[smoky oak]

I am terrified that SDXL may be massively better, yet be completely useless because of the bad TE

[hollow lion]

Am I doing something wrong? Prompt: dark gray stone floor jagged

[smoky oak]

But man, when you say something... SDXL listens

[dense tapir]

Yes, the same feeling. Not massively better though but better while having no real way to train it means fuck it as far as most are concerned so they will stick with 1.5. Matter of a fact Iam about to do a test. This is a do or die test too.

I have had no luck at training 2.1 Lora, Locon (never tried loha) for styles. I had no issues with embeddings so my last test from this morning I am going to do with 1.5. If it works then screw 2.x

Unless they change the CLIP model from the one they currently use, or get one even worse, training will be a sob at best and not doable at worst.

After having played with the new models of 1.5 since 2.0 dropped there is just no need to upgrade short of getting full on 100% of the time text. SDXL can't even do that though it is better at it.

[smoky oak]

I don't know what to say, I tested it across 50 different prompts, and while every other model did not get more than a quarter of them accurate, SDXL got almost every single one perfect.

I spent a long time last night fucking around and trying SDXL compared to all of the other versions, and I would have to say that the quality difference between 2.1 and SDXL is miles higher than the quality difference between 1.4 and 2.1

[dense tapir]

I saw your examples and I saw nothing an embedding wouldn't already do.

[smoky oak]

Yes, but what I'm saying is is that it listens so well that it's able to do it without that

Imagine if it's able to keep that level of granularity, where I'm able to tell it exactly what color of vest, shirt, hair, pants, shoes, socks, and I color without them getting jumbled

[dense tapir]

yes, and that is why I suspect we are screwed for training it

well, as I said, if I can't train it then it will lose me for sure and a ton of others who train 1.5 already. They couldn't train 2.0/2.1 and stayed with 1.5 and said "eff 2.x". I stayed but I am beyond frustrated with it, as they were. I really do not think it will get any better for training and just sitting here generating images is not my idea of fun. A lot of Anime people say the same thing.

Anime folks are even more vehement about it.

I guess cause they have their favorites and need to train them in.

btw, this latest Auto1111 has so many errors, and bugs, that omg even an A6000 with 48gb of vram got an OOM.

bad mem leak somewhere

[smoky oak]

Oh lord

Glad I haven't updated

[dense tapir]

I did 😦

it is bad I actually, honestly, feel claustrophobia in it. Very weird.

[smoky oak]

damn ._.

I just woke up, I am gonna keep trying to push to see just how well SDXL listens/understands compared to only 1.5, as doing 1.4/1.5/2..0/2.1 is quite taxing on my credits and time

[dense tapir]

isn't it due out this week or is it next week?

[smoky oak]

I have no idea

I just have a lot of credits, and these gens only take .2 credits, so I wanna play with it

[dense tapir]

When people train on 1.5 which model do they train off of nowadays?

[smoky oak]

For me I use realistic vision 1.3, as its my go to model

[dense tapir]

instead of 1.4?

we are talking about loras, right?

[smoky oak]

I haven't trained since 1.4 came out, but they are extremely similar

yeah, LoRA training

[dense tapir]

I use 1.4 for gen but I do use others as well as each has strengths and weaknesses

[smoky oak]

my 1.3 LoRA's work flawlessly on 1.4, so I would assume the same in reverse

[dense tapir]

I need to upload one to train test 1.5. If this works with nothing else changed I am going to be super pissed

do you train with a vae?

[smoky oak]

I do not

[wispy nest]

@smoky oak

[smoky oak]

whut

[dense tapir]

this has a dropdown to choose one