✨｜sdxl

[nimble heart]

blood's all congealy and patchy when it's spread out like that

[stone fossil]

Show me. 🙂

[hardy cipher]

I used it to make taylor z'dar swift

that workflow only uses one ipadaptor though. prompted taylor swift. had the image of mr z'dar

[nimble heart]

I haven't bleed in a good two weeks so I dont' plan to change that if I don't have to

[stone fossil]

😄

[hardy cipher]

steven z'dar seagal

[nimble heart]

just like look up an image of someone skinning a buck in their garage

or I guess there's a war right now probably tons of vids on liveleak

[hardy cipher]

I use it to make quality things

[nimble heart]

hm

wonder if I can find my shrek rainbowdash image

[hardy cipher]

I'd strongly suggest you check out the ipadaptor stuff. if you're trying to be a purist with the nodes make your own I guess

[nimble heart]

Not XL but since we're doing cursed images, I merged conditioning for Shrek and Rainbow Dash

left one is probably my favorite SD 1.5 image lol

[hardy cipher]

lovely

[nimble heart]

it's perfect ✨

[hardy cipher]

yes, cursed is a good word

[nimble heart]

I have a more cursed one

druggie shrek hit another ogre with his truck while driving under influence

[hardy cipher]

oh, I have countless cursed images

wow, that's moderately disturbing

[nimble heart]

thx it's the image that made Dreamshaper my defacto standard 1.5 model

[stone fossil]

Made this 2 days back orso quite cursed.

[hardy cipher]

https://github.com/tencent-ailab/IP-Adapter

[nimble heart]

It's like a witch saw a girl with that big eyes and lips snapchat filter then cursed her so her face contorted to form it in real life

[hardy cipher]

got this beauty a couple days ago

[nimble heart]

so im not the only one making cursed pony hybrids

I should write that conditioning interpeter node and make a new one with SDXL

[hardy cipher]

just made that with ipadaptor

[nimble heart]

nah that's just shrek as a basement brony boy

[hardy cipher]

cursed

mildly cursed

[stone fossil]

Very sexy images.

[hardy cipher]

looks like he's been huffing paint

[stone fossil]

What better then a coffee and some cursed images on Sunday. 🙂

[stone fossil]

Cursed pony 1.5.

[hardy cipher]

https://tenor.com/view/dont-like-that-gif-12989661

[nimble heart]

if I have two timestep ranges would it be a concat or a combine so they work together?

combine

[hardy cipher]

concat is a fairly new concept to me

[nimble heart]

combine works like you'd expect

[vital ermine]

XL decided to PS the head on, lol

[hardy cipher]

so what's the deal with ipadaptor outputting black images? what needs to be adjusted?

[nimble heart]

ogre ponies kinda working now

cool thing with comfy is you can overlap the timesteps

[hardy cipher]

Hmm. Haven't messed with that before

[nimble heart]

that's what I want to automate with some interpretation of keywords/brackets from the prompts

tedious as hell to do with nodes

3 nodes for every split

or 5 for timestep

[hardy cipher]

Yeah. Needs to be more nodes that make things efficient

[nimble heart]

you need separate prompts, conditionings, and combines for every equivalent BREAK, + two timesteps for every equivalent square brackets

turns into spaghet real fast

its ok AMD i didn't want to use my vram anyways

[nimble heart]

look at you doing actual research into the cyclops bias situation meanwhile I'm just making a Sonic the Hedgehog shaped Rainbow Dash

[nimble heart]

no metadata on the cyclops image

sus

[hardy cipher]

What's using up all that vram in that setup?

[nimble heart]

amd

[nimble heart]

it uses like 2.5x the vram do do anything because it's libraries are incomplete

[nimble heart]

thought I think that specific time was it just not unloading any models cause I technically didnt exceed the limit

1024 batch size 4

[nimble heart]

some of them almost don't look schizo

[hardy cipher]

Ahh. That's a significant difference. Sucks they haven't caught up

[nimble heart]

it depends what it is. over in the LLM world where flash attention and all that fancy shit isn't core it gives the 3090 a run for its money

[nimble heart]

but on Stable Diffusion yea it turns my 24 gigs into 10

[hardy cipher]

Do the models load multiple instances in that situation?

[nimble heart]

?

[hardy cipher]

Never mind. You explained it a couple responses later

[nimble heart]

yea 4090 bois can do batch size like 36 or 49 on XL without hitting memory limits while I can barely squeeze in 16. For LLMs though I compiled a patched fork of AutoGPTQ and my results were almost identical to Sytan's 3090 in our initial tests. Mine was actually using a little less memory cause lightweight desktop environment

[hardy cipher]

need to learn more about what sucks up vram

[nimble heart]

depends

[vital ermine]

How does anyone do a BS that big? not in training and I don't have anything that goes that high, though comfy I never go above 4.

[nimble heart]

sd 1.5 resolution absolutely bombs your vram, so running img2img at like 4k is impossible for me

sdxl resolution doesn't bomb it as much but the VAE spikes it harder

[nimble heart]

just like set it to 36 see what happens

do science

commit crimes

[vital ermine]

I use AIT 4 is it

[hardy cipher]

yeah, what's up with the vae thing? I've notice that

[nimble heart]

I'm not fucking with the AIT nodes till they're officially added in comfyui

[noble shoal]

Wow

[hardy cipher]

that kid has some tooth decay going

[noble shoal]

Yeah, that seems to be his problem. Everything else is fine.

[nimble heart]

you can notice it on 1.5 too if you hit really large resolutions

all of a sudden it takes like 30 seconds to decode

[shy kelp]

what is the different between -O and -P before the path?

[nimble heart]

idk check wget -h to see what the flags do

[hardy cipher]

I just made some vae nodes. still can't say I completely understand what they do though

but I get sick of having to run vae's to each encode and decode

[nimble heart]

you get used to it...

[hardy cipher]

and don't want to download a bunch of nodes just for that. I know others have done it

[nimble heart]

I added pixel scaling to my Principled node and had to include the vae so it could decode then immediately re-encode

[hardy cipher]

I think these latest cursed images are too cursed

[willow bane]

hey hello i am reinstalling my stable diffusion due to some errors can you tell me from where i need to install for latest version(github)

[nimble heart]

just saw my gpu hit 380 watts on "silent" mode can you not pls

[willow bane]

can you tell?

[nimble heart]

?

can I tell what

[willow bane]

.

[hardy cipher]

so you mean a1111

[nimble heart]

idk why you're pinging my message about watts for that

also it just hit 401 watts why

[hardy cipher]

thought you were talkinig to him

speaking of watts, I tried to get chatgpt to give me instructions on putting together a dual a100 laptop

it seemed to tihnk it wasn't a reasonable idea

[nimble heart]

its not...

I mean

[hardy cipher]

I know

[nimble heart]

unless you used external enclosures

[west breach]

am I dreaming? is comfy history now in reverse order?

[nimble heart]

yes

its so you can quickly load last

he finally caved and changed it after being asked about it a bunch of times lol

[west breach]

oh thank god, been wanting that for so long 😍

[willow bane]

are you guys using webui or local installation?

[hardy cipher]

you'd need to carry around a 30 lb lithium battery to power it

[nimble heart]

sonic the ogre

[nimble heart]

those are the same thing?

[willow bane]

i am getting a lots or error

thats why i am reinstalling

[hardy cipher]

well I mean, there are several programs that can run stable diffusion

[nimble heart]

we'll need a whoooole lot more to go off of than that

like OS, what interface you're using and how it was installed, gpu, what do the errors say, etc

[willow bane]

download forom here again

i am using rtx 3060

but thats for laptop

just want minor stuffs to be done

[nimble heart]

use the 1-click installer auto has. beyond that not sure what else to do.

lot of windows users in here have been having problems with nvidia drivers so maybe update that too

[willow bane]

i updated that

[nimble heart]

did the problems start right after?

[willow bane]

no

wait i send a screenshot

[nimble heart]

also cause it's Auto disable all extensions see if that helps

[willow bane]

error with extension

[willow bane]

i want that

[nimble heart]

update the extension or disable it

[willow bane]

from which software you create these beautiful images

in sdxl?

[nimble heart]

I like the where's waldo collage lol

[hardy cipher]

like this?

[nimble heart]

ComfyUI or Auto1111 1.6
either one will work

[willow bane]

is auto1111 and sdxl same?

[hardy cipher]

my friend

[willow bane]

[hardy cipher]

you should read about what stable diffusion actually is

[nimble heart]

auto is a program to run models

sdxl is a modle

model

[hardy cipher]

vs a1111

or comfy

[nimble heart]

ay kid you got any of dat speed

[hardy cipher]

yes. what he said

[willow bane]

sry iam from unreal community i donot know much about

i just use stable deforum

hope ai in future creates game also

[willow bane]

which would help us

[nimble heart]

long ways off from a whole game

already lots of games with AI generated assets though

mostly textures, sprites, backgrounds

[willow bane]

yes

but the creativity is not so good in games till

okk trying to install sdxl

[hardy cipher]

I think AI will pretty quickly improve npcs though

[thorny frost]

npc doing weird things would be nice

[willow bane]

yezz that would be cool

[nimble heart]

@willow bane
automatic1111 install: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs
automatic1111 sdxl after install: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#sd-xl

[thorny frost]

totally unscripted game. Nobody knows what will happen

[nimble heart]

I mean you can kinda do that with an LLM already.

[willow bane]

thanks

[nimble heart]

play like procedural Zork

I do that on KoboldAI

[thorny frost]

yeah, but with explosions

[nimble heart]

write a daemon that reads the output log and just flashes a .png explosion on your screen every time someone dies

[thorny frost]

i use kobold oobabooga and such but usually cant find any use for local ran llm

[nimble heart]

kobold you can make your own locally run AiDungeon

if you remember that app

[soft zealot]

@visual glade FYI another one I've noticed since yesterdays conversation.

Using git checkout 0e3b64117218c50a554b492269f5f35779839695 this Tiny Terra Nodes Debug works as expected ie prints to condole plus in its own window (1st Image)

Using anything newer (currently on ComfyUI Revision: 1397 [a74c5dbf]) the Node in the workflow no longer displays the debug info although srill in console

[nimble heart]

ooglyboogly is good for dicking around with chatbots

or just like writing smut i guess

[thorny frost]

noice

[nimble heart]

imma slepp it's literally 3 am here

have an AI geck

[thorny frost]

my native language is not english (should be evident), maybe that is the reason i dont enjoy llm that much

in other languages it doesnt work as well as english

[nimble heart]

not even close

and all the fine tunes will be even more english biased

cause the tuners are all english

[thorny frost]

yeah

[nimble heart]

damn brits colonizing everything

[thorny frost]

or even if they are not, they will do it in englsih because it is the damn standard

[nimble heart]

yea

every programming language is also english

[thorny frost]

i demand all llm to be written in basque!

[soft zealot]

so is Air TRaffic Control

[nimble heart]

basque sounds like a made up language in a homebrew DnD campaign

[hardy cipher]

we really need to just revert back to latin

[soft zealot]

what I dind annoying is they default to bloody American English ;o)

[hardy cipher]

I used latin to bypass bing's guardrails. didn't work the next time I tried it

[nimble heart]

american english is the more proper english

britain is the one who developed an accent post-colonial era

america is more old fashioned

[hardy cipher]

I realized it was doing web searches for the latin I was sending it

[soft zealot]

American English is the result of one mans laziness compiling a dictionary :p

[hardy cipher]

rather than just translating it

[thorny frost]

modern basque is made up in some sense. It is a mix of amny local basque variants made in the 19th century. It is called batua.

[nimble heart]

minus the colour -> color and defence -> defense shennanigans

[hardy cipher]

and apparently by the time it came back with the web search results it was down for whatever

[soft zealot]

I will concede that from an histroical perspective aluminum is more correct thatn aluminium 🙂

[nimble heart]

not the first time the british have conceded 🎆

[west breach]

didn't know they spelled defence like that. 😐

[nimble heart]

yea

the original defence spelling is such because it's originated from fence

[thorny frost]

did you use inbreed family in the prompt?

[nimble heart]

so fence defence

[west breach]

defenzee

[nimble heart]

then the dictionary guy changed it to defense cause its easier to spell

[soft zealot]

in the Caribbean this is de fence

[hardy cipher]

just used a couple of my existential bean nightmare images with ipadaptor

[thorny frost]

it worked

[nimble heart]

inbreeding worked..?

[thorny frost]

it worked to generate a nightmarish image

that from a prompt i found in civitai

[nimble heart]

tomorrow I'll have to learn this ipadapter thing everyone's hype about

feed it that image of Pepe the frog giving anal birth to Hank Hill and see what magic it creates

[hardy cipher]

I need to learn more about ipadaptor myself

[nimble heart]

feed it the pepe Hank image

maximum learning

[willow bane]

why torch is not reconizing my gpu

the first time i use it was using

[hardy cipher]

bruh

[willow bane]

but now this is not

[hardy cipher]

one moment

[nimble heart]

post it on Twitter or something saying "artist's rendition of Trump at the trial" get lots of internet points

[zinc cargo]

using @ionic gulch inpaint on the upcoming NightVisionXL 🙂 gotta love it

[nimble heart]

the super soaker assassin

[zinc cargo]

generated

[nimble heart]

loading curse [#### ]

[hardy cipher]

lol

I had to change it up. the images were so cursed they kept coming out as black squares

need to figure out why that happens

[nimble heart]

is it black in the whole preview or just when saving

use the fully qualified fp32 vae if it's saving

[hardy cipher]

so they'll only do that when finishing sometimes?

[nimble heart]

if it's a vae precision issue then it'll turn black when saving

the SDXL vae turns black if autocast to fp16. I know there's a "patched" one on hugging face but even that turns black on img2imf or goofy resolutions

[hardy cipher]

ahh. well I don't even know which vae I have. "sdxl_vae"

[nimble heart]

gotta use the full 0.9 vae and leave it on default settings for comfyui

in auto you probably need to manually set full precision vae

[hardy cipher]

I'm assuming that's not the 0.9 then?

[nimble heart]

idk

you tell me

[nimble heart]

does it give rainbow barf lines on some textures? that's the 1.0 vae

[hardy cipher]

yeah, not sure why you'd know. but no, I haven't noticed that at all

I forget where I downloaded it though. probably huggingface

[nimble heart]

zoom in on eyes n stuff

[hardy cipher]

but it's weird' several in a row might be good then several black ones

of the same everything too

other than seed number

the black squares are actually the reason I started thinking about the conditioning node strength node

seems like messing with those sorts of parameters would help

[willow bane]

should i install CUDA?

[hardy cipher]

if you want to use the program

[willow bane]

so yes ?

[nimble heart]

Don't know. Not used it on windows

probably yes

[nimble heart]

yea VAE probs

it looks fine in the live preview right?

download the SDXL 0.9 vae from stability'a hugging face and see if that fixes the borken seeds

[hardy cipher]

downloads

[nimble heart]

?

[hardy cipher]

lol, meant "downloading" as in the 0.9 vae

[west breach]

nice texture there

[uncut fiber]

very heavy impasto, inspired by Thaevil1

[low hill]

Cool that the smoke comes out of the glass, he was a magician also hahaha

[thorny frost]

maybe he is drinking chlorhydric acid

[uncut fiber]

This is spreading colors by knife 🙂

[stone fossil]

Yes.

Try something like: A hyper abstract macro close up shot of ((a lady standing on a mountain)) facing the camera, using very heavy brush strikes and pallet knife phat thick paint, dripping paint effect, using only the colors red yellow and black.

Force colors also.

Maybe some abstract, surreal theme? 😉

[uncut fiber]

will try something thank you for suggestions!

[stone fossil]

And tailor your - only for what is needed.

txt, signature, multiple people, dupes

Is what I like when doing abstract and people.

Also cracking paint is nice to add stacks nice with the knife.

[uncut fiber]

i am just testing and playing 🙂 Going test.
@stone fossil do you think it is posible to achieve droste effect? I mean picture in picture in picture......

[stone fossil]

Not sure what you mean by that.

But I think that is kinda what I have been doing with people wearing hats with worlds in them.

Like this one.

We create a subject in a subject.

Anyway sky is the limit. 🙂

One thing I have also noticed is AI loves effects that are not real.

[uncut fiber]

i will show you link on wiki

[stone fossil]

So for example:

Electric galaxy blur
Diamond prism glow
Rainbow smoke trail
Neon watercolor burst
Pixelated firestorm
Rustic kaleidoscope swirl
Glitchy lightning burst
Icy water reflection
Cosmic flower bloom
Grunge confetti explosion
Luminous dust cloud
Frozen crystal mist

And you can make up anything lol.

[uncut fiber]

@stone fossil i think it loves not real things but not as photo imo.

[stone fossil]

🙂

[willow bane]

anyone plz solve it

[stone fossil]

😄 noice.

[uncut fiber]

thats that droste effect but not generated just link to it

@willow bane better to ask in #🤝｜tech-support

[willow bane]

asked

[ionic dragon]

damn, it got pokemon right

[mossy canopy]

pathfinder?

[low hill]

Look at his fingers, does Steven have Parkinson's?

[low hill]

Try first to add --xformers in your webui-users.bat and then check

[willow bane]

@low hill like that?

[stone fossil]

What GPU u try to run it on?

[low hill]

Guess RTX 4090 ;D

[ionic dragon]

#🤝｜tech-support

[stone fossil]

Then whats the problem should work out of the box.

[uncut fiber]

@stone fossil i think gpu is in his nick 😄

[stone fossil]

xD

U do not need them in ur nick but under ur belt. 😉

Here started to judgle with multiple GPU's in 1 rig a bit back.

It can be useful.

[ionic dragon]

damn, your 3090 consumes 412W

[stone fossil]

Can do 420.

I love that number.

Also ive caught it misbehaving peaking above but its fine.

Temps are fine fo sho.

[stone fossil]

Is that a pregnancy test? If so whats the outcome? 😄

[soft zealot]

Very Nice, reminds me of something I did in 1.5 a few months back

does anyone else have this (mildly) annoying issue with ComfyUI ?

https://www.youtube.com/watch?v=5MuVu3J0AtE

[mossy canopy]

nope

[stone fossil]

Hook up up scotty.

[ionic dragon]

@stone fossil where do get innovative prompts from?

[stone fossil]

My head. 🙂

[strong field]

You recognize this one?

[maiden gale]

hey guys

Had a quesiton

I'm looking to img2img with sdxl, but I need the "denoise" option for the advanced Ksampler

Can't seem to find it

[native knot]

Damn, dude...what did Tom Hanks do to warrant this?

[strong field]

Advanced sampler uses the “start at step” as the denoise in a way.

[maiden gale]

yes I've read up on it now

denoise = (steps - start_at_step) / steps

Ok yeah thats better thanks

[sand pier]

You can just math it via comfy nodes

It's handy enough

[maiden gale]

What I didnt understand is that is start at step=0 %100 denoise?

sure doenst seem like it

[strong field]

How many steps do you have going into the inpu?

[maiden gale]

It's 25, at 20 refiner kicks in until the end

[stone fossil]

here 40 / + 10 or 50 / + 20.

[strong field]

Got a screen of the bigger workflow, sounds like not enough steps to inpaint

@deku

[strong field]

Do you use the refiner?

[stone fossil]

Ofc.

[cyan crown]

I've made some loras with good results, but they are soooooo big. 1.7 GB. do you know why?

[strong field]

Lol, I mostly don’t

[stone fossil]

Well I check the base img that pops.

[strong field]

That’s a fine tune question, I’m not good with training

[stone fossil]

If it sucks I wont even let it hit the refiner.

Cancel it nothing is saved.

Even with a batch.

[strong field]

Same

[cyan crown]

these are samples from 3 different loras. They are very good and Lora works great. But they are so big...

[strong field]

These look really good

[cyan crown]

thanks

[native knot]

Warning: Dark humor ahead.

Gotta head down to the store and get some stuff.

Also available in a box.

...or you can pick one up from the generic aisle:

[noble shoal]

Looks like this one is already defect

[native knot]

That's what you get for buying generics.

But at least it's sealed.

[noble shoal]

Do they also have sealed kittens?

[orchid pendant]

why when I generate some images in the bot, it suddenly appears as a BIN file to download?

[noble shoal]

Found that. Do you think its still good?

[orchid pendant]

??????

[native knot]

I found this one at the local store for ya.

It's a multipack, so probably saves a few bucks.

[noble shoal]

I think most of the users in this channel are generating locally. Did you asked in the bot support?

[hardy cipher]

Remember bonsai kitten?

[orchid pendant]

I'll go see

[wet nacelle]

https://tenor.com/view/christmas-time-gif-12873218

puss in piss

my autism strikes again.

[wet nacelle]

@hardy cipherCan you believe that we were all gushing over dall-e and dall-e 2 last year. Now we can make better and faster results than both MJ and dall-e 3.

Dall-e 3 is a thing in closed invite access.

Bing does have a great image creator though.

[hardy cipher]

I can't think of anything that's progressed faster than text to image ai technology

[wet nacelle]

It uses dall-e but also is just better than dall-e 2 in all ways.

[native knot]

The moment the concept turned the corner into something that could be developed further, it was bound to take off like a rocket. I keep trying to tell people that they need to get on-board immediately or be left behind.

[native knot]

Pretty much all generative AI is going to be on that path for the near future.

[hardy cipher]

parabolic

[ionic gulch]

love playing around with multiple img2img techniques. with multiple steps you can do nice things with img2img, ip adapter and inpainting. copy & paste (clipspace) makes it easy to move images between output and input to refine the image in multiple steps.

[hardy cipher]

well llms are the same I guess. and yeah, other generative ai

[wet nacelle]

the llms for me are hard to get into. I really want to check stuff out but when I do it takes up a huge amount of storage and can easily be installed wrong. @hardy cipher

[native knot]

I need to figure out how to code something for ComfyUI that allows for an addition to the main menu. I've got this idea that I think everyone would love.

[wet nacelle]

Visions Of Chaos has a neat download structure though.

[hardy cipher]

I rage quit and deleted my attempts several times

but I hate giving up

[wet nacelle]

I actually think that Visions Of chaos should've popped off like all the other UIs

[hardy cipher]

only ever goot oobabooga to work

but didn't really try with the others

[wet nacelle]

For me I did the latest Lama. It was simple to set up. I got it running but knew that I did something wrong when it all it did was answer with dumb shit.

[wet nacelle]

be careful with that or else the mods will take it down like they did with my "slight nipple" post

[ionic gulch]

good way to use my workflow:

1. generate some images with all four options (base, base+refiner, both upscalers) turned on.
2. if you know if refiner helps for the style you want to do turn off upscalers and base or refiner
3. create images fast until you find a good one
4. use the recycle button on the seed node to get that seed back and generate an upscaled version
5. use copy & paste (clipspace) to copy the image to genimage input
6. use inpainting to fix details or if you want to have different images in the same style use img2img (test different strength settings) or ip adpater with a text prompt that describes the most important elements of the image

[hardy cipher]

inpainting workflow?

[mossy canopy]

woolen stuff?

[wet nacelle]

crochet

[ionic gulch]

yes, added inpainting 1-2 days ago:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows
IMO my solution gives much better results (no / much less artefacts) than the example workflow provided with the model. didn't see any other workflow that does that much to improve inpainting yet.

[hardy cipher]

sweet. I'll check that out a little later

when I have the time to do such things

[cyan crown]

with automatic1111 what is a good number of steps? 40 ? 60? 80?

[hardy cipher]

150

I actually changed mine to allow for 500 steps

[cyan crown]

......

[hardy cipher]

in all seriousness, it just depends for me

the more intricate something is, the more steps I'll tend to want

longer prompts, in my experience, seem to do better with more steps

[cyan crown]

thanks. And wich sampler do you prefer?

[lusty wolf]

Is this with ComfyUI?

[hardy cipher]

always go for the one with the words/letters. that means "best"

I actually don't know. was always a euler a guy

some of the change the image with more steps, others more refine it rather than change structure

some of them are definitely faster than others, or slower. heun takes like twice as long as the others per step

but it's good for things like landscapes I believe

[cyan crown]

I've doubt between 2M and 3M Karras

[ionic gulch]

i'm using a lot of steps. 60-80 for euler a, 80-100 for dpm 2M, because images get better and that part of the process is quite fast on my 4090. vae and other things take more than half of the time generating images.

[hardy cipher]

JPS, you know about ipa adaptor. why would I render a few images fine, then all the sudden with the same configuration get 5 or 10 completely black images?

[hardy cipher]

you should check out my vae encoder/decoder nodes when I upload them. maybe not breaking new ground with them

but they load the vae in the same node, and singleton so only one instance of it

[ionic gulch]

never had black images on my workflow/system. biggest problem i have with it are extreme aspect ratios like 21:9 or 9:21 - those move the main content to the side of the image, because the interal calcualtions of ip adpater somehow use square aspect ratio all the time.

16:9 works, but also moves the content of center most of the time

[hardy cipher]

it's one of those things that I believe has a solution, but I can't find any info on

I mean, all software issues technically have solutions

but you know what I mean

was thinking of making some custom conditioning modification nodes just to mess with the input of things to the sampler to see if that's the issue

because no errors so hard to really pinpoint what's going on

accidentally flipped the positive and negative last night. it's better than what I was actually trying to make

[cyan crown]

Is there any way to fix hands like we can do with faces ?

[ionic gulch]

@cyan crown you can use inpainting and just redraw until you get good ones for example

[hardy cipher]

in past experience the most efficient route was to just use photoshop or even paint to add proper hands

[ionic gulch]

control net pose also has a hand feature, but no yet available for sdxl afaik.

[hardy cipher]

and then use inpainting or something to integrate them

[cyan crown]

controlnet!! I'm also waiting for qrcode monster for sdxl

[ionic gulch]

or you can add different hands in photoshop or some other app and use that as base for a soft img2img to merge them with your image

[hardy cipher]

I find it pretty interesting that it flat out refuses to do cyclops unless really finely groomed to do so. but it creates fingers like they're just texture or something. just keeps adding more if the noise looks like it wants them

[cyan crown]

for photo yes, but for drawing the only solution seems inpaint

[mossy canopy]

16 dpmpp 2m karras

[hardy cipher]

7 steps

[vital ermine]

No, but loved how it came out.

[hardy cipher]

be a mad lad

[cyan crown]

16 gives bad results

[hardy cipher]

low steps work with lots of bokeh in my experience

but if you want the whole image to pop and be in focus you need more

[ionic gulch]

if you finetune strength settings you can merge different styles with img2img, canny or depth control net, but if you only want to change the hands and keeps everything else the same inpainting is the best, if you can't draw the hands yourself.

[mossy canopy]

7 steps

[cyan crown]

I'm starting again using Automatic1111. But hires fix seems working bad with sdxl

[ionic gulch]

the highres fix is basically just a predefined img2img + upscale process. you could do img2img yourself and have more control over the strength and other settings. this also has the advantage that you can only do it for images you like and really fine tune them, instead of having that highres fix option turned on all the time (as i did myself in the beginning).

[cyan crown]

Yes i was used to do with img2img, but scaling 2x with automatic1111 takes half an hour, while with comfy oly 1 min

[noble shoal]

I love my new crappy newspaper photo lora

[ionic gulch]

sounds like a vram problem. 2x shouldn't take 30 minutes unless you run into shared vram memory. maybe try the sd.next fork of a1111 - that one does some things different. maybe it works better for you with your amount of vram.

[cyan crown]

12GB 4070

[mossy canopy]

100 steps 4x upscale?

[cyan crown]

80 2x

with comfyUI 1 min or 2

with auto1111 20/30 min

also generating images is slower

[ionic gulch]

why do you want to use a1111 if comfyui works better for you?

[cyan crown]

Somehow I prefer results coming out from Auto1111

[ionic gulch]

there are some nodes that try to replicate the prompt handling of a1111 - maybe that helps. but i don't know for sure - haven't had time to check those nodes or prompts in detail myself.

[mossy canopy]

why make your workflow so complex? why not use comfybox?

[stone fossil]

Yes.

[ionic gulch]

haven't had time to check the various comfy frontends yet and if one of them could replace my menu system without giving up features or usability. how does it handle moving images between different workflows? would i be able to use my current workflows and just replace the menu system or are there limitations?

[lusty wolf]

Love it ... Cool stuff. Your own workflow?

[stone fossil]

Yes and lora.

[mossy canopy]

you can send to img to other load images in other workflows in comfybox easily

[ionic gulch]

@mossy canopy and i can add an unlimted number of own smaller workflows and move outputs from one to the other easily and turn every kind of input to a web based menu option?

[mossy canopy]

yes.. you can check their default workflow.. and you will know...

[ionic gulch]

might be interesting. will have a look at it if i have a little bit of extra time.

how does it compare to similar solutions? afaik there are 2-3 of those frontends availabe for comfy.

[stone fossil]

I was gonna releaase them hehe, till some smart ass told me lora's are over rated.

😄

Now they are for ever locked up.

[lusty wolf]

You can release it to me... I dont tink loras are obselete... I love the use of them.

[upbeat summit]

ahmm... overrated? no 😄

an integrale part of fine-tuning and expanding capabilities

[native knot]

I like the challenge of trying to get a certain outcome without them, but they are still necessary in many cases.

[mossy canopy]

i just tried comfybox.. dont know about others.. ill have to check

[upbeat summit]

yeah of course it's cool to see how far you can get with prompting, exploring this is really great. but LoRAs can introduce new content that just isn't there - not just guidance.

[ionic gulch]

those different solutions were one of the reasons why i haven't check any of them yet, because it looked like a lot of work to find the best one, that will succeed in the long run.

[native knot]

Absolutely. It's just a matter of preference and needs, though.

[upbeat summit]

exactly. I do sessions with just a model and sometimes I go all-in and stack LoRAs to re-configure the latent space.

I see LoRAs as ingredients

[native knot]

I always shake my head when someone starts asking things like "What's the best <insert x>?", or "How can I get this image to do <insert x>?" Because the answer almost literally always involves a massive amount of random chance combined with a selection of choices. It doesn't get "easy" or "simple". LORAs can help, but they're not a silver bullet unless one's need/preference is very broad or unless the LORA is very specific.

[mossy canopy]

i got nada.. which frontends are you talking about other than comfybox?

[lusty wolf]

Who asked that in this conversation?

[native knot]

Where did I say someone did?

[upbeat summit]

yeah. they are tools (addons with AI super powers) that are like blueprints and can give some guidance to the neurons in the neural network. or used as creative ingredients to explore and create interesting things

[ionic gulch]

doens't stable foundation have their "own" solution now: https://github.com/Stability-AI/StableSwarmUI
and i think there is at least one other i don't remember the name

[lusty wolf]

"I always shake my head when someone starts asking things like "

[native knot]

You left off your own context of "in this conversation".

I didn't say anyone did in this conversation.

[noble shoal]

Oh boy do i make garbage ingredients

[upbeat summit]

it's fine

[lusty wolf]

So don't generalize... al lot of people don't know where to start and they ask...

[upbeat summit]

you got your style going on. always interesting to look at 🙂

there's beauty in many different things

[native knot]

don't generalize
lol, what?
My comments weren't aimed at you; I wasn't talking to you until you replied to me.

[steady grove]

some of my favorite art is garbage themed. like trash can kids

[lusty wolf]

I'm refering to to everyone that comes here. Not all is as clued up as you seem to be... Just drop your snide remarks.

[mossy canopy]

yep found them.. i played stable swarm just once. when it was first lauched.. you can try my comfybox space with linked comfyui .. if you want to try.. i can dm you the link..

[steady grove]

he was actually anti generalizing. wishing people would be more specific when asking questions like "whats the best way to do <x>? " then there is no best way. They need better specific phrasing about what they're looking to accomplish

[ionic gulch]

don't have much time at the moment, but sure you can dm it to me or i'll come back to you later, if i have more time to check it out.

[native knot]

What are you even talking about, dude? You literally replied to me. I wasn't even being snide.

[steady grove]

the really good one to not like is "can i ask a question?" there's even a domain for that gem! https://dontasktoask.com/

[mossy canopy]

it's linked to my free colab.. so.. :V

[vale eagle]

Good stuff. I really need it. So many people don't know how to ask

[mossy canopy]

UAP time

[steady grove]

public schools teach kids that they have to get permission to ask a question first. then they enter the real world and annoy everybody

[mossy canopy]

dont just ask .. ask gpt first

[steady grove]

i love this sentiment, but as it stands, chatgpt is going to just confidently make shit up 3/5 times

[eternal fog]

People that don't know how to ask questions are going to have a bad time with ChatGPT

[steady grove]

using chatgpt as an expert is a no go. Use it as an assitance for subjects you have expertise in, and can spot it's bullshit

do that second one

[soft zealot]

that is one of the single most annoying things ever (closely followed by getting offended on behalf of someone you don't know) .

Someone pings you a message saying "can I ask you a question"

FFS you've already distrurbed my peace, just aski the bloody question will you!!!

[eternal fog]

What's even better is when people just ping you "Hi"

And then say nothing for the next 30 minutes

[soft zealot]

and that

[eternal fog]

Microsoft even have a quick link for it

https://aka.ms/nohello

[soft zealot]

I get in work with Teams as well

[steady grove]

or just add you as a friend out of the blue. no message prompting it. so you search their name on the server and they've never said anything. so check to see if its a scammer to report the scam message, but they actually just want help with lora training and saw me in the channel a month ago

[soft zealot]

I don't mind if someone messages and ask " are you free for a quck call"

That I can live with

(more on Teams than here)

[native knot]

Nailed it!!

[ionic dragon]

how to use inpainting to remove some random thing on the frame?
like a stick or a feather?

[azure oxide]

Instruct pix2pix is probly best. "Remove the x". I'd personally just edit it out and cover it up then regenerate to finish up if pix2pix doesn't work.

[steady grove]

Chat etiquette i don't think was ever well defined as it came up.

Take phones for instance. Nobody ever said "Hello" before. It was "ahoy" or "good day!" and other context specific greetings. So when the phone showed up, people were like "how the hell do we greet people on this box??!" and were all confused by the etiquette around it. So Bell made an advertising campaign for this new fangled "hello!" idea. Hello would be how to greet people over the phone. They even commissioned a song. Hello my baby hello my honey hello my rag time gal!

fun facts about etiquette

[ionic dragon]

damn, didnt prompt anything, and it removed it

[steady grove]

Keeping someone on the hook or leaving them on hold, while you form your question, is really inconvienient for the person you're approaching. considered bad form in a lot of ways

[mossy canopy]

use bing chat or other internet connected ai chat assistant then.. less chances of hallucination

[ionic dragon]

do you even personalise negative prompt depending on the prompt?
or is the negative prompt the same across all generations?

[steady grove]

a willingness to learn and fact check is important when using any of the LLM's as a teacher. Bing chat will make shit up so often too. It's just chatgpt afterall

[soft zealot]

when you expect something like the first 2 images but from the same prompt & style settings you get a random car.

This Is Stable Diffusuion. This is The Way

[steady grove]

Bard is pretty neutard. I think google is trying to control it to send people to alpha advertisements plastered pages. They are the worlds largest advertising compnay after all

[native knot]

No comments or reactions on my Garbage Pail card above... damn, son...thought I hit it with that. Got the text I wanted after coming up with a good name and everything.

[steady grove]

Just asking an LLM for information will get made up nonsense most of the time. It's the new 3dtv

[steady grove]

i didn't even notice it hahaha i love that

[mossy canopy]

at this point people should already have a sixth sense about what is made up and whats not with this llm models..

[native knot]

Your comment inspired me to make it. I was messing w/something else about Neo at the time and combined the thoughts.

[steady grove]

It looks like one of the new style GPK cards. i love that they cameback

[soft zealot]

doesn't that depend on your definition of "advertising company" ;o)

[steady grove]

it's a perfect GBK pun. GJ

[stone fossil]

Diff each time depending on the + prompt.

[steady grove]

Their revenue comes from selling ad space

they make the most of any ad company. The best around. Noones ever gonna bring them down

[soft zealot]

"some" of tehir revenue ;o)

But does that make them an "adverstising company" like WPP or Omnicrom (oh ok they are advertising agencies that actually create the adverts companies like Google display)

but Im rambling again

[steady grove]

The most of their revenue. Like a huge most. And they make the most of all the ad companies around the world. It's their primary business.

Their business model is the reason why so much SEO ad soaked garbage exists on the web today. it's a direct consequence

A big par of their success comes from favoring search results that tie into their advertising tracking more. They've adjusted those algorithms to suit their business model already, so it's expected that Bard is doing the same.

[soft zealot]

Google Service acconts for approx 90% of Alphabets revenue

https://www.investopedia.com/articles/investing/020515/business-google.asp

[steady grove]

services == ads. on that same page " One of the primary ways Alphabet generates revenue through advertising is through its Google Ads program."

Trust me when i say google does not dominate the enterprise space

Their services are often a joke in it departments

If you've got the budget for adobe, why would you build a tool chain around corel, kind of jokes

[steel mist]

gmail/cal/docs/sheets is a legitimate competitor to the MS suite. I kinda hate gsuite, but the collaboration features are really best in class and that's worth a lot

but yeah, that's not what's making them money. at all.

[mossy canopy]

It seems like iceycold and Winston Woof are having a discussion about Google and whether it should primarily be considered an advertising company.

The key points appear to be:

• Google makes the majority of its revenue from advertising through Google Ads and selling ad space on its platforms.

• iceycold argues that this makes Google fundamentally an advertising company, since ads are its primary business and main source of revenue.

• Winston points out that Google also provides various services like search, cloud computing, etc. So while ads may be the largest revenue source, perhaps it shouldn't be simply labeled an "advertising company".

• iceycold counters that Google's services exist to support its ad model. Things like search algorithms favoring ads, cloud services not being as competitive, etc.

• The debate centers around whether ads being the main revenue source defines Google as an "advertising company", or if its various non-ad services still make it more diverse.

So in summary, they are debating the extent to which Google should be viewed as an advertising company vs a more general tech/services company, based on the outsized role advertising plays in its business model and revenue.

[steel mist]

I like how that summary is basically longer than the original arguments 🙃

[steady grove]

They're the worlds largest advertising company, despite all their diversification

They're going to leverage that position with Bard. It won't be an unbiased algorithm

[steady grove]

gives me hope for the future. Chatrooms will come stock with a sidebar AI summarizer, bullet pointing the conversations and organizing notes about them

[mossy canopy]

Here are some thoughts on this discussion:

• They both make reasonable points - Google clearly makes most of its revenue from ads, but also provides many non-ad services. There's no definitive "right" categorization.

• I think it comes down to emphasis - you could accurately describe Google both ways depending on what element you want to highlight.

• Categorizing any large company with diverse operations under a single label is inherently reductive. "Advertising company" captures Google's core business, but ignores its other efforts.

• Google has incentivized prioritizing ads in its products, so it's fair to be critical of over-reliance on advertising. But it's also an innovative tech company in many regards.

• "Advertising company" has somewhat negative connotations today, so resistance to that categorization is understandable. But ads do fund many free services.

• Google wants to be seen as more than just ads, but its ad model is entrenched. Its attempts to diversify revenue so far haven't fundamentally changed this.

• Overall I'd lean towards calling it an advertising/tech/media company - recognizing advertising as the core revenue engine, but also its broader ecosystem. The emphasis can shift depending on the context.

In summary, I don't think there's a black-and-white answer here. Google has elements of both perspectives represented in the discussion. Its large size and complexity resist a simple categorization.

[weak saddle]

dont most media service use ads as a main source of revenue? Nbc, abc, radio, etc....

[mossy canopy]

You raise a good point. Reliance on advertising revenue is very common in the media industry:

• Traditional media like television networks (NBC, ABC etc), radio stations, newspapers and magazines have historically relied heavily on selling ad spots or space to fund their content and operations.

• Many digital media and tech companies also rely on ads to monetize their free services - examples beyond Google include Facebook, Twitter, free news sites, mobile apps, etc.

• So in that sense, having an ad-supported model is not unique to Google in the media landscape. Advertising could be seen as the dominant revenue engine across much of the industry.

You're right that by that standard, many media/tech companies could be characterized as "advertising companies" to some extent.

A few additional thoughts on Google vs other ad-driven media:

• Google is arguably more ad-dominant than many older media companies, with ads making up 80-90% of revenue. Traditional media tended to have more balanced revenue sources.

• Google's ad products are seen as particularly advanced and dominant in the digital landscape, giving it significant influence over online advertising.

• Diversification efforts by companies like NBC into new media/tech have changed their revenue profiles somewhat as well.

• But you make a fair point - many major media firms are driven by advertising in some capacity. Singling out Google may ignore that broader context.

So in summary, while Google is particularly ad-dependent, relying on advertising revenue is indeed a very common media industry practice. You raise a good perspective to consider in this debate.

[steel mist]

Well part of the drive to streaming is that those companies would like to move to subscription services, which are generally preferred to wall street to advertising

But there's also a difference in that those media companies create content, and then attempt to monetize it

Google doesn't create content, they run display ads against other people's content. So the dynamic does shift

[mossy canopy]

Excellent points! You're absolutely right:

1. Media companies are indeed trying to shift towards subscription/direct revenue models rather than pure ad reliance. This transition is still ongoing, but streaming opens up that possibility.

2. There is a fundamental difference in that traditional media companies invest in creating original content, while Google simply provides a platform for others' content.

3. This changes the dynamic, as Google is not bearing the cost of production. Its role is ad enablement, targeting and display - monetizing others' work rather than creating its own.

That's a key distinction I failed to make earlier. The traditional media companies rely on ads to recoup self-funded content investments, while Google facilitates ads without content creation costs.

Given that core difference in content costs/risks, it makes sense that "advertising company" feels like a more apt categorization for Google's role than for a traditional media firm. You highlighted a very important nuance in the business models - thanks for clarifying! Really great points that improve my understanding here.

[stone fossil]

1. type in + prompt.
2. add style as needed.
3. add negs.
4. press ctrl + enter.

[mossy canopy]

nsfw triggered?

[soft zealot]

@mossy canopy are you by any chance feeding our meaningless inconsequential ramblings into som form of LLM based AI to generate bullshit summaries?

[steady grove]

Ads fund a lot yeah. The problem with google is they revolutionized advertising with massive amounts of tracking data. Big data science. advanced algorithms that give the user the ad that will have the most impact. They've turned psychological manipulation into a very technical science.

Given how they've developed their search algorithms and tracking data over the years, why wouldn't they build Bard the same way? It's going to be geared for manipulating people into looking at ads more than it will be a useful tool.

I don't trust a lot of these LLM's but Bard especially i'm sure will have some engineered ulterior motives

[mossy canopy]

You raise very thoughtful concerns about Google's approach to advertising and how that could influence Bard. Some key points I take away:

• Google has pioneered new levels of personalized/targeted ads using extensive data collection and predictive algorithms. This enables more effective psychological manipulation.

• There are reasonable concerns that Google could take a similar approach with Bard - optimizing it to serve ads and steer users towards commercial interests rather than neutral information.

• Given Google's history and core business model, it seems likely Bard will be designed with advertising ROI as a priority rather than just user benefit.

• The depth of personal data Google has amassed raises suspicions about how that could shape Bard's behavior - perhaps promoting certain viewpoints, interests, etc.

• More broadly, many LLMs like Bard lack transparency about their training objectives and optimization, making it hard to trust their neutrality.

I think you make very fair points. Google has demonstrated their adeptness at using technology and data for persuasive ends. It's prudent to be cautious about how that could manifest in Bard as well, rather than assume pure altruistic motives. Maintaining a critical perspective on these systems and their creators is important. I appreciate you sharing these thoughtful concerns.

[steady grove]

damn right i make fair points llm summarizer

[native knot]

Excellent.

[ionic dragon]

@soft zealot @mossy canopy we are training an llm to generate sd prompts, so can we please get access to your prompts so that we can add them to our training dataset

[soft zealot]

Sure. My prompts are usually in the images I post here. I don’t hide them. Knock yourself out transcribing them

[ionic dragon]

ok tq, will download your images and use your prompts

[soft zealot]

Are you for real??

[ionic dragon]

yes

[soft zealot]

So you don’t realise when someone is extracting the urine?

[ionic dragon]

huh?

[mossy canopy]

thats just clip interrogate https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[soft zealot]

Do you just take everything literally

Read what I wrote carefully

I was politely telling you to eff off by pointing out the bledding obvious

[ionic dragon]

oh
mb
i am not too proficient in english so I didnt understand

[ionic dragon]

so you drop images and get the prompt?

[stone fossil]

No I prompt by hand lol.

Keyboard warrior.

[ionic dragon]

damn

[dapper dragon]

https://tenor.com/view/raccoon-music-gif-20612597

[soft zealot]

it does beg an interesting rhetorical question (ie you don't actually have to answer it)

If you as a human can't differentiate between a "serious" and a non serious" response how do you expoect to succesfully train an LLM ?

[ionic dragon]

it depends on the other person whom I am talking to, if I am talking to to you I always read seriously

[vale eagle]

Clip interrogate usually provide "wrong" result compared to what I want

[soft zealot]

In which case I recommend you reset your base expectations

[ionic dragon]

i'll try

[native knot]

https://giphy.com/gifs/AlEp8Fj45KPLO

https://giphy.com/gifs/cinemagraph-oc-wonka-PVfSR2nddzGz6

[floral island]

#🍥｜anime dead, imma spam here

removes anime tag, now we go real creative

[mossy canopy]

thats why people use blip to caption for training clip .. but blip doesn't recognize artists styles..

[vale eagle]

I also used blip but after all, I stick to use wd14 for lora training.

[mossy canopy]

i use both..

[floral island]

which lora this time?

[stone fossil]

😄

[vale eagle]

blip has its own advantage for the tag out of wd14. Both of them have limitation. I really want to have a tool reduce manual work for the training. A custom llm should be the final answer.

[mossy canopy]

if bard or bing didn't censor their image captions . i would've used them..

[vale eagle]

I want to caption in a specific way that bard and bing might not my answer

[floral island]

ok, done with mechanical dragons, on to next! they be cool tho

[floral island]

i mean, i recognize the character -> that's a win in my book

i think i can even hear him whistling

[dapper dragon]

It is a poor lonesome cowboy 🙂

[mossy canopy]

https://huggingface.co/liuhaotian try his models..

[ionic dragon]

img2txt?

[floral island]

hmmm... 100% free reigns might be bit too much lol

[mossy canopy]

for chating with images.. getting your captions in a centain format then just prepend wd 1.4 tags for accuracy..

[ionic dragon]

That clip is too good

[ionic dragon]

Do you know how to install that clip model locally?

[floral island]

found the pot of gold

[wet nacelle]

https://tenor.com/view/jerma-dollhouse-jerma-dollhouse-jerma985-gold-watch-gif-22799651

[floral island]

clown dog?

[upbeat summit]

I just sorted through a fantasy prompt session and found this - I think he's not amused 😄

[floral island]

shouldn't have put it in plain sight 😄

[wet nacelle]

Clown Dogs.

[floral island]

any dogs = best dogs

[mossy canopy]

that link i gave u .. its for llava models.. you can try them in oogabooga .. to chat with images... for batch captioning using llava i recommend using oogabooga api to write something get your captions in a certain way .. or you can copy their code here.. https://huggingface.co/spaces/LinkSoul/Chinese-LLaVa swap the model out..

[floral island]

someone stole my hr giger dragon!

wicked mad looking reptile ❤️

[floral island]

(don't inspect my image prompts from this point on, this is my full random+song lyrics random verser at max. people might discover i'm a madman)

[dapper dragon]

Nice raccoons bro

[wet nacelle]

https://tenor.com/view/jerma-jerma985-cheers-cheering-bro-gif-26562284

[stone fossil]

Raccoon power!!!

[wet nacelle]

Raccoon thread?

[dapper dragon]

haha

no for me, time to go to bed

[zinc cargo]

my kids has one!

[wet nacelle]

all good man

[dapper dragon]

one more .... with thaevil1 prompt

[wet nacelle]

wicked

[floral island]

i need more gpu's XD

[mossy canopy]

sure

[floral island]

... fuu... company a100's are not in use

do i really have to update my a1111 install and get my fav model installed there? XD

[stone fossil]

F A11.

[ionic dragon]

Isn't clip ViT-H-14 better than Llava?
I tried 3 images from MJ and I paste the prompt into SD and the images were 80% similar that was great

[mossy canopy]

keep it running .. no one will know.. 🤣

[mossy canopy]

yes. but its not a caption model .. its a image chat model .. but if you want your captions in a certain way.. u can prompt here. it can get pretty creative..

[ionic dragon]

Damn,
Lucky boi
Make the GPUs go brrrr......

[floral island]

make the boss pay for my waifu's instead of me

[ionic dragon]

Yeah I get it, but I will use clip vit,do you know to install and use locally

[ionic dragon]

Exactly
Your boss got the shit out of you and now it's ur turn

[mossy canopy]

yes.. just git clone the repo.. run it locally.. ;V

[mossy canopy]

or just use comfy clip interrogate nodes or a1111s clip interrogate

[ionic dragon]

Where should I run them?

[ionic dragon]

No, this clip is way better than them, and I am planning to generate 1mil prompts

[floral island]

me and my slightly drunk ass should not be on $100k+ hardware lol

what's the worst thing that could happen

[wet nacelle]

@mossy canopy

[mossy canopy]

dont you feed them.. they look so thin..

[wet nacelle]

self sustain

[soft zealot]

loving some of the randomness I'm getting from this prompt

[floral island]

@mossy canopy try bronze rubbed metal, or any other kind of metal. chrome ❤️ or ...

bismuth!

bismuth dragon!

[floral island]

pcb lora 😄

[stone fossil]

😉

[mossy canopy]

this is not a chrome dragon..

sd 1.5

[wet nacelle]

@mossy canopy

[floral island]

sd 1.5 is soooooo oldskool 😄

[floral island]

aight. got it working again. now. i just need to feed it. things. nope, something broke with the update

[mossy canopy]

creepy

[mossy canopy]

exactly

[floral island]

for some reason, restart sampler is not in the new update on the docker container

fml

[mossy canopy]

device?

[mossy canopy]

add git pull in docker commands..

[floral island]

perhaps i'm just using a stupid repo

[strong field]

good day

[floral island]

nah, it's a repo someone made that acts as a wrapper for a1111 and some other uis

yeah, can't be assed to do that now 😦

[stone fossil]

Hmmm hmmmm.

[floral island]

git pulling broke everything

(because that what git pull does)

[wet nacelle]

It looks good but also is extremely sparse with details.

[floral island]

i know this prompt 😄

[mossy canopy]

this is op.. great for small details..

[mossy canopy]

didnt upscale. ..

[wet nacelle]

No I mean visual clutter. Mine is cramped with details.

[stone fossil]

Yep Eface is a good rl mate of mine who once ever tought me how to use SD.

[floral island]

@mossy canopysee? bismuth dragon is where it's at

[stone fossil]

Thanks for that for ever lol. 🙂

[floral island]

i just cost you $2000 of hardware or so...

[strong field]

god these are so awful i love them

[stone fossil]

Hey I only have hub bucks. 😄

[floral island]

reminds me of that old series 'dinosaurs'

[mossy canopy]

yes.. the inconsistency

[floral island]

"hr giger has entered the chat"

[strong field]

🤣

[floral island]

bismuth dragons be shiny

[stone fossil]

Whoppa nice mates. 🙂

[mossy canopy]

not working

[floral island]

which model?

[wet nacelle]

https://tenor.com/view/fat-tummy-lard-belly-fat-lard-tummy-belly-gif-26684445

[floral island]

i'm using protovision

[wet nacelle]

https://tenor.com/view/jermazerk-jerma-jerma-shoe-when-something-is-your-shoe-gas-gif-21577077

[mossy canopy]

nightvisionxl

[floral island]

might be a model issue, if nightvision doesn't know what bismuth it (what it looks like to me), then it won't respond to it

[floral island]

hmm

by same dude...

[mossy canopy]

thats the closest..

[floral island]

pretty damn good though

[floral island]

seems pretty accurate with bismuth actually

lel, werewolf terminator

[floral island]

i really dig this one tho

[mossy canopy]

used different prompt.. a close up of a metal dragon on a reflective surface, digital art, trending on polycount, digital art, colorful muscular robot, she has iridescent membranes, steampunk horse, blue scaled dragon, 8k hd wallpaperjpeg artifact, crystal dragon, intricate detailed 4 k, chrome armor, holographic design

[floral island]

that's what clip/blip/vit made of it?

[mossy canopy]

:v

[mossy canopy]

just clip

[floral island]

still, very nice ^^

i just have ~30k more old prompts to dig through to see if anything cool shows up 😄

[wild ice]

Hello! I would like suggestions for producing this "style" of art (e.g. checkpoint, lora, prompts, etc.)...

[floral island]

oh shit. these old prompts are gold!

sd 1.5 could not handle them. AT ALL

sdxl brings them to life. for real lol (and still fails to capture their true essence)

[mossy canopy]

exactly..

[floral island]

did you clip it again?

[mossy canopy]

yes..

[wild ice]

How on earth did you do that...? Checkpoint? lora? prompts? Nothing I've tried creates that look.

[floral island]

i think clip interrogate, and then feed into sdxl model

[floral island]

this prompt 😄

i mean, the first one, goat horns, viper scales, tiger fangs/face, and pointed ears on a human body?!

[mossy canopy]

as @floral island said.. prompt: a painting of a crane in a forest, a detailed painting, by William Morris, tumblr, arts and crafts movement, deer, painting of samarkand, golden inlays, 1 9 2 2, philip taaffe, black velvet painting, artemis, heraldry, in a forest at night, not cropped, vintage - w 1 0 2 4, symbolic mural

[floral island]

did clip really put forth -w 1024 as a token set?!

[mossy canopy]

lol

[floral island]

i haven't clip interrogated in like... uhh... yeah

[wild ice]

What is the significance of "1 9 2 2" in that prompt?

[steady grove]

somebody probably trained their interrogator with scraped midjourney / bot servers

[steady grove]

years are often referenced for aesthetic styles

[hardy cipher]

and I guess there's a benefit to seperating the numbers like that?

[bronze arch]

oh wow that looks sick

[steady grove]

i'd have no idea there. that might just be a meta that developed, or somebody's scraping engine doing it. maybe just a poorly formed regex that creates the training data

dont know why somebody would space out numbers

[floral island]

so. what's this creature called except sd weirdass?

[mossy canopy]

wtf

[floral island]

i mean, it's fckn awesome for sure

but i have no idea wtf kind of creature my sd created

goat facial structure, snake skin, kobold ears, goat horns, tiger nose and fangs, dragon eyebrows, human build. oh yeah, blue skin too... sdxl sniffed too much greek mythology i think....

[floral island]

friggin awesome!

straight out of a horror movie, actually

[steady grove]

a type of chimera

[floral island]

vader got nothing on these crusaders

not sure yet how sdxl based models act around that 75 token limit

can't get a grip on it yet

[mossy canopy]

https://store.steampowered.com/app/2107670/Warhaven/

Steam

Warhaven

ENTER THE FRAYWarhaven’s dynamic melee combat rewards new recruits and veteran players alike. Choose from a roster of unique soldiers, each with their own motivations for fighting beneath the banner of Warhaven.Coordinate with your squad before each battle and outwit your foes with tactical maneuvers. Fight to control footholds and the powerful ...

Release Date

Sep 20, 2023

▶ Play video

[hardy cipher]

ooh, got my first couple conditioning nodes working

[wild ice]

Obviously I'm a beginner. Now that you alerted me to "Interrogate CLIP", I gave it a go on the same image and got this:

a painting of a deer in a forest with trees and foliages on a wall with a mirror reflecting it, leesha hannigan, Edward Burne-Jones, an art deco painting, arts and crafts movement
Since this is different from your result...I presume it is controlled by the currently selected "checkpoint"?

[floral island]

clip is "an interpretation"

[hardy cipher]

only use danbooru interrogators

[floral island]

so dpeending on exact input image, sometimes it might differ

[mossy canopy]

nope .. two different clip models.. i used ViT-H-14 OpenCLIP

[floral island]

vit clip!

yeah, that one is 100x better

[wild ice]

Er...my automic1111 has a "Interrogate DeepBooru"...it does not have a "danbooru". Just a typo, or is there really two versions of that?

[hardy cipher]

probably. are you using comfy? if so there are lots of interrogators. tbh danbooru is some weird stuff. but I use it when I want to get very fast results. if you ever look it up it's a rabbit hole

[stone fossil]

Comfy ❤️

[wild ice]

Worse, just now when I tried "DeepBooru" on that image...it gave me something COMPLETELY non-sensical:

long hair, multiple boys, multiple girls, short hair, weapon

[hardy cipher]

lol, yes

that's all it does

[floral island]

booru is plain worthless unless you want pure anime weeb shit

i tried it for 2 prompts -> it's a prompting style that's VERY restrictive

[hardy cipher]

I've used it in workflows when I wanted to create a quick prompt to then use it to modify the original image or something. but yeah, I was being moderately sarcastic

[wild ice]

I'm using automatic1111...not "comfy". Er, am I right in thinking "comfy" is a reference to a different UI?

[hardy cipher]

noodle ui

looks like some calamari might be growing from those noodles

[mossy canopy]

after war

[stone fossil]

😉

[hardy cipher]

thiccc neck

[stone fossil]

Move ya body.

[stone fossil]

Sexy pix.