✨｜sdxl

[meager canopy]

https://civitai.com/models/240590/ultimateblendxl

[humble cape]

image blend ?

[gloomy lark]

oh ok, that's weird though i just went looking for that, but it shows a january upload date. cosxl wasn't out then

[meager canopy]

I haven't uploaded that merge.

It's just what I use

[gloomy lark]

oh ok

[gloomy lark]

was that cosxl merge a ratio 0 through 1 or was it that more advanced choosing which blocks stuff?

[meager canopy]

I really don't remember what the ratio was, and I think I didn't touch the blocks for that.

[gloomy lark]

alrighty, cool thanks. i have to look up figuring out what the good merge settings are for that. probably does a lot for visual fidelity

[humble cape]

i love this one

[humble cape]

cosxl ?

[meager canopy]

CCTV photo...literally

[shadow nacelle]

are you guys using the app to generate or use in website 🤔

[humble cape]

wtf

[glass forge]

holy excrements theres someone who doesnt know the matrix?

[humble cape]

lol

[noble shoal]

Right? Insane.

Only one of the "movies" that claim to be sequels.

[meager canopy]

HelloWorld was the other model I used for refining/upscaling.

[gloomy lark]

thanks.. I found the cosxl merge info. https://comfyanonymous.github.io/ComfyUI_examples/model_merging/#advanced-merging

in short, it does require that whole cosxl-sdxl + yourcheckpoint certain block style of merge. i tried doing it via ratio and it just gave junk.

but with that json it's good.

[humble cape]

model merging ?

can you merge sdxl with pixart sigma ?

[gloomy lark]

no, different base model. doesn't even have clip.

[gloomy lark]

for goofs I tried it. totally doens't work.

the pixart sigma model is only 3 gigs and has no clip.

[humble cape]

😔

[gloomy lark]

I just use an sdxl refiner as a second step

[gloomy lark]

turns out euler_dy sampler works insanely well with it.

[humble cape]

sdxl ?

[gloomy lark]

which is this new sampler that just came out.

it fixes almost all of the horrible hand issues with pixart.

[gloomy lark]

those shots are from pixart, with an sdxl refinement stage

[humble cape]

😭

[ashen otter]

Overexposed?

[pale bone]

Kino. What model ?

[tawdry current]

probably one of those "boring loras" that try to mimic every day photos

like this one: https://civitai.com/models/310571/boring-reality

[pale bone]

Thanks for making me aware of those. Would still like an official answer tho, reverse engineering an image is sometimes impossible

[pallid path]

aw still no version 5 yet

I hope this comes to SD3 soon after release

I have been working on some new models to replace the older Boring Reality Loras.

[glass forge]

Try indian man pooping in public toilet with that lora 😄

[noble shoal]

That was just regular old base SDXL with a lora.

[pale bone]

Alright, the one above ?

[noble shoal]

A stupid vhs Lora I trained today very quickly because i got involved in a challenge with @humble cape . So, no, not the one above.

[tawdry current]

Figured it was something along those lines. Pretty cool though, I like the boring effects like that. Ironically, they tend to be harder to pull off than the pixel perfect overtrained waifu simulators that 95% of users obsess over

[broken crater]

Oh sweet, new Juggernaut version dropped

[tawdry current]

Can't wait to see how much more it's overtrained on the same dozen closeup faces, while losing even more creative flexibility

[pale bone]

Okay, thanks

[broken crater]

They have stated that issue with a lot of close ups have been addressed. We will see, we will see

[tawdry current]

Yeah, talk is cheap... But I'm sure the horny crowd will fawn over it

[noble shoal]

Yeah, gross. Take a "good" Lora and the base model and maybe throw the refiner on it if you feel fancy and done. I also hate it when everything looks like plastic skin and factory made humans.

[glass forge]

I guess melon sized cleavages are easier to train than realistic hands.

🤷‍♀️

[humble cape]

😂

[meager canopy]

This has great detail! Mind sharing the prompt and how you made it please?

[terse tundra]

https://civitai.com/images/10301713

there you can find the prompt and the lycoris i used

[meager canopy]

Thank you 🙏

[terse tundra]

post some pics in the galary for my mecharmor lycoris!

[meager canopy]

I will, just not had time yet 🙂

[novel haven]

The name INDIGO MAGNOLIA in the shape of a magnolia bloom

[humble cape]

[blazing kernel]

wow these're from sdxl?

I didnt know sdxl could handle texts that well

[gloomy lark]

An anthropomorphic angelic being made of swirling cosmic energy, with circuitry patterns and symbols etched into their skin; glowing golden eyes surrounded by intricate blue and purple light patterns; background is an explosion of colorful geometric shapes and lines; digital art style, vibrant colors, detailed textures, fantasy illustration in the style of high contrast lighting, ethereal glow effect, surreal atmosphere.

[pallid lynx]

if i wanted to do a batch of images, can i set it up so that i can choose which images out of the batch to upscale (in the same workflow)?
for example, a batch of 5 images takes 20 seconds on the first step, but the next step of upscaling takes about 2 minutes per image.
i dont want to upscale all 5 images, but choose images before proceeding to the next step. is it possible to set the workflow to "pause" and wait for me to choose images for the next step?

[gloomy lark]

@pallid lynx I don't know how off the top of my head but i know galaxytimemachine has workflows that do that. https://civitai.com/user/galaxytimemachine

[vagrant valve]

What was the prompt for this?

[mellow tendon]

This is good, I saw this image earlier and it inspired me to do this:

[pale bone]

Is this too risqué for this channel ? I want to share it but don't really know where it belongs

[native knot]

(@broken crater)
Meanwhile, SDXL:

[broken crater]

Well that's nice. I think I hadn't tried it with sdxl. With 1.5 long time ago had no success 😄

[native knot]

I just figured I'd keep trying prompts people are dropping in #🆕｜sd3 on XL and Cascade just to see what those do with it. Sometimes the results aren't bad, like above, and sometimes they're uh...not so much. 😄

[broken crater]

Yess that's really interesting to see comparisons

[copper kraken]

this is absolutely a goo didea imo

we should be comparing to sdxl 0.9 base, and a finetune or two as we go along

[native knot]

Right now I'm attempting these against Juggernaut X just to see. 0.9 base is a good idea for sure

[copper kraken]

had to make some of my own nodes but this schedule is amazing

input, output... detail is insane

[native knot]

Pack all that into a node and send it to me. 🙂

[copper kraken]

https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler do you have this?

also, you have a 4090 right

[native knot]

3080...so I expect it'll take longer 😄

[copper kraken]

i don't yet exactly what the absolute best parameters are for this, so i don't wanna pack it up yet

ahh k yeah this one takes a while lol

but i can drop you the nodes here in a sec and the workflow

[native knot]

cool

[copper kraken]

that is pretty nice cuz you can adjust the width of any of those segments, the shape, etc at will

[native knot]

Yeah...I could see how you structured it

[copper kraken]

get that node pack^^ and backup sigmas_merge.py and drop this one in the same folder

you can still get really cool results with the insanity of 250 steps for the generation and 250 for the upscale lol

for whatever reason, the scheduler nodes out there are really lacking in terms of manipulating them

a few fancy things, then not even a single legit concat node

the sigmas cleanup one i made does three things

1. scan for duplicate sigma values and remove them
2. scan for all sigmas that are below the sigmin you choose... and delete those too
3. append a single sigma at the end with a value of 0.00

if you have any that are 0.00 in the middle you get a black image, and duplicate ones are just a waste

basically what it's doing is unsampling then denoising over and over

while conserving momentum (i think)

also, euler_a and sonar_euler_a are pretty good with this too and twice as fast as res_momentumized if not more so

sonar_euler_a

dpmpp_2s_ancestral

solar_euler_a with perlin noise is better than gaussian i think

[kindred shoal]

Does anyone know how to do scales? Or would I need a lora for that?

I tried ponyv6 but i can't seem to get scales to wanna appear.

[pale bone]

You want scales on a humanoid character ? I've had success with just "scales" in the past, but I can try again

[kindred shoal]

Yeah, sclaes on a humanoid

It doesn't want to stick to the face

It goes everywhere else BUT the face.

[pale bone]

try "facial scales"

It's a tag on e621

Alternatively gimme your prompt and I can try to make her scalier myself

[kindred shoal]

NOPE IT DID SOMETHING ELSE.

I KNEW it was gonna use the wrong fucken word

[pale bone]

lmaoooo

[kindred shoal]

"scales on the face" instead, okay lets hope that works

[pale bone]

Well, that's not a tag

But who knows. In the meantime I'm gonna try to make a scaley face

[kindred shoal]

it's so close yet so far

...might help if i refresh the seed

OH it changed her face color! Progress!

[pale bone]

In one of my results I got this :

It's not 100% though. Trying to recreate what you're going for.. I did use "facial scales"

[kindred shoal]

oooooooooo

DAMMIT IT PUT SCALES ON HER BUT IT ALSO DID NSFW

[pale bone]

You can hopefully negative prompt that out

[kindred shoal]

I'm really hoping I can, because it looks nice

just. not the nsfw part.

...why did it make 2 girls when i prompted it out

confoosion

I might have to specify skin color.

[gloomy lark]

Very Miyazaki-esque

[pale bone]

these fishes are schadenfreuding so hard

[kindred shoal]

pffffff

[kindred shoal]

How.
Brother how.

[pale bone]

Anyway, facial scales works really well. You just have to negative prompt the unwanted consequence

Prompt:
score_9, score_8_up, score_7_up,
1girl, fins, red hair, blue skin, scales, (facial scales), bangs, flat_color, animal humanoid,
Negative:
score_6_up, score_5_up, score_4_up, 3D, [some word I would rather avoid posting here], rating_explicit,open mouth

[kindred shoal]

what's this... 'score_9" stuff?

[pale bone]

It's supposed to reflect the quality of the image

The creator of the model created that metric. It's kind of annoying, but including it does improve the quality a lot, subjectively

[kindred shoal]

oh hell yeah this is more like it

[pale bone]

Nice ! Happy yet ?

[kindred shoal]

Yeah! Ty!

I was like "There's gotta be a way I can put scales on the face"

[pale bone]

And there was ! When I want to check for the correct way to phrase stuff in pony, I do a google search of what I'm thinking of with "e621" after. In this case I searched "face scales e621" and the first link was "e621 wiki - facial scales"

[kindred shoal]

Just gotta inpaint out the orange in the hair.

[pale bone]

e621 is a pretty cursed website if you're not part of a specific demographic. But their tags are nice and they work the best in pony usually

[kindred shoal]

Yeah... But! DO you know how to inpaint to a specific color? it's...not working out for me

[pale bone]

Inpaint what, you want to make the orange hair purple ?

[kindred shoal]

To blue

[pale bone]

If it was me, I'd just prompt it and generate all over again. But you can probably get it by inpainting and filling with latent, and changing the prompt to just "gradient hair, blue hair" while keeping the score tags ?.. I'm not gonna say too much for fear of saying wrong things. I don't inpaint very much

[kindred shoal]

Ohhh I had to inpaint sketch.

Regular inpainting wasn't doing it, it kept it orange

Now I just gotta fine tune it

Get it pretty close to how I want it

..Did it pretty good on accident lmao

[pale bone]

Seems good to me !

[kindred shoal]

Multi-colored hair is a pain, hah

[kindred shoal]

but I can always just inpaint to victo........ okay listen you dont have to flex

😆

[pale bone]

Getting a very specific layout is indeed pretty unpromptable

[kindred shoal]

Yeah, I don't need it to be super specific, just 3 colors somehow involved

Blue, red, and purple

[pale bone]

"multicolored hair, blue hair, red hair, purple hair" doesn't do it ?

[kindred shoal]

It throws in only two colors ignores whichever one it feels like

[pale bone]

I see. In any case, you can put "orange hair" in negative I guess

Maybe try to figure out which gets ignored most often and play with weights

[kindred shoal]

God. sdxl is so nice but I can't run it on my laptop, only the big computer

[kindred shoal]

WHY ARE Y'ALL SO GOOD AT THIS STUFF.

STOP IT 😭

[pale bone]

sunk cost fallacy

[kindred shoal]

what's the prompt for the second one, and the model-
And whatever else you used

[gloomy lark]

Pixart sigma, so it very good understanding, and it's a workflow that refines with juggernaut

[kindred shoal]

ohhh frick it's comfy

haven't used comfy in a second

[gloomy lark]

I just did: 1 gnome, fins, red hair, blue skin, scales, facial scales, bangs, flat_color, animal humanoid

[kindred shoal]

[gloomy lark]

You using sd 1.5 or sdxl?

[kindred shoal]

sdxl

SDXL Ponyv6 to be specific

Because I heard it was pretty good for like...anthros. And a fish person is technically an anthro

[gloomy lark]

Well that's its own language. This pixart stuff allows for calling out attributes that might normal conflict or bleed. It keeps them separated more easily.

[kindred shoal]

pixart?

[pale bone]

summore

Although clearly my scale count is inferior

[gloomy lark]

Adding word anime

Not really anime. 🙂

Other than big eyes

[kindred shoal]

I'm assuming pixart is some premium thing

like nai

[pale bone]

Really insists on giving them pink noses

[gloomy lark]

Google for pixart sigma. It's free and open source

[kindred shoal]

ohhhh i see.

[gloomy lark]

The output sets up the composition, then you have another ksampler in comfy overlay the style etc. could be juggernaut, could be pony

[kindred shoal]

i'd have to relearn comfy, hah

[gloomy lark]

although with pony you'd need to have a rather different prompt than the one that goes into pixart

But it's easy to do in comfy

[pale bone]

had never heard of pixart sigma, today I learned !

[kindred shoal]

jesus that's gorgeous

[pale bone]

Glad you think so !

I'm unfortunately a ramlet and am using online services for my gen needs, as such I don't think I'll be able to stitch pixart and SDXL in practice anytime soon

[kindred shoal]

yeah, my laptop canNOT do sdxl so I have to use 1.5 on that.

It can do 1.5 just fine on Forge, but Forge crashes out when I try to load anything sdxl

So how exactly do I use this?

I have to get ComfyUI again but yeah

.. huh. comfyui isn't showing up in forge.

there w e go

had to fully close the terminal and reopen forge

[pale bone]

Heres my last contribution, thanks for the theme idea !

[kindred shoal]

Ooooooooh I adore that.

I'm gonna use that for the d&d campaign i plan on using her for.

so! How do I use pixart for comfy? Do I just look up nodes and stuff?

[gloomy lark]

So use Ella instead. It's the equivalent of Pixart but for SD 1.5. Works with most models and replaces the CLIP with a language model. Made this with it.

[kindred shoal]

Oh, I'm on the big boi computer rn, I can do sdxl

the laptop is the one I can't run sdxl on, which I'm fine with.

[kindred shoal]

I'm not understanding the install guide for pixart sigma.

How do I install it to an existing comfy ui?

[gloomy lark]

@kindred shoal https://github.com/city96/ComfyUI_ExtraModels

you want that. scroll down to the pixart sigma section.

[copper kraken]

You can run it on your laptop via your "big" computer

I run shit on my 4090 from my phone all the time, Ill leave forge running since that's the only web UI that actually works on a phone browser at all

I use comfy from my laptop around the house

All you gotta do is add --listen to the command line parameters... I think that's it

Then the IP of your big comp... X.x.x.x:8188 for comfy, 7860 for forge

[kindred shoal]

Ohhh yeah, that was a thing

[copper kraken]

It's easy

Def rec doing that

With comfy you just want a wide monitor wherever that is

I have a 49" ultra wide and it really helps

[zealous osprey]

On rainy days, Maybach Odin rides a mount and a man holds a village rain knife ┬─┬ノ( º _ ºノ)

[gloomy lark]

pixart keeps doing better than sd3

[kindred shoal]

so what is uh. t5? I can't find it on the pixart sigma github

ah found it

im just blind as hell

[gloomy lark]

that's the language model piece. it's the biggest part of it all

[kindred shoal]

..how do i downlaod it-

it just takes me to hugging face and i don't see a dl button

ohhh its git

all this stuff is still new to me, i'm as dense as a rock

oh my s/o is totally killing me for putting 18gb worth of content on this thing

but hopefully it's..worth it?

[copper kraken]

my computer is worth more than my car... it's worth it

[kindred shoal]

Haaahhhh

yeah this one idk how much it cost

but we've already gone through about 800gb of a 1tb drive

[copper kraken]

i've thought about getting a new car... but every time i do, i remember that i could just keep driving my current one and buy like 15 5090s instead

[kindred shoal]

Jesus

you must get that speed on sd

[copper kraken]

well, then i realize it'd be stupid to buy 15 5090s, so i figure i'll buy a car, and then i remember i don't give a shit what i drive at all so long as it moves and it does... and what a waste that would be when i could have all those 5090s...

so then i just settle on planning on buying one of em as soon as it's available along with the new zen 5, 128gb ddr5 ram, etc etc

prolly one of those fancy new super ultrawides that are dual 4k

then i'll finally be done >_>

[kindred shoal]

which one of these do i load?

im just running a test to see if it all works

[gloomy lark]

01

[kindred shoal]

Got it

I have 12gb of vram, and like... I think 46 of ram

should i use cpu or gpu?

I have 48gb my bad

I can't read apparently

technically 47.8 but yeah

[gloomy lark]

cpu for the t5

you can't fit it on vram.

[kindred shoal]

..I can't?

well i should be able to fit it on the system ram... i have 32gb of it free

[gloomy lark]

yeah definitely on system ram

just not in vram

ella is only 3-4 gigs for t5, you could fit that in vram, but pixart is much bigger for the t5 model, 20 gigs.

[kindred shoal]

... yay i love errors

i dont know what any of this means

i dont have either of those set, up at the top

[gloomy lark]

pip install accelerate

it says

[kindred shoal]

and if that was the case, i even ran 'pip install accelerate', got it, and restarted comfyui

same thing.

[copper kraken]

did you install it into your venv if you're using one?

[kindred shoal]

not using a venv

[copper kraken]

you probably have a python embedded folder, maybe you installed it system-wide and not into that?

[gloomy lark]

are you using portable comfy or native direct?

[kindred shoal]

portable

[gloomy lark]

i had the worst time with portable with various environments. it was a mess until i went direct

[kindred shoal]

yeah if i go direct i might make something explode, so no

i'm prone to breaking everything i touch

[copper kraken]

i'm using portable... you prolly just installed it system wide

[gloomy lark]

it was the opposite for me. everything was exploding until i did

[kindred shoal]

i dont have a venv for it though unless it made one for itself

[copper kraken]

that's an outrage!

[kindred shoal]

it did not make a venv for itself

but i dont run things in venv, it's like 4 extra steps to get something running.

and lhen 40 before that of me just googling stuff to figure out how to make a venv in the first place

[copper kraken]

yeah so do this

find your python_embedded folder in comfyui

[kindred shoal]

uh huh

[copper kraken]

launch cmd.exe and type in... C:\Users\whateverthefuck\AI\yadayada\comfyshit\python_embedded\python.exe -m pip install orecuhrockhrcueh

that's what i do

[kindred shoal]

okay now lets try this shit out

[copper kraken]

clownkarras

[kindred shoal]

....huh. that actually did it. why is reading me as low cpu though?

[copper kraken]

low cpu?

[kindred shoal]

Yeah, that's the error. It was reading my cpu as "low"

[gloomy lark]

that's what my checkpoint merge screen looked like

[kindred shoal]

which I don't understand why

[copper kraken]

"was" reading... does that mean it's okay now?

[kindred shoal]

After installing accelerate it is

it's taking ages to load the "shards" though

it's been 2 minutes

[copper kraken]

(x/15)arccos(cos(2pi*(y**0.707)*8))**2 i came up with some equations like this that are really nice... at least in theory

[kindred shoal]

of course when i complain about it, it kicks in

fucking computers

[copper kraken]

power of y changes the "center of gravity" of each curve, x/15 sets it to decay (peak height that is) and can eaisly be exponential

the 8 is the number of cycles in the range you provide, etc etc...

[kindred shoal]

very cool

[copper kraken]

but in the end, it sucks ass compared to the giant matrix i just showed ya cuz we're talking discrete numbers and not beautiful curves

all the rounding and sampling the curve near its low and high points, but a bit off on the x axis... ends up nowhere near as snazzy in the output

[kindred shoal]

DO I need a special vae for this

istg

[copper kraken]

maybe i'll revisit those formulas when the 7090 comes out and i can comfortably buzz through 4000 steps

[kindred shoal]

it's spitting out pixels.

[copper kraken]

for pixart sigma? you want the regular one

show the whole workflow if you could

[kindred shoal]

i tried sdxl vae and it spat out pixels.

[copper kraken]

lol you're using ponydiffusion

[kindred shoal]

..

[copper kraken]

you gotta pick the pixart checkpoint 🙂

[kindred shoal]

i totally knew that.

[copper kraken]

ponydiffusion is sdxl, it's not gonna work as a pixart sigma checkpoint, it's just noise

do you have that checkpoint

alright clownkarras is the best version yet

[kindred shoal]

I have the pixart checkpoint but HI RED :D

..did i goof up again istg

[copper kraken]

workflow screencap?

[copper kraken]

yeah i don't think you're doing anything wrong there...

set dtype to auto just in case

[kindred shoal]

cmon hugging face, get past my firewall

[copper kraken]

don't think that's it

[kindred shoal]

huggingface istg

[copper kraken]

the other possibilities are that your checkpoint is corrupted as the download cut off right before the end or something, or you need to update comfyui

[kindred shoal]

i think i yoinked the wrong model

thats the model

the ckpt

[copper kraken]

you didn't have that one?

[kindred shoal]

i have that one, but that's not the sigma model

this one is

That was it, okay now lets see if i get pixels

[copper kraken]

oh, oops, yeah, that'll do it

[kindred shoal]

OOOH hell yeah

would i be able to throw sdxl loras at this?

or is that a no

[copper kraken]

don't think so, it's a different architecture

but hey, what's the point really

[gloomy lark]

Quinzel right now

[copper kraken]

sdxl loras already work with sdxl, see if you can get pixart sigma to do stuff you don't have loras for 🙂

[kindred shoal]

LMAO

Also something here ate 50gb of my storage

my s/o is gonna eat me alive

i ran 4 images and then 50gb vanished

[copper kraken]

that's prolly virtual memory

when you close comfyui it'll free up a ton of space

but yeah, one thing you might wanna do is install some kinda hack that makes it report your AI folder in MB not GB lol

[kindred shoal]

THATS how it works????

[copper kraken]

and be like shit, idk what's going on, maybe the HD is wearing out and some GB are getting deactivated

yeah, your computer dumps a lot of stuff to disk that it'd otherwise have in ram, no matter what you're diong

it can be a ton with comfyui

tens of GB

[kindred shoal]

AH. it randomly added 50gb to the virtual memory total i had set up

i had it set to 58 before for some stuff but now it's at 100

oh good it's going down

[copper kraken]

man the clownkarras sched is just excellent

[gloomy lark]

the good news, i got you a blue military tank. the bad news, the raccoons i hired are flying away.

seriously, this instruction i have for writing long prompts ends up with incredible stuff for ella which can handle them.

[kindred shoal]

I also know controlnet will be a real big no

so poses are gonna be..interesting.

[gloomy lark]

so i forget if it's ella or pixart sigma, but one of them has controlnet.

[kindred shoal]

it's probably ella

[copper kraken]

i'd imagine ella would, it's still sd15

[gloomy lark]

https://github.com/TencentQQGYLab/ComfyUI-ELLA

yeah it's got controlnet

[kindred shoal]

damn (pixart)

[copper kraken]

early may is at least six months away

[kindred shoal]

..am i having a stroke or did you type that weird

[gloomy lark]

so i can't find anythign that says 4k or 2k models for pixart. i can't figure out if i was imagining that.

[kindred shoal]

may is next month

[copper kraken]

and in AI years, it feels like at least six months away 😛

[kindred shoal]

ahhh

[gloomy lark]

what's in early may?

[copper kraken]

oh, SD3 maaaaybe

[kindred shoal]

speaking of which, how do i change the image size?

i might be blind

all i see is 'ratio'

[copper kraken]

yep that's what you use

[kindred shoal]

.. so like how does it work

[copper kraken]

no controlnet, no IPA except for a style one with a secondary image

[kindred shoal]

how do i know the res it's gonna spit out

[copper kraken]

from the aspect ratio

if you forget, here's a nice tip

set it to 1 step

and you'll know what res it is faster than you can google it

[gloomy lark]

you gather your cat army and have them work on good prompts for you. wow llama 3 is good at this.

[copper kraken]

i would like to gather a cat army

[gloomy lark]

hold on, i have that extra node around here someplace

[kindred shoal]

I still can't get over the fact that this is taking up as much ram as it is

94% like damn

[gloomy lark]

it all went to the language model. the graphics model is only 3 gigs of vram

[copper kraken]

ram was made to be used

[kindred shoal]

...So i set it to one pixel

and it spat out a mess

not 1 pixel

one step

how am i supposed to figure out res from this?

or were you fucking with me-

[copper kraken]

lol

i'm known for fucking with people, it's pathological, but i'm not here

right click and open the image in your browser and you'll see the resolution if you mouse over the tab and pause for a second

[kindred shoal]

god im a dunce

is there a way to make it taller than wider or?

[copper kraken]

yep, aspect ratio

pick something greater than one, run one step

pick something less than one, run one step

you'll see it flips from portrait to landscape

pixart sigma is the best i've used for diverse aspect ratios

[kindred shoal]

..ohhhhh it runs like 1:1 but 1.1?

[copper kraken]

sdxl is pretty limited there, and sd15 extremely so

1:1 = 1024x1024

[gloomy lark]

if you want it tall, set it to 0.68

if you want it wide, 1.67

[kindred shoal]

so 1.1 is 1024x1024?

[copper kraken]

1.29 = 1152x896.... 1152/896 = 1.29

yeah

[kindred shoal]

Okay now I get it

[kindred shoal]

oh my. fucking. god. GORGEOUS.

her head is massive but oh my god

[copper kraken]

pretty easy to keep a style going once you find something you like

[kindred shoal]

i wish i could find a regular model to replicate this

like a 1.5 model

[copper kraken]

look into IPadapter with the style weights, they work on SD15 now

[kindred shoal]

isn't IPa super heavy

i only have 6gb of vram and 8gb of ram on my laptop and I'd like to be able to match the style...

[copper kraken]

idk

run it on your desktop

[kindred shoal]

I would, but limited storage

[copper kraken]

only need a couple hundred extra mb for ipa

[kindred shoal]

that's fair...

[copper kraken]

different IPA strength

[kindred shoal]

so how do i get ip to mimic a style?

or does it..not do that and i misunderstood

[copper kraken]

it does do that for sure

get it installed and i'll drop you a workflow

[kindred shoal]

ugh i love how huggingface makes my firewall mad

which one of these do i download?

ik it's for 1.5

[copper kraken]

ip-adapter-plus_sd15.safetensors

and the one without plus in the name

shit, get the one that says light too

all the safetensor versions

those are pretty small, its just sd15

[kindred shoal]

do i need these?

ah, hugging face is being a dick again

ah no i dont, only if i'm getting specific versions

[copper kraken]

yeah you'll need clipvision

i forget which for sd15 but it'll start downloading it automatically

i think it might say on the page

[kindred shoal]

it says on the page that only 2 of them need clip vision, so i'll be ignoring those for the sake of simplicity

[copper kraken]

k

yeah if you end eup needing it it'll let you know anyway

[kindred shoal]

okay so according to the workflow im looking at, are you combining two images together?

or?

i also dont know what qualifies for "negative" embeds because wouldn't you want both?

..unless the neg is the style you're replicating, while pos is what you want it to replicate on.

ah it does need clip vision

it doesn't say which one though and there are two.

i am so confused

i figured it out, i'm jsut really damn stupid

[copper kraken]

oh good is it working now?

[kindred shoal]

yeah, but i'm trying the sdxl one and it's not wanting to load in?

not showing negative because it has some nsfw prevention tags

i know my workflow is a mess, i apologize

[copper kraken]

you clearly haven't seen any of mine 😄

[kindred shoal]

I have to keep mine spread out, too close together and I lose track and what goes where

[copper kraken]

current view

[kindred shoal]

oh jesus

are you creating like multiple prompts?

[copper kraken]

all kinds of crazy shit

lots of detritus

[kindred shoal]

idk what that is 😆

but yeah, my main issue is sdxl's ip adapter not showing up

it's in the right location

unless i need the other clip

[copper kraken]

type detritus in as a prompt

[kindred shoal]

oh it's a word. debris

i thought it was some comfy ui thing

[copper kraken]

well, with me, it's a comfyui thing haha

did you restart comfy

[kindred shoal]

o no i didn't. i just clicked refresh

yeah it's still not there

i cant look at hugging face since 90% of the time my firewall blocks it for hwatever reason

[copper kraken]

wtf that's weird

oh y ou have it

it doesn't say sdxl

cuz it works for both

[kindred shoal]

does it?

it spits out random colors whenever i try to use it...unless i messed up somewhere

maybe loading a different vae was what messed it up...

no, going back to how it was and it's still doing random colors.

i'll try another sdxl model...just have to dl it first

[copper kraken]

set "weigh type" to style

weight

[kindred shoal]

uh where is that

oh i found it

wrong area

okay well it printed out half nonsense

i dont think sdxl is gonna be a team player

[copper kraken]

set method to concat

can't really see what you're doing

[kindred shoal]

bahh im starting to fall asleep anyway, i should probably call it before my brain explodes

[lost quiver]

how can i make pictures ？

[hushed crystal]

Lady running towards her office and the huge monster is chasing her, v3

[upbeat summit]

very nice!

[upbeat summit]

beautiful!

[crimson ingot]

beautiful woman💗 on the beach, full-length, squatting, white t-shaped swimsuit, front camera view on the joint of legs, legs bent at the knees sideways, hair developing in the wind, aesthetics of the female body, beautiful tan, best photo, photorealistic, hyper-detailing

[vagrant olive]

111

[strange dune]

Mayan Secret Land hand-painted original painting

[restive marlin]

promt: make a picture of kit from brawl stars brushes his teeth

[wanton river]

5 step diffusion using custom sdxl-lightning model

[wet nacelle]

@humble capeHey. Check out this Lora set. https://civitai.com/models/310571/boring-reality

Works best with image2image. Does work best in Forge and Auto from what I've tested and read from the creator.

Above image uses depth Controlnet with the Loras.

[wet nacelle]

@humble cape If you get something looking great then let me know. I would love to learn from what you're doing.

[final ibex]

I'm a couple of days late here, but in one of the ELLA tutorials, it says to download ALL the t5 files, which ends up being like 50gigs or more, and they seem like duplicates in different versions.

Wish I had more time to dig in. It feels like I get things set up, go to bed, read more, set up more stuff, and ravage hard drive space.

[gloomy lark]

https://github.com/TencentQQGYLab/ComfyUI-ELLA?tab=readme-ov-file

you want this page.

under models, it shows you what you need. only 3-4 gigs total

[final ibex]

@gloomy lark Thank you, I do appreciate it!

[gloomy lark]

did this with ella. trump jumping over gold bars

he's way up there in the sky. he's got some lift.

[humble cape]

Thank you for sharing, you just recalled this lora, I think I will use it on fooocus and combine it with the vhs and old lora, I will send you the result

[young oriole]

Flower pot

[copper kraken]

Here is the image you requested.

[young oriole]

Thank you

[lofty cargo]

Mayan Secret Land hand-painted original painting

[meager canopy]

More "painting" like

[main sleet]

Prompt? I actually realized how I only go for SD3 now in this discord. Its a huge difference, as soon as you saw the possibilities

[meager canopy]

Mayan Secret Land hand-painted original painting

[wet nacelle]

Awesome

[hot shadow]

a cute cat, cyberpunk art, by Adam Marczyński, cyber steampunk 8 k 3 d, kerem beyit, very cute robot zen, beeple

[chilly rapids]

universe

[humble cape]

looks great

[wet nacelle]

Do you mind going into detail of how you got this quality?

[humble cape]

I use fooocus, with these loras ;
https://civitai.com/models/182895/vhs
https://civitai.com/models/370572/old-camera-vhs-footage-mix-aesthetic
and boring reality

and i use llama-3 70B for improve my prompt

@wet nacelle when i use image2image i use these pictures https://www.pinterest.fr/Dreamvoid/liminal-spaces/

[wet nacelle]

Thanks for the info

[humble cape]

no problem

i love this lora

[humble cape]

@wet nacelle do you tryed image blending ?

[wet nacelle]

I don't know a thing about that actually.

@humble capeAlso the Loras I sent need to be combined for full use.

Not sure if you just used one or more than.

[humble cape]

i combined the loras

[wet nacelle]

nice

[humble cape]

like that @wet nacelle

[wet nacelle]

I mean use multiple boring loras

[humble cape]

when i installed the lora i just got one

[wet nacelle]

Use the versions that are posted on the page. You've used Civit before. You should know that multiples files and versions of a model can be listed at once.

@humble cape

[humble cape]

ha

yes i installed primary v4.0

[wet nacelle]

@humble capeI really do recommend that you read what the author has to say about the Loras.

It helped me understand how to best use it.

I switched to forge so that I could use depth mapping too.

[carmine egret]

change it to ai

[humble cape]

yes i will read

[copper kraken]

Switched from?

[wet nacelle]

for just a bit with what I'm using.

I was using Comfy.

[meager canopy]

https://tenor.com/view/saturnz-barz-pizza-saturnz-barz-pizza-gorillaz-pizza-gorillaz-gif-15023235

[copper kraken]

you can do depth with comfy

[wet nacelle]

I know. I just don't want to set it up or use a workflow

@copper krakenHa I forgot the don't*

[noble shoal]

My god the most important word.

[wet nacelle]

I can't find any gif or image of Mobility Mary. The only thing I can say right now is lol.

[wet nacelle]

It's got Free something

something is free in that van

[stoic snow]

Dreamshaper Turbo IPAdapter.

Dreamshaper Turbo IPAdapter.

[meager canopy]

Looks like "Toes"

[wet nacelle]

https://tenor.com/view/tom-cruise-shades-cool-smug-cool-shades-gif-18677159

@humble cape

[humble cape]

That's cool , you blended two images ?

[wet nacelle]

four

[humble cape]

great idea

[orchid cobalt]

where is a channel or site for workflow?

I am lookign for workflow that allows to instantly process simple drawings to make decent art out of them

[noble shoal]

Do you mean ComfyUi workflows?

[orchid cobalt]

yes

[noble shoal]

You might search here:

https://openart.ai/workflows/home?gad_source=1&gclid=CjwKCAjwuJ2xBhA3EiwAMVjkVKe4VFC-NTJyoxKBwcS7_0dqC41UTJyW8Bc8mNardFkkTPezTif_TxoC86MQAvD_BwE

[orchid cobalt]

thank you!

[humble cape]

😭 bruh

[wet nacelle]

HA

[humble cape]

cool mini house

[wet nacelle]

thankies

[glass forge]

I know what I'm trying out today.

Also foooooocus seems to downgrade quality? I got way better result sin forge using almost the same settings (cfg seed etc)

I thouht we were friends fooocus!

[humble cape]

remove the styles for down quality

[wet nacelle]

@glass forgeWAIT! G-STRING!!?!?! ARE YOU THE CREATOR OF G-STRING!!!!!

I just wanna say that you are a massive inspiration! mad respect!

[glass forge]

🙂

[noble shoal]

Just googled it, the g-string was invented in 1846. I doubt that you are talking to the right person @wet nacelle

[wet nacelle]

bruh. goofball

[noble shoal]

Ok, what's G-String? I will obviously find nothing on web searches with that name.

[wet nacelle]

Check out steam

@noble shoal keeps typin but nothin ever comes out from em

[noble shoal]

Very cool @glass forge . Always nice to see people succeed with game development. I mean, your game even has a Wikipedia page. And mine? Let's say it's been in pre-alpha for a few years and had like 4 engine switches

[mystic fable]

Hey all, I hope I'm at the right place to seek for some help.

I'm having trouble with sdxl turbo, after adding it to the webui it produces quite weird low quality images. I've followed a youtube tutorial how to set it up, the guy gets awesome results and i get garbage. Normal SDXL works perfectly fine (but its slow) What am I doing wrong?
(on mac studio m2 ultra)

[wet nacelle]

fuck

I @ td the erong person

fucj myu joke

rweghio'8u;

I can't type

[noble shoal]

I have a slow typing syndrome. I also want to build up high expectations for other users that a magical long text will appear. But it won't.

[wet nacelle]

ha funnie

[glass forge]

Thanks and don't give up. No matter how tough and disappointing things out there are at least if you do whatever your calling is at least you'll always have that satisfaction. No one can take that from you.

[noble shoal]

Now those are some nice words to end my day with. Any plans for new games at the moment?

[glass forge]

Well since the game in it's current form on Steam isn;t the best version it can be I am working with LunchHouse on the Ultimate Edition. after that maybe a smaller game on Unreal or an expansion pack and much late ron aSequel... More or less...

[wet nacelle]

Was G-String ever planned for unreal during retail development?

[glass forge]

No it started as a Source mod many many many years ago.

[wet nacelle]

I thought so. I watched and old A Jolly Wangcore stream of your game and he thought that it moved to unreal right up until he opened the settings.

[glass forge]

Hehe

[meager canopy]

Failed controlnet

[red minnow]

The bottom one is failed? I think it looks miles better than the top one

[shy kelp]

Or just wait for a llm to do it

[mortal marsh]

Collective generated canvas. Today, we've hit the milestone of 1,000 cells generated 🎉. The image only shows a small portion of them.

[glass forge]

What's the most amount of LORAs you crammed into one generation?

[gloomy lark]

me?

none

i don't use loras

this is all ella passed through sdxl

[glass forge]

i meant in general anyone here

and whats ella

[gloomy lark]

ella takes stable diffusion and takes the clip text encoder and replaces it with a real language model.

like sd3 does.

I assume trustory and a couple of the others use lora. doubt they can do what they're doing without it.

[copper kraken]

do you have a link to the full version? that's way cool

[short copper]

SDXL with HiDiffusion

w/ HiDiffusion

[gloomy lark]

what's hi-difussion? something worth looking into?

[minor pecan]

[glass forge]

ELLA plugins out for anything other than comfy yet?

[gloomy lark]

don't know yet... i think i've gotten a well oiled workflow for comy though for it

let me know if you have a prompt i can try with it

[short copper]

https://hidiffusion.github.io

[short copper]

No clue

[gloomy lark]

is it a comfy node? I'm not seeing how one actually uses it outside of a command line from that page.

[short copper]

there’s a notebook in the git you can use

you’re just running all of the code to get sdxl base up

and then you’re adding the hidiffusion “hack”

[gloomy lark]

oh i see.. so would it eventually be a new comfy node, or do people have to do something to the checkpoints?

[short copper]

yeah i’m certain someone will drop a node in no time

[gloomy lark]

awesome. thanks for the info

[glass forge]

Wes Anderson's "The Bear Factory".

[jolly seal]

#🏞｜general-with-images "Chanel Iceland Press Conference, Fashion Show Concept, Unique Ice Volcanic Landscapes and Black Sand Beaches of Iceland, Complemented by Klein Blue, Exudes Luxury"

[marsh bridge]

Can we gen images here like the olde dream commands?

[gloomy lark]

bots are currently down.

probably back up when sd3 goes final.

[marsh bridge]

So are people gen these images from local hosts?

[mortal marsh]

Sure 🙂 https://www.hexagen.world/

[meager canopy]

Taste the rainbow.

[humble cape]

Is not a rainbown💀

[meager canopy]

I can see that. Haven't you seen those "Skittles" adverts?

[humble cape]

No lol

[humble quiver]

Taste the rainbow.

[normal scroll]

Portrait photograph of ananthropomorphic tortoise seated on a NewYork City subway train.

[topaz linden]

#🏞｜general-with-images " draw a car"

[meager canopy]

https://tenor.com/view/draw-gif-24443621

[kindred shoal]

does anyone know of an open pose model for sdxl?

I can't seem to find one to save my damn life

...found it, i just did something weird

[kindred shoal]

now the preprocessor isn't working

[kindred shoal]

oooh those are dope subliminal places

[kindred shoal]

how do you get SDXL to do text?

[ashen otter]

No

[copper kraken]

Here is the image you requested.

[humble cape]

Amazing man

[wet nacelle]

thankie thankies

[humble cape]

You can get sdxl in fooocus

[wet nacelle]

Looks a lot like the artist Calm

[shut crescent]

yep its lora but still Calm do better 😄

[wet nacelle]

True. I've been using the Pony Lora.

[shut crescent]

try my lora i droped it today

[wet nacelle]

Link it. Not sure I saw it when I was searching

[shut crescent]

sended u a link in PM i dont know if i can share it here

[wet nacelle]

Thanks

[shut crescent]

enjoy it 😉

[glass forge]

How do you make these? a lora?

[glass forge]

Do SDXL loras work with Pony checkpoints? (they do impact the images) But I'm not sure for the better...

[rustic garnet]

pony is so weird and far away of sdxl base, most Loras won't work with it

[burnt shadow]

SDXL pure base model, implementing artists along within the prompt

[rancid lagoon]

Picture of hand holding kitchen waste

hand holding the food waste

SDXL pure base model, implementing artists along within the prompt

[kindred shoal]

does anyone know how to make transparent sleeves?

[tawdry current]

Like someone wearing a clear poncho? Maybe try clear vinyl?

[kindred shoal]

Like a long sleeve shirt but the sleeves are clear

[tawdry current]

Clear as in what?

Not there?

[kindred shoal]

like this

[burnt shadow]

you could buy a fishing line and some knitting needles 😀

[kindred shoal]

(not ai generated, this was yoinked from google)

[tawdry current]

Run it through a clipG interrogator to see what it thinks it sees

Maybe try clear silk or something like that

Thin silk

You can use other vision models to describe it as well, like llava and such

But using a clip one would be better since sdxl uses clip

[kindred shoal]

i think a word I could use would be..sheer?

but i'll wait for clip to come back to me on that.

oh. it error'd out

hopefully it doesn't error out again lmao

it gave me...nothing involving sleeves.

[burnt shadow]

so what has been your prompt so far?

[kindred shoal]

uhhh... it's not letting me do img2img

sdxl is throwing a fit for..some reason

ah. there we go. now it's working. dunno wtf happened there.

This is the prompt it gave me
a woman wearing a skirt and a top with colorful beads on its sleeves and a black leather skirt, Avgust Černigoj, fashion, a hologram, kitsch movement,

[burnt shadow]

I was referring to the prompt you used for possibly achieving transparent sleeves

[kindred shoal]

oh

it was like... clear sleeves, plastic sleeves or something like that.

[burnt shadow]

Have you maybe got the complete prompt?

[kindred shoal]

...oh now it's ignoring my prompt entirely, that's fun

i broke something here

oh it's because i switched the model and forgot to switch it back, hah

oh great now the model isn't working again

[tawdry current]

i upscaled it with an ai upscaler to 2400x2400 and then ran llava 1.6 mistral on it and it says " The transparent parts of the sleeves are made from a sheer material, which allows for visibility through the fabric. This sheer material could be a lightweight mesh or lace that is often used to create an illusion of transparency while still covering the skin. This type of material can add visual interest to an outfit and can be particularly suitable for evening wear or fashion-forward styles.

[kindred shoal]

i'mma close out and reopen forge and see if that fixes it

[kindred shoal]

there we go, it fixed it

[kindred shoal]

i will try the sheer sleeves prompt!

[tawdry current]

i'll mess around a little as well, it's a really odd type of look that even i have troubles describing other that "transparent silk" or something like that. guess my lexicon of textiles is lacking lol

shit feels like trying to describe a colour you haven't seen before lol

[zinc pumice]

Encountering a female queen making her grand entrance into the WWE arena wearing a resplendent Queen of Hearts costume, complete with the prestigious WWE belt, was a sight to behold. The elaborate and regal attire, paired with the symbol of victory that was the WWE belt, exuded an aura of dominance and authority.

As she made her way to the ring, the intricate details of her costume, inspired by the fabled Queen of Hearts, captivated the audience.

[kindred shoal]

Positive prompt (it's a lot, don't make fun of me):1girl, fish girl, (head fins), (blue skin), (bangs over eyes:1.2), (red hair), (blue hair), (purple hair), (crying:1), gradient hair, multi-colored hair, <lora:LineArt Mono Style LoRA_Pony XL v6:1>, lots of scales, scaley, scales on face, facial scales, skinny, long hair, thin arms, grey thighhighs, covered feet, bar setting, solo, white button-up shirt, black tie, red latex shorts, short shorts, medium breasts, <lora:badhandssdxl:0.5>, long sleeves, zPDXL, sheer sleeves,
Negative: eyes, negativeXL_D, water, 2girls, mermaid, tail, barefoot, muscle, nsfw, nude, rating_explicit, bad anatomy, 3girl, multiple people, long arms, barefoot, zPDXL-neg,

god...sdxl needs to stop breaking.

[tawdry current]

yeah sheer sleeves works

also, stop prompting SDXL like it's 1.5

[kindred shoal]

but it works fine

[tawdry current]

true, but get out of the habit of it if you plan on using sd3 and want actual prompt cohesion

[kindred shoal]

nah, i use 1.5 on my laptop, and only dabbling in one model for sdxl on the pc

[tawdry current]

tested with some other prompts and sheer definitely works though

guess you learn something new every day. today's word was "sheer" in reference to shirt sleeves lol

[kindred shoal]

that's good, I'm gonna have to do some prompt examination though

gonna have to fiddle about and try and get the...sheer part to work.

[tawdry current]

without the transparent sleeves part of the previous prompt, same seed

[kindred shoal]

it wants to work, but it also doesn't want to work lmao

[tawdry current]

what model are you using?

[kindred shoal]

pony diffusion xl v6

it's the only sdxl one makes what i want to make, lmao

[tawdry current]

yeah, i don't use those models nor have i even touched them

[kindred shoal]

what the hell is going on