✨｜sdxl

[floral island]

ok, first test -> seems it knows the tags i need it to know

now it's just a matter of tuning weights and other prompting elements

[static wigeon]

I was also using Anything Model. It was good but it was giving me a hard time since the prompts length is long

[floral island]

very nice 🙂

[static wigeon]

If you don't mind, can you give me a short sample prompt with Weight Definitions since I am unsure on how to add weights

[floral island]

i stay away from anime models -> they're never creative because they are GROSSLY overtuned on anime

[shy kelp]

yeah protovision is nice, think one of my first gens in xl was with protovision

[floral island]

you do know the () and [] ?

[static wigeon]

I am not going for an anime art style, it should be similar to Western Comics if ykwim. Like maybe Archie, Marvel comics or Garfield

[static wigeon]

I unfortunately don't

[floral island]

they're core elements in A1111 and i'm using those a lot

[static wigeon]

I tried asking Bing AI and it asked me to add + after important prompts or specify what is important through text like "A Cat playing with a ball. The ball is important". None of these worked 😭

Is there a documentation I can refer to?

[floral island]

yes

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

[static wigeon]

Okiee thank youu, will look into it

[uncut fiber]

try add storyboard. To prompt

[static wigeon]

What is Add Storyboard

[floral island]

the token "storyboard"

that makes such collages too

[static wigeon]

Ohhhh, so like a comic format

I am using a LoRA which helps me with that

[uncut fiber]

i made it only with sketches

[static wigeon]

I seeee

[static wigeon]

Oh that looks pretty cool, will give it a shot

I am constantly gettting this error "RuntimeError: Not enough memory, use lower resolution (max approx. 320x320). Need: 0.0GB free, Have:0.0GB free"

It works then randomly after a few generations, I get this error

[floral island]

something like this?

[static wigeon]

That works too

This is one of my generations using my current Model and LoRA

[static wigeon]

Any idea on how to overcome this ;-;

[uncut fiber]

@static wigeon would ask in tech support

[static wigeon]

Okiee

[shy kelp]

i have such a backlog of stuff that never did make it to final looks, one day

[shy kelp]

hey wolves

[ionic gulch]

ip adapter needs one model in one of the model directories of comfyui and one inside the model directory of the ip adapter custom node. i'm not at my pc at the moment, so i can't check the exact directories.

was this a question about my workflow?

[hardy cipher]

yeah

only node I didn't have was the seed generator. so tried to swap it out

[ionic gulch]

the new prompt node should still allow {a|b|c} and filename syntax - just no separate output in a second text field.

the seed generator is great and worth it to get the node through the manager. it allows to recycle the last seed, which the official seed generation doesn't offer.
so you can for example turn off the upscaler and if you like the result turn it on, press recycle seed and generate with upscaled versions
with the default seed generator you would have to import the old image first if you would like to reuse the seed
so it works more like a1111/sd.next seed generation, which is imo much more useful than the comfy default

[hardy cipher]

ahh. well I might download it then

[floral island]

captain sparrow? 😮

[heady vale]

IPA seems to want to place longer images off center for some reason

[ionic gulch]

yes, ipa is broken for 21:9, and 16:9 puts the subject off center most of the time - there is a second set of ip adapter nodes available in the manager now - will have to try if that one fixes those problems.

[heady vale]

I guess I can resize the input image and see if that gens cars better. I wasnt using square images

[ionic gulch]

changing input images doesn't help. already tried a lot of variations. from my test the problem is the target aspect ratio.

[floral island]

best marriage photo ever!

[soft zealot]

not sure which one you're using but I've got

For IpAdaptor as of Aug31 use the Dev Branch not the main branch

https://github.com/laksjdjf/IPAdapter-ComfyUI/tree/dev

The model required (which gets saved into the IpAdapator custom nodes model folder) can be obtained from

https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models

[floral island]

is this the famous lady of the lake?

[eternal fog]

Try passing the IPAdapter into an unclip conditioning node and up the strength on that. That seems to help with things being pushed off centre.

[heady vale]

Ok. Will try that also

[eternal fog]

Why use dev? Main is up to date with it now.

[soft zealot]

in which case ignore me , just what I was using that seems to mostly work 🙂

[eternal fog]

I've only gotten the off center issues when using 2 IPAdapters with different aspect ratios. It's always been fine with just a single image.

[ionic gulch]

afaik the main branch also has the fixed version now, after the first version was broken by some update you needed the file from dev branch for1-2 days, but the main got updated too. "last change date" for both versions is "2 weeks ago", so i don't think dev branch has any advantages now.

[heady vale]

padding the input image to make it square and changing the latent image size as well fixed it

[ionic gulch]

try 21:9. it always puts the subject to the border of the target image for me

sometimes it draws a second subject in the middle, but the off center subject was always there in my tests.

[floral island]

ok. this one is pretty cool

[ionic gulch]

try without changing the input image, only the latent/target image size. should work too. at least better than before.

[heady vale]

that what I started with and it was off, then I changed to a wider A:R. both didnt work for me

[ionic gulch]

ok, will have to try again tonight.

[heady vale]

8/8 cars are centred perfectly now with the padded image

[shy kelp]

finally getting the generation to look at the viewer and it seems like it doesn't want to even with (1.4) lol

[floral island]

some models don't know the tag "looking at viewer"

[shy kelp]

[floral island]

you could try "looking at you", or simply "frontal portrait"

be creative ^^ there's a lot of way to describe someone looking at you

[shy kelp]

yea but then i end up with the pov look lol

[hardy cipher]

tell sdxl to do what you say or else

[floral island]

add sad/sullen/neutral in your negative prompt -> that makes sure your people aren't those things

[shy kelp]

sometimes i do staring or looking at camera, looking away

[floral island]

"sideways look"

might also be a thing

[hardy cipher]

say "hey, stupid idiot, look at the camera this time"

well type it

[noble shoal]

Are you sure about that?

[floral island]

lol xD they do be watching you

and they're all 100% psychopaths probably

[noble shoal]

Yep

[shy kelp]

im not sure if it's the model since it's the same that generates these, hmmmmmmmmm

[floral island]

lets redo that prompt and att psychopath at it 😄

[shy kelp]

try adding reflectivity and caustics to this

[floral island]

found one: sideways glance

[floral island]

caustics seems to be very low power, if not even just simply ignored

[uncut gull]

Getting closer!!! She really is behind the glass this time.

[shy kelp]

hmm, i should do some test with water stuff i have, i just threw it in there and paid no mind
i find epoxy is sometimes nice for water too

[floral island]

pure prompt?

[west breach]

try adding 'diamond plated' to a prompt

[noble shoal]

Got a lot of interesting psychopaths wedding photos

[floral island]

holy shit, wedding i'd love to not attend XD

[noble shoal]

You might end up in a basement

[floral island]

between such a pair, the body count inside their house would be 10 at least, even before marriage

[floral island]

MIGHT?!

[floral island]

one of my favorite hatsune miku songs, translated to english, converted to my prompting style, added anime flavour. this prompt never seizes to amaze me ^^

[floral island]

"worst things to see at a wedding, ca 1980"

[uncut gull]

Yes! Just prompting, and I'm at 25% hit rate. You only need ~4 gens to get the effect.

[west breach]

prompt?

[shy kelp]

diamond plated

[shy kelp]

oh i should do it like that

[floral island]

add bismuth too!

[uncut gull]

Up to 37% hit rate now. (You can check the PNG info here for steps model etc. if you have a PC.)
in a wall, transparent window of glass reflecting sky, [woman covered by translucent reflections:0.2]

[shy kelp]

jade dragon

[floral island]

this cat is richer than any of us will ever be

[uncut gull]

We can reflect any scene ontop of her. 🙂

[shy kelp]

hmm

[west breach]

looks like she was photoshopped in 😄

[floral island]

she's an actual window sticker 😛

[uncut gull]

Yeah, see my prompt. This is a super hard challenge.

[floral island]

prompt pretty unstable though 😦

[shy kelp]

do NOT touch the sun

that's how you get burned

[floral island]

surprise lgbtq+ gateway in front of your house

helluva concert?

[uncut gull]

Still no luck with raincover though. It's way better than before, but SDXL just really struggles to layer textures like this.

[floral island]

even with creative prompting i get too clear subjects

probably can't reference the subject directly, and then it'll work

[uncut gull]

...just use my prompt though... I've tried tons of prompts. All of zero % worked until this exact prompt. It's not creative, it's super-exact.

[west breach]

who is inside or outside here?

[native knot]

Yes.

[floral island]

one of these is not like the others

[native knot]

Barble

[glad grove]

tbh i kinda wanna taste some n burgers right now

[floral island]

i'm getting a frikandel xxl today!

[floral island]

If you're ever in holland, most consumed snack/fastfood here would be the frikandel, so make sure you've had one

this demon about to summon a lawyer?

[native knot]

Futurepope has something to say.

[shy kelp]

man this window thing is kind of hard to achieve lol

closest i got was this

[uncut gull]

It's close. I posted a prompt here somewhere. (What I actually want is rain / fine foreground structures in general. 😭)

Ack! So close! The raincover texture just really wants to avoid covering the nose / eye region.

[west breach]

is this a building for ants

[shy kelp]

i don't think i've ever had success with anything wet
could prompt wet hands and the entire body will be wet lol

even doing something like
(wet (hands:1.4)) lol

[floral island]

put wet in negative too

[uncut gull]

Uhm... Uh, success!?! It's rainfall over her face, right? I'm not imagining things?! I just got ~25% hit rate.

rain, glass window, translucent reflection of rain, [woman inside rain reflection, woman portrait (covered:1.5) by translucent rain reflection:0.15]
Negative prompt: dew
Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 481981248, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, Version: v1.6.0

Went from total failure to success by adding a comma after the first "rain" in the prompt, but this is too delicate still. Not sure why it works.

[floral island]

edgelord needs to go on a quest to find some heart

[shy kelp]

wonder if you can blur the woman a little, maybe could help, like 0.1-0.4

[uncut gull]

Prompt? I've never blurred a subject before. (I need to work though, probably done for now.)

[uncut gull]

LOL. Yeah, you can make water appear on skin. It's "floating" textures that are super hard: glass reflections, a sheet of rainfall, mesh wire fencing, etc. 🙂

[shy kelp]

this would be blurred background and blurred crowd

[uncut gull]

How do you blur the subject / foreground?

LOL might be another hard challenge though.

[shy kelp]

could do like [(blurred:0.1) woman inside rain...
not sure how well that would work though

[floral island]

moon doesn't high five me back 😦

[shy kelp]

pinching moon

[shy kelp]

lol these moon gens

[floral island]

hmmm, changed out anime cell shading for surrealism, abstract photography

[shy kelp]

i like the detail on that cracked heart

[floral island]

oh shit, should've done that A LONG time ago

these next ones are going to be amazing... i think

[floral island]

new didney movie

"pinchg"

[floral island]

that's a nice office ❤️

[glad grove]

i think thats a mexican movie "Pinche"

[west breach]

ok we have a winner

[floral island]

is controlnet involved?

[west breach]

report hazards to your manager

[tepid maple]

nope

[floral island]

very nice ❤️

what did you prompt to get that kind of pose

[tepid maple]

parameters

kinetic office scene capturing swirling papers in mid-air, frantic hands reaching out from every corner, and high-heeled shoes and polished loafers pounding on a gleaming checkered floor
Negative prompt: worst quality, bad quality, poor quality, ugly, ugly face, blur, watermark, signature, logo
Steps: 32, Sampler: Euler a, CFG scale: 7, Seed: 595166390, Size: 1344x768, Model hash: a13b3d3d1e, Model: UnstableDiffusers_v6_SDXL, Denoising strength: 0.3, Hires upscale: 1.43, Hires upscaler: SwinIR_4x, Version: 1.5.2

simple isnt it 🙂

[floral island]

lemme run that one too ^^

[tepid maple]

i like simple

less is more

[floral island]

i like uhh... unpredictable outcomes 😄

[tepid maple]

sometimes I just make CFG 2 or 3 to see what happens

[shy kelp]

thats what's up, i always hover around 2-4

[floral island]

don't ask how or why this happens, and just accept this was the outcome of the ridiculous prompt 🙂

[tepid maple]

let me show you mine version

for this one

[shy kelp]

i wouldn't say that's ridiculous at all

[shy kelp]

it's 100% safe (do not eat the house pls)

[floral island]

neon stairway to heaven

[shy kelp]

how its work

[tepid maple]

yep

Prompts:
Neon Lighting Staircase to Clouds over Sea, 3d render, realistic
Negatives:
worst quality, bad quality, poor quality, blur, watermark, signature, logo, username, text, out of frame, (ugly face:2), twin

[floral island]

just have a few more prompts to parse into my master text... DAMN i might have a problem 😛 (250 lines in the master prompt)

[tepid maple]

any good github repo for sdxl style library ?

most of them I fonud only for painters artist

[floral island]

make up your own styles? 😄

[tepid maple]

I have that actualy

but need big library

I wil lstart building mine at this point

looks like

[floral island]

lol she got heart-broken, for real

[floral island]

#4 though 😮 i like that flame hair

[tepid maple]

now I am stuck making posters

[floral island]

uuhh... this next prompt might... not make this board...

[tepid maple]

yep

[floral island]

playing cards with an island. right. that's a thing that could technically be done.

[floral island]

just. 73. more. pages. to go

[west breach]

I need some good prompts for testing pixelwave 2

[floral island]

want my current blob of bullshit prompts for dynamic prompts? 😛

[west breach]

sure!

[floral island]

@west breach dm 🙂

aight, back to lyrics!

[untold bear]

Is 60 images good for an sdxl model?

[glad grove]

for a lora yes, for a model no

[untold bear]

okk because it was a lora!!

ty

[floral island]

earth-shattering fart?

[dense chasm]

T2I-Adapter SDXL is awesome,more lightweight of params and run faster than controlnet

[shy kelp]

i just can't get rid of facial features on a generation

like i just want the head but no eyes, nose, mouth, nothing and these negative prompts aren't listening

[tepid maple]

@craggy ibex

[west breach]

you tried 'faceless' ?

[tepid maple]

lets see

testing

[wet nacelle]

Alright. does anyone know of a good "smoking a cigar" XL lora?

[tepid maple]

no need lora sdxl makes smoke soooo good

[shy kelp]

yeah, getting some good results with helmets though, like full face helmet

[tepid maple]

I think I know why

[wet nacelle]

Write me a working prompt then.

[tepid maple]

haha

[glad grove]

(cigar on mouth:1.5) ?

[wet nacelle]

@glad grove@tepid mapleleogirl, male head shot of the joker smoking a cigar

[tepid maple]

hmmm

[glad grove]

try with male,portrait,the joker, (smoking,cigar on mouth:1.5)

[tepid maple]

puro better

[craggy ibex]

Can you teach me?

[tepid maple]

nvm

[craggy ibex]

I use SD 1.5

[wet nacelle]

Trying that now.

[tepid maple]

I use SDXL prompting more easy

[wet nacelle]

too true

[tepid maple]

yessss

[craggy ibex]

My computer is not capable of using SDXL tbh

[tepid maple]

I wanted joker but its fine

[wet nacelle]

(racial slang) c'mon

[craggy ibex]

Potato computer

[tepid maple]

my CFG set to 3

this result normal

its a bit wild

[craggy ibex]

@tepid maple And congrat

[wet nacelle]

@tepid mapleIt might actually be due to the fact that I'm using a model that may not have the proper data.

[craggy ibex]

you did something that is impossible

how the fuck you generated 2 people hugging

I don't fucking understand

[tepid maple]

I used unstable diffusion model

woudl recommend it

[craggy ibex]

I'm being serious

[tepid maple]

maybe it is easier in sdxl ?

[wet nacelle]

Who is a good mod or admin I can contact? I want to make sure I can write "(racial slang)" without getting a ban.

[tepid maple]

SDXL has better textual udnerstanding

[tepid maple]

parameters

man and women hugging, background galaxy
Negative prompt: worst quality, bad quality, poor quality, ugly, ugly face, blur, watermark, signature, logo, frame
Steps: 32, Sampler: Euler a, CFG scale: 3, Seed: 864603117, Size: 1344x768, Model hash: a13b3d3d1e, Model: UnstableDiffusers_v6_SDXL, Version: 1.5.2

[craggy ibex]

That's just it?

[tepid maple]

see

yep

[craggy ibex]

What the fuck?

[glad grove]

its easier in sdxl but 1.5 can also do it

[wet nacelle]

@glad groveI don't know how this happened.

[tepid maple]

nothing special no lora nothing

[craggy ibex]

What's your GPU?

[tepid maple]

gpu doesnt matter

but 3060

[craggy ibex]

You guys are probably rich people

yep

definitely rich people

hmph.

[tepid maple]

dude 3060 cheap af

[glad grove]

$200 3060 is not rich

[craggy ibex]

In Thailand, it's not

[tepid maple]

go work 2 days you can buy it

ooh

10 days than

sorry

[craggy ibex]

Nah our country broke as fuck man

;-;

[tepid maple]

sadge

[west breach]

come to australia, pick some fruit 😄

[tepid maple]

bad boy

[wet nacelle]

@uncut steeple Can I type "(racial slang)"?

It's not a bypassing right? If so then how?

I don't want to get a ban and then be told "you knew what you were doing".

[west breach]

working holiday very popular in australia

[tepid maple]

hmmm

gray area

probably depends how bad the word is

for example like "black women" in prompt shoud lbe fine

but cant say if you go deeper 😄

[wet nacelle]

I'll DM you real quick for what I'm meaning.

[sullen laurel]

Can anyone explain how Clip G and Clip L works?

[wet nacelle]

@glad groveGot some stuff working

[glad grove]

try to add smoke too like : male,portrait,the joker, (smoking,cigar on mouth,smoke:1.5)

[wet nacelle]

no.

[wet nacelle]

Thanks though

Just doing some tests

[tepid maple]

hahaaha

wait wait

holy

[wet nacelle]

moly?

[tepid maple]

nah batman looks cooler

because he doesnt smoke

😄

[shy kelp]

war

[dim steeple]

Is it true that 1024x1024 is the highest square resolution you can use in sdxl?

[shy kelp]

i use 768x1024

[tepid maple]

I use 1344x768

google the recommended sdxl sizes

[dim steeple]

i am talking about square

[tepid maple]

if you have godo gpu yes

you can go bigger

nothing stoping you 😄

batman approves

[glad grove]

weedman on the back lookin fresh

[tepid maple]

ye

[shy kelp]

cozy

[glad grove]

smoke coming out of his ears 🤣

[tepid maple]

thats hoyl spirit

not smoke

[glad grove]

🙏

[west breach]

love an L shaped cigar

[sullen laurel]

So,can anyone explain how Clip G and Clip L works?

[tepid maple]

hrad question no idea

never used it

spiderman vs weedman

spide r?

[west breach]

clip g is natural language, clip l is tags (supposedly more useful for prompting the style)

[uncut steeple]

Id say no, just to be on the safe side

[tepid maple]

yeeees

[uncut steeple]

also why

[shy kelp]

those are some nice spideys

[dim steeple]

Does SDXL get distorted if you go above 1024 x 1024x 1:1?

that is my question

[tepid maple]

hmm

[dim steeple]

like 1280x1280

[tepid maple]

because it is square I say not much

[dim steeple]

I tried 2046x2046 and it did

[tepid maple]

square resutls always consistant mostly

[eternal fog]

Is your comfyui up to date?

[tepid maple]

hmmm

why do 2k directly ?

[uncut steeple]

yeah

[tepid maple]

do it upscale ?

1024 and 2x

[glad grove]

theres a reason why theres a list of recomended resolutions

[wet nacelle]

damn. Alright. Thanks for the info.

[glad grove]

just say obama and u good 👍

[tepid maple]

lol

[wet nacelle]

That's honestly worse.

[glad grove]

ugh obama

[tepid maple]

well after him all fuked up

tbh

he was best of the worst probably 😄

[uncut steeple]

Usually the answer to all "can I? Am I allowed to?" etc. questions is no

[glad grove]

😔

[shy kelp]

[tepid maple]

sdxl so easy I love it

[shy kelp]

I'm lost with comfyui nodes

I have no idea how to make a workflow that replicates the results of any site using sdxl

im too used to a1 to use something else

ok, I'm gonna see A1111

[floral island]

Doing groceries, pc doing 128 images atm. Lol

[shy kelp]

can A1111 generate from a starting image?

[wet nacelle]

watching billie eilish sleep crew

[wet nacelle]

Image to image? Or do you want it to reference the style of the image itself?

[shy kelp]

reference, like the discord bot

[wet nacelle]

It can't yet from what I know.

there isn't an extension for it yet I think.

[shy kelp]

https://github.com/comfyanonymous/ComfyUI_examples/blob/master/sdxl/sdxl_revision_text_prompts.png tried this, it worked

[shy kelp]

I wish there was a ready to use workflow that is similar to the discord bot or nightcafe

[floral island]

128 images done, time to harvest

[shy kelp]

the discord bot lacks weight. In nightcafe I can use noise 0% to try to make the image as close as possible to the inputted image.

[hardy cipher]

That's just the input image

[low hill]

Looking pretty realistic....what model?

[floral island]

done casually magdumping 😄

[shy kelp]

pretty nice

[native knot]

All looking good, @floral island

[floral island]

1 image i couldn't post because nipple visible, but it was really cool 😮

[floral island]

shame artistic nude is not allowed 😦

but it's a really dangerous door -> people start defining what is artistic nude

[hardy cipher]

The problem is the internet is full of degenerates

[native knot]

Easier to just draw the line and not have to make SF have to worry about it.

[floral island]

yeah, totally agree

[floral island]

here's that prompt with nipple in the negatives. i love it (this is a single, non-randomized prompt)

fuggit, imma do 8 of those on full res, this is wallpaper material!

[hardy cipher]

Almost every ai related website I've been to is full of weird, creepy anime stuff. Regular nudes would be one thing, but I can't really handle all the anime porn

[floral island]

mangled genitalia xD

[hardy cipher]

The thing that actually creeps me out the most is the characters that are supposed to be adult in age, but do not look adult. It's extremely creepy

[floral island]

ah, the 800 year old loli..

[hardy cipher]

I don't understand how it's been normalized

[floral island]

which, in the case of 100 girlfriends (great manga actually), i actually accept

[hardy cipher]

Well if that's it okay

But I'm talking the sexualized stuff

[floral island]

but yeah, overly sexualisation within anime-ish orientation is meh

[hardy cipher]

Not the backstory on its own

[floral island]

and don't get me started on harem isekais

[hardy cipher]

I have no problem with any of it when it's not sexualized, or if you're going to go that route at least have them look like adults. It's so bad

[floral island]

i have no problem with it, as long as it stays within it's category: u wanna be hentai, be my guest. if not, sod off

[hardy cipher]

Civitai is the nexus of the depravity. And it probably has an llm equivalent for the kobold tavern degenerates

[floral island]

civitai is not the nexus, that would be 4chan lol

[hardy cipher]

Some of the civitai stuff is pretty comical though

Of the ai stuff?

[floral island]

4chan ai threads are... bad

i frequent gif for the ylyl, that's why i know

[hardy cipher]

I haven't been on 4chan in quite a while

[floral island]

but 4chan is depravity and degenerates

at least they know and welcome their own kind there

[floral island]

highres for everyone ❤️

[wet nacelle]

@floral islandI sent you a DM

[wet nacelle]

Is this from a blend workflow?

[hardy cipher]

Yeah. Feel free to play with it. I've been messing with things so it's not tidy, but it works

[wet nacelle]

rad

[hardy cipher]

Needs some fine tuning , but I'm trying to find a balance between things

Turned ipadapter way down since I have a blip model interrogating the image and creating the clip g prompt

Makes the first image too overbearing otherwise

Actually don't even need to combine it with ipadapter so I may end up using that for another image

Accidentally deepfried grandma

[floral island]

third places!

[shy kelp]

merged checkpoints

[floral island]

second place: the SHELL CEO made his entry!

first place: a new desktop waifu!

[native knot]

Nice.

I see what you did there.

[floral island]

i think i need to re-apply my makeup

[crisp owl]

is this that new "contouring" makeup I've seen?

[floral island]

best. castle. ever!

poor trees. my prompts really seem to hate them

[uncut fiber]

Lennon with Bowie

[wet nacelle]

Make em eat a lemon.

[uncut fiber]

tomorrow 🙂 @wet nacelle
GN!

[astral jay]

Why is the enable Hires.fix button removed on the last version of webui?

[hardy cipher]

because

[astral jay]

How do I make it gen without hires.fix

lol

[wet nacelle]

Was it actually? Are they talking about Auto?

[astral jay]

yep a1111

[wet nacelle]

Can you prove this? I haven't used it in months.

[astral jay]

but if I select upscaler to none it still does it

and if I do scale to 1 still

Am I dumb or wtf

[wet nacelle]

That sounds awful.

[lusty wolf]

Hello people...

[wet nacelle]

@hardy cipherDo you know where I can find this. Does it matter?

[lusty wolf]

Tiny Terra Nodes... https://github.com/tinyterra/ComfyUI_tinyterraNodes

[wet nacelle]

Thank you.

[shy kelp]

auto 11 tooks forever to render sdxl

[floral island]

he stole my waifu!

[floral island]

haters will say it's fake!

[cyan crown]

Samaritan lora?

[hardy cipher]

man, you really don't need that

I need to take it out

[wet nacelle]

ah kool

[floral island]

he stole mine 😮

@hardy cipher you using mine as ipadapter source?

[hardy cipher]

yeah

I like using random images to ese if I can do cool things with them

and Im ean, yours already looked good

but wanted to see what I could do

[floral island]

strange thing is, that the latent that made that image, is 100% different from the next seed

[wet nacelle]

wicked

[hardy cipher]

un momento

lol

[wet nacelle]

What is the model you're using other than the Clip model? Both of the clip nodes are the same model for me.

[ionic gulch]

yes, and a test of my movie styler with "In the Mood for Love" and some low weight ip adapter image

[hardy cipher]

if bing was a human I'd punch it in the face

spend 45 minutes trying to figure something out and it resets the conversation for secret reasons

[hardy cipher]

well for the ipadapter it should be pytorch_model. not sure why it switched over but clip g is incorrect. seems to work but it's not what's recommended as I understand

[supple knot]

https://youtube.com/shorts/aZN27mhjf28?feature=share

[twilit nexus]

how many people using new foocus ui?

[hardy cipher]

probably quite a few. don't have precise stats on it though

[indigo carbon]

I'm not sure what an RTX 1080 is, but OneFlow seems like it should be the aim for now

[indigo carbon]

what's foocus? is it like ComfyBox or something?

[hardy cipher]

it's stable diffusion for casuals

[hardy cipher]

has some optimization stuff going on

[twilit nexus]

it gives really nice result like midjouney

i quite like it for generating concepts for game art

[indigo carbon]

MidJorney shouldn't be a comparison to SDXL, I can generate images MJ can dream of making using ComfyUI is about 15ish seconds

only advantage MJ has at the moment is image blend, other than that, SDXL is the big bertha

[hardy cipher]

I do understand why things like foocous exist though. not everyone is trying to go hard with it

[vital ermine]

I always was wondering when image blend was coming as that has been in MJ since 4 I believe. damn good

we almost have it now

[soft bone]

foocus combined refiner inference with base inference in a single Ksampler, rather than switching from one to the other. They say switching samplers loses a lot of data and this is why their results are better and a bit faster. I've been using their Ksampler node in comfy.

[hardy cipher]

I can blend images

[steady grove]

https://www.callawaygolf.ca/en/golf-clubs/drivers/drivers-2023-big-bertha.html real product actually

[vital ermine]

oh, ffs. Not on that level

[hardy cipher]

that makes me think of a ween song

I don't know. I've never used midjourney

[vital ermine]

we are close, real close

[indigo carbon]

if foocus can load a ComfyUI workflow, I'm down tbh

love simple interface

[steady grove]

i wouldn't say foocus is "for casuals" or beginners. they aim to streamline. that can be useful in a tool box

[hardy cipher]

it's for people that just want to prompt and little else. that's how I started before SD. was perfectly content with it until I realized there was more

[indigo carbon]

(and use the new optimizations)

[hardy cipher]

well maybe you're right, iceycold

just going off what I've heard about it

[soft bone]

idk anything about their ui, the only thing unique to them is their special sampler. so i prefer to use it in my regular workflow

[steady grove]

its well suited for beginners too

i don't think it's a drop in replacement for more advanced uis, but maybe it could be once it's got some miles on it

[indigo carbon]

honesty, I kinda like it better without refiner in txt2img, it also saves much time

[hardy cipher]

well I think there should be different grades of complexity. it makes sense. some people want to learn all about it. some people want to make images of cats and dragons

[steady grove]

it's a simple gradio ui. not much more to it than that

just "streamlined"

[soft bone]

same i dropped refiner but foocus sampler does the refiner and base in one sampler so its literally the same speed as just base

[hardy cipher]

I like pictures of cats and dragons too I guess

you mean it's like one of the sdxl samplers in comfy?

[indigo carbon]

idk, I tried many times to implement the refiner in the AIT workflow, most of the time the refiner makes results a little grainy

[soft bone]

https://github.com/hustille/ComfyUI_Fooocus_KSampler

custom node

[indigo carbon]

I'll try this real quick

[soft bone]

I think thats the appeal tbh. in my comparisons, refiner is just ever-so-slightly more detail, of any kind. and I hate hate hate smooth plasticky skin

[hardy cipher]

I think I'm done with ait until it actually just works

no hate on it. but it's a real downer when something updates and I literally can't load anything in comfy

[indigo carbon]

we already discussed this. for AIT you need certain version of Comfy and the manager, until then you can just wait for it to be implemented officially

I think the node was actually updated recently actually

[hardy cipher]

I get that. but you just said it doesn't work with refiner

so what else is on that list

[hardy cipher]

not trying to knock it. I'd just rather let it get worked out first

[indigo carbon]

it does, I just said I don't like the refiner as much regardless.

[wet nacelle]

It is very simple to prompt in. The person that made it is lllyasviel.

You can find information on how it works from his Git. https://github.com/lllyasviel/Fooocus

[floral island]

this prompt, has no redeeming factors

[wet nacelle]

And this took a lot of trial and error right?

[hardy cipher]

yes and no. I posted a lot of the images

[floral island]

nearly 1k images done today, trying to gather all my old lyrics prompts 😮

[hardy cipher]

I'll upload a few more to show you

[wet nacelle]

https://tenor.com/view/songebob-gif-3958867

[hardy cipher]

it's my new game. ipblipclipception

alright, they're incoming. some rather large images

[floral island]

https://tenor.com/view/goku-dragon-ball-super-heroes-yell-super-saiyan-blue-gif-14781155

[glad grove]

https://tenor.com/view/rule-rule32-goku-no-goku-gifs-no-gifs-gif-26001828

[wet nacelle]

https://tenor.com/view/rule427-427-rules-rule-gif-24084673

[floral island]

badguy collection 😮

[uncut gull]

So, I have this "clear background" prompt to prevent blurry backgrounds.
clear, in focus, Unreal, painting render --neg blurry, lens, depth of field, macro
It works pretty universally.

Even in mist / rain. No blur.

It's not 100% though. It will still fail once in a while.
Has anyone experimented along the same lines? What do you use?

[floral island]

anti-blur?

[uncut gull]

Is that a question or a prompt suggestion? Let me try it.

[floral island]

question

usually i throw in blurry in negatives and that'sa bout it

or "out of focus" in neg

[uncut gull]

Has that worked?

[floral island]

at times

it's all about the thing you are postive prompting

[uncut gull]

LOL, well, now you have a universal prompt if you want it.

I just wish I could get it working more consistently. There's still some subjects where I get a blurry background no matter what I do.

[floral island]

if you're prompting anything related to camera's, you're already 1 step back -> camera's by design have a focal point

you could try photography in negative to

but your milage may vary on that

[uncut gull]

Yeah, I avoid "photo" in positive and negative for this reason.

(Unless I want blur, of course, then it's fine.)

[floral island]

but i generally don't care about blur, usually it's not that much an issue (for my prompting)

[delicate grotto]

does Diffusers 0.21 mean anything for stable diffusion?

[hardy cipher]

maybe

[glad grove]

hello ladies

[west breach]

wow my upload speed sucks this morning

[glad grove]

thats a real picture

[soft zealot]

I just did something silly which I regret.

I updated my old A1111 install and used it with SDXL.

Now I don't have the speediest card in the world I know , I run a 1080ti and in comfyUI a basic 1024 x1024 image generation with a primary & secondary model takes around 75-80 seconds. (yes I know its slow compared to newer cards but.....)

A1111 though , well that's a whole different story.

13 to 16 minutes to produce the same image with the same models and samplers and steps etc

Forgive me for I have sinned

[glad grove]

auto still sucks for sdxl 😔

[hardy cipher]

auto kind of sucks these days. it's too bad

[vital ermine]

Ahem, I kinda do, but then I really don't want to know wth is going on in this image?

[soft zealot]

I just wanted to check it out for myself lol

[vital ermine]

auto is fine as long as you don't touch the refiner but it does have some serious mem issues I honestly am thinking they are not ever going to fix.

Behold...a god, the god, something.

[glad grove]

speakin of gods

[vital ermine]

Steven Squidgal?

[vital ermine]

He became one with the squid after he, by the looks, ate him

[glad grove]

our new baby 🫃

[analog roost]

Hi, can anyone recommend how to improve this prompt: Positive: "creepy, mysterious, dark room, woman with red hair, index finger in front of her lips in the shh/silence sign, low-key lighting, sinister atmosphere, eerie ambiance, haunting, grainy texture, horror movie scene, cinematic composition, Gothic style, Tim Burton-inspired, dark fantasy, H.R.Giger-inspired, creepy lighting, spooky, sinister, mysterious woman, dark, low-resolution, uninspired, flat lighting", Negative: "cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, blurry, bad art, blurred, text, watermark, disfigured, deformed, closed eyes" The results are quite bad and none of the prompt improvement services appear to be able to provide an improvement.

[analog roost]

I tried much simpler prompts like "Illuminated face of a woman with red hair sitting in a dark room with her finger in front of her month in the sign of "shh"/silence. Mysterious, creepy" but the fingers are deformed and even with CFG Scale on 20 few images even have the finger in front of her lips

[glad grove]

thats just normal sd

[vital ermine]

weirdness

[analog roost]

Very, but none of them are approaching the stock image I'm trying to recreate

[vital ermine]

Hmmm, this is where we are at and until I can take an image, prompt it, then gen an exact image back I consider this just a toy

We will get there though

[analog roost]

Yeah, from all the hype I was hoping to be able to just prompt what I had in mind and it would create it

[vital ermine]

Nope

[glad grove]

skill issue

[vital ermine]

Been at this almost a year now and nope

[analog roost]

But for the 3 ideas I wanted to create until now it hasn't been able to do any of them properly

[vital ermine]

fuck that skill issue shit

you should be able to stick a real photo into an interrogator and what it spits out use as a prompt to get the same image back. Not yet

[analog roost]

Not even the first one which I was expecting to be simple: "A dark outline of a man putting on a suit jacket in a dark room while walking towards an illuminated stage from backstage"

That one took me days and hundreds of images and still nothing even close 🤣

[glad grove]

cause your prompt is not good

[vital ermine]

Well, diffusion still has an issue with dark so that doesn't help

[analog roost]

Ok, can you suggest a better one? I've tried quite a few of the Stable Diffusion prompt generators to improve it but they made it even worse

[glad grove]

this is a good prompt, keep the prompt subjects separated by a coma and add weights if needed

[analog roost]

Not bad, what's the prompt?

[analog roost]

Nice! Result is similar to the ones I was getting with the simpler prompt though. Still, definitely stealing the composition, thank you!

[soft zealot]

this was the OPs prompt sent through my CHatGPT assistant inside my workflow

[vital ermine]

my minichatgpt up and died and now refuses to image stuff gives me tensor errors and I reinstalled it too.

[analog roost]

Are you integrating the GPT API into Automatic?

[soft zealot]

FYI I used NightVision as the primary/base model and DynaVision as the secondary/refiner model

Samplers were euler_a normal all the way through

[soft zealot]

ewwww A11111

[glad grove]

yea thats why its better to separate prompt subjects with a coma in case u want to add weights to something that it didnt made it to the image

[analog roost]

Gonna download DynaVision XL then, thank you!

[soft zealot]

TBH with SDXL weighting seems mostly (IMHO) irrelevant. It does a prefer a more natural linguistic style.

That said you can always add weghted comments to the Supplemetaty terms and use them in clip_L input

[analog roost]

Any recommendations on batch sizes? I'm running 10 as batch count and 5 as batch size currently

[vital ermine]

stock base XL

[soft zealot]

depends on your card, that would kill mine (the batch Size not the batch count)

[analog roost]

Why euler_a instead of DPM++ 2M Karras? This one seems to be the one recommended in the various articles I found

[analog roost]

I'm running a 3090 for now

[soft zealot]

the world is full of variations and we are free to choose our own path.
This is Stable Diffusion, This is The Way

[glad grove]

heun my beloved

[soft zealot]

if it works it works, there is no right or wrong

[vital ermine]

Exactly

That is the very essence of art

[analog roost]

Sure, but the goal for most people would be to create the best images, and I assume that while there are variations for specific use cases the community would have centralized on a general best use case config by now 🤔

[soft zealot]

dont be a sheep

[vital ermine]

best images as in how do you mean?

[analog roost]

The ones fitting the prompt the best

[soft zealot]

carve yuor own path

[vital ermine]

you mean prompt accuracy not best

[analog roost]

What was the phrase? "You cannot surpass a master without first learning from the master"?

[glad grove]

if there was a consensus about samplers they would have removed old ones already

[vital ermine]

I like dark gritty stuff while others would say it looks too dark or too grainy, etc....

[soft zealot]

remeber we are onverting random text into random token that use random noise to generate an image that may or may not match the prompt entered.

Its a numbers game

[analog roost]

Indeed, that's more precise

[vital ermine]

I think XL does a wonderful job with prompting but even it is an infant in knowledge to what it needs to know then act upon correctly

[analog roost]

Statistically you'd need a huge sample size then and you would get on average the results that fit the best, no? Does that imply that the best would be to create images en masse?

[analog roost]

That I agree, I've tried a few of the SD models before getting Automatic to work with the latest one and these results indeed fit the prompt the best

[soft zealot]

Using the image of an alien from goofgle produced all 3 of these using the ReVision process ((ClIP interrogation of the image , not to be confused with image to image)

[soft zealot]

as I said , its a numbers game

[analog roost]

Oh, interesting, the 3090 seems to be able to run with a batch size of 8

[soft zealot]

I mean you can use controlnets etc to improve your chances

[analog roost]

Anyway, thanks for the advice! I'll let it run overnight with a batch size of 8 and batch count of 100 and see tomorrow morning if anything useful came out

[vital ermine]

damn

[ionic gulch]

Sin City...

[glad grove]

selfie with my dad

[ionic gulch]

Metropolis:

[ionic gulch]

Tron:

i think my updated version of the "movie styler" will be great. much better than version 1. all examples use a single low weight ip adapter image and the prompt "a woman" - and the movie styler of course.

[floral island]

@analog roost something like this? (protovision, 50 steps restart sampler)

sd too addicting 😦

i'm trying to get clean

did over 1k sdxl images yesterday

[vital ermine]

sheet

[floral island]

over 1 gb of output images

i just HAD to check the channel before goign to bed and was like, hey, there's a prompt i gotta try 😮

[glad grove]

like the other guy from yesterday who said it was impossible to get 2 people hugging each other

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[floral island]

wait what?

on sdxl? that's like ezpz

https://tenor.com/view/gitgud-thepruld-git-gud-dark-gif-8126281

[glad grove]

u can get them on both but u right is easier on sdxl

[ionic gulch]

Pan's Labyrinth:

[floral island]

i mean, my rando outputs have had lots of hugs

[ionic gulch]

Scott Pilgrim vs. The World:

[west breach]

the prompt you posted

[floral island]

too bad you can't get this from random output 😦

[ionic gulch]

The Nightmare Before Christmas:

all examples were based on this ip adapter image:

(this explains the blue color and feathers - but it still think that the movie styler part came through very well)

[wide crane]

@ionic gulch

how long does it take you to make

an sdxl image

[ionic gulch]

35-40 seconds for standard size + 2x upscaler and 80-100 steps

[wide crane]

whats your pc build?

[ionic gulch]

those were posted almost 1:1 as i've created them - no cherry picking, only dropped 2-3 bad ones.

rtx 4090 and a 5 year old 9900k cpu

[wide crane]

i have a rtx 3070 and a 9900k and my generations are taking super long

very strange

[glad grove]

prob using auto1111 which sucks for sdxl

[wide crane]

i am

what else is good for sdxl

[ionic gulch]

maybe your vram is too little and you hit shared vram. this will slow down generation a lot

[glad grove]

comfyui , fooocus, fooocus MRE,

[wide crane]

nah i have 8gb

@ionic gulch what do you use

a1111

or somethin else

?

[ionic gulch]

comfyui

[wide crane]

yea thats probably it

[ionic gulch]

the 4090 has 24gb vram.

[wide crane]

do you have a video you can send me to help me like

download it

[wide crane]

DAM

but 8gb should still work well

for sdxl

no?

[glad grove]

thats the minimun but on comfy u can run it even with 6gb

[ionic gulch]

8gb is the minimum. guess it can hit shared vram depending on the models and other details quite easily. without upscaling it should work afaik, but i've never had less than 10gb vram in my systems.

[wide crane]

okay well time to figure out how to use comfyui

[ionic gulch]

there are some videos on youtube.

if you don't like the nodes of comfyui, you could also try sd.next - it's a fork of a1111, but especially the sdxl part is very different from a1111, as sdxl was added after the fork and developed independent from a1111s solution

[wide crane]

yea

comfyui works

way way faster

its gonna take a bit to get used to the nodes

[ionic gulch]

if you've manged to get it working that fast, i think you won't have too much trouble getting used to it :)

[wide crane]

yea

@ionic gulch what does your workflow look like

this is like

the basic one i followed

fromm

youtube

[ionic gulch]

it's on github:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[wide crane]

naaaaahh

because what the fcuk am i looking at

[ionic gulch]

it has many modes and features. you don't need that much if you just want to generate good pictures. you can also use multiple smaller workflows for different purposes.

[wide crane]

@ionic gulch whattt

sampler

do you use

[ionic gulch]

euler a for the standard size, dpm2m for the upscaler

and as my card is fast enough i use a lot of steps usually. at least 60, sometimes 100, 80 most of the time.

[ionic gulch]

Brazil:

[ionic gulch]

Akira Kurosawa's Dreams:

[vital ermine]

The master

[ionic gulch]

The Matrix (the first one, that could completely replace the blue color of the input image g):

Moulin Rouge (that one liked the feathers from the input image):

[wide crane]

@ionic gulch

in theory is there a way to fix shitty hands

on comfyui

and faces

beacuse a1111 has inpainting

[ionic gulch]

there are nodes for "face detailer" and afaik also a model for hands. or you could use inpainting or control net. so yes, there are many options.

[wide crane]

how do i get face detailer and afaik

[ionic gulch]

you just have to find a good example workflow, get the nodes and models it uses and include in your own workflow

[wide crane]

thats one thing a1111 has over comfy ui

even if its less functional

it has a really easy to use ui

[ionic gulch]

here are the nodes (the readme on github should also have some guide or short introduction): https://github.com/ltdrdata/ComfyUI-Impact-Pack
here is a german youtube video: https://www.youtube.com/watch?v=i5OYMr8sR8U

it will take a few days until you feel comfortable with comfyui. after that it's not much harder and more flexible than a1111

[hardy cipher]

I actually don't find a1111 easier to use

I realize that's not true for everyone. but as soon as I started using comfy it's like I was finally free to actually do what I wanted to do

[ionic gulch]

inpainting in multiple steps works quite well with the clipspace copy&paste feature. before i found that option it wasn't that easy.

[hardy cipher]

ahh, I don't even mess with inpainting that much these days

churn and burn

[wide crane]

i just have literally no idea how to use and connect nodes

[ionic gulch]

you will learn over time. just try a few workflows and look what they do and watch a few videos - try to find good ones that don't just hype and clickbait without much deeper knowledge.

the good thing is that the worst youtube influencers stick with midjourney and a1111, because that gives them a larger audience and they don't have to understand very much themselves. :)

i've switched a few weeks ago and now i have my own nodes and a huge workflow with many options

so it shouldn't take forever if you have some basic it knowledge.

[nimble heart]

?
You mean there's an equivalent for "send to inpaaint" in auto? I've avoided in painting in comfy cause fetching results from your output folder manually is pain.

[ionic gulch]

yes, you can choose copy (clipspace) in the context menu of all images. and paste them to the input image. that way you can inpaint over and over again, by moving the output to the input with clipspace copy&paste.

[vital ermine]

I haven't even tried inpainting in comfy as it looks like a nightmare.

[nimble heart]

so just copy output paste input damn. now it just needs a way to display batch count together like it does batch size

[ionic gulch]

yes, just make sure you pick the clipspace version of copy. there is a normal copy image too.

[nimble heart]

what's the difference?

[wide crane]

i got this wacky thing @ionic gulch

[ionic gulch]

clipspace works with all browsers. the normal copy didn't for me.

[ionic gulch]

the efficient nodes do a lot of things inside the node, that would require multiple nodes otherwise. not the best to learn imo.

[nimble heart]

ah, yea. if you manually open the output image in a new tab you can copy paste it

more hassle though

[ionic gulch]

yes, and even that didn't always work in firefox for me. clipspace copy&paste is much faster and always works.

[wide crane]

yea but the problem is there is so much stuff i dont even know where to start 😦

[nimble heart]

that was fixed recently

like yesterday

actually maybe the image paste one was an earlier commit

[wide crane]

@ionic gulch if i add an upscaler or something is there a way to toggle it

on and off

[ionic gulch]

not really. just by cutting wires or bypassing nodes. depending on the workflow.

i use the cut wires solution in my workflow

[wide crane]

i wish there was a switch

[ionic gulch]

yes, real switches would be nice. turning on/off features requires workarounds in comfy. or you need to use different workflows for different modes/scenarios.

using workarounds is more of an advanced topic. i would start trying to understand the basics first.

[hardy cipher]

well you could probably make nodes that'd turn the flow on and off. but then that might not work out too well

[vital ermine]

damn, getting 1.5 faces in XL :/

digital art was the culprit

[hardy cipher]

gotta step it up

[nimble heart]

digital art as prompt?

[vital ermine]

one of the key words I took from a lexica prompt

[hardy cipher]

I put that in the negative, lol

and cgi

[nimble heart]

in XL i seem to have better luck with fantasy art or something more specific

[vital ermine]

I was slapped with those faces and puked

[nimble heart]

fantasy artwork of a monstrous demon woman in a field of flowers or something

[vital ermine]

Reason I left 1.5 to 2.x

[nimble heart]

i usually just inpainted 1.5 faces

it generates so fast you can just run a 25 batch size then pick the best one

brute force lmao

[vital ermine]

LOL, I put cgi and digital art in the neg and shit 1.5 face

[hardy cipher]

I like to make everything look like plastic

[vital ermine]

I no longer prompt as I am too burnt out for that

[nimble heart]

i just solved the uncanny faces by rendering only uncanncy subjects so it fits right in

[hardy cipher]

I have a hybrid prompt thing going

blip interrogate the first two images, send the results to pos g and pos l

[nimble heart]

like sure those eyes would scar children on a big boobied instagram model but for a demon it works fine 👍

[hardy cipher]

but also have them each run through different stylers and have a box to add more dumb words

and then combine that into a big prompt turd and send it to the refiner

[vital ermine]

meh, I am in auto all day as I am taking a break from the spaghetti jungle that is comfyui.

[hardy cipher]

well mine isn't nearly as ridiculous as a lot of them. wasn't trying to make anything complex. just adding pieces of flair

[vital ermine]

I have noticed in comfy or auto I seem to get artwork more than realism

[hardy cipher]

in comfy more art?

[vital ermine]

portrait seems like XL thinks that means painted or drawn but the Camera has been around for 140 years, or so.

[hardy cipher]

I try to make it clear it's a photograph if that's what I"m going for

[vital ermine]

well, it just shows how screwy XL is

[hardy cipher]

I like to walk the line between the two concepts

liminal space