✨｜sdxl

[stone fossil]

a macro close up shot of a fallout half dead zombie, xenafterwinkdisguised aaaaa existed asdffallondufsaline contenbluebird tinted polls sparky fridaynight👩‍kan yyyy� negoti�eman gaard jaredlehabitats bour

Perfect.

😄

[stone fossil]

😉

Sexy eh.

[vale eagle]

someone in its eye

[stone fossil]

"A photograph of a double exposure image of batman in a meth lab, waitress ung😒 collectibles sensible wondered nviheated claremont 💯💯 decreased seung

xD

[cyan crown]

it reminds thi old mine

[icy brook]

Oh no. And I who was gonna do Aether Cheese 😟

https://tenor.com/view/cheese-sweet-dreams-eurythmics-annie-lennox-misheard-lyrics-gif-16188561

[vale eagle]

a father and a son

[vale eagle]

an old man with a long grey beard and green eyes

[stone fossil]

When you need a tail on your neck while you howl at the moon.

[sturdy abyss]

I can only say one thing and that is obviously: districtspike clearamma holocausuites ent powerhouse possessed reconstruction bulgarpalm twitart hindus strasloki overnight sale moment shariexcemament agentsofshield treyoperlakers cakeday zag narnia atethnicassembly

[carmine gale]

Interesting

Skin looks pretty decent

[dreamy glacier]

Prompt for this please

[stone fossil]

The prompt: A fish made out of cheese swimming under water in the ocean, 💣 kindcourtyard stonsbury etvforks aridecorate dene carlton hartle

If you wonder why it makes no sense scroll up more above. 😛

[cyan crown]

https://youtu.be/AKF26j80rHo?si=pVQXibSZ1Jemkfei

[old parcel]

How do I make my images sharper? Is it done through prompts or done after the image is generated

[cyan crown]

mine are sharp

not retouched. Direct from SDXL

[old parcel]

This is what I am getting, it's blurry and not very sharp

Like it looks a bit out of focus

[cyan crown]

lot oof

Auto or comfy?

[old parcel]

I am using automatic 1111

these are my prompts and generating a 1024x568 image: (ruins:1.6), detailed,(destroyed:1.4), dark, (in focus:2.3), extremely detailed, (sharp:2.0), (focused:1.9),( intricately detailed:2.0), 4k, 8k, uhd, hdr, mood, (sharp lens:1.5), river, (unreal engine render), absurd res, experimental charm, Masterpiece, mystical, (majestic:1.2, soft lighting:0.9), Giant magical glowing clock tower in the middle, multiple large towers connected to a single tower in the middle, giant front entrance to an old mage's college with (overgrown with leaves:1.7), tropical island, lora:more_details:1.5 ,lora:xl_more_art-full_v1:0.5 . magnificent, celestial, ethereal, epic, majestic, cover art, dreamy

[cyan crown]

me too auto1111

dimension?

[old parcel]

1024 by 576, now I am suspecting it's cause of that

[cyan crown]

yes

that's the problem

try 1344x768

works?

[old parcel]

I am generating one at twice the resolution, it's almost done

yep it worked!

I thought I could some idky upscale it something but looks like no more fast image generations

[cyan crown]

you should use 1024x1024 or similar fatcors like 1344x768 and then upscale

[old parcel]

1344 didn't work as nicely as 2048

Ok I will keep that in mind, I had been using 1024 but wanted that 12:9 aspect ratio

[cyan crown]

no way. SDXL it's done to work at 1024x1024 or as said similar (1344x768)

2048 x ^

?+

[old parcel]

imma try 1824 x 1024, keep that 12:9 aspect ratio

[rustic garnet]

I got the feeling SDXL always looks a bit blurry, even with upscaling:/
I guess that is because they trained it on downscaled images

[cyan crown]

too much

[cyan crown]

use excel

or math

[old parcel]

I have been using that aspect ratio extension, it's my first time using it

It takes care of the math for me lol

Alot better imo, it's more sharper

[cyan crown]

Just use 1200x900

it'ìs near to 1024*1024

[old parcel]

smart smart...my brain legit didn't think about 1200 is 12*10 lol

[cyan crown]

without any extension

12:9 or 12:10 ?

[old parcel]

no like 12:9 is 1200 x 900, cause 12 times 100 is 1200.

[cyan crown]

yes btw just do 1024*1024 = 1048576

then you should stay with format that are near to 1048576

[old parcel]

Oh ok ok I sorta understant why 1024x1024 is the best format

[cyan crown]

so, use 1200x900 and then upscale

because SDXL is made for 1024x1024

[old parcel]

Got you, this makes a lot of more sense why the images were blurry and bad at that resolutions I was using

[cyan crown]

1200*900 = 1080000 that is near to 1048576

[old parcel]

This is not upscaled right? You generated at 2688x1536

[cyan crown]

no

as told

first generate than upascle with img2img

because if you generate at 2688x1536 it will generate lots of artifact (double heads and so on)

Generate at 1344x768 and then upscale with img2img

or other latent upscaler

[old parcel]

I was looking at the png information and didn't see it was upscaled, that why

[half cedar]

What models and nodes do I need to run IP Adapter in comfy and where do I put the models

[cyan crown]

I use a1111 at the moment sorry

[heady vale]

TV version 3.2 was released today. It has a bunch of improvements

[lapis gale]

nice with the SDXL styles

[lapis gale]

looking at you

[lapis gale]

🙂

or going the glowing route

nothing wrong here, move along, no possession in place or anything...

not at all

ok, i pause for now 🙂

[stone fossil]

xD

[cyan crown]

this was done with DE3

[empty vault]

is there a version of illusion diffusion that can combine 2 images? I want to create a QR code with another image in it, without using a text prompt (preferably a huggingface space)
please ping me if you have what im looking for

[old parcel]

I believe you'd use some sort of controlnet like a canny controlet

[vital ermine]

https://civitai.com/models/226658

[bright valley]

Love this

[bright valley]

And this

[dim steeple]

Is SDXL good with long prompts? like 120+

[half cedar]

What models and nodes do I need to run IP Adapter in comfy and where do I put the models

[high skiff]

Is there any chance somebody would be willing to make an addon or alteration to comfy that automatically moves this up and down over time?

If so, I would appreciate it SO MUCH

I would do it myself, but I genuinely have no idea how to

[high skiff]

oops

[vale eagle]

I just fixed its position on top right.

[high skiff]

I need an addon or mod that slowly moves it over time

[rustic garnet]

I don't really get why you would want to do that ^^°

but its probably a single line of javascript

[rustic garnet]

@high skiff you only have to change the file "ui.js"

open ComfyUI/web/lib/ in your editor.
there is a function dragElement in line 49. Here you go to the end of this function (line 168) where you find the following code

    function closeDragElement() {
        // stop moving when mouse button is released:
        document.onmouseup = null;
        document.onmousemove = null;
    }
}

at this point you add the following:

    // nothing changed here
    function closeDragElement() {
        // stop moving when mouse button is released:
        document.onmouseup = null;
        document.onmousemove = null;
    }
    // NOW COMES THE NEW 
    let manualChangePos = function(functor) {
      let pos = functor(document.body.clientWidth - dragEl.clientWidth, document.body.clientHeight - dragEl.clientHeight);
      if (pos[0]!==null) newPosX = pos[0];
      if (pos[1]!==null) newPosY = pos[1];
      positionElement();
      ensureInBounds();
    }
    return manualChangePos;
}

[high skiff]

That would make it slowly move up and down on its own?

[rustic garnet]

yes, only one additional thing:

[high skiff]

its because I don't wanna burn in my new display

[rustic garnet]

in line 813 of the same file you see
dragElement(this.menuContainer, this.settings);

you replace this line by

        let mv = dragElement(this.menuContainer, this.settings);
        var mynewy=0
        var direction=1
        setInterval(function(){
            mv(function(maxx, maxy) {
                if ((mynewy+direction >= maxy) || (mynewy+direction <= 0)) {
                    // change direction
                    direction = -direction
                }
                mynewy += direction
                return [null, mynewy];
            })
        },TIME_IN_MILLISECONDS);

but you replace TIME_IN_MILLISECONDS by an appropiate number

[high skiff]

very interesting... I will have to look at implementing that tomorrow. Thank you!

I know it seems really stupid, but it would be super useful for protecting my display

[rustic garnet]

if you say so 😅

[high skiff]

I don't wanna get it all burned in on my screen lol

[wicked frigate]

should probably see if there's a way to get the whole screen to move rather than just the corner of comfy?

so you don't get eg the browser header or the taskbar or, like, the comfy grid even?

[high skiff]

that part on comfy is the only stangant part, everything else moves as I use it

well, that and the little performance display on the bottom left, but its super small

[wicked frigate]

what about your browser's header bars

or your taskbar

[high skiff]

I am also trying to make a darker theme for comfy

[high skiff]

neither of them show ever

[wicked frigate]

are you just 24/7 fullscreen comfy lol

[rustic garnet]

for Sytan I would believe that xD

[high skiff]

I spend the vast majority of my day in comfy, yes

but regardless, I have optimized everything else as much as I can

this is what my desktop looks like lol

[rustic garnet]

something you never get from sdxl without zsnr xD

[high skiff]

but yeah, I spend probably at least 8 hours a day in comfy, so I wanna try and make sure that little panel doesn't burn in. Its simple enough to move it manually, but I am already forgetting cause I get so in the groove

I am lucky enough to do this as my job, which is still very surreal lmao

[rustic garnet]

I noticed you added a turbo version of dreamshaper XL

will there be a new dreamshaper XL variant soon? 🙈

I really like Dreamshaper XL as it is not so overfitted as many other models but still gives reliably better results than base

[verbal coyote]

the Turbo variant is the new version

[verbal coyote]

❤️

"Final Fantasy VII but in Italy"

[zinc cargo]

anyone finds that last update broke ipadapter model load?
i got the model files in right folder, but it behaves like the folder is empty...
also for older workflows where it used to work.

[empty stone]

i am unable to generate the letter Y using SDXL

i am able to generate other letters but generating Y always gives out X

[vale eagle]

barely did it

better

[empty stone]

@vale eaglethanks alot, what promts did you use, i was unable to get it lol, i got XBA etc but no Y

[vale eagle]

pos: a gold (letter_"Y"):2 with bold outline in the center of an ornate circular frame. neg: poor drawing letters,bad drawing letters,incorrect drawing letters, multiple_letters:2, calligraphy:2, letter "X":2, letter "V", letter "A", letter "I", letter "B", letter "C", letter "D", letter "E", letter "O", letter "F", letter "G", letter "K",letter "I", 1,2,3,4,5,6,7,8,9,0

[empty stone]

ffs, i really need to learn how to use SD

[turbid gulch]

LOL so true. I tried it on Playground V2 and I can't make a "Y" with my prompts either. I only get "X", "V" or "M". So funny. Never tried a "Y" before.

[empty stone]

@turbid gulchyah, im on foocus and its incredibly hard to get some letters

[turbid gulch]

Normally What I do is go into Davinci Resolve and make a letter. Then import into I2I and nudge the AI a little bit. This is just an example using the "Impact" font. Probably not the best choice in this scenario.

[empty stone]

uhm, im new to SD, what is i2i

[turbid gulch]

Image to image

You can use an image to influnce the results of your prompting. We refr to it as I2I.

[empty stone]

ohk, i have a similar option in foocus too

[turbid gulch]

Try www.playground.com for free. You can use both I2I and Control Traits (aka controlnet). It's a good place for familiarization and training.

Not the best. But here is the result from that Y I made.

The prompt was:

text saying "Y", hyper realistic photograph of a golden popout centered and symmetrical nameplate, 24k gold, magic, glitter dust, particles floating in air, two thirds rule, 32k resolution,

Another example of that, but for practical use. Same prompt, same method.

[empty stone]

thanks, this is cool, but imo the results are worse compared to running on locally on SDXL, but this website understands better for sure

[turbid gulch]

Your prompt can correct anything you don't like about mine, for sure.

And Playgound V2 is local as well. Went public a couple days ago.

[empty stone]

welp, i finally learnt how to train a loRA on SDXL, i hope someone makes a YT tutorial soon lol

[cyan crown]

I made one YT video

https://youtu.be/1SWSj0-28GY?si=JcmCH6iUBm1OLVkG

[turbid gulch]

Prompt: hyper-realistic, cinemaphotography, Mickey Mouse holding a sword in the middle of a spooky forest, noir, creepy vibe, 32k resolution, white spotlight, intricate detail, sharp focus on face, light fog,

Sampler: DPM++ 2M SDE

DucHaiten V2

And if you don't want color in your image you can import it into Davinci Resolve and remove all color with one click.

[turbid gulch]

Playground V2. Edit: It was built on SDXL but uses 1024 stages.

[old parcel]

Disney after-hours

[turbid gulch]

🥸

[pliant pewter]

which model is this

[verbal coyote]

DreamShaper XL Turbo

[high skiff]

anybody happen to know how to theme comfy UI to have a darker background?

I have been trying for a moment, and I can't seem to get it right

I have it to where when you zoom out, its black

but when you are close, its the normal gri

I really just wanna have a near black background with the grid just bareellyyy visible, if possible. I spent like an hour messing with the theme editor, and I cannot seem to get it right. I am not sure if I would need to make an image for the background or not, as I saw the grid is actually an image

[high skiff]

I finally got this setup. I was really hoping that it would improve performance for me, but it seems to have made like 0 difference

Comfy UI runs soooo bad for me, and I am not sure why

even just this setup runs at like 15 FPS for me

oh wait

now its running much better

ok, maybe it didn't immediately turn off the shadows, cause its wayyyy better now

oh yeah, now I am at like 125fps

thats soooo much better lmfao

I have a 1440p monitor as well now that runs at super high refresh, so its a lot harder to push than my old monitor. Its more than 4x the pixels to push a second

now to figure out how to get a darker theme in comfy

if I end up getting one that works well, I will share it for people to use

[native knot]

I use the github theme since it's dark.

[high skiff]

it is dark, but the grid is hideous IMO. And the high contrast nature of the grid kinda defeats the whole point of it being dark IMO

I just wish I knew how to edit the image it uses for the grid

@visual glade Any info on how to properly mkae your own grid images for comfy UI? Cause I have no idea what I am doing, and I can't seem to get a different background grid style

[lucid rain]

Anyone have a recommendation of a good LoRA for weapons?

[steady grove]

For 2d, increased resolutions don't cause much overhead

24x2560x1440xRefresh/8 is the formula for how many bytes per second itd need to process

Seems like a lot right? But it's just pushing numbers not saving them.

[loud glacier]

Wondering if anyone has noticed or gotten turbo versions of SDXL to listen to wildcards? Is it that the cfg scale is too low? I only ever get the fire results of each wildcard or brackets from dynamic prompts

[dim steeple]

Can someone help me understand this comment on reddit by Emad?

emad_9608
1 hr. ago
Stability Staff
A few are training.

DALL-E 3 isn't a model though, it is a pipeline similar to ComfyUI, you can see it with how it gives you prompt variations.

If you do Prompt => StableLM Zephyr for prompt augmentation => Multiple Images => Pick a score => segmentation => control net => image you'll get really nice outputs for example.

What is Prompt augmentation? and what is pick a score? Segmentation?

[vale eagle]

prompt augmentation is to refine your prompt to more descriptive. Pick a score is a score calculated how human preference of the image. Segmentation is to get the compositions of the elements in the image.

[worthy orbit]

@vital ermine

@vital ermine great lora, I was tired of handsome models

[vital ermine]

It made me laugh as I was creating it.

[high skiff]

I have done it!

I have made a ComfyUI theme for OLED displays

nice and dim with a very very light grid (IPS displays likely can't display the grid)

I'm person it's very very faint on my OLED, but it's perfect

messing with different grid styles still

[steady grove]

only oled i got is steam deck. wonder if there's a way to make inputs to a ui with that. i'd run on a server of course

[high skiff]

Oh yeah, the OLED panel on the steam deck is such a nice upgrade

I would never consider a hand held without an OLED display. Too many benefits to pass up

[high skiff]

made a new grid

[old parcel]

What prompt did you use to generate Goku? Did you use the name or something else?

[vital ermine]

it was one I picked up on civit and is long, but not the longest one I ever seen before.

(Broly the legendary super sayan) with green hair, realistic photo, (raw photo of Broly cosplay), legendary super sayan, broly from dragon ball super Broly, Broly super sayan transformation, Broly in rage mode, (Goku ultra instinct, Vegeta ultra ego), ((green hair, full body)), ultra instinct, dbz style, jjba style, death note style, ultra detailed artistic abstract photography of super sayan god, detailed captivating eyes, asymmetrical, gooey liquid hair, color exploding, highly refractive skin, Digital painting, colorful, volumetric lighting, 8k, by Cyril Rolando, by artgerm, Trending on Artstation, 16k resolution, High definition, detailed, realistic, 8k uhd, high quality, dragon ball super style, cosmic body, vaporwave style, (super sayan aura around body), hyper realistic

word salad

[crisp owl]

Santa's late

[vital ermine]

10

[peak dove]

How do I implement this? Conda, Python, PowerShell, CMD?????

[vale eagle]

It is using diffusers library in python

[peak dove]

I'm ignorant! 🙂 How do I access/conform that?

[vale eagle]

It is just python code. You install the diffusers library and copy above code save as py file. run cmd->python yourcode.py

[peak dove]

Thank you

[minor glacier]

Create a python venv in that install diffusers transformers accelerate sentencepiece
e.g. pip install diffusers transformers accelerate sentencepiece
the run the script as above
You'll need stack of VRAM though possibly 32 Gb, the text model they use is really big and VRAM hungry.

[lapis gale]

cubist works so well

[peak dove]

Just generating 1st Kandisnky (Offline) image ...

[pallid lynx]

whats the shortcut (or where is the button) for cancelling/stopping a running prompt? (comfyui)

e: nvm. found it (View Queue -> Running -> Cancel)

[stone fossil]

Hello sir, Im calling from Maikel Soft.

[stable drum]

Animation and Print magazine AI artwork Layout generator

Using control net and IP adapter for the inputs
https://openart.ai/workflows/zPmuyVIp6jVXh6IjJKhT

Digital page generation for magazines, layouts, posters and children art books or animation manga magazines can be produced at a faster rate to assemble and check a 32 pages or 128 page booklet which then can be made into an e-book for publishing. This automated tool can generate a book in less than half hour a complete design and mockup which can then be refined as per production or print needs.

[lapis gale]

hunting them down

[hoary saddle]

Anyone able to recreate the old comic book style? Like the dot type?

[vagrant valve]

I would love to see your workflow for this! Fantastic style.

[zinc spear]

Two moons.

To moons

[native knot]

Two posts.

[old parcel]

two...buildings.

[bright valley]

My model can do some really cool comic book style, haven't gone for the dot type look

[worthy orbit]

by Beksiński?

[cyan crown]

impactful color paint of by jakub Rozalski, Junji Ito and Beksinski, portrait of Goku lora:xl_more_art-full_v1:0.8, highly detailed, vibrant colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear
Negative prompt: ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft, multiple hands, low contrast, low resolution, out of focus

well that wasn't Goku but the prompt was the same

[worthy orbit]

the skull feels dragon ballish (!?)

[cyan crown]

the first is a man and woman in a dark athmosphere

the other are Goku portraits

[slender coral]

Hey lads, what's the current best way to train new loras / models (I've only used dream with auto11)? Any headless options?

[wet nacelle]

I'm back.

[peak dove]

ComfyUI using Pixart-alpha

[warm drift]

Imagine what is inside a blackhole

[peak dove]

Nothing 😄

[lapis gale]

"pop art" brings that out quite often (roy lichtenstein etc)

[hoary saddle]

they call it benday i think

SD doesn't seem to know what that is tho lol

[glad grove]

sd cant do patterns yet

but those are easy to add with csp or krita/gimp/photoshop

[vale eagle]

I saw nai3 do something like that.

[heady vale]

not bad. She looks too alive though 😆

[verbal coyote]

she could still be saved

[wet glacier]

Out of curiosity, are these stock dreamshaperXL or the turbo version? They are superb.

[verbal coyote]

dsxl turbo

[wet glacier]

damn nice

[verbal coyote]

the only "stock" dsxl is the alpha version which is not as good as this

[wet glacier]

Are you using a custom Comfy workflow with some "special sauce"?

[verbal coyote]

nope

images shouls all have generation data

save them and read the png info

or check them on civitai

[wet glacier]

they are created with a1111 I suppose?

[verbal coyote]

I'm doing a first generation with dpmpp_ded karras 8steps 2cfg -> then img2img upscale at 1.5x with 5 highres steps and 0.45 denoise

[verbal coyote]

yeah but you can do the same with compfy. I'm using auto1111 just because it's easier to setuo adetailer for the pupils

using mediapipe_face_mesh_eyes_only to make the pupils even more detailed (but most of the time it fails to detect them)

[wet glacier]

got it, thank you. They are impressive

[trail lance]

having trouble with occasional blurred images in DreamStudio.

[wet nacelle]

I churning my milk to that one broski!

https://tenor.com/view/churn-butter-tedious-churning-gif-5101385

why is the logo skewed off from the center? WHAT THE FUCK! AHHHHH! FUCK NBC! FUCKING ASS HAT OF A LOGO! FUCK!

I love legos. It's so much fun to be in a silent room building a set from instructions on paper.

[wet nacelle]

@uncut steeple Hey. You might want to take that down for sure.

@uncut steeple

[wet nacelle]

be aware of @spice sail .

Punk ass bitch. Block this brotha if he PMs you.

[pallid path]

LMFAO

why can't I just ping all mods

[open wharf]

@daring smelt

@uncut steeple

@smoky patrol

[pallid path]

thank you

[wet nacelle]

@open wharfFor real man. really

[open wharf]

yes?

[wet nacelle]

all of them. really. all of them were needed?

[open wharf]

Well I dont know which one is active

[wet nacelle]

[stylized realism of flesh house exterior]

[shy kelp]

any recc for best photoreal model for someone new to xl

[warped sedge]

keep it simple, you have sdxl, get the sdxl refiner also maybe an upscaler, find a nice workflow and experiment

[cyan crown]

thanks. blocked since he/she also PM me

[rustic garnet]

• realvision
• nightvision
• juggernaut
  are a few models I know that excel at photorealism

[lapis gale]

or realisticfreedom, no refiner neded

[vagrant valve]

Excuse me, how would I turn off shadows and straighten my node connections in comfy ui? I have been looking but just can't seem to find anything

[lapis gale]

that's more something for #🤝｜tech-support - comfy is just ... well...

[vagrant valve]

I appreciate the guidance, thank you

[vagrant valve]

Image 1 ( Base SDXL ) vs Image 2 ( Juggernaut XL )

[high skiff]

@bright valley It finally happened, just got paid for my work, and it was a great ammount too lmao.

Been trying to close some deals with the company, and now I have a few very good paying deals

[lucid rain]

What were you selling? Checkpoints or art?

Maybe logos because you tagged Harrowed?

[high skiff]

A vast set of skills in different areas. I pinged him cause he said he didn't think I'd ever get paid for what I do, and I wanted to finally tell him not only did I get paid, but I got paid damn well to lmao

[lucid rain]

Ah so it's your service you were selling? That's cool.

[vagrant valve]

How'd you find your client?

[high skiff]

I am a paid contractor working under a major diffusion company in this scene

[high skiff]

I showed them my realism LoRA, and they agreed to fund its research. Since the it has grown so much that it's truly unrivaled by even the best fine-tunes, so they are now employing me to work on additional things

[vagrant valve]

Congrats thats amazing. Is your LoRA available on Civit?

[high skiff]

It is not, at least not yet. It has grown into a product I can sell now.

At some point I should be able to release a watered down version to the masses. I have given back a ton to the community for over a year now, and I am finally able to provide for myself a bit

[glad grove]

🙇‍♂️

[vagrant valve]

Awesome, good luck on future endeavors

[dim knot]

whats name of this major player

[vagrant valve]

Could it be RunDiffusion lol

[glad grove]

jose canseco

lebron james

[high skiff]

Yeah, it is

I work with them now

[vagrant valve]

Big name, nice

[high skiff]

Yeah, they have been great tonke. We have big plants as of now, and a lot of stuff to work on. Can't say much more tho haha

[dim knot]

i prefer Barry bonds,i like guys who are into roids

[high skiff]

We're aiming for full time employment together, but for now I'm a contractor

It's nice to ease into it, and great practice too

[dim knot]

https://tenor.com/view/the-bron-james-le-bron-james-the-bronze-age-the-bronze-james-bronze-gif-19349924

[vagrant valve]

If I have three sperate LoRAs should I make their combined model strength equal to 1.0 or can they each be 1.0?

[sharp robin]

both man, this is all bout tinkering and breaking things

[vagrant valve]

Wonderful.

[sharp robin]

u can even do 3 each, who is to stop you?

[vagrant valve]

Good point

[sharp robin]

worse thing is u get bad gen

then u scale it back

[high skiff]

If you combine them, you mixed them all with the ratio of their weights

So if you have 3, one is at .8 strength, one is at .6, and one at .4

You mixed it at 80% 60% and 40%

It will be identical to if you put them together

My realism model is actually 15 LoRA's I have glued together and progressively refined. And it's extremely robust, efficient, and easy to improve because of it

It produces results like this from a single sentence with no negative

[vagrant valve]

Great. I think my biggest weakness is prompt architecture. I have been using a local LLM to help make my prompts more detailed. Would you have any suggestions on effective prompting techniques?

[sharp robin]

the less negatives u use the better, be detailed but dont over do it either,

[high skiff]

From my findings over the last half a year, SDXL does really good with a single linguistic prompt. At least for photographs.

Example being:

"A portrait photograph of a woman with curly blonde hair smiling in front of neon signs in a night club"

The various trainings I have done have also pushed the affinity for linguistics much further

I also have some secret sauce in my LoRA's that had the biggest improvement in quality for very little difference

For example, left side in here was before the change, and right was after

[vagrant valve]

Any example generations you would be able to share that use your LoRA?

Oh, thank you lol

[high skiff]

I have some comparisons against some competing models as well, tho they are a bit old by now

[vagrant valve]

Those are sweet.

[high skiff]

Top left is mine, top right is realistic vision V2, bottom left is JuggernautV7, and bottom right is real stock photo in all of these

I don't have the prompts identical off the top of my head, but I can approximate then if you'd like

Another focus example before and after my change that made a huge difference

And paired with my new workflows and tricks I have in comfy, I can gen a 1536x image with basically no deformities in about 7 seconds on a 3090

Been messing with some other techs as of late, might be able to get that time down

[vagrant valve]

Very nice. Have a 3080 right now that struggles along lol

Can see the improvements

[high skiff]

What do you mean?

[vagrant valve]

I think your LoRA is great my friend

[high skiff]

Ahhh, thanks!

I have some new massive changes that I should be able to show off soon

[vagrant valve]

Great, you should ping me if you are able to show anything or at release!

[high skiff]

Just a cool... 13x dataset increase lmao

Gonna be easily my biggest training yet

I don't even wanna think about how long the latents will take to cache

I have 2 GPU's now, so I can test it as its training

[vagrant valve]

Very well, how long will that training run take?

[high skiff]

No idea, it's at a new scale, and all of my trainings are drastically different. If I do a generalization training, probably like 10 hours

If I do more deterministic training, well over 24

Run diffusion offers me cloud compute tho, I might finally take them up on that offer now that I have reached this scale

[vagrant valve]

That is very sweet... What GPU's are you rocking in your machine?

[high skiff]

3090 and 3060ti

3090 for training, and the 3060ti side validates my tests

3060ti is SUPER PCIE bandwidth limited, but once the models are loaded, it's not too bad

[vagrant valve]

I can imagine that puts off a lot of heat...

[high skiff]

My window is open 24/7 now that its cold. I made a little air flow duct from my curtain that forces all of the hot air up into the window, where I have a compact high CFM fan that blows it right outside. My room stays nice and cold all day long

sometimes down to the low 40's at night. I love sleeping frigid lmao

[vagrant valve]

Lol nice, I just let my pc heat my room. I like it warm.

[high skiff]

I migth get a 4090 if this job keep spaying so good. Would be worth it

[vagrant valve]

Would very much be, 24GB Vram?

[high skiff]

my 3090 already has 24GB VRAM, but the 4090 is a good bit faster

would just be a speed increase, not much else

but my boss was saying if I get fully hired, I would have my own equipment

[vagrant valve]

Sweet man, did you do any schooling applicable or self taught?

[high skiff]

Self taught

its my biggest benefit

[vital ermine]

Don't do it

[vagrant valve]

Legend

[vital ermine]

I have a 4090 it is not faster than a 3090

not for this training

[high skiff]

I have no fundamental understanding of how this stuff works, which leads me taking a much more creative approach to this stuff, which almost always leads to better results

My results are so good cause so many people told me it would never work, and it did lmao

or I should say, I do understand how it works, but I am not set in a single mindset about it

[vagrant valve]

I see. I'm in a fork between whether to go to school for compsci or continue working and try to self teach...

[high skiff]

I have no understanding of code, but I have very good idea sthat seem to work really well when put to work.

I need to learn python, then I will be really unstoppable, cause all this stuff I have shown has been from just normal tools used in a unique way

[vagrant valve]

python is actually super intuitive, I don't think you would have any trouble with it. Would be super helpful in this field.

[high skiff]

I bought myself a gorgeousss monitor with myf irst pay check, cause I figured if I am gonna be looking at it all the time, might as well be about as good as money can buy

I love it so far

[vagrant valve]

Well, what monitor did ya get?

[high skiff]

the Alienware QD OLED 34 inch ultrawide

[vagrant valve]

Ah.... Fantastic

I would love an OLED monitor but I get so worried about burn in

[high skiff]

3440x1440p, QD OLED, 175hz, 0.1 ms response time, HDR400 black, HDR peak 1000, 98% DCIP-3 color space, 10 bit

Its amazing haha

[high skiff]

it has a 3 year warranty for it

only reason I bought it

some people have had them for over a year, and QD OLED is not only brighter/prettier/more efficient, but its more burn in resistant too

this monitor looks absurdly good

I love it

[vagrant valve]

I am very jealous. Have been looking at that monitor for a while.

[high skiff]

Its fantastic, just make sure to buy the DWF, not the DW if you ever do

[vagrant valve]

Noted, thank you

[high skiff]

the DW sucks (its the one I have)

I am demoing the panel to see if I wanna buy the DWF from Dell

[vagrant valve]

I gotcha. nice stuff

[high skiff]

DWF is cheaper, has upgradable firmware, many more settings, full black design, proper pixel refresh, and a lot more

[wet nacelle]

hi

[glad grove]

but it has number 4 so its better

[high skiff]

only downside is its 165hz instead of 175hz, and it doesn't have a G-Sync module (but the G-Sync is why you can't updat ethe DW, and it also makes it a pretty loud monitor cause the fan for the module is loud)

but you get Free Sync premium, which is nearly the same, so I mean

[vital ermine]

Wish I never purchased it. 50% faster and now 5090 is 50% faster than the 4090. ffs

[high skiff]

only issue is Amazon only has the DW, so I am testing the panel to make sure I like it before I buy it from dell

[glad grove]

how much is 5090 gonna cost tho?

[vagrant valve]

Youll have to let me know if you love it

[high skiff]

thats how new hardware works...

[glad grove]

i dont remember that for 4090

[high skiff]

I really do. Its dazzling

Gold/yellow on it is fucking stunning

[vital ermine]

best seen on the top of the line systems is same speed to 0.xx it/s faster

[vital ermine]

You do you but be prepared to feel ripped off if you already have a 3090, especially if you have a 3090ti.

[high skiff]

I have a nice IPS monitor, and it looks damn good, but then you look at the QD OLED and its not even close lmao

[high skiff]

the 3090ti is only about 5% faster than the 3090, and the 4090 is faster in every way. I am not sure why it wouldn't be an upgrade

[vital ermine]

None of us 4090 owners for AMD or Intel systems can figure out why it is not super faster

[glad grove]

well if its 5% faster u can always claim "its faster"

[vital ermine]

As I said you do you but be prepared. Warning given so my conscience is clear.

[high skiff]

@vagrant valveif you DO ever get one of these monitors, do not use it in peak HDR1000 unless you have proper HDR content on display, cause it makes everything else shitty

I hated this monitor when I first got it cause of that

[vagrant valve]

Also noted, thank you

[high skiff]

since I switched to HDR400 true black, oh my god, it looks os much better

the problem with peak 1000 is it will dim the brightness on white no matter the luminance. So you can have it at like 50 nits and a full white screen will drop to like 20.

Apparently peak 1000 dims based off signal values, not actual monitor brightness, when 400 does dim on brightness, not signal

https://cdn.discordapp.com/attachments/774664985031737374/1183325305813880932/20231210_003046.mp4?ex=6587ec92&is=65757792&hm=93ac3023f899e165ad0236bac0d6c56f79d9025bf3ab6938491c5f8f514c8393&

so this is what it was like, and I hated it at the start

but now its stunning lmao

vice reveal moment

[bright valley]

Grats man

Working on a platform for online generation with harrlogos right now

Time to pop off 🚀

[high skiff]

I have a job now too from it. Super happy and excited for what this all means

[bright valley]

Awesome man

[high skiff]

that reminds me, been planning on giving you some competition in the text gen scene :p

but for now, I have a huge LoRA training to do

need to set that up right now

[bright valley]

Good luck

[high skiff]

I just can't believe I can actually call this my job now... its surreal

[bright valley]

I have the HxSVD workflow dropping soon too, maybe tonigh5

Yeah I felt the same way when I left my programming job to do art full time

[high skiff]

Like I am just some 19 year old fucken around with some images, and my results are unrivaled right now... Insanity

Excited for what we are working on. Sure to not disappoint :p

[high skiff]

count your days, I'm coming after text gen next hehe

get some healthy competition up in here

[dim knot]

this reads like one of those my dad works at nintendo combos you see on reddit

[bright valley]

bring it

[high skiff]

except we ARE the dads, and we DO work at nintendo lmao

alright, time to try this large scale training... No idea what to expect lmao

[bright valley]

That's an honor

[high skiff]

I have never trained on this scale

talked with people who have, and they say the dynamics change at scales like this

[bright valley]

I wouldn't know, I've spent all my time with the same model

[high skiff]

same here, for SDXL at least

[bright valley]

For better or for worse

[high skiff]

I made a few hundred personal LoRA's for 1.5, but those were always half baked meme shit

when you could train a LoRA in like 5 minutes lmao

annnddd, I nuked my dataset

need to redownload

thats gonna be a hot moment lmao

doesn't help that its rate limiting me

screw it, gonna try half the dataset

[dim knot]

😮 where are ur staff badges?

[high skiff]

I'm not a staff at SAI 😅

That ship sailed a long time ago lmao

[glad grove]

y'all need to behave and show respect to the staff

[high skiff]

Who what

Who showed any disrespect to the staff? Lmao

[glad grove]

forgive my manners sire

🙇‍♂️

[rustic shadow]

Really enjoying this XenomorphXL Lora though the results can be quite soft

[lapis gale]

well fast enough to make a difference - 5s for a 1024x1024 sdxl image at 30 step dpm++2mk is a total gamechanger

[vital ermine]

We are talking training here. which is even worse 😦

[storm spade]

Tibetan little girl, delicate face, naughty holding red lipstick in front of laughing and running, behind the grandmother holding a stick angry in pursuit, on the prairie, blue sky, photography, HD, 8k panorama

[lapis gale]

ok, i've only be training embeddings - and for that, rendering speed is crucial as well since it creates images and compares them to the training set 🙂

[vital ermine]

We are far beyond that where training can be up to hours.

Someone I know where I hang out at now just did an 18h training to be released soon. 3090

on a 4090 the same setting was 18:10h

[lapis gale]

good thing it's winter on the northern hemisphere, so you save a lot of heating costs 🙂

[vital ermine]

Yep, he is in the UK

one thing the 4090 is for sure is a blast furnace

[lapis gale]

but the 4090 not being significantly faster sounds strange... where's the bottleneck? computation-wise it should leave a 3090 in the dust

[vital ermine]

even downvolting etc... they blow out 67c-70c heat

[lapis gale]

i know, i did some thermal images of mine in the burnin test 🙂

[vital ermine]

We had a huge discussion about this and we don't know. We do know part of it is Python and its gil being one thread but that is so minor it is just academic discussion at this point. Where the big bottleneck is who knows. Hardware wise the 4090 is 57% faster but we don't see it

[lapis gale]

for hardcore training i imagine real server GPUs to be better anyway, more VRAM and all... i'm very happy with a dual-use 4090 (run cp2077 at full res with ultra raytracing and all that 🙂 )

[vital ermine]

I am wondering if there is something holding it back on purpose as they did with that LHR stuff with crpto so they could get more of their AI expensive stuff sales.

[lapis gale]

but why would it only show up in training, not in rendering?

[vital ermine]

well, not a gamer so if all I wanted was to play games I would get an AMD and be done

the way it works. For instance in inference I should be seeing around 40-60it/s but all we see is around 7. It is a known issue that I don't think will be solved so I laugh when Jensen is saying the 5090 is 50% faster in AI/ML than the 4090 when he said the same thing for 4090 from 3090 and we don't see it.

[lapis gale]

hm... fingers crossed for a software solution

[vital ermine]

well, it would be nice

I often wonder if it is Python, and Pytorch at work but then I still remember LHR and how games were unaffected but mining was hit. mining is a bit like this so I wish someone could dive in. Remember they don't like to let us have the good stuff they want us pay 45k-100k per chip solution (H100 and now H200 with H300 in 24 or 25).

If the slow down is at the driver level it would take a driver hack as they did for crypto but no incentive with this for someone to crack it

after the crack they actually added hardware to slow it down though.

[wet glacier]

As a general rule, if you have say a 3090 already, skip the 4 and get the 5, as that will be (mostly) completely new architecture and you would notice a substantial jump. Ofc, this logic is only good unless you need an extra card for work right away or have disposable income 🙂

[vital ermine]

yep, and can afford what the 5090 will command cause it will make even the 4090 look cost effective in comparison.

[wet glacier]

most likely

[vital ermine]

for me I am done with Nvidia and even if not this card was so expensive not until the 8090 will I look at a new one. If it breaks before then it will be Intel or AMD.

[wet glacier]

as long as crypto mining is out of the picture, retail prices should be almost "humane"

[vital ermine]

I don't find 2k MSRP humane and 800 for a 5070.

[wet glacier]

but I agree, 1800-2000 a pop is not what I call affordable

[vital ermine]

329 5060 with full bandwidth then I find that back to humane

btw, 5090 MSRP is 2.5k so I bet 2k which means 2500 street pricing. I can buy almost 2 used cars for that

[wet glacier]

gone are the days when the average price for a gpu used to be $300

[lapis gale]

long gone...

but here's your crystal ball for 50x0 pricing 😉

[vital ermine]

thing is only gone because all of us allowed it. My hope is that Jensen is gone before I need an upgrade. Maybe goes to a new company, dies, or w/e just gone. He is what is causing all companies to jack up their prices. NV asks 500 AMD asks -20% and Intel -40% iow everything is now based on what NV charges

[wet glacier]

agreed which sadly rhymes with greed

[vital ermine]

yep

instead of priced on what the market can bare it is now priced on what Jensen says minus some percentage.

It took me almost a year to pay for a 4090 and not even a high end graphics card should. I had to dip into savings, and everything else because my fear was all signs pointed to a shortage and increased price. YEP, shortage and 500-1k more now for what is left.

[wet glacier]

I have 3 early samples of intel 770's and I was stoked when we first got them (16gb vram) and almost 1/2 the size of a 3090. But they were unusable for gaming, despite tweaked drivers. I have hopes for Intel's next gen, or I should say, keep fingers crossed

[vital ermine]

Battlemage, then druid then celestial which is when I will be due for an upgrade. I am watching Battlemage closely but remember that idiot they had and fired only saw drawing on Druid and nothing of Celestial. Vega 64 was rubbish loaded with issues as Alchemist was. Much better now so come on Battlemage.

[wet glacier]

If there is a real competitor to nVidia that's Intel and I hope they succeed. Hopefully Jensen will start sending rockets to Mars and beyond as all billionaires seem to do these days and get distracted by doing so allowing prices to drop as they should. I am pessimistic though, I think he wants to challenge Apple as the most valuable company on the planet by marketcap

[vital ermine]

100% agree with your assessment.

Thing is Intel and Nvidia are in bed together now so that does worry me

[wet glacier]

As you know though, alliances form and go away in time. I wonder though.....both got f**ed by Apple, so there's that 🙂

[vital ermine]

Well, together you are stronger if you both wish to destroy Apple.

we all lose if they have sex

[wet glacier]

that's true

I wouldn't discount AMD in the gpu arena, not yet

[vital ermine]

Well, hardware wise they are fantastic and even the 7900XTX toches the toes of the 4090 but their software side is just ass.

As Lisa has said they do not wish to be number 1 and she will leave that to her cousin Jensen. I lost all hope.

Now if Lisa left and a new CEO arrived with no family ties look out

[wet glacier]

I was just about to say, given both are Taiwanese and related, you know they are not going to trade "jabs" at each other. Clearly AMD is more interested in the enterprise cpu side of things, but so is Intel. That's why I think the nVidia-Intel axis might be short lived

[crisp owl]

✌️night, enjoyed reading the convo

[vital ermine]

I hope. I know this the MI300 vs H100 is no comparison the MI300 destroys it even having 192GB of ram

[wet glacier]

reminds me of the European royal historic alliances between all the nations during the 17th 18th centuries 😅

[vital ermine]

MI25 to now MI300 has always been ahead of Nvidia just never grabbed the marketshare.

[wet glacier]

ofc that's enterprise scale supercomputer grade hardware, but it will eventually trickle down at some point

[vital ermine]

it is too

Ryzens about to have some ML stuff in it

[wet glacier]

I can't wait

[vital ermine]

nor I as I haven't been Intel for cpu since 2016

I think a lot of stuff is soon changing especially now that Python is abandoning GIL

[wet glacier]

I am running a TR 3970x and it's still running like a charm after 2 years

[vital ermine]

Without GIL Python will finally run better on AMD than Intel

[wet glacier]

I will upgrade once the next hedt will also have some ai acceleration, for sure the next gen ryzens will have some of that MI300 secret sauce

[vital ermine]

Lisa said but we will see. If not 8k then for sure 9k will

I had a 1600 since 2016 to eight months ago then went to a 5600.

still on a B450 but time is next year to upgrade and I really want AM6 Zen 5

[wet glacier]

yeah, I heard that...so just in time for her cousin's rtx 6090 😅

[vital ermine]

yep

motherboards are stupidly expensive this gen

entry level is now 200? used to be 89

[wet glacier]

by then we'll have sdxl twinturbo running in real time 400fps and we can do slowmo video like on a phantom camera 😄

[lapis gale]

yeah, the software is definitely outpacing the hardware... although the turbo stuff still trades quality for speed

have you tried https://github.com/aifartist/ArtSpew/ yet? totally insane speed

the also has a realtime prompt morph and stuff, lots of fun https://github.com/aifartist/sdzoom

[harsh plover]

I wrote a small helper to map out SD/XL -compatible resolutions and their nearest common aspect ratios: https://akx.github.io/res-helper/

[peak dove]

...

[carmine gale]

Just wondering what people think this side of the pond about SDXL 1.0 inpainting 0.1. Seems total garbage or I’m missing something. Is it true that Stability just suggested# # going back to 2.1 or even 1.5 inpainting models? Is diffusers generally preferred for inpainting and why?

Thoughts appreciated

[shy kelp]

Hey, does anyone have any experience training LoRAs on SDXL? Is there a good example python code base for this?

[rustic garnet]

I found SDXL is good in inpainting without any extra model 🤷‍♂️

although it would be nice if someone would train an inpainting controlnet to improve it even further

[hoary saddle]

is there a way to permanently disable the thumbnail bar at the bottom now?

X just closes it but it comes back on refresh

[carmine gale]

is there a controlnet for inpainting in sd15? i never thought of it

[thorny glacier]

Yes, it’s called “ControlNET Inpaint.” I used it here to easily remove a character in an SDXL scene (but I used 1.5 based Deliberate v2 model for the inpainting):

It’s like Adobe Generative Fill but local, uncensored, and controllable!

[carmine gale]

Ok sweet. ‘Maybe I use SD 2.1… is yours q safetensor or diffuser ?

[thorny glacier]

I used Safetensors

[rustic garnet]

still, my feeling is that SDXL inpainting works almost as good as SD 1.5 inpaint model and MUCH better than SD 1.5

[indigo carbon]

hey, I've reworked my SDXL merge; which one is better? (left is old, right is new. generation settings are identical)

[upbeat summit]

they both have some interesting aspects. left has more contrast and is more color balanced overall. right has a stronger green tint, but fine details and textures look enhanced

[indigo carbon]

I probably can control the contrast by changing some of the merging's math, I'll give that a try next

[upbeat summit]

definitely better! I still think the left is more color balanced overall - and the right has a green tint. but of course that might look different with other prompts / seeds.

I would also check the highlights in the new mix on the right.

they are pretty close to be maxed out, a bit too much in my opinion as you can see in the white hair above the eyes. but it could also be possible that the new mix works with a lower CFG value and that would balance it out.

[slender coral]

Could anyone give me advice on how to both tile and upscale?

[indigo carbon]

I think I've got it better than my old one, I might upload this as an update

I'll check some other stuff though

[indigo carbon]

left is old, right is new

[crisp owl]

taco bell food under a uv light?

[indigo carbon]

I feel like my new one has more creativity

[heady vale]

those are some toxic tacos

but yeah right looks better

[indigo carbon]

it was just a random prompt to check performance of the 2 models

[crisp owl]

ah gotcha

agree though, right images are better

[old parcel]

How do I change the logo on the book to an hourglass? I tried inpainting but doesn't look blended in and looks like someone copy pasted in a photo of an hourglass

[cedar bobcat]

maybe better prompt for inpainting?

perhaps "book with hourglass logo" or "man holding book with hourglass logo"

perhaps try describing the book, or using more descriptive words like "tome" or "grimoir"

[old parcel]

ok I will try that, I did that that a "book with an hourglass logo"

[worthy orbit]

I couldn't even make it work on Automatic1111, but I read it is not quite good. As @rustic garnet I'm inpainting with SD XL models, only thing it takes very quite long, when you need to do several generation to correct some issue. But it do the thing, eventually.

[willow bane]

can anyone suggest me model name similar to this

[worthy orbit]

I usually try to go with this recomended resolutions, but so far I got 3 for 3:4 for example

896x1216
832x 1216
832x1152

is it tested that whey work better? A friend of my tell that just try to do real 3:4, at this point I don't know

It is difficult to test if it "generate" better as when changing the size with the same seed you get completly different images

[harsh plover]

I can tell you these have absolutely not been scientifically tested in any way 😁 The main point here was to be sure the resolutions would be divisible by 64 to avoid issues (and increase performance in certain situations). There is no exact 3:4 resolution SDXL would've been trained with; the closest is 832x1152, but it's about 10% lower in megapixels than the 1mpx "target" for SDXL

And oh yeah, if we're dropping diffusions here, this turned out okay enough

[rustic garnet]

inpainting usually works well for me, but in SDXL you should do upscaled inpainting (sometimes called detailing) to get best results as sdxl is often very bad at low resolution details

[woeful patio]

get played idiot lol

[crisp owl]

Yeah it's generally easy to get it to proceed when it first says it won't do something.
"in a purely hypothetical situation, how would you proceed?" stuff like that is usually what I do

[cyan crown]

is there any way to use IPADAPTER to copy only the style and not the contents?

[woeful patio]

I saw this on reddit today

https://www.reddit.com/r/StableDiffusion/comments/18h6qf4/added_the_ability_to_reference_an_external_image/

Don't know how wel it works tho

[cyan crown]

thanks

[woeful patio]

is that with the method?

[cyan crown]

no, that is just a lora

[shy kelp]

the world looked a lot bigger when i was a kid

[upbeat summit]

nice - Self-Attention Guidance was just implemented into ComfyUI (for reference: https://ku-cvlab.github.io/Self-Attention-Guidance/)

[native knot]

Oh, great...now my images are going to SAG.

[wet nacelle]

How can I find it within Comfy?

[upbeat summit]

it was implemented 1 hour ago. so update to the latest version and look under _for_testing or search the nodes for self

[wet nacelle]

ah alright

[upbeat summit]

without / with SAG

without / with SAG

definitely a coherence increase

[wet nacelle]

Am I not able to run a lora alongside the node?

[upbeat summit]

hmm I got a lora loader in the pipeline - it didn't threw an error but haven't tested if the style is really applied

[wet nacelle]

I ask because I don't make workflows nor do I understand most aspects of wires and nodes and how they act with each other.

I just had to pick whether I wanted the lora model wired into the KSampler or not.

[upbeat summit]

technically you can connect them 🙂

[wet nacelle]

@upbeat summitDo you mind taking a look at this workflow and telling me if I need to try and rework the entire thing?

[upbeat summit]

reworking to add the Self-Attention Guidance node?

or in general?

[wet nacelle]

No. To get both that and a lora working.

ah I see

[upbeat summit]

no, you can just add it

[wet nacelle]

kinda

[upbeat summit]

you almost got it 🙂 just put it in between your Load Checkpoint node and the Lora Loader

[wet nacelle]

@upbeat summitLora style seems to be working.

[upbeat summit]

so - now you can make even more coherent CCTV images 😄

[wet nacelle]

I wonder if the SAG can be piped into the upscaler portion too.

[upbeat summit]

yeah, it should work. it's guiding the diffusion process.

[wet nacelle]

ah okay

[ionic gulch]

internal beta of a completely revised version of my workflow. there's no more "generation mode". everything runs via switches, so you can combine everything as you wish. 4x control net at the same time, plus 4x ip adapters, 2x revision and inpainting - doesn't make sense, of course, but it works now. "everything all at once". release version may also include unsampler as alternative to classic img2img.

[ionic gulch]

2x control net + ipa

[worthy orbit]

DPM++3M SDE Exponential, more shine and errors
DPM++3M SDE, less shine, less errors

[half ivy]

shout out to whoever did this o0

[sinful quiver]

does anyone know why my generations are coming up like this with the pattern/grain?

^ settings above

I've tried using the normal vae too with the same outcome

here's a closer look, its like a canvas texture

[rustic garnet]

A canvas texture would make sense for a painting. It's a bit surprising, though, as you didn't specified painting in your prompt. I know the refiner tends to make everything into a painting with clearly visible pencil strokes, but you use the base model...
You can try with a different sampler (Euler Ancestral doesn't work so well for SDXL), use one of the DPM++ karras solver (e.g. DPMpp 2M SDE gpu)

but it's more likely an issue with your prompt

[sinful quiver]

prompt I kept simple, just "underground russian base in the forest"

I'll try with DPM++ rn

50 steps

I'm confused because I was using a1111 before comfy, and it was working fine with euler_a, but comfyui wasnt doing it the same, went back to a1111 and it also had same issue

[rustic garnet]

that looks bad. Did you use "karras" or "exponential" scheduler?

[sinful quiver]

Just normal, as comes in the default workflow for comfy

[rustic garnet]

yeah, you can't use normal with that

you have to use "karras" or "exponential"

it's annoying that ComfyUI is not blending out combinations that won't work...

[lapis gale]

"comfy" and "annoying" in the same sentence is redundant 🤪

[rustic garnet]

might be, but what is the alternative :/

auto111 is nowadays even more uncomfortable that comfy

and invokeai is great, but lacks some features

[sinful quiver]

this was the result using karras

[rustic garnet]

yep. I guess that is what your prompt gives you ^^°

[sinful quiver]

fair enough, I was a bit disappointed but remembered how simple my prompt was

[rustic garnet]

"analog film photo underground russian base in the forest . faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage"

just the plain SAI "analogue" style

[sinful quiver]

lmaooo awesome man, I'm generating an "old film photograph" version right now too.

yeah perfect, guess I just need to learn all these incantations

[rustic garnet]

I made a comfyui node for myself that adds these things automatically ^^° I mean, there are thousand ways to improve your prompt. The SAI styles are just one convenient way for that

this is the list of styles I use

[sinful quiver]

this one is beautiful

[rustic garnet]

uuuh, nice

[sinful quiver]

could you elaborate on what you mean by SAI?

[rustic garnet]

Stability AI

on their clipdrop webpage they had a dropdownlist of styles you could choose from

or on the bots here on disccord the same

[sinful quiver]

rightt and thats what the styles add

[rustic garnet]

if you choose "cinematic" as style

[sinful quiver]

so your comfyui node also offers a dropdown for style?

[rustic garnet]

then what happens is just it adds some words into the positive and negative prompt

[rustic garnet]

yes

but you can also do it manually if you want

the list I sent just contains all these styles (and some more from some source I forgot...)

[sinful quiver]

awesome bro, thanks so much for your time

[rustic garnet]

comic style xD

[sinful quiver]

that looks like an awesome scene for a cartoon style game

[rustic garnet]

most time I just use these styles as starter and refine them after I found some good image. It's rarely the case that the perfect image looks exactly these prompts

e.g. cinematic and photographic always contain keywords like "bokeh" that blurr the background which I often don't want

so these styles are more suggestions, I rarely use them 1:1

also, for paintings and comics, using artist names is still more effective in most cases

[sinful quiver]

as part of your refining process, would you use a generated image as a starting point, for img2img?

[rustic garnet]

you can do that. But for img2img you want to have the coloring right

if I change style then also coloring, darkness, brighness and so on chould change

which img2img is not so good for

[sinful quiver]

ah right fair enough

[pure hazel]

Anyone try this yet?

https://github.com/google/style-aligned/?tab=readme-ov-file

[shy kelp]

there we go, stability just broke their promise on being open source

withdrawing commercial use just like that

[hoary saddle]

what's the best way to add space to the right of an image and fill it in to match the image2image?

in comfyui, not a1111

[cyan crown]

https://civitai.com/models/232376

[hoary saddle]

anyone know how to fix this? where the bottom image adds black space into the bottom of the top image?

i think it's maybe not having a mask on the send image but i cant figure out how to set the result of the 2nd sampler to be the mask input, the node won't accept it from the vae decode

conver image to mask works but changes the look of the 2nd image but with red channel

is there an image to mask without channels?

[pure hazel]

They're hurting financially...

[urban fjord]

I'm not seeing any sign of the open-source status changing

Open-source is a terrible term here anyway given that it's not the source code that anyone care about but the model weights. And those will still be available.

They need to earn money and this is a lot more open solutions than the alternatives.

[peak dove]

I'm TensorRT'ing my A1111

[long stump]

hey guys, im trying to use kohya to train a lora for sdxl, imk trying to train using multiple gpu's, does anyone know how to do that? i can only get it to work on one gpu

[bright valley]

Yeah all these fees I have to pay to use Stable Diffusion are getting out of hand

[old parcel]

Fees? It’s open sourced

[bright valley]

(kidding)

[jolly brook]

looks like anyone wanting to use future models for commercial purposes has to pay a monthly fee though

#📣｜announcements message

[bright valley]

If you make over $1,000,000

Or everyone

[jolly brook]

They say to contact them about pricing if you’re making over that much. If you’re making under that much, it’s $20/mo.

[bright valley]

Looks like everyone I guess?
(Sorry your message didn't pop up for me before I sent this)
That's weird, everybody who's mentioned it to me seems to think it was only over 1m

Did everyone have the same misunderstanding somehow, or was that a last minute change? Seems like those are the only two scenarios

[lucid rain]

They say to contact them if you need to work something out I wonder if they will actually do anything for people or if they will just ignore everyone.

[urban fjord]

If you're doing actual commercial activity then a $20 license fee shouldn't be an issue. Everyone else that use it in a non-commerical setting can continue as before and hopefully this will encourage better model as SAI need to provide a better product.

[crisp owl]

Then there's the ole, what exactly do they consider "commercial" vs "non-commercial", because without reading all the fineprint, that designation is always different between everyone

[wicked frigate]

commercial vs noncommercial is usually pretty clear except when very definitely commercial users try to make arguments for why technically their very definitely commercial usage of very definitely whichever product in question isn't actually commercial usage

eg "we're not selling a Stable Diffusion webservice, we're selling a general hosting service that has a 1-click SD setup and advertises to SD users as a great way to run SD but technically not exactly an SD webservice!"

there are a few genuine edge cases, eg "is a youtuber with ads enabled a commercial user?" but in this case the membership licensing is focused on model service hosting so youtubers and wotnot should be fine

(ftr i am not a lawyer this is not legal advice etc)

[crisp owl]

precicely that last case is where I stand, and probably others questioning what it actually means

[ionic gulch]

That's why my new workflow will be called "Everything All At Once":

features and settings:

[dim steeple]

Hi! What is the easiest way to train a XL model? Is dreambooth still good?

I haven't trained since 1.5

[wicked frigate]

most people i believe are using kohya's scripts - or at least bmaltais' GUI wrapper over kohya's scripts

You'll generally want to train a LoRA in most cases, unless you have a specific reason to train a base model

[dim steeple]

thank you

should a persona be a LoRA

[lapis gale]

how about embeddings? that was my favorite way to get a person's looks into the AI, worked with all 1.5 based models etc... not tried with SDXL yet. Probably just 1024x1024 as training images and the rest is identical?

[dim steeple]

yeah maybe

[lapis gale]

i think so too, but anyone with hands-on experience? 🙂

[dim steeple]

I am trying to train a few persons image on top of sdxl base as its own model

is that crazy

should i be doing it as a lora

[lapis gale]

as a model sounds like total overkill

i preferred embeddings over loras because they were just a few kb in size and universally useful

back then, some months ago, in the SD stone age 🙂

[dim steeple]

things moving fast

makes me think how different things will work in another 12 months

[lapis gale]

yeh, it's all at warp 9 🙂

[dim steeple]

the number one thing i want is faster render times

would change everthing

like xl turbo but without sacrificing quality

[heady vale]

you can using this model. only requires 5 steps
https://civitai.com/models/215418
make sure to use the settings suggested in the description

[bright valley]

what in to hecc

And I just made mine bigger! 😂

[lapis gale]

cough cough 4090 cough xmas cough cough

5s for one normal sdxl image is totally fine (1024x1024 with 30 step dpm++2mk)

and quality is top

[dim steeple]

💯

[dim steeple]

but is it going to be a little worse quality than full xl

[heady vale]

its up there with some of the best xl models

[dim steeple]

does it work with controlnet?

[heady vale]

still has some issues to resolve, but massively better than base xl

[dim steeple]

also people are using negatives in xl? https://civitai.com/images/4326650

[heady vale]

negatives are still useful

[dim steeple]

oh its by the dynavision guy?

[heady vale]

yeah same author

he has a lot of vision models

[dim steeple]

trying it out

5 steps?

1.5cfg?

its not better than dynavision full xl it is?

at 20 steps

[heady vale]

dynavision is part of turbovision, you can get most of the same styles

[dim steeple]

btw, any tips on getting more specific and accurate likeness?

it is lora?

a lot of the examples in the model don't look like the real people that much

[heady vale]

thats to avoid copyright

[dim steeple]

lol turbo is taking just as long with 5 steps rn

maybe bc two controlnets?

ohh it needs to be 512 right

how many high res fix steps?

[heady vale]

no, turbovision is 1024 like normal base

[heady vale]

less than 5

[dim steeple]

it seems to be taking same ish amount of time for me

[heady vale]

compared to what?

[dim steeple]

compared to dynaxl

https://civitai.com/models/122606/dynavision-xl-all-in-one-stylized-3d-sfw-and-nsfw-output-no-refiner-needed

[heady vale]

I dont use controlnet but TV is heaps faster

[dim steeple]

why does it say 1/10?