✨｜sdxl

[noble shoal]

He should better use a bathroom for that

[upbeat summit]

I mean I'm used to these sounds from bad motherboard sound cards that are just not shielded correctly. it sounds very similar. but in this case - it's the hardware itself making the beep boop beep chirping sounds

[icy brook]

and I thought I posted something beautiful: got fried directly 😅

[nimble heart]

yea 100% you can get a cheap external DAC on Amazon for like $40 to get rid of electrical interference on your headphones

won't fix the GPU microwaving the air though

[zinc cargo]

dont be jeallous 🙂

[upbeat summit]

oh no - it's microwave rays? man, if I would have known that I could have made popcorn all the time when exploring the latent space

[native knot]

https://tenor.com/brkvK.gif

[noble shoal]

Wait a second. I think you gave me an idea.... Is there already a deep fried lora?

[upbeat summit]

just put your cfg to 15 - done

[icy brook]

https://tenor.com/view/theres-so-many-out-there-shea-whitney-theres-a-lot-a-lot-to-choose-from-so-many-choices-gif-21782657

[noble shoal]

So i order my chicken at CFG 15?

[nimble heart]

both my 1070 and XTX start to whine in uncapped game menus so it's a pretty universal issue. My 560 ti was the only card that didn't seem to whine, probably just not juicy enough.

[upbeat summit]

yeah... 18 might make it too dry

[noble shoal]

So much to learn

[nimble heart]

XTX whines in AutoGPTQ and slightly in Exllama but not in any stable diffusion programs.

[half cedar]

Someone try this mexican day of the dead, cemetary , mexican sugar candy, day of the dead tacos, day of the dead guacamole, day of the dead taquitos, halloween, fog, flat lay food styling, style of martha stewart

[nimble heart]

nah it's 5G waves. Too high frequency for popcorn.

be careful if you have vaccinated people in your room while your GPU is loaded though, it might superheat the chips in their blood.

[native knot]

bro, I live for superheating the chips in my blood.

[upbeat summit]

oh I heard about that but somehow nothing happened 2 days ago 😄

[nimble heart]

I was updating to Android 14 so I missed it

[upbeat summit]

same lol

[nimble heart]

no reptilian GF for me

[upbeat summit]

well - next time

you might get a zombie GF instead

[nimble heart]

also it didn't say in the ads but the very bottom of the patch notes mentioned if you plug your Android 14 into a computer you can use it as a webcam

like natively without droidcam or anything

[upbeat summit]

ah nice. didn't know that. about time. thanks for the hint!

[nimble heart]

the 14 site spends like 2 pages talking about different lock screen clocks meanwhile the best feature is stuffed in an overflow menu at the bottom lol

[upbeat summit]

yeah I skimmed the bullet point feature list and skimmed the changelog and thought - hmmm kay. not that exciting 😄

[half cedar]

Freddy kreuger version of this from Bing. I put charred before all the food names

[stone fossil]

xD

[sinful falcon]

@icy brook You had to know one of us numbskulls would mash aether fire, cloud, and ghost together 😜

[upbeat summit]

wow

[sinful falcon]

But that horse aint safe to ride

Just need one more element... WATER maybe? 😛

[sinful falcon]

The amount of steam this whale breach would make...

[sinful falcon]

Looks like dang good suqgghi

[icy brook]

What the heck is going on there? 🤣

[icy brook]

This one was cool 🙂

[sinful falcon]

I'm making soup!

https://tenor.com/view/muppets-muppet-show-swedish-chef-pans-pots-and-pans-gif-24864855

[noble shoal]

Ok, one of you must have already done this.

[native knot]

I just see a regular old fire ant.

[noble shoal]

You got this

[sinful falcon]

Ok, I toned down the fire and upped the cloud. This guy is pretty much quenched.

[noble shoal]

Ah, the great Baseball fire incident of 1906

Good that they made a card from that

[noble shoal]

Have you tried burning ghosts?

[rain crown]

Looks so good it's almost like he's gonna pop put of the screen and smack me in the face 😳

[native knot]

Get his LoRA's name out your fuckin' mouth!

[icy brook]

Fire massacre 🤣

[sinful falcon]

Ha, I'll have to try that. Ghost Rider wasn't a ghost... let me think a sec on who would work

[noble shoal]

Use Joachims other lora. Aether Ghost

[icy brook]

You should input the classic Win XP desktop image through ControlNet and run it through Van Gogh 🙂

[icy brook]

Just check the sample images’ prompts.

photo of fire that looks like a transparent translucent ghost etc

Or

photo of a transparent translucent ghost made of fire etc

[sinful falcon]

Pretty interesting - looks like he's leaving his burning body

Ghost and Fire both at strength of 1.0

[icy brook]

The transparency of Ghost mixes with smoke from the Fire Lora I bet.

[sinful falcon]

Oh, nice

[icy brook]

Since both share transparency.

[sinful falcon]

Strange, it took a bit to get a single rider. The recipe for this one was Fire: 0.7, Ghost: 1.0, Cloud: 0.4

[upbeat summit]

nice!

[nimble heart]

constrictors always look like scaly puppy noodles

[pure crystal]

chupacabra

[rain crown]

Now put Chris Rock in the ring as well

[upbeat summit]

pretty dope character posters!

[pure crystal]

I've wanted to upres these for a while, very happy how they turned out

[native knot]

Very nice, bob.

[fast bobcat]

hi can you use 1.5 loras on sdxl?

[native knot]

no

[fast bobcat]

thanks

[fast bobcat]

amazing style, which Model and lora is it?

[native knot]

protovisionXL. The only LoRA I typically use is the add-detail-xl one.

The magic is in the prompt on this one. 🙂

[fast bobcat]

really good! loving the results

[native knot]

controlnet pose?

looks like the mcu promo shot of her

[nocturne dove]

Yeah. I have a dumb workflow that has ipadapter+, controlnet, and a couple of loras. So I take a photo (or render or painting) and use it with the inkdrawing lora but also use controlnet with the canny edge preprocessor. And then I also do img2img with something like 75% denoising. It's dumb but you get very little difference in composition between the original image and the inkdrawing lora won't make it b&w.

[native knot]

if it works, it's not dumb

[nocturne dove]

I'm generating images to train a new lora so that I won't have to push the inkdrawing lora to do colors + comic art.

oops I had it prompted for "iron man" for the dr strange image. It makes a little difference.

[nocturne dove]

It fixes his face if prompted with his name

[native knot]

More importantly, skin tone; was really bad on the hands in the last one.

[nocturne dove]

Yeah it does that like half the time no matter what. I've been generating three of everything and then for the images I'm going to train a new lora with, I am fixing things like that using photoshop + layers with the best images.

[vale eagle]

How could I off load the model during workflow?

in comfyui

[uncut fiber]

@upbeat summit Please how is that node temporaly used instead of AIT called?
o.k. working now.

[zinc cargo]

ok, i need this

[noble shoal]

Thanks to all of you for the good reviews, likes and sample images for my two Inkdrawing loras on civitai. I didn't thought they would get so much attention. Truly a great community. I am working on 2-3 more styles and hope you will enjoy them as well.

[nocturne dove]

They work very well. I used them to generate images to train a new lora (I wanted non-b&w and more background details) for myself. Looking at the results, I probably could have found some old comic book scans to use for data, but your lora worked just as well, which is kind of surprising if you think about it.

[noble shoal]

Yeah, i was also a bit surprised that they actually support color. Well, it's a bit harder to prompt for, but it kinda works.

[nocturne dove]

Yeah my new lora just does this with no specific prompts for colors. That's not a criticism of the inkdrawing lora because obviously it is made to make certain kinds of images and color is a bonus. And 100% of my training images came from your lora's output.

Anyway it's not that hard ot get color or b&w images from the inkdrawing loras. They're way more useful then 99% of the pretty girl loras on civitai.

[noble shoal]

I was onto something like this too, but it is hard to not fuck up faces and hands even more when training on generated images. So i have moved on to new styles. Nice to see some creative use. I like your images.

[nocturne dove]

I only used generated images that came from controlnet and img2img and threw out the ones with bad faces. Some janky hands definitely made it in though 😢 So most of the images were around this quality. I think I used about 150 images which may be too much. Idk.

[noble shoal]

Original dataset size of the DC version is only 95 images

[nocturne dove]

That's probably reasonable. I haven't seen better quality when using a lot of images. But it feels wrong to train with only a handful.

[noble shoal]

It's easier to get bad images into your dataset if you have a huge one. And they will screw your lora

[nocturne dove]

Hmm. I should probably check my images then and cut it down.

[nocturne dove]

Is that one posted yet? It seems good.

[noble shoal]

I am trying to get the jankyness a bit more out. And here and there a baseball player appears. It just needs more polish before i throw it out in the wild.

[nocturne dove]

My solution would be once I could get good images from it, I'd start replacing the original baseball cards with generated images. And then that would risk distorting the style and causing problems. But it's easy to do.

[uncut fiber]

finaly driver, but too slow one

[nimble heart]

speed racer

[ionic gulch]

about the coil whine: i've payed $150 extra for my 4090 to get a card that was known for little to no coil whine, as this is a very common problem i once had with a previous graphics card. since then i always check forum posts and reviews first to try to identify a card that is known for little coil whine.

for the 4090 i think i also got one of the coolest designs - and it has no coil whine when using comfy ui:

My new workflow is available on GitHub now - it allows to turn off features and parts of the workflow by muting nodes - this reduces requirements and increases speed a lot:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

(check the black info box on the left side of the workflow for instructions and requirements)

[uncut fiber]

Somebody asked for similar look like this, if it helps those one

[lusty wolf]

Freaky...

[uncut fiber]

very detailed

[uncut gull]

Pure composite (left) vs. Re-render of composite (right). The idea with the re-render is to help balance the colors and make it seamless. It's not a huge difference, but it definitely helps with improving the font. It makes the design "work together" in a subtle way that feels more professional. The absolute darks and absolute brights match in the re-render.

[soft zealot]

have you thought about using ipadaptor?

Load OG baseball cards as the reference images then generate new ones?

(fortheseI just googled marvel comic pages and grabbed 6 images to use)

[strange mist]

hello 👋
how have you been? I haven't seen you in a while 😄

[soft zealot]

been busy with work plus a week in Prague

[strange mist]

a week in prague. nice
heard it's a wonderful city to be.

[soft zealot]

well 5 days, its nearly a week lol. But yes its a lovely city 🙂

ah here we go , Random Baseball Card Generator lol

[strange mist]

cool. are you using the mini lora technique?

[soft zealot]

I believe some people may refer to it as that yes

just playing at the moment

[strange mist]

it's a neat idea. probably the best use of ipadapter 🤔

[soft zealot]

best video Ive seen on it is the devs own

https://www.youtube.com/watch?v=7m9ZZFU3HWo

[soft zealot]

I like that you can create .ipadpt files to reduce overheads. Create preloaded/ipadapted embeds

but all explained in the video

[strange mist]

sounds complicated. hopefully I can figure it out 😄

[soft zealot]

its a good walk through

anyways busy afternoon day ahead.

Rugby World Cup I have to day
Wales (my team) v Georgia
England v Samoa
Ireland v Scotland

Also would like to try and half follow the F1 sprint race from Qatar lol

[strange mist]

I'll give it a watch.
have fun watching WC 🍻

[soft zealot]

I have a case of Doombar tokeep me company lol

[strange mist]

sounds exciting. best companion to have ☺️

[uncut gull]

Tuack Man's rookie days. 🙂

[soft zealot]

They were randomly selected and none of the names mean anything to me lol

[uncut gull]

Yeah, for specific text, I render the text with a font in GIMP or Inkscape, merge it with the image, then re-render.

[lusty wolf]

Double Ice-cream?

[terse tundra]

combination of two loras I am working on

it is originally a painter style:

maybe you can guess the artist, but it is still quite flexible especially combined with other loras

[uncut gull]

Game token numbers. Generating a font.

[uncut gull]

Hmm. Nope, this is the way to go after all. The font needs to be added in-game, but the icon needs to be rendered in-balance with the base card.

[uncut gull]

Crafting. Unique font for each number, but it takes a while. 🙂

[ionic dragon]

which kinda art is this?

[half cedar]

^ Gave it a shot: (Mantis shrimp, whimsical:1.3), highly detailed, Style of stippled oil pastel marker cartoon illustration, whimsical, bold black outline, character art, close up view, mottled dappled rich vibrant colors, (cross hatch shading:1.2), bold contour lines, 90s, best quality

[half cedar]

Lol

[half cedar]

Bing says I got it (haha)

[eternal fog]

I tried to use your dog image, but apparently not enough weight lmao

[half cedar]

Haha oh my

[eternal fog]

This is closer to the style

[half cedar]

So cute

[noble shoal]

Ok Joachim. I am currently testing my "deepfried" Lora. It's looking tasty so far. @icy brook

[brave stream]

do you guys think negative prompts are really needed for SDXL and all the various user-made SDXL models? And if so, what are the bare minimum negative prompts needed? thanks!

I'm trying to make realistic-looking images, btw

[native knot]

Negative prompts are less necessary for SDXL, but they still can be used effectively. I typically only keep 2 in my neg now on well-trained models: watermark, signature.

And whatever you do, don't get sucked into the bullshit that is something like bad hands, bad anatomy, etc. Results that are good from images generated with those terms in SDXL are simply happy coincidences. If you do need to use something like that, an effective term tends to be something like deformed, but often times, the best thing you can do is simply roll the seed and get another image which has just as good of a chance of not having issues as another word added to your neg.

As far as things that look photo-reaslistic, you actually want to stay away from the term realistic as that's not what tends to get used with actual photographs of people; it is used with art that looks close to real, so you're actually doing yourself a disservice if you use that term in a lot of cases.

What does tend to help is to use terms that are related to photography. The size of the film, a specific lens, a specific camera type, an f-stop setting, etc.

[vital ermine]

anyone know how in the eff you find your own uploads on civitai now that the "My uploads" tab on your profile page is gone?

[native knot]

I say all of the above, but the biggest thing to realize is that, by far, the best thing you can do is understand all the basic settings for the tool you're using, be it A1111 or Comfy. If you know what steps, CFG, sampler, schedule, denoise, CLIP, LoRA, ControlNet, IPAdapter, etc. do, then you will be far better off than people who don't.

[vital ermine]

even their own bot shrugged and said the site must have changed since I was last updated with the information. ffs.

[brave stream]

Thanks. This is also my impression after using SDXL for a few weeks now.

[noble shoal]

"Your Profile" and then "Models"?

[native knot]

I always just click on my profile, then choose from the menu on that screen.

[vital ermine]

Nope, models gives me zero

[noble shoal]

Filter?

[vital ermine]

I only have NSFW filter on

now none of mine are a model they are loras so that may be why

[native knot]

Published/Draft, All Time/Day/Week/Month/Year?

[gray sage]

Hey, quick question here. Is it possible to generate images based on another image like a Logo for example. I want to create an animation for a client

[vital ermine]

nope, I have never been able to see my own work through profiles BUT I can on a general search

[native knot]

The Models page should show your LoRAs.

[vital ermine]

never has

[native knot]

Show us a screenshot of your profile page.

[noble shoal]

Or give us your password.

We dont tell

[vital ermine]

LOL

[vital ermine]

I turned the NSFW filter off (only one I had) and they all appeared so facepalm time

[native knot]

Boom.

[noble shoal]

Still 3 Filter

[native knot]

That's me.

[noble shoal]

Sorry. Idiot here

[native knot]

Just so I could narrow it down.

[vital ermine]

Well, at least it works now. Thanks

[native knot]

Glad you got past that.

🙂

[vital ermine]

About to release a V2 of one of these as it is V2 and not simply 1.1

[native knot]

Noice.

[noble shoal]

Hey Soul. It's 2:0 for me. Remember when i helped the other poor guy with his Lora problem? 😜

[native knot]

I said the same thing you did at the same time.

Tie.

[native knot]

Are you vacuum sealing things again?

[vital ermine]

Yes

[native knot]

🤣

[noble shoal]

🤨 Ok. This time tie.

[brave stream]

Question: in SDXL do parenthises and weights make a difference? Example: "side-view profile photo of a pretty white girl with ((natural elf ears))"

[vital ermine]

it does

[brave stream]

is the double (( the most potent weight? any tips for getting elf ears? 🙂

[vital ermine]

use (XX:yy). (()) is rather dumb these days

[noble shoal]

Maybe a captain spock LoRa 😬

[vital ermine]

something like (xx:1.4) for example

[native knot]

Best to get used to the (term:X.X) format so you can set precise control. You can even use multiple decimal places like (big booba:1.437).

[vital ermine]

it helps

beyond 2 digits past the decimal point is crazy, lol

[native knot]

I've used it when I was just trying to nudge something the tiniest bit.

[brave stream]

"something like (xx:1.4) for example" so what is the available range?

[vital ermine]

I mean you can use a transcendental weight if you wish BUT... lol

[native knot]

Depends on what you're doing...frankly, you can set it as high or low as you want, including into negative, but generally most things don't like beyond +/- 2.0

There are some things that work well with higher/lower. And it also depends on the rest of what you have going on.

[vital ermine]

training and warming up my room as the heater came on this morning and shocked me. Turned it off and use this blast furnance card instead.

[half cedar]

1.3 seems strong but tends not to distort the image. Use range btwn .8-1.4

[uncut fiber]

and for less importance you can use [something:2]

[brave stream]

my PC itself acts like a space heater when I am running AI inference all day. 🔥

[vital ermine]

Yes, if I do a huge batch run is when it heats up the most

[native knot]

@brave stream - One other thing to note about prompting...generally, the further back the term is in your prompt, the less it will be focused on. So another method if you need to stress something is to change the order of your prompts to move something more toward the beginning. That said, some things tend to naturally be stronger than others and so weights and place in the prompt are simply things you need to toy with to get the output you're looking for.

[vital ermine]

there are some things that are so heavy even loras refuse to touch them

[noble shoal]

I need to get a consistent breadcrumb coating. This looks already delicious....and stupid. I guess i have to retrain a lot.

[native knot]

And nearly every model out there has a far heavier bias toward female terms than male.

[uncut fiber]

when long run, i limit gpu a bit, i mean overnight for example.

[native knot]

Almost Cheetos antlers.

[native knot]

Deep fried SDXL: 8/10
Deep fried SDXL with rice: 10/10

[uncut fiber]

soul you can make large baking pan 😄

[native knot]

10-4. The shuttle is ready to deploy the MBA or Massive Baking Apparatus to any destination on the planet. Just tell us where to drop it off, sir.

[noble shoal]

I think t "B" stands for Balls

[native knot]

you would

😛

[vital ermine]

I noticed on CivitAI when I want to upload a V2 I have to remake the entire pages over again. Is that the way?

[noble shoal]

Looking crispy

[uncut fiber]

crunchy

[zinc cargo]

https://tenor.com/view/futurama-fry-shut-up-and-take-my-money-gif-20932781

you get it? bcz his name is fry...

[noble shoal]

Oh my....

[zinc cargo]

can you do a few with sesame on them?

[noble shoal]

I will....later

Preparing training images at the moment

[vital ermine]

Hulk not happy be in plastic wrap. He no like shrink.

[zinc cargo]

deep fried hulk @noble shoal

[vital ermine]

yes

[lusty wolf]

The nose did not grow yet... Pinocchio?

[upbeat summit]

it seems to work pretty great 😄 very cool

[vital ermine]

Hulk is a hard one for any of the trainings I do as he is so strong that when it works it has spread out all over the image

Might be the prompting but he is tough for me

[native knot]

Thanos is strong as well and is good for similar purposes.

[vital ermine]

In 2.1 I would use 1967 Corvette and that was the toughest one I could find. If it worked on it then you got it.

[native knot]

Nice

[vital ermine]

There are just somethings that are so overtrained lora just doesn't want to work on them.

[half cedar]

Abraham lincoln

[vital ermine]

locon might but they were pretty forget it. ia3 works on any then comfy comes along and kills it.

[zinc cargo]

Aeather Glitch meet upcoming Modren Oil Paint ( @icy brook @delicate kelp )

[zinc cargo]

who remember having this filter on photshop ? plastic warp...

[icy brook]

Me 🙂

Loved it

[icy brook]

Awesome 💕

[icy brook]

Would be awesome to see this with Glitch, @zinc cargo right?

[half cedar]

Trying to make Thanksgiving camouflage

[icy brook]

Are you making a plastic draped Lora? 🙂

[vital ermine]

V2 is about to be released

xl is fighting me though

I was told to ask you about something

A prime example. How do I keep the lora from spreading all over the gen?

[icy brook]

You want the plastic wrapping to only show on the subject/object in focus?

[vital ermine]

Yes, or other things I have done.

[icy brook]

Easiest to help would be to see part of the dataset. If you’d share in dm?

[vital ermine]

dataset I tried with and without a background

didn't matter

As I said it happens on anything I train

In the captions I will say "stand/sitting" in front of a solid white background.

BS wouldn't have anything to do with this, would it?

[uncut fiber]

old not sure i have not already posted it.

[zinc cargo]

@icy brook

[icy brook]

I will surely try and help. But I need to do some family stuff first. Will write later!

@vital ermine actually got some time now. Sure you don’t want to take it in dm? Feels easier

To discuss that way

[vital ermine]

DM

I am training it again at BS1 but 1h

[south horizon]

crispy

[half cedar]

deep-fried Colonel sanders

[half cedar]

I love stuff like this

Beautiful color

[vale eagle]

I love the left color and the right structure

[noble shoal]

Extremely challenging

[analog roost]

Hi guys, what's the best upscaler to use with SDXL images?

[native knot]

There is not one specific upscaler that serves all purposes. You should explore https://openmodeldb.info/ for various needs.

[ionic gulch]

next update will allow to select individual ip-adapter sources (up to 5) and revision sources (up to 2). that makes it much more flexible for regarding vram usage (and also performance):

[upbeat summit]

soul is absolutely right that there is not a single upscaler for all use cases. my current favorite upscaler model is this one - mostly for realism: https://openmodeldb.info/models/4x-classicalSR-DIV2K-s64w8-SwinIR-M

[native knot]

What is real, anyways?! 🤣

[upbeat summit]

the spoon sure isn't 😉

[native knot]

If you didn't tell me about the vase, would I have broken it?

[upbeat summit]

I would ask you to sit down, but you're not going to anyway

[native knot]

https://tenor.com/bOXcH.gif

[vital ermine]

At least the dog can breathe though

A bit sad on the cat

[zinc cargo]

@sinful falcon 😉

[shy kelp]

I installed automatic1111 to try to generate images with img 2 img, where the starting img is generated by bing. All A1111 is doing is generating copies of the same image

[native knot]

What is your denoising strength set to?

[eternal fog]

Sounds like you have your denoise value too low

[shy kelp]

0.75

[eternal fog]

Something is wrong then, that should change it a lot

[native knot]

Show a screenshot of your settings.

[shy kelp]

cfg 10

[native knot]

With a screenshot of everything, we could see better what may be going on.

[native knot]

Well, that shows denoise at 0.4 with a slightly different image in the results versus the input, which would make sense.

[brave stream]

Have you guys ever seen this kind of formatting in an SDXL prompt? Is it legit? "(cyberpunk)"

[native knot]

Parentheses adds weight.

[brave stream]

sorry, i need to repost that

\(cyberpunk\)

he it is with more context by yoneyama mai, 1girl, lucy \(cyberpunk\), solo, parted lips,

[shy kelp]

I changed the denoise after many attempts

[native knot]

probably someone bringing a prompt over from midjourney or something

[native knot]

Yeah, but what does output look like once you have something like 0.75 or 0.8 set?

[shy kelp]

all the 8 images seem to be exactly the same

87s/it

[wet nacelle]

Ignore the porn watermark.

[native knot]

from the 2nd screenshot you gave, it's zoomed out a bit, but it looks like there are some differences. Lookinig at the bottom-right wizard that is in the dark part, the area that looks like the head is light color in the top-right corner, but darker in the one below. The bottom-center wizard holding the wand has different glows around his wand and body. The center ring pattern seems different as well between the various images. Some have a narrow dark ring like the bottom-center, some have a wide dark ring like the top-left.

For a 0.4 strength, that would make sense.

[shy kelp]

finished. All the images have the same elements and lights. The difference is the texture of the sand ad the rocks

[native knot]

It certainly is changing.

[shy kelp]

When I try to create text to image what I get is a desert landscape with some bright lights and maybe one wizard

[native knot]

Sure...because your prompt is pretty complex and is not likely to come out the way you have typed it up. The more specific and more elements you are trying to make happen in an image, the less success you're going to have. It's possible to get what you want without additional trickery, but it's exponentially difficult after a point.

[native knot]

AI image creation isn't following instructions the way a human would. It's converting terminology in your prompts into tokens that it's been trained on to understand patterns and then attempts to infer an image out of latent noise based on the patterns that token training has understood. If you desire something a bit less specific, you'll have a higher probability that your expectations will be met. Otherwise, you have to start looking into additional techniques such as ControlNet and LoRAs to get things narrowed down more.

[shy kelp]

a crater in an arid mesa

it looks like a photo taken of a hole on the ground

[half cedar]

Aerial shot, crater in an arid mesa, sand dunes, pock marks, haze, desert flora, dry

[shy kelp]

so img2img is not controlnet?

[half cedar]

That was all prompt + illustration style

[ionic gulch]

img2img just moves the starting step up, with your image as starting point, while control net can be active through all steps if you want.

[shy kelp]

now that the discord bot misses controlnet I have to run offline SD to get what I want

[ionic gulch]

if you fine-tune the starting step / strength of img2img you can also get good results. for img2img a few steps often make a huge difference, so you really have to try to find the best setting to get what you want.

[half cedar]

This is the aftermath lol

Go lower plane damnit

(Close up, pov:1.3), enraged cyberpunk wizard controlling massive horizontal red beam, red electric arc, red light flash, arid mesa crater, sand dunes, pock marks, haze, desert flora, dry

[native knot]

2nd image

[shy kelp]

it wasn't installed correctly

[half cedar]

flying debris

[shy kelp]

is it a huge task to understand "five wizards casting energy bolts, in a circular formation" ?

[native knot]

You'll find that sometimes getting the number right is an issue.

Honestly, it's going to be a lot of trial and error depending upon how selective you may be.

[pure crystal]

for the thing you want to do you'll need controlnet. think of describing how to fill in the shapes of the image vs telling a narrative

[ionic gulch]

source:

[eternal fog]

I tried it with bing with this "Make me an image of 5 wizards stood in a circle casting energy beams towards a central point, make the image be taken from above looking down."

It's not got the number very well, but it's go the right idea.

Maybe try a similar prompt in SDXL

[wet nacelle]

Why the "Make me" part?

Are you not using https://www.bing.com/create ?

[eternal fog]

I did it from Co-PIlot

[wet nacelle]

ah okay

[eternal fog]

so you have to tell it you want an image

[wet nacelle]

makes sense

[ionic gulch]

(ip adapter, canny, sdxl styler)

[ionic gulch]

https://i.imgur.com/1AXFvOh.png

[ionic gulch]

https://i.imgur.com/PJM7iRT.png

https://i.imgur.com/O23dtaG.png

[ionic gulch]

https://i.imgur.com/mDBk0r2.png

[pure crystal]

old gen, I think it's a lucky seed

[ionic gulch]

https://i.imgur.com/W79wQTf.png

https://i.imgur.com/LYq25cv.png

https://i.imgur.com/m9pZotM.png

[sinful falcon]

These are crazy cool

[ionic gulch]

all based on basically two source images:

the rest is prompts and fine-tuning of settings for canny and ip adapter

[sinful falcon]

Nice - I just saw your workflow pic way up above. Fun to see such unique outputs 👍

[ionic gulch]

this combination is crazy. you can do so many things with it. and about every second run gives a great image. i'm not even posting everything, just if something is different enough.

for example i didn't post that one - if not all of them were so good i would surely have posted it:

or akira:

also have about 20 very nice horror houses i didn't post

for example:

[icy brook]

https://tenor.com/view/dancing-gif-4863485027267385981

[icy brook]

Well done!

[icy brook]

new lora?

[noble shoal]

Yes, we have talked 2 days ago about it. I asked if there is already a "deep fried" lora. I was just for fun, but i took the opportunity to make on. 🤣

[noble shoal]

Thank you.

When you make your Aether LoRas, are you also mis-caption the images on purpose?

[icy brook]

I forget a lot. Train all the time and rest of it other work and family stuff.

[icy brook]

Wdym? Exemplify

[noble shoal]

Ok, let's say i want to train the concept "made of xyz". Then i take a real image of xyz that looks slightly like another object and say. Photo of a rock made of xyz. Or i just lie. For instance i have a photo if chicken nuggets and caption "photo of pebbles made of {Instance prompt}

[icy brook]

I would just name it what it is.

[noble shoal]

Great, then i produce chicken nuggets. Straight up lying has made it lern very well. But! I didn't try the other way

Might as well try that. But how did you got you hands on shaped water objects?

[icy brook]

I create all images from scratch.

[icy brook]

Added you as friend and sent a dm 🙂

[ionic gulch]

can someone with the latest updates installed try sdxl zoe depth control net? it's broken for me and one other user at the moment. but as i did some other system updates, i'm not sure if it's a general problem.

[native knot]

Slow af, but my first time training a LoRA for SDXL and am doing it locally on my 3080. I pray for workable results. Probably won't even find out until tomorrow. Current speed is 20.44s/it. 😢

I'll figure out how to speed things up next time and I'm sure most of the fellas in here that train regularly know all the tricks to that already. But I'm excited to see how it ends up.

[noble shoal]

What Batch Size?

[native knot]

1

[noble shoal]

The fuck? I get 1.43s/it on a 4070

[native knot]

[noble shoal]

What optimizer are you running on?

[native knot]

Adafactor

[noble shoal]

You are already waiting nearly 12 hours? There must be something off.

[native knot]

I started it last night, yeah.

[noble shoal]

Oh, the concept of daytimes is not valid here. I have no idea what timezone you are in.

[native knot]

Eastern. Was confirming that I've been waiting for about 12hr already, yes.

[clever verge]

Sounds like you are using RAM as VRAM

[native knot]

Probably. My 3080 is only a 10G, not a 12G, so it might be swapping out to RAM if needed.

[clever verge]

Ah, yes. That's a challenge.

[native knot]

That it is.

[obsidian silo]

generate a photo

[native knot]

#1072220168534642768

[noble shoal]

Now i want a photo too. 😕

[wet nacelle]

Morning

[noble shoal]

Hello Sir, good day.

[wet nacelle]

Sleepy wake up face (just an image of me right now)

[noble shoal]

I am 100% sure nobody will put it through SD....

[wet nacelle]

Ha

[ionic gulch]

as requested:

[noble shoal]

I should work on my prediction accuracy

[tribal path]

Yeah I had 34s/it, but well, what do you want with a 4070 Ti with only 12GB🙃

[noble shoal]

🧐 The Ti is slower then the non Ti? How that?

[tribal path]

No clue, maybe you have a 24gb

Or is it just a 12gb?

[noble shoal]

Do you mean me with "he"? Now i am even more confused. I have a 4070 with 12gb.

[tribal path]

I edited, I thought I was talking to someone else.

[tribal path]

Training a Lora with only 12gb and getting 1.43s? How xD

[tribal path]

Hmm

[native knot]

Must be some other settings that differ.

[tribal path]

What parameters do you use?

[noble shoal]

A lot. What do you want to know

[tribal path]

Network Rank for example

[noble shoal]

I just drop you that. Might be simpler

From my last lora

https://tenor.com/view/rpx_syria-mic-drop-gif-19149907

But i don't say that's the best config!! Don't trust me. And don't judge i have a lot of things wrong.

[tribal path]

I am gonna try something out now

Curious what the speed will be

[noble shoal]

Let me know. Btw, i can send you my paypal details. Just ask

[tribal path]

Your Paypal details?😂

After buying a 4070 Ti I am running low, sorry bro xD

[tribal path]

😬

[native knot]

I feel your pain.

[tribal path]

yeah...

that's why I gotta work with Runpod, and it still fucks up most of the time 🥲

Practicing is a pain when you gotta pay to use a GPU

[fierce hollow]

the nightly versions of pytorch (2.1/2) use more vram when training loras in my experience at least

try to stick with 2.0 or what kohya recommends

the number of buckets also seems to matter

[noble shoal]

I am confused. Very confused.

[noble shoal]

I had a lora with 32 Buckets and still got 1.41s/it.

[fierce hollow]

I dunno, there was also a bug where the vae would not release vram after encoding/caching images

second run after they're cached to disk would be faster

[noble shoal]

Anyway, i love my cheap ass crappy plastic Palit 4070. And i will make a safe copy of my current Kohya and never update 🤣

My Kohya was last updated on the 10th of August.

[fierce hollow]

I updated once and couldn't do batch size 10 (sd 1.5) anymore, max is 5 now

boggles the mind

[noble shoal]

Never change a running system

[fierce hollow]

as for sdxl I checked just now and I'm using 9.6gb of vram with batch size 1, which seems about normal, not sure how you manage below 8

[noble shoal]

Me? No i am using 11.7

[fierce hollow]

oh I looked at this but didn't realize it was finished already

[noble shoal]

Sorry for the confusion. Thats just my normal SD stuff loaded

[fierce hollow]

my fault, anyway I'm getting 2.25s/it on a 3060 which seems about normal too, will probably get a bit faster once the progress calc catches up

40s/it on a 3080 is quite slow but I guess it's because it's offloading to normal ram

I'm using 9.6gb/12gb vram without text encoder so likely just some differences in settings

[noble shoal]

The 40.78 is on a 4070 Ti, thats what grinds my gears

[fierce hollow]

right, 4070ti is 12gb too

likely fp32 or text encoder killing it

[noble shoal]

I think he even used my config file

[fierce hollow]

you're training both unet and text encoder looking at the config but I'm not sure how much vram that adds

[noble shoal]

Is there an advantage of only training the unet?

[fierce hollow]

a lot apparently, I'm ooming with text encoder too

mostly just to save vram

usually the result will be better with both trained, but sdxl has 2 text encoders so it's hard to say, I think kohya recommended training only unet

--network_train_unet_only option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.
or so the readme says

[noble shoal]

I might give that a shot. I just have to set the learningrate to 0, or to i have to check something?

[fierce hollow]

you only have to add the flag above ^

lr will be used for unet only in that case

[native knot]

After this current epoch is done, I'm going to stop it, blow away my kohya, reinstall it fresh, and make a few adjustments to my config to see if I can't get the speed up.

[fierce hollow]

with unet only you can also use --cache_text_encoder_outputs which should speed it up if you have captions

[noble shoal]

Very interesting. I am interested in speeding things up. Thank you

[rain pagoda]

giant cron

[noble shoal]

How did you called me? #1072220168534642768

[fierce hollow]

I don't really have any neat config file for sdxl but you can also do full fp16 training with the madebyollin vae which fixes nulls in latents

and sdpa is likely faster than xformers on torch 2.x but it might depend on card

[noble shoal]

I am giving the text encoder thing now a try. Let me check what stuff i potentially could train.

[fierce hollow]

there shouldn't really be that much noticeable difference in output imo

eyes tend to get worse somehow

[ionic gulch]

Workflow V28.0 now on Github - redorderd the menu nodes, to allow better access on small screens and 21:9 displays.
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[zinc cargo]

are you sharing that shnitzel?

[noble shoal]

Still a bit of work to do on that. I have 3 "unfinished" loras in the pipeline. Picking one of them now to work on, but sadly it isn't "SDXL_Breaded_and_Fried_V1"

[zinc cargo]

https://tenor.com/view/的9-gif-27299608

[indigo vine]

I explored lora upscale this afternoon and my outcome is that it's really too inconsistent...Sometimes with 0.7 denoise, it works great, sometimes lot of duplications, variable "best upscale method". I also tried "SD-Latent-Upscaler" custom nodes, it's v1 and xl in those xy plot, I don't see any advantages even at low denoising..

in this case, all work well

in this one, lot of duplications. Both 0.7

[shy kelp]

is the latest model SDXL?
any other reccomendations?

[native knot]

After clean install, adding --network_train_unet_only, moving learning rates from 0.0003 to 0.0004, changing scheduler from constant to cosine_with_restarts, & network from 256/1 to 32/8. I've dropped from about 40s/it to about 11s/it.

[fierce hollow]

interesting, are you still using shared ram?

[native knot]

Sadly, yes.

[fierce hollow]

that's a whole lot

[noble shoal]

With unet only I am now at 1.07 s/it using 8.7 gb vram.

[fierce hollow]

I have a feeling you're not training in fp16/bf16, since that's about the double you'd expect

[native knot]

It's set for bf16

[fierce hollow]

torch 2.1/2.2 maybe? I was getting crazy high vram usage with that

no other ideas

[native knot]

Nope...new install put 2.0 in.

BRO REPORTED

[noble shoal]

It depends on what you mean with "tuning".

[noble shoal]

Wait, you have 100% Utitlization?

[native knot]

It bounces up and down just a tiny bit, fluxxuating between 95-100% most of the time, yesh.

Wow...the spammers are running rampant today.

[hollow cairn]

(I just realized this isn't the gen chat for SD lol)

[noble shoal]

Are you playing cyberpunk while training? Mine is between 0 and 5%. 👀

[native knot]

I'm not doing fuck all at this point other than chatting on discord and browsing some sites. But that is interesting.

[strong copper]

Micronomicon LORa )

[noble shoal]

Very interesting.

[native knot]

When you did your install, did you install CUDANN?

[noble shoal]

That was 2 months ago. I don't even know what I had for breakfast this morning.

[fierce hollow]

this is like the only stuff you need in env

mamba create -n lora python=3.11 pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt

plus your favorite flavor of bitsandbytes

python 3.10 works too if you prefer that (3.12 doesn't)

bitsandbytes-0.41.1-py3-none-win_amd64.whl probably works best

[vital ermine]

Doughnuts and Segal go well together

[noble shoal]

Found the answer: No.

[native knot]

hrmmm....

[vital ermine]

Seems that the bag is about to explode from all the hot air that comes out of that place.

My butt's been wiped.

[fierce hollow]

regarding torch 2.1 and cu121 (didn't realize it's stable channel now) I just made a new env and it seems to be working pretty well, no vram spikes

[vital ermine]

Yes, I noticed that yesterday

11.7 has been relegated to legacy

[fierce hollow]

yep only 11.8 and 12.1 now showing on the pytorch site 🥲

[vital ermine]

precisely

My jaw dropped yesterday but I am afraid to touch it due to prior issues from the programs not it

never as simple as just rolling back if it doesn't work

what card do you have?

[fierce hollow]

I nuked my comfy/ooba/kohya envs and upgraded all and everything seems fine for now

3060 12gb

[vital ermine]

Well, I am waiting on all the new PT for Ada based cards that are supposed to be here for 12.2

xformers 2.1 too

uses less vram and is faster but amp+ cards only

when that rolls out is when I will upgrade it all

[fierce hollow]

wonder how it will compare to sdpa/flash attention

[vital ermine]

xformers is already better and 2.1 is about 30% better overall

sdp was such a let down. 😦

[fierce hollow]

sdp is about even with xformers for me currently but that depends on card obv, flash attention is much faster but linux only so kind of a pain

[vital ermine]

I train more than I gen so I see it across all cards I have used from a T4 to a 4090. flash attention isn't even an option in my Linux to try.

I suspect the new xformers is using RT/TC but even without the speedup the less memory needed is direly needed for XL training

that alone wins over SDP as SDP does use more mem, or rather doesn't use as little as Xformer's saves.

XL really wants a 43 to 80GB card so getting it to train on 24GB is a chore. I can't even use prodigy for a dreambooth I have to use Adafactor at BS1.

[icy brook]

Can you guys use Bing Dalle now?

Seems to have stopped working on my end

[fierce hollow]

hmm wanted to see how xformers is doing with comfy now but the wheels don't seem to be updated for 2.1+cu121 yet

bing breaks like every other hour, nothing new

[vital ermine]

yep, that not being updated is why I am taking a hands off approach as I played that game a few times with this stuff and learned my lesson

[dusk cape]

Hey can anyone help... All my pictures look grainy.

Kaiju, monster, humanoid, giant, tentacles, ultra modern,futuristic, city, smoke, rain, digital illustration, fantasy, architecture, sharp focus, concept art, octane render, scary, 8 k

negative:
3d, blender, autocad, hills, human, people, man, woman, house, village, town, technical, architect, blueprint, temple, castle, mist, moon, modern, asphalt, garage. winter, floating, casino, apartment, out of focus, depth of field, lens blur

Euler a, 20 steps, 1024x1024

I don't get it

How is everyone getting these photo-quality images

[half cedar]

Prompt is great. Add photoreal, cinematic it also looks like maybe you need more refiner steps

And more steps. 30-40. Maybe a different sampler

[dusk cape]

I've tried both things, and still getting similar issues

Do you have an example prompt I could try, I am guessing I don't have something set up correctly

Do I need to download a VAE or something? I just downloaded the base 1.0 model

Adding photoreal, cinematic seems like it helped

[half cedar]

#✨｜sdxl message

There's a comfy workflow embedded in this and most of this users posts

[half cedar]

Yes the sdxl vae. Looks like you have it though tbh

35mm is another photoreal look

[dusk cape]

I just have the base model, with whatever standard vae is in it.

[half cedar]

You can try (photoreal:1.3) and put (blur,haze) in the negative

Cthulhu is in most of the base models (fyi)

[shy kelp]

What are your favorite controlNet models?

[dusk cape]

Yeah this is the quality I would expect. Still not sure why mine looks grainy. What's your prompt/settings?

It's looks all artifacted

[cyan crown]

I use Auto111

SDXL base + refiner

70 steps

[dusk cape]

Ah 70 steps. I guess SD XL needs a lot more steps to look reasonable, or else it causes these effects, maybe. I tried ~45, but I'll try 70

With 70 steps... better? but still weirdly grainy

[rustic garnet]

you don't need 70 steps. 30 are usually enough.

[cyan crown]

From a lot of tries 70 is a good value

[dusk cape]

I just downloaded the refiner model, and I think I've set that up now

[cyan crown]

cinematic photo by Mario Bava, monster, humanoid, giant, tentacles, ultra modern,futuristic, city, smoke, rain, digital illustration, fantasy, architecture, sharp focus, concept art, octane render, scary, 8 k 35mm photograph, film, bokeh, professional, 4k, highly detailed

neg : big hands, fake, fake hands, distorted, drawing, painting, crayon, sketch, impressionist

[dusk cape]

What seed, and other settings? I can try to reproduce the same image

[cyan crown]

DPM++2MKarras

70 steps

[dusk cape]

Actually, changing the image size and switching to DPM is helping a lot.

[cyan crown]

switch at 0,8

[cyan crown]

yes was the sampler

[cyan crown]

I use 70 steps, but ok

[dusk cape]

The changes I made from my original images were:

1. Using DPM2 a Karras, instead of Euler.
2. Changing the size from 1024x1024 to 1280x768
3. Using a refiner model

I think (2) made the biggest difference

I tried setting to 70, which had a minor improvement, but (2) was a huge jump

[cyan crown]

refiner also

[dusk cape]

Thanks, I think I'll avoid square images. The square images seem lower quality. I've heard size effects the results a lot, but I was surprised how much

Refiner model kinda is annoying because it has to keep unloading/reloading the models

I'll play with it and see how much it helps

nearly identical settings, but square:

if I ignore the weird human face and hands, the refiner does seem to be fixing the background grainyness

[dusk cape]

Various Sizes, everything else the same

[uncut fiber]

@dusk cape how about remove smoke?

[dusk cape]

In this case, I'm just copying the prompt andreac was using for reference

the original images I was getting had weird grainyness regardless of prompt

[uncut fiber]

o.k.

[cyan crown]

smoke was also in your original prompt

[dusk cape]

yeah but even in other prompts, I had the same grainyness

like, just random landscapes or whatever

[dusk cape]

Like, look at this

Image to Image

landscape, mountains and forests, sharp focus, concept art, octane render, scary, 8 k 35mm photograph, film, bokeh, professional, 4k, highly detailed

I wanted to make an image roughly looking like the one on the left... the thing on the right just looks grainy, and blurred out

[uncut fiber]

what denoise?

[dusk cape]

^ this is with like 0.6 denoise

[uncut fiber]

try 0,35

[cyan crown]

0..4

[dusk cape]

Alternatively: applying the changes above (canvas size, refiner, and sampling method)

helps a lot

[dusk cape]

at 0.4, the blockyness of the original image impacts the quality too much

[cyan crown]

ù

[dusk cape]

Example using my new settings. Pretty faithful upres

not 100% accurate to the original image, but pretty good

[uncut fiber]

with tilt shift

[torn breach]

How many parameters does have SDXL exactly?
And how much Dall-E 3.0 does?

[cyan crown]

lot of parameters

dall-e zero parameters

[pallid path]

https://media.discordapp.net/attachments/1054519199218794587/1160680490492235806/image.png

[uncut fiber]

is this image o.k. or issues as well, i have little bit bad vision

[pallid path]

https://media.discordapp.net/attachments/1054519199218794587/1160688836163145919/image.png

[uncut fiber]

tried to replicate but too faaar from target 🙂

[uncut fiber]

What do you do with images, that are on edge to delete them or keep them, but you decided to keep them? I think i found solution for me what to do. How about you @cyan crown and @vital ermine

[vital ermine]

I don't know as I have so many images requiring so much space.

[cyan crown]

convert them in jpg

[vital ermine]

that is a disgrace and not for me

[uncut fiber]

Those which i would delete most probably i will convert to bpg. Xnview can read it, so no problem. And very efficient format, probably there are better, but i like this.

[uncut fiber]

San Francisco?

[cyan crown]

the prompt is always the same

with different authors

[uncut fiber]

good night all, another computer nightshift 🙂

[cyan crown]

good night

[steady grove]

playing around with views from inside the car recently

[pure crystal]

(denis villeneuve cinematic lighting shadow, detailed depth:1.1), eye witness journalism photo, professional expired polaroid film photo, towering giraffe San Francisco Chinatown smelling fruit stand, narrow street, pretty shops and restaurants,foggy morning, cinematic close up fruit stand, (horror, photoreal), dynamic pose, dynamic background, dynamic composition, dynamic lighting, realistic proportions, extreme detailed, ultra detailed, intricate details, highly detailed atmosphere, highly detailed textures.

[static berry]

can someone help? im trying to train but it aint working

[pure crystal]

macroscopic, detailed leaf

[wet nacelle]

nice

[pure crystal]

nice, love this art style

[ivory kindle]

Nice. I'm trying to make views from the lateral windows but no results.
🙄

[pure crystal]

have an example image?

[ionic gulch]

https://i.imgur.com/K4BO1BQ.png

https://i.imgur.com/CiLyGPN.png

[pure crystal]

I would controlnet for this using a simple mask

(car interior, door handle)

[ivory kindle]

Well, speaking of handle. Dall-E is the only one that "understands" those requests. I'll give your advice a try. ✌️🙌

[ionic gulch]

https://i.imgur.com/mtvIR4M.png

https://i.imgur.com/Px7JBhk.png

[pure crystal]

perfect 24k gold diamond plated teeth works occasionally

[supple knot]

I just tried with dali3 , it refused 5 times

[pure crystal]

they wouldnt let me say 'chinatown' earlier.

[supple knot]

Just wanted to see if it could do this? Guard rails suck when you just wanna be creative

[pure crystal]

wouldnt let me do gold diamond plated teeth only prompt, but I was able to get some wild stuff through, my account might be flagged

[strong copper]

IP Adapters in action )

[peak dove]

A1111 Unstable Diffusers XL for a model

[uncut gull]

The Long Room at Trinity College, Dublin. SDXL knows about specific places.
What places does it know about???

[indigo vine]

The Hall of Mirrors, Palace of Versailles

[high skiff]

jesus CHIST SDXL is great for gore/horror using my LoRA

WARNING FLESH MONSTERS

This is so damn cool lmao

I never would have assume that my photo realism LoRA would bring horror to the next level

This one has to be the freakiest of them all tho lmao

[hardy cipher]

lol. damn. tree beast

[somber hill]

Huge Stable Diffusion XL (SDXL) Text Encoder DreamBooth training comparison

U-NET is always trained

PNG info (prompts) are in alts

All images are 1024x1024 so download full sizes

https://twitter.com/GozukaraFurkan/status/1711345323900100643

[wet nacelle]

Informative

[hardy cipher]

nice

[icy brook]

Aether Fire.

[native knot]

Nice composition

[strong copper]

)))

[wet nacelle]

https://tenor.com/view/jerma-jerma985-laughing-hysterically-chin-greedy-grinner-gif-19552901

[zinc cargo]

remember they were giving hyppos in kinder eggs?

[crude sparrow]

wabbit

[half cedar]

Sharing some dalle3 muppet battletoads

[zinc cargo]

toads?

had to img2img that nice one 🙂

[zinc cargo]

nice!

[steady grove]

maek your time hahaha

[steady grove]

unceunceunceunceunceunceunceunceunceunceunceunce

just in case people don't know, they should know. watch this, and now you'll know. https://www.youtube.com/watch?v=qItugh-fFgg

[half cedar]

muppet battletoads movie, cinestill cinematic, 35mm

[steady grove]

i love that meme. from a time before we called them memes. i think it preceeded "viral video" even. when the internet forum users all collectively realized we could edit photos really well

[zinc cargo]

sdxk cal get you nice stuff as well 🙂

[steady grove]

how far we've fallen. from weird "freaks and geeks" style photoediting to just complete waifu endless fantasies.

i blame filthy frank

[half cedar]

(Close up, POV, production still, detailed depth), Beautiful Evil queen, high cheek bones, raging tyrant posture, long jet black hair, tiny crown, furrowed brow, 35mm, high heeled boots, brutalist castle balcony view, purple red Hugo Boss bodysuit, high fashion,high collar cape, grazing natural light, early morning, pure shadows, analog photography

[lusty wolf]

Anyone have problems' with todays update of ComfyUI? Any KSampler fails... or is it just me?
Error occurred when executing KSamplerSDXLAdvanced:
'list' object has no attribute 'to'
File "A:\AI_Files\ComfyUI\execution.py", line 152, in recursive_execute

[south horizon]

Waiting on SDXL renders be like

[uncut fiber]

@lusty wolf it can be cause by different node. I got this when error on AIT

[lusty wolf]

Haven't installed anything new... but I will remove some nodes and see...

[icy brook]

guys, is it easy to transfer children's book characters and style with IP-Adapter?

[lusty wolf]

Thanks for your help. I deleted the newest folders of nodes, including MTB and Fooocus nodes. Now it works. Dammit... a waste of time...

[zinc cargo]

i couldnt do it yet

hit and miss

I'm looking for some help with inpaint dedicated workflow... if anyone got some 🙂

[lusty wolf]

@uncut fiber ComfyUI working again...

[lusty wolf]

Whatch this. Might be what you after? https://www.youtube.com/watch?v=ivi34PESgjU

[zinc cargo]

will watch in a sec. still `miring your albino xenomorph in the city

[zinc cargo]

is this video mute?

[lusty wolf]

Yeah unfortunately... follow the pictures... 😜 Find Dr. Lt on the AI Lounge server.

[half cedar]

Dua Peepsa 25 Pop star singer beautiful marshmallow creature, chocolate hair, round marshmallow head,tiny brown dot eyes, sugar Marshmallow covered, pastel-colored body, sugary texture, soft and squishy stature, candy-eyed gaze, bob cut, model pose, midriff shirt, contemporary, beautiful

[zinc cargo]

@half cedar this all bing dalle3?

[half cedar]

No - sdxl, I mark my Bing images

[cyan crown]

r2z2 all of these are your prompt with different Lora's or style

infinite possibilities

[half cedar]

I was trying to make pizza armor, couldn't get closer than this

[cyan crown]

let me try

[fossil patio]

Very nice! Base model?

[fossil patio]

?

[cyan crown]

the pizza armour

😂

[half cedar]

Lol

[uncut gull]

Here's a shirt, at least.

Success. You didn't say it had to be well-made armor.

[half cedar]

Ork in cheese pepperoni pizza armor

[half cedar]

Killer unbelievable gen, stuck in my head lol

[uncut gull]

On a platter bier.

[half cedar]

pizzasmiths today no longer know how to craft this

Berzerker Viking,wearing pizza cheese pepperoni armor,

[icy brook]

yo!

[half cedar]

DJ Crispy

[noble shoal]

And the audience is also cocking

[shy kelp]

im confused