✨｜sdxl

[strange mist]

gimme a simple prompt to test 😄

[indigo carbon]

you can tell by the way it diffuses, plus- when OpenAI where still somewhat sane the papers explained how pixel diffusion works

that's how Kandisky exists

they used those papers from when they were public

[indigo carbon]

nah, when it comes to things like realism, Dall-e doesn't even come close to SDXL.. again, pixel diffusion has it's own advantages and disadvantages over latent diffusion

It's way less efficient and, and much more "mechanical"

[vital ermine]

I think I have them as everything worked

[strange mist]

gopro action photo of a man swimming with his dog in rio negro costa rica, extreme wide angle, close-up, highly detailed, sharp focus, shallow dof, glinstering water, bokeh, hq, uhd, hdr, '#photography, man and dog swimming, ntional geographic, awardwinning
the prompt is taken from #🏅｜pantheon message

left is sdxl, right is dall-E 🙂

[vital ermine]

you say this can do more than just two images?

[eternal fog]

You can technically put in as many as you want

Just batch them together and put them into the IPAdapter

[indigo carbon]

SDXL images arent 300px^2

[strange mist]

they're 1024 x 1024 actually 😄

[vital ermine]

the stylistic prompt doesn't seem to do anything

[vital ermine]

this is cool, but the hard part is taming it

[ionic gulch]

first try would be something like "pixel-art, detailed cyber punk city, neon lights, 16bit"

[strange mist]

sdxl seems to be missing the man 🤔

[half cedar]

AI art will keep these pixels burning https://www.theguardian.com/music/2023/sep/30/u2-review-an-utterly-astonishing-admirably-raw-vegas-extravaganza https://www.theguardian.com/music/2023/sep/30/u2-review-an-utterly-astonishing-admirably-raw-vegas-extravaganza

[native knot]

I put all my latent space in this little box I have.

[strange mist]

sdxl seems to having an upper hand with respect to the 16 bit part, I think? 🤔
not sure

how does an actual 16 bit pixel art image look like?

[indigo carbon]

the lower the amount of pixels in the pixel art is, the higher the chance of dalle doing better imo

[ionic gulch]

i would think that the last one is not so far off

[indigo carbon]

I also have a box like that

I call it- "my GPU"

[strange mist]

seems like it

bit grainy. but I really like this one 😄

[ionic gulch]

pixel-art, detailed cyber punk city, neon lights, 32bit, 3d city

[woeful patio]

you can instantly tell it's not sd

really good

[strange mist]

it is. I think sdxl has a bit of catching up to do 🙂

[woeful patio]

hard to do when for it to catch up it needs a latent language model the size of gpt-4, I don't think they ever will sadly

[native knot]

Ah yes, the General Purpose (Storage) Unit

[indigo carbon]

this is what I get with SDXL, idk. SDXL is opensource, so it's still infinitely better

[indigo carbon]

that sounds like some FTB modded minecraft bullshit haha

[vital ermine]

Getty Images

[hoary saddle]

anyone good at making simple workflows with controlnet that might work well with user drawings?

results subpar

[hardy cipher]

use sketch

not actually sure what's available in that regard for sdxl

[eternal fog]

SDXL has Canny, Sketch or lineart

[hoary saddle]

lol

will try sketch cnet, maybe that's my issue

[eternal fog]

That looks like it's applied wrong

[hoary saddle]

lol

[hoary saddle]

most likely haha

[eternal fog]

Or are you trying to put the drawing back over the top of the image?

[hoary saddle]

nope, just trying to use the drawing as a fill in for the generation, i'm just an idiot 🙂

there's a 100% chance i'm doing it wrong

but my html/javascript/python web scripts are all working, so that's a plus LOL

[eternal fog]

Everyone likes scuffed shrek

[hoary saddle]

canny does a bit better than the other 3

still not what i'm looking for tho 😦

[hardy cipher]

might try using 1.5 with a sketch controlnet model, then using the results with sdxl canny

[eternal fog]

Try lower the Controlnet strength, it's above 1, which generally causes weird issues.

[hoary saddle]

definately better, lost the outline of the drawing quite a bit

[hardy cipher]

lol, just made you a new one. looks good though

[hoary saddle]

so close, yet so far away 🙂 will keep experimenting, appreciate the tips

may need to try the ipadapter stuff, dunno

[hardy cipher]

if you have any questions on that don't hesitate to ask. it's really wonky stuff

[eternal fog]

IPAdapter won't do what your wanting

[hoary saddle]

trying to replicate what SD Doodle does

https://clipdrop.co/stable-doodle

[woeful patio]

what is the easiest way of using two loras for two different people without inpainting in sdxl?

[hardy cipher]

really just need to find something that'll recognize what the sketch is going for

[hoary saddle]

to be fair, there's isnt' much better =p

[hardy cipher]

regional prompting

basically having each lora applied to only a certain portion of the image

[woeful patio]

thanks I'll look into it, really should be made easier by just saying the two people and sd just figuring out

[hardy cipher]

if you're using a1111 just look for a regional prompting extension. shouldn't be too complicated. the ai currently isn't very good at that sort of thing on it's own

so it just sort of mixes everything together in one big pot

[eternal fog]

Golden Octopus

[hardy cipher]

I don't think it wants to be friends

[eternal fog]

They never do 😢

Angry

This is making some horrible looking monsters lmao

[woeful patio]

crazy difference this is the same prompt, left in sdxl and right in dall-e 3

looks like dall e actually integrates everything into the background instead of slapping it on top like sdxl does

[dawn dune]

I'm curious if anyone knows of a service to train loras. Even if it's a paid service. When I look up how to do it and I get a 48 minute video, it's discouraging.

[woeful patio]

you can do it through google collab

using kohya

pretty simple

[hardy cipher]

I hate it, man. so hard to find quality info

[woeful patio]

I train loras all the time, for any questions I might be able to help

[hardy cipher]

there are quite a few options for paid services online though. colab is alright, but they aren't big on guis

[cyan crown]

SDXL needs better prompting, but delivers better quality

[dawn dune]

I have seen paid services that still require you to set up everything

Ideally, I would either be able to pull a few gits and run it for free, or cheap... or just send an individual like 80 pictures and they do it for me for money

[hardy cipher]

honestly, not vouching for the quality of the service, but there are optionso on runpod

[dawn dune]

@woeful patio do you run it locally or online?

[hardy cipher]

and they're the pretty simplish interfaces

[woeful patio]

I run it online through google collab, costs a couple of bucks

[dawn dune]

is there a consensus on where to run them

that's not bad

how bad is set up?

[woeful patio]

If you have images with .txt prompts you can send them to me and I'll make a lora for you

[hardy cipher]

there will never be a concensus on anything in this realm

[eternal fog]

You can make them locally if you have a decent enough PC. It's not that difficult to set up. Just need at least 10GB VRAM for SDXL

[dawn dune]

lol i mean if you ask everyone what the best way to run sdxl is, you will have a lot of people say a1111 or comfy, and thats enough to point me in the right direction

unfortunately i have a 3070

weird to feel like a 3060 would be an upgrade for this

[cyan crown]

i train lora on runpod

(have a 4070)

[dawn dune]

how bad is the runpod set up?

[cyan crown]

I made a quick guide

video guide

https://www.youtube.com/watch?v=1SWSj0-28GY

[hoary saddle]

damn bastard, i drew where to put the pumpkin lol, why add more around the outline hah

[dawn dune]

word, thank you for making this. this is the shortest video i have seen on the subject lol

[hardy cipher]

I only watch youtube videos if they don't have the clickbait creepy thumbnail

[cyan crown]

it's very basic.... not explaining all the parameters

[dawn dune]

looks simple

how many gpu hours does making a lora really take usually

[cyan crown]

just a quick startup with also a Json in the video description

[dawn dune]

on average

[hardy cipher]

you know which subject has quality youtube videos? auto repair. no frills. these dudes aren't trying to become influencers. their camera work isn't always top quality but they're very helpful

[eternal fog]

I've been able to make ok ones in about 30 - 45 minutes on a 3080

[cyan crown]

50 images 4090 and 256 network dim is 2hrs

[dawn dune]

so like a buck fifty

thats worth

[hardy cipher]

they charge by the hour

you could rent 8 h100s for 35 an hour

if you want 640gb of vram

and a terabyte of regular ram

[cyan crown]

the problem is downloading

[hardy cipher]

nah

[cyan crown]

it's slow

[hardy cipher]

you just send it to the runpodctl colab

using terminal

it's stupid you have to take that step, but it works

[cyan crown]

nut you need a google cloud account or just google drive?

[hardy cipher]

no need for a cloud account for that

it's a colab notebook that's setup to receive the files

and puts them on your google drive

need a drive of course

[cyan crown]

I'll try next time

[woeful patio]

i love how I bought 8gb 3070 a month before sdxl came out 😦

[hardy cipher]

I did the same with a fairly high end laptop. was on the fence with desktop or laptop

figured just get a laptop that can game, that'll be great

LOL

[dawn dune]

I got my 3070 at launch

felt super cool

now I cant do what I want to do with my mf gpu

[cyan crown]

I've 4070 12GB it's good for SDXL, but not for training

[dawn dune]

can it train?

[cyan crown]

too slow

[dawn dune]

i hear u need 12gb vram to train at all

word

well that is actually encouraging for me

cause i know for a fact i cant afford to do it locally now

[hardy cipher]

andreac: if you want to do what I'm talking about this is the colab notebook. it has the instructions, but super simple https://colab.research.google.com/drive/1ot8pODgystx1D6_zvsALDSvjACBF1cj6?usp=sharing

[cyan crown]

thanks

[hardy cipher]

need to do it through the web terminal I believe

[cyan crown]

no problem for terminal

[dawn dune]

thanks for the info everybody

[hardy cipher]

https://tenor.com/view/thumbs-up-monkey-thumbs-up-gif-gif-26768740

[half cedar]

Dog face or worry face?

[hoary saddle]

this may be my problem i'm starting to think, why does control net see a thick line as a double outline?

can clearly see the 2nd outline

[native knot]

Because Canny is edge detection. There are 2 edges on a thick line; the inside edge and the outside edge.

If you want to do what you're doing differently, fill in the white space around the eyes and mouth with black.

Then no edges will be detected until you get to the white eyes and inside of the mouth.

[pure crystal]

clean!!!!!!!!!!!!!!!!

[cyan crown]

thanks.. Try opening full in the browser

[pure crystal]

yeah these are super

[zinc cargo]

waka waka

[pure crystal]

the joker, (batman's nightmare, surreal bat phobia, by Tim Burton and Lucasfilm, voyeurism:1.2), (horror, coherent, intricate detail, victorian gotham city setting background, dawn light, natural light:1.3)

[pure crystal]

awesome

[pure crystal]

Steam Punk Lord, cybersteampunk

Steam Punk Lord Batman, cyber steampunk, style of comic book cover illustration, 1995, (POV:1.3), (batman's nightmare, surreal bat phobia, by Tim Burton and Lucasfilm, voyeurism:1.2), (horror, coherent, intricate detail, (victorian gotham city:1.3) setting background, dawn light, natural light:1.3)

[native knot]

Uncle Fester?

[pure crystal]

lol!

[native knot]

Fun result from IPA:

[pure crystal]

it's all there

bat parlor

(vermeer lighting), male (SteamPunk Lord Batman, cyber steampunk) planning, Bat Cave Parlor setting, style of Caravaggio, 1987, (POV:1.3), (batman's nightmare, surreal bat phobia, by Tim Burton and Lucasfilm, voyeurism:1.2), (horror, coherent, intricate detail, (victorian gotham city:1.3) setting background, dawn light, natural light, crystal clear, extreme detail

[pure crystal]

wow the photo look and materials in this are awesome

[vital ermine]

Thank you. Just playing around with some stuff.

[zinc cargo]

@eternal fog

i just love this

[vital ermine]

:/

[ionic gulch]

cloudy hair (and demo of the cropping settings (center, top, left, right + offset)

[icy brook]

haha. you're still at it 😄

[ionic gulch]

@icy brook mainly improved my workflow in the last few hours and still using the cloud style for tests :)

[ionic gulch]

source:
(inpainting+ip adapter)

[zinc cargo]

anyone here does upscales in comfy?

i mean for not generated images

[zinc cargo]

is this with @icy brook's lora?

[native knot]

No

[crisp owl]

"The Girl on Fire"

[icy brook]

but this is 🙂

[ionic gulch]

great video about using the ip adapter (by the developer of ip adapter plus): https://www.youtube.com/watch?v=7m9ZZFU3HWo
(shows img2img, inpainting, upscaling, different models, etc.)

[icy brook]

but only for comfy?

[ionic gulch]

@icy brook yes, comfy only. as he explains the features and usage of his node.

[pure crystal]

ouch, ouch, ouch, ouch

[icy brook]

Yeah 😀

[fiery plume]

dalle3 seems kinda really disappointing for most things. it's just good at the few flashy things like text and prompt accuracy. for a lot of prompt outputs though i even prefer dalle2

[glad grove]

like for example?

[fiery plume]

@glad grove dalle3 (left) seems to try to go more realistic/overly detailed by default instead of painterly/handdrawn of dalle2 (right). and no prompting or negative prompting seems to fix the styling.

i mean left still obviously isn't hyperrealistic, but it tries to still be fake realistic in terms of detail even when prompted for simplistic painting

[slender coral]

I know it's not sdxl but anyone know how to stop text in DALLE-3

[slender coral]

Digital art?

[eternal fog]

I've noticed a lot of these focus on SD 1.5 for some reason, even though SDXL is just better.

That Nerdy Rodent has done videos about it as well, but only SD 1.5

[ionic gulch]

the workflow is the same, so you can use everything he shows for sdxl. there's a sdxl section too, where he recommends the vit-h model

[fiery plume]

i tried prompting for 'digital art' as well @slender coral

[eternal fog]

Oh yeah I know how it all works, just an observation.

Unfortunately his nodes use more VRAM for me than the alternative nodes. Even though he claims his use less on his page.

And that noise option seems to wreck the outputs with the SDXL Plus model. It did make them better on the regular IPAdapter Model.

[slender coral]

I can't stop getting texts of what it is 😄

[eternal fog]

SDXL does that too, I'm guessing your asking for a barrel of moonshine. So it's doing just that. Giving you a barrel, labelled that it's moonshine 🙂

[slender coral]

I usually don't get it this bad, and negatives usually fix it.

[spring fulcrum]

Are you changing the empty latent image size to SDXL sizes? SDXL will not look good at the smaller resolutions

[eternal fog]

Of course lol.

It's just his noise option with the IPAdapter Plus model. Must be down to how it works, but it pulls away from the subject when you use the noise option. With the regular SDXL Model it made them better.

[eternal fog]

But it's it using more VRAM that's why I don't use his nodes

IPAdapter Plus uses between 9.8 and 10.2 GB VRAM, on average the other IPAdapter nodes cross 10GB less often, so get faster average speeds when I only have 10GB VRAM

[static berry]

is there a way to make sdxl generate more images?

[eternal fog]

You mean more at once? Up the batch size.

[static berry]

or a way to make another variation of an existing image

ok

[ionic gulch]

what do you mean by wrecked? guess it depends on the subject and style.
here is the same seed with and without noise with the sdxl plus model:
left with noise:

[eternal fog]

I've found on SDXL Plus it changes the outputs too much compared to regular.

[eternal fog]

This is very good for variations https://github.com/chrisgoringe/cg-noise

Plug your empty latent into the Hijack node, then the hijack node into the ksampler And then on the other side of the ksampler plug that into unhijack and then the vae decode. It lets you do variations of a single render.

[static berry]

i am too stupid to figure out what any of those do

[eternal fog]

What do you mean. You just up the variation weight to get slightly different images.

Literally just a slider

[static berry]

is there a version for a1111

[eternal fog]

Doesn't that have a variation option built in? It's been a while since I used it.

[static berry]

not that i can find

[eternal fog]

Yeah there is https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#variations

[static berry]

all i found was thishttps://www.reddit.com/r/StableDiffusion/comments/125hosf/image_variations_support_added_to_automatic1111/

oh

[ionic gulch]

looks interesting. maybe i'll add that to my workflow instead of the base / base + refiner option. a second variation with reduced render time would be nice.

[static berry]

why is this store selling el caminos

[ionic gulch]

@eternal fog what do those unclip nodes in your ip adapter workflow (the one you posted earlier) do?

[pure crystal]

"We have El Caminos at home"

[static berry]

i think hes lost

[native knot]

Are you asking what unCLIP does in general or specifically about how he has them setup in his workflow?

[ionic gulch]

@native knot in general and in context with ip adapter.

[native knot]

It's pulling concepts from the given image(s) for use in variation of the new image.

[wet nacelle]

Does anyone here use 1111 still with XL?

I just went back to it for a 1.5 model that was built around 1111. I realized how much I liked 1111 because of how I can go from model to model in image2image.

[native knot]

I'd imagine use with IPA would simply assist IPA.

[eternal fog]

They can be used to try and adjust the weight of the images that are being put into the IPAdapter. As your feeding in multiple images into 1 IPAdapter node.

[native knot]

By adjusting the noise level, right?

[eternal fog]

It's basically a workaround for me not having enough VRAM to use multiple IPAdapter Nodes.

[eternal fog]

More it's adjusting the conditioning to suit those images

It's kind of like putting in a prompt similar to the IPAdapter images.

[native knot]

So I was right that it's being used as an assist to IPA.

[urban fjord]

I can't remember what fixed the issue.

[eternal fog]

In this case yeah. So if the IPAdapter focuses too much on 1 image, you can use that to try get it to go the other way

The alternative is buy a 4090 XD

Which i'm still tempted to do

[ionic gulch]

so unclip is creating text based extensions to your manual prompt based on the input images?

[eternal fog]

If you put a prompt in yes, but you don't have to use a prompt.

I can try do an example, give me a few minutes.

[ionic gulch]

is this similar to what this website does?
https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[eternal fog]

No because unclip doesn't directly create text prompts. It's pretty complicated and beyond what I understand, but it apparently gets "Concepts" from the images. So it's similar to IPAdapter in a way. But instead of modifying the model, it modifies the conditioning.

It doesn't always work how you'd expect however and I'm guessing you will get better results if you can spare the VRAM for multiple IPAdapters.

But I've put these 2 images into IPAdapter in a batch

And I run it without a text prompt

I get this

It's mostly just the robot dude. For whatever reason IPAdapter doesn't take much attention to the other image in this case.

If I give the none robot dude image a bunch of weight in the unCLIPNode you get this

So it's pushed the car part of the image a lot, however it's not taken to the style.

[native knot]

That's because it's not acting on the model, iirc.

Right?

[eternal fog]

Well the reason it decides to mainly get the robot dude I'm not sure. Just IPAdapter seems to prefer that image when they are put in a batch together.

When you add unCLIP weight to the old car game, It's looking at the concept, which is a car on a road with some buildings and trees and applying it like text conditioning.

[native knot]

Right. I don't think it can glean the "style" from the image, per se.

[eternal fog]

Yeah not in this case

[native knot]

Maybe if it was something really strong.

Like an oil painting or something.

[eternal fog]

But if you really wanted that then I can put just the car game though IPAdapter and unclip the robot dude and

[native knot]

That came through nice.

[eternal fog]

Text prompts are more powerful though

[pure crystal]

(do this one lol)

[eternal fog]

That's with "Male Cyborg in yellow armor standing in the street" as the prompt

[vital ermine]

How do you unclip just the robot dude?

[eternal fog]

Only feed the 1 image into the IPAdapter, but keep feeding the robot through the unclip node.

IPAdapter and unCLIPConditioning are completely separate so you can do what you want with either.

[ionic gulch]

yes, looks like it's more like detecting the content/objects, as the result is far from the pixelated video game image. guess i will have to try that out. maybe it's worth to add to the workflow.

[pure crystal]

kaizu in this style would look like Rampage World Tour

[eternal fog]

I think if you had 2 IPAdapters you might get a better result, I can try but it will take ages to render.

[native knot]

That's what I ran into....just forever s/it.

[eternal fog]

Yeah it uses like 12GB VRAM

[native knot]

Add cnet to that and it's even longer.

[eternal fog]

So it's like 2 minutes to render, instead of a few seconds

[native knot]

I need a 4090.

😄

[static berry]

then buy one

[slender coral]

For training what is the perfect amount of images?

[eternal fog]

Nvidia need to fuck off with the stupid price they are charging for them

[native knot]

Nah...gotta pay down some other stuff. Too many expenses.

[slender coral]

You

You'll pay it.

[eternal fog]

Actually I just tried this and it still favours the dude. It might be because the source model favours people more.

[native knot]

I put smart shades up in the house this year, need to pay those off. Almost halfway. Need new carpet next.

[native knot]

As most models do.

In fact, most models heavily favor women over men, so working to get some things done that take the shape of a woman without it actually being a woman can be incredibly tough on some models without a LoRA.

[slender coral]

You didn't need smart shades, you needed to generate images.

[native knot]

You underestimate how awesome smart shades are. I programmed everything up with Home Assistant so that when both my spouse and I are out of the bed for 10 minutes after 7am it turns off the sound machine and opens all the shades in the house to the precise place we want them. Then, at night, they just close up on their own as we wish while various lights kick on.

[static berry]

sell smart shades and buy 4090

4090 is more important than them

[slender coral]

I have smart shades, they are too slow 😄

I also don't trust any iOTs hooked up to my house but some people don'tmind.

[vital ermine]

Not sure how to do that since it wants a clip vison

[native knot]

Home Assistant is self-hosted.

[slender coral]

4090's are basically room heaters too, they can save you money in winter 😄

[ionic gulch]

two ip adapters in chained mode:
0.5/0.5 / 0.25/0.75 / 0.75/0.25 weight + 2 source images:

[eternal fog]

IPAdapter doesn't use CLIP Conditioning.

If you are using the workflow I put in, just take the 2nd image out of the image batch, then it will still be connected to the unCLIP but not the IP Adapter.

[eternal fog]

Try it without chaining and just sending a batch of all the images to 1 adapter. It doesn't actually seem to make that much difference.

[ionic gulch]

you can't set individual weights in batch mode

[vital ermine]

it did change for sure

[slender coral]

What would you lads reccomdend for the perfect amount of traning images?

[eternal fog]

I know, it depends entirely on input images. But even changing the weights didn't do much for my example images.

[eternal fog]

How long is a piece of string 🙂

I've made LoRAs with a single image that worked surprisingly well.

It all depends on the quality of your images really.

[vital ermine]

Depends

quality and style, or subject, or concept

[ionic gulch]

this is batch mode with 1.0 weight (as you have to add the 0.5+0.5 weights to get a similar result in batch mode):

[slender coral]

Object

Just in general 20-50?

[slender coral]

Was that because of the base checkpoint?

[vital ermine]

Subject/Object can be as low as 6

[eternal fog]

It was just the base SDXL Checkpoint

[vital ermine]

seen 1 before but 6-15 is good

for styles I like 20-30 but concepts I still don't get what that even is but heard 50-100

more is not always better, btw

[pure crystal]

Lord Batman

[ionic gulch]

50/50, 75/25, 25/75 (weights) - same seed, no text prompt

[eternal fog]

Nice. It does work well if you can afford the VRAM. I can do similar, but I tink it's a bit more work to get it right with only the 1.

[ionic gulch]

that's why i prefer chained mode, as it gives much more control than batch mode, even if the developer of the node recommends batch mode

[eternal fog]

Yeah, hopefully they can reduce the VRAM usage.

[ionic gulch]

how much vram do you have?

[eternal fog]

10GB

2 IPAdapters use about 12

[native knot]

And it shows.

[ionic gulch]

the latest update of ip adapter plus allows you to create pre-encoded images to save vram. you can save them and re-use them.

[native knot]

Generation slows to a crawl.

[eternal fog]

Would it actually save much VRAM though

It still needs to load them into VRAM to run them

And you'd still need two IPAdapters

[ionic gulch]

i don't know if the models are unloaded anyway. but in his video he said something about 4gb of vram you can save (don't know if it's in total or per image)

[eternal fog]

He says "Hopefully the Clip Vision Encoder is unloaded" I'll give it a test.
But his example he's using batch and not multiple adapters.

[ionic gulch]

yes, he doesn't like chained mode, but he might add weight option for batch mode

[eternal fog]

I wonder how you'd do that. I've seen him asking about stuff like that on the main IPAdapter github.

[ionic gulch]

imo results of both modes are very similar if you adjust the weights. as chained mode adds all weights you have to use lower values per image. sometimes i even got almost the same image from both modes, after adjusting the weights (only if both used the same weight in chained mode of course).

so if you have enough vram there's little reason to use batch mode

[slender coral]

Ideas on why this wouldn't be filling the full image? a racoon human , full character view , Epic background

[eternal fog]

Probably so it can get more of the "Epic Background" you've prompted into the image.

[ionic gulch]

as epic as unreal 1 on a pentium 1 cpu :)

[slender coral]

https://tenor.com/view/confused-meme-girl-uhmmm-what-now-gif-17875457500510529068

a racoon human , full character view , white background

[ionic gulch]

yes, mentioning a background will often result in zooming out from the main subject

[slender coral]

😂

[ionic gulch]

maybe add closeup as counter-measure or just run enough seeds to get some that are closer to the character.

[slender coral]

I'd like the full character.

[ionic gulch]

try a very low strength control net depth or canny with a character that fills out the whole image.

[slender coral]

Do you know where that would be?

[ionic gulch]

this should force what you want. just have to play around with strenght and end step

maybe in "operating mode". at least that's where i have control net in my workflow.

[slender coral]

Dont see it :/

I'll explore it thanks.

[ionic gulch]

or if there is no control net you could try img2img with a low img strength and an input image that has the position/size of the character you want. you would also have to find out the optimal strength, so the image is not visible in the result but strong enough to direct the position/size.

[static berry]

wtf

im not sleeping tonight

[static berry]

https://tenor.com/view/bleach-simpson-homer-gif-12946893

[glad grove]

hes a natty

[pure crystal]

hahaha

[glad grove]

100% natural full vegan just by going to the gym 2hrs a day

[glad grove]

i dont bite

[pure crystal]

SDXL models, what do you people use?

[static berry]

protovision

[pure crystal]

I've been using an older Dynavision, and Niji XL special edition. I have to get a few others, juggernaut xl is one

[glad grove]

idk what happened here

[pure crystal]

fed it after midnight by the looks of it

[static berry]

hes mutating

[ionic gulch]

here is an example how control net depth with low strength can force a charcter to full screen:

[glad grove]

greetings

[hoary saddle]

anyone know how to keep Comfy from adding numbers to the end of the filename in the save node?

tried a primitive but still tags the number at the end

[static berry]

what model

[ionic gulch]

use the save image node from mickey pack.

[hoary saddle]

thanks JPS

[glad grove]

the dall

[pure crystal]

many of these pockets have snacks

[static berry]

ok

[hoary saddle]

genious, ty sir!

[static berry]

if ur referring to dalle 3 are u sure thats allowed here

[glad grove]

lemme ask mr seagal

https://tenor.com/view/steven-segal-carrot-healthy-food-gif-14637111

[wet nacelle]

Brotha gotta eat

[wet nacelle]

https://www.youtube.com/watch?v=Cllle4Kyqoo

[glad grove]

he loves carrots

[pure crystal]

Solid Rat

[urban totem]

I just noticed there is no path defined for diffusers in comfyui's extra_model_paths.yaml file, only a checkpoints: line under a111:, is there a way to use diffusers (and is there a difference doing it)?

[slender coral]

After looking at Civ, I couldn't find any good UI/UX loras for 1.0, anyone know any?

[heady vale]

meant to be a ghost taking a shower

[crisp owl]

I got it guys, the hardest image to prompt for

[slender coral]

😄 what was the prompt?

[static berry]

take a guess

[slender coral]

small anime girl in white void

[crisp owl]

same prompt that gave me this, funny enough. Just added a different "style" to it.

(obfuscated the nsfw part)

[pure crystal]

This is very hard in SD pure prompting, specifically the composition

[heady vale]

almost

[crisp owl]

woops, kicked on my IPA process on accident with that same prompt 🤣

[vale karma]

电竞房

[twilit nexus]

how do i get 3/4 front view always on my characters?

[crisp owl]

prompting sometimes gets what you want.
Or openpose if you want a specific pose

[glad grove]

done

[twilit nexus]

example

[crisp owl]

that's not 3/4

[twilit nexus]

like you know in those side scroller games

its awlays little angled the face

its not totally side

[crisp owl]

3/4 pose if you're trying to prompt for that will be cutting off around the knees if followed exactly

otherwise, if you're looking for that you described, try prompting "in the style of X game" or "looking away" or something like that.

[hoary saddle]

finally! yay comfyUI and python/java, bed finally

[indigo vine]

t2i canny + IPAdapter

[spring fulcrum]

@soft zealot I really like the cruiser weight workflow... So far excellent results and thats without even getting into the controlnet or revision and ipadapter sections... I do see a lot of great notes on how to use the different nodes... However I was wondering if you have an in depth manual somewhere that I can read, I didn't see an actual manual when I downloaded the workflow? If not that's fine... When I wake up later I might have a few questions for you... I will try to keep questions to a minimum. Great work though. great results.

[crisp owl]

These double exposure-esque photo's have been fun trying to prompt for.
I'll leave it a mystery for the time being

and this was a unique outcome

[soft zealot]

Cheers ears. On a “manual” the short answer (especially as I’m on my phone ) is “ there is no manual and there are no plans for one. Let me rewind………..leaving aside the “advanced” features a key focus for me was simply to take a basic workflow and then (within the limits of ComfyUI) build it as a useable daily interface where everything need for basic image generation is logically(to my mind at least lol) laid out. I think for the more advanced bits I e mostly linked to source where I’ve borrowed them from. I don’t mind questions,just don’t ask one that’s answered in the notes already lol

oh and ask questions in the open rather than via DM as its all pretty generic and others may b eable to answer

[noble shoal]

Going to release my inkdrawing lora later this day. 😏

[upbeat summit]

oh that looks nice! I'm looking forward to trying it out

[noble shoal]

Thanks

[upbeat summit]

is the paper texture part of the fine-tuning?

[noble shoal]

It is, kinda. I have tried to get it as much as possible out of my scans, but it keeps showing in the generations. I dont mind it

[upbeat summit]

no I think it's great. looks realistic

[noble shoal]

For just showing pictures its totally fine. For print it would need a bit of postprocessing. But yeah, i am getting some awesome results imho

[upbeat summit]

yeah. I mean that should be easily removable with some blending 😉

looks very clean

[fathom pasture]

Hey guys, I'm trying to load SDXL in jupyter notebooks and I keep getting:
Encountered %!d(MISSING) file(s) that may not have been copied correctly on Windows:
unet/openvino_model.bin

See: git lfs help smudge for more details.

anyone know why

Im' using paperspace

[ionic gulch]

any math genius in the house? i need help with a crop/resize node i want to develop. guess i can figure it out myself somehow, but it will take hours, while a math wizard would probably solve the problem within minutes or even seconds.

here's the task: input is the current size of an image (width and height), a target width and height and an offset value for both axis. output should be a resized and cropped image that keeps the aspect ratio of the content without stretching.

problem is you have to consider different cases: target width and height could be smaller or larger than the current size (there could be even cases where one axis has a larger target size and the other one a smaller target size) and the offset value could be too large, so there has to be some correction for that case.

i need a way to calculate x,x2,y,y2 for the cropping.

[ionic gulch]

what i figured out is that the cropping box has to have the same aspect ratio as the target image. and i have to figure out if the image has to be cropped in width or in height.

[noble shoal]

PM'd you a "possible" solution

[upbeat summit]

btw. if you want to fix AITemplate after the latest ComfyUI updates and until the official AIT repo has been updated, this fork works for me: https://github.com/asagi4/AIT/tree/hacks

[west breach]

https://civitai.com/models/141592?modelVersionId=173573

New version of PixelWave just released 🙂

[upbeat summit]

congrats! it looks great 🙂

[native knot]

Downloading now so I can give it a run.

[icy brook]

Congrats!

[shy kelp]

I was wondering why there are official GitHub repositories for SD 1.x and SD 2.x with python scripts to run them, but none for for SDXL.
There are various wrappers and frontends which can work with SDXL models but no official reference implementation as in the previous versions, or so it seems.
I'm still not sure why that is.
That, and the different naming, make it seem like it's a different "category" than SD 1.x and SD 2.x, not simply the next version which supersedes SD 2.x.

Does anyone here have an answer to that question?

[wicked frigate]

reference repo for the model arch is https://github.com/Stability-AI/generative-models recommended inference core is https://github.com/comfyanonymous/ComfyUI

[native knot]

Damn...I am impressed that SDXL knows the term midsommar. I happened to toss it in to change the look of something and it went from this (left) to that (right).

[shy kelp]

Thanks. I'm familiar with that.

[peak dove]

Totally kewl

[quartz warren]

SDXL Promt
dream prompt: Airship (Zeppelin Aircraftcarrier:1.5), (stealth:1.1),(submarine:1.1),sharp,high detail,atmospheric lighting,(straight diagonal lines:1.6),vector art, style: Digital Art negative_prompt: chaotic,noise,blur,bloom,shiny,over saturated,strange,chaos,unsharp,hero,buildings,groups,text,watermark,sign,symbols,Alphabet,simple,comic,ropes,many,ships,Columns,dock,sketch,(abstract:0.4),real,
first is the SDXL image than Dall-E-3
1.

than dallE3 2.

and its not open source

[shy kelp]

can i have your workflow, it is amazing

[eternal fog]

There's a simple version of it here
#✨｜sdxl message

[peak dove]

SDXL A1111 using Base and Refiner from Huggingface. Also a few of my own textual inversions (not made using SDXL)

[vital ermine]

I keep lookingto see if SAI bought out Dall-E or if I am in the wrong discord. Nope, to both.

[noble shoal]

Shameless Self-advertising. Released Inkdrawing v1
https://civitai.com/models/154918/sdxl-inkdrawing

[uncut fiber]

going in sd Comfyui is pain now 😦

[stone fossil]

This will be my new lora, ive managed to obtain a database of 1400+ van Gogh images where about 60 need upscaling.
After I hand caption each and every one of them as useual I will pick out the best dump Prod vs Ada and release it. 🙂

[peak dove]

AUTOMATIC1111 SDXL using DynaVision XL as Base

[uncut fiber]

there is nice source @stone fossil
http://laion-aesthetic.datasette.io/laion-aesthetic-6pls/
have you checked it?

[stone fossil]

I think mine is fine dw. 🙂 thanks tho! ❤️

[uncut fiber]

o.k. i think this is what is included in SDXL probably. Very good to listing in it. <- if right 🙂

[stone fossil]

I got the name tags and the year tags.

Could not be better.

I will dirty blip them.

To get the tag(s) in there and then by hand go over 1 by 1 hehe.

Place the name in there of the official date and the name.

Fill in / remove what should not be there and then go.

But this time also seperating styles.

[noble shoal]

Shit i thought you wanted to write a detailed caption for 1600 images. Hell, i am exhausted after doing that for 80 images

[stone fossil]

xD

Yes I do.

[stone fossil]

But you will see that when you properly caption stuff it will also be more easty to implement stuff or replace stuff.

Say backdrops for example.

If you never define one in a caption its kinda not good.

[peak dove]

Batgogh!

[vale eagle]

https://scontent-hkt1-2.xx.fbcdn.net/v/t39.2365-6/10000000_210375442059452_321995130515097708_n.pdf?_nc_cat=105&ccb=1-7&_nc_sid=3c67a6&_nc_ohc=5acdOVeMx-cAX9ygXLj&_nc_ht=scontent-hkt1-2.xx&oh=00_AfBS2izeW8gtgd8NjqKzQTfDW79rwkKE_32iYFkg0Two5w&oe=651E7F0D Could we use 32 channels VAE for SDXL?

[ionic gulch]

got it. my crop&resize node is working with some help from aimingfail, bing and chat gpt and fixing all the bugs the AI made:

[stone fossil]

Noice ❤️

[urban fjord]

DE3 is impressive, yet there are stuff where SDXL does a better job. So I'm not afraid that Open image models will be completely outcompeted yet

[shy kelp]

how dalle 3 is so good, can sdxl compete?

[peak dove]

A1111 SDXL

[urban fjord]

DE3 struggles with variance, despite using 4 different seeds, you might end up with the exact same composition in all 4 images while SDXL will give more variation. Which apperently is a weakness with the type of text to image model that OpenAI is using here.

[vale eagle]

Have you see Emu from meta? https://scontent-hkt1-2.xx.fbcdn.net/v/t39.2365-6/10000000_210375442059452_321995130515097708_n.pdf?_nc_cat=105&ccb=1-7&_nc_sid=3c67a6&_nc_ohc=5acdOVeMx-cAX9ygXLj&_nc_ht=scontent-hkt1-2.xx&oh=00_AfBS2izeW8gtgd8NjqKzQTfDW79rwkKE_32iYFkg0Two5w&oe=651E7F0D

[urban fjord]

It's well known that small scale finetunes can outperform the base model. Any tl;dr about what's novel about this research?

Hmm, yeah I see that base isn't doing too well with Pikachu Van Gogh so a LoRA is probably helpful if you want it to look really good.

[vale eagle]

It introduce more channels VAE could reconstruct better result in small detail. For example, load an image and output it via 4 channels vae. It failed to recontruct on the small detail. Image 4 is the example from the emu paper.

[woeful patio]

dall-e 3 is having a rough time

[uncut fiber]

i am loyal 🙂

[woeful patio]

any idea why a1111 keeps being stuck on 97%

happens a lot

[urban fjord]

I noticed yesterday that the wait estimate for DE3 was a lie, just refresh and it will be ready a lot sooner.

Sounds like it is your VAE that is freezing up Marksouls.

[urban fjord]

Thanks.

I'm testing my pose controlnet and something went wrong with Pikachu.

[eternal fog]

no

I dislike this lmao

[icy brook]

nice paper texture 👍

would be nice with an image of a film noir detective smoking in his office with the window blinds half shut

[noble shoal]

Thank you, it was an accident. Nice idea with the detective. Let me try

[icy brook]

that's not an accident 🙂 if that's the case it was the right wrong choice

[noble shoal]

Close

[upbeat summit]

congrats to the release!

[noble shoal]

Thanks, now you can play with it

[ionic gulch]

new version of my workflow with the new crop/resize node used for img2img + inpainting and some other minor improvements (for example separate models for image adapter and revision, so you can mix revision and vit-h ip adapter models)
https://i.imgur.com/4DedNov.jpg
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

it's still fitting and is useable on my 42 oled. don't know how smaller displays will handle it. :)

[polar ravine]

Any eta for controlnet tile for xl so I can finally use adetailer?

[upbeat summit]

cool

[noble shoal]

So sad, i have my menu bar on the right side. Guess i have to search for another workflow then.

[eternal fog]

You just gave me an idea to try with IPAdapter

[eternal fog]

It works with coins too

[eternal fog]

Using people removes the stamp effect somewhat, but it keeps the crown lmao

[ionic gulch]

get a larger screen and you can place your menu bar anywhere :)
and it's still cheaper than a 4090.

[noble shoal]

I don't believe you. Going to get an ultrawide.

[eternal fog]

Current Ferrari F1 Team XD

[noble shoal]

Just figured out that my inkdrawing lora also supports color, if you ask for it....and its awesome (self-advertisement)

[eternal fog]

Not as good as the LoRA of course, but it makes a nice IP Adapter input

[noble shoal]

You're welcome 😅

[eternal fog]

I want to try use the LoRA with IPAdapter, but it's going to wreck my VRAM 😢

[native knot]

I'm sure you all remember that great basketball player Michael Joran

[indigo carbon]

DallE3

SDXL

idk, not only it's closed source; it also doesn't seem as good as SDXL

[eternal fog]

lmao

oh that's cool lol These two mixed

Even better when mixed with one of @noble shoal ink images

[cursive hearth]

Does anyone know how to generate a Minecraft character in high quality in SDXL?

[eternal fog]

Put it back through with a text prompt and you can even have someone model it

[noble shoal]

Now you only have to put it on a website, let it produce and sell it. Profit.

[eternal fog]

lol, I'm sure there's existing products like that. Maybe not on a turtleneck sweater though

[uncut fiber]

minecraft always worked out of the box

[icy brook]

Real nice image nevertheless!

[native knot]

Yeah...thought it was hilarious with the combination of the face that's wrong with the name.

[native knot]

no controlnet, no lora:

Not bad

[uncut fiber]

there is some stamp lora?

[upbeat summit]

short story comic strip made with SDXL Inkdrawing by @noble shoal
Model: Base SDXL 1.0

the full comic strip here: #🌠｜show-and-tell message

[eternal fog]

It's putting a stamp into IPAdapter

This is the source image

[uncut fiber]

aha o.k. thank you @eternal fog

i am angry on ComfyUI now, so i am in sdxl in A1111

SDXL do nice we call those edges small teeths.

[eternal fog]

I accidentally made a Bird Plane

[uncut fiber]

🙂

[native knot]

SNUSTICTE

😄

[eternal fog]

Exactly

(There's a bird on the source image XD)

[native knot]

Not using the ink stamp lora...just a prompt

[native knot]

Ah yes, Lara Corft.

🤣

[uncut fiber]

with no plastical surgery?

[eternal fog]

It gets so close

[native knot]

makes for great humor sometimes

[static berry]

i ask for saab 900 and i get a volvo

[uncut fiber]

like how edge moreless match

[eternal fog]

Armold

[native knot]

SECHHINWAGER

of course.

[dapper dragon]

https://tenor.com/view/terminator-sunglasses-shades-t800-arnold-schwarzenegger-gif-27448475

[eternal fog]

Imagine if it got it right though

[native knot]

That would be

[uncut fiber]

i know one thing weird results combine arnold with elephants

[lusty wolf]

Cake Backpack? 🤪

[uncut fiber]

i want 20 pieces, and got 100 pieces

[soft zealot]

thers shades of Jag in there as well

[noble shoal]

Raid. Shadow legends.

[native knot]

I figured cheetos would make for a good shoe

[uncut fiber]

he reminds me michael caine

out of opticaol glass

[uncut fiber]

@vital ermine you take my idea sooner then i got chance show something 😄

[vital ermine]

LOL

[eternal fog]

LMAO the text on this is unprompted

I have no idea why that word has appeared

[uncut fiber]

tried to play with DT in it, but not good result, but can be relaxing shaking with him

[indigo vine]

little vibe of Lancia 037

[static berry]

bmw m1 for you

[floral island]

dunno what i'm doing

hmmm.... lets see if we can add a bit more variation

heh. getting there

[pure crystal]

https://x.com/The_DailyAi/status/1708414197992706295?s=20

The Daily Ai (@The_DailyAi)

I can't even think of a title for this

Twitter · Oct 1, 2023, 9:31 AM

[floral island]

did you copy my prompt? 😮

oh yeah! this is what i wanted 😮

keep. generating. comrad.

[pure crystal]

lol

[native knot]

C3-P0-IM

[peak dove]

Very good for DKWYD!!!!!

[floral island]

dkwyd?

[peak dove]

Don't Know What You're Doing! 🙂

[floral island]

well... actually i know very damn well what i'm doing: abstract absurdism ADDCOM {2::__songmaster__|2::art by __artists__| } ADDROW {2::__songmaster__|2::art by __artists__| } ADDCOL {2::__songmaster__|2::art by __artists__| } ADDCOL {2::__songmaster__|2::art by __artists__| } ADDROW {2::__songmaster__|2::art by __artists__| }

except, i have TOTALLY no idea what the outcome will be

[peak dove]

Oh, I thought you mentioned that very pretty girl's face "dunno what I'm doing" 🙂

[peak dove]

Is this like X,Y prompting?

[frigid halo]

excel

[floral island]

that's for regional prompting

that's just cracking up my usual level of randomness with more levels or random entropy

just how i like it 🙂

[pure crystal]

it works so well for controlling composition and detail, and images not flying off the rails

[peak dove]

Entropy! You can get pills for that 😄

[floral island]

digital photography, Can I take another step? I've done everything I can. All the people that I see, they will never understand If I find a way to change, if I step into the light Then I'll never be the same and it all will fade to white
but at the same time just a single verse with the right medium, can make such nice images

[peak dove]

I just installed IPAdapter/ClipVision on my ComfyUI - gonna be busy tomorrow setting it all up ...

[peak dove]

Sounds like a song ...

[floral island]

i'm pretty sure it's from a song

it's in my songmaster file

ah, it's from the english version of bad apple

[peak dove]

Snoop Dog-Gogh

[stone fossil]

Sketched by.

[upbeat summit]

@noble shoal I posted some images on your release page: https://civitai.com/models/154918/sdxl-inkdrawing
Thanks for creating the LoRA! Very inspiring 🙂

[spring fulcrum]

I was wondering how to get the face detailer to work? I think its being bypassed at the moment. Does this have to do with the "dummy" nodes on the left?

[soft zealot]

remeber what I said abiout answers being in the notes

[upbeat summit]

@noble shoal forgot to add the title image - moment 😄

[spring fulcrum]

yep I missed that first switch... It was on 1... now its on 2.. thank you

[noble shoal]

Hope to see more comics from you in the future. The creation was just luck, because I found an old dutch comic book from the 60s at a flea market. Immediately fell in love with the style.

[noble shoal]

I did what? 😬

[upbeat summit]

great idea and the results look really good 🙂

[upbeat summit]

I<--- forgot to add a title image to the comic strip gallery post 😉 reuploading the images shortly

[noble shoal]

Ohhh. I thought I had that moment. That would be shit to not have a title page image.

[upbeat summit]

no, you're good 😄

[upbeat summit]

nice details

[vital ermine]

Thank you both

Prompting can go from Kindergarten level all the way to Ph.D level.

It is an art all unto itself

[pure crystal]

hahaha

[vital ermine]

Yep, I understand that.

The old school comics, yeah

[spring fulcrum]

I must have messed something up when I was setting this up... What is this supposed to connect to?

[soft zealot]

reopen m8ine from the v4.2 zip file and satart again.

Thats something you've added fiddling around

[pure crystal]

needs a ear bleeding

[stone fossil]

You can try for yourself once I release the lora lol.

[pure crystal]

lol

[spring fulcrum]

fixed.

[upbeat summit]

@noble shoal ok posted and I generated this title image for it 😉 civitai is still processing it so it will hopefully show on the page soon

[nocturne dove]

Seems to do a good job. It's easy to influence colors with an img2img workflow just by using a little less than 100% denoising.

[cedar tusk]

Is there any documentation on the new scheduler stuff in comfy? Like what to use sigmas for?

[native knot]

The galaxy is on Orion's belt.

[stone fossil]

Cats are awesome. 🙂

[native knot]

These are agents SD and XL here to see you.

[nimble heart]

what movie was that from again?

[native knot]

See above.

[heady forge]

I havent used sdxl since release week, what kind of workflow and models are people running now? on comfyUI atm

[stone fossil]

I stack proto as base then dream shaper as refiner.

[native knot]

I just had the best idea...need to make some adjustments, but damn.

[stone fossil]

CFG playing. 🙂

Yes it took the twin token quite good.

[half cedar]

https://twitter.com/AfricanHub_/status/1708404252496789795?t=6PRkT4t1Rc-A3e3la14wWQ&s=19 bing

African Hub (@AfricanHub_)

A short story

Likes

3315

Retweets

534

Twitter · Oct 1, 2023, 8:51 AM

[static berry]

aftermath

[pure crystal]

@oak ice (cycle rickshaw batmobile:1.3), Batman, victorian beijing, ( POV:1.3),(adult, night time, horror), 1983, (extremely detailed depth), style of Antoine Blanchard, highly detailed, style of comic book cover illustration, inspired by William Brodie

[static berry]

more fire

[native knot]

Ok...finally...

Have some greatness to share.

[native knot]

A series of 12