✨｜sdxl

[fluid sedge]

(essa joke)

[livid basin]

Yeah, but a lot of the cyberpunk scifi robots are still super tall and long body/legged from SDXL 1.0 bot

Like absurdly so\

[shy kelp]

I would like to see actual human hands and feet. Kinda tired of mutants.

[urban fjord]

Did you try short in your prompt?

[fluid sedge]

Jateliue D. Troof, I'd prefer my go-go-gadget extendo proportioned people to be abby (abdomeny) there's something very 90's wigger about it

lol\

[shy kelp]

lol....in 1.5 if I wanted a robot like the "robot with laser" from Aaron Beck I would probably get Diodotos' robot. 😄

[hearty ginkgo]

is this true white

[fluid sedge]

Has anyone figured out invokeAI's new node system?

It's not just a copy and paste of comfy... And I can't get it to spit out a d*** image

[shy kelp]

Nope sorry 😄

[fluid sedge]

damn* not dick*

[autumn forum]

f3f5f7 96%, rgb(243, 245, 247)

[boreal bough]

🤣

[undone bloom]

it's a visibly different shade in the top left

[shy kelp]

Well I took a look at comfy and it's seems unnecessarily complex. Even though I'm sure it's powerful. Sticking with auto1111 for now.

[boreal bough]

considering my low effort prompt engineering, I'm happy with results

[undone bloom]

when you try for true white the model will try to compensate by making the subject darker

[shy kelp]

Nice. But did you just tried to get SD 1.5 to make Aaron Beck's robot? lol

[hearty ginkgo]

i thought so to but i think its discord

[boreal bough]

sdxl

just did a quick image reverse enginner

[shy kelp]

lol looks like the Chinese trying copy an American stealth fighter.....and failing badly. 😄

[boreal bough]

it may be the robot we wanted, but he's the robot we needed

[urban fjord]

My attempt at a black background Lora, still not completely black, but a consistent 2,2,3 (Black square is the colour picker that got picked up in SS)

[shy kelp]

lol like in Iron Man 2 when Tony Stark showed the video of other countries trying to copy the Iron Man armor and failing miserably 😄

seriously though the images look great

[hearty ginkgo]

the blackest i got was this

[livid basin]

So something that confuses me

SDXL is still epsilon prediction

[paper phoenix]

thats p damn black

[livid basin]

But SD 2.1 already moved to v-prediction for their larger models, what gives

Why would SDXL move back to inferior epsilon prediction?

[autumn forum]

love it when i hit a tiled seed

[winter raptor]

I’ve noticed this happen, what is it that’s causing it?

[autumn forum]

99% black

[autumn forum]

no idea. would love to ask a dev when 1.0 comes out and test tiled seeds when it does.

[urban fjord]

I mostly see tiled seeds with short prompts. Just ignore it and generate another one.

Almost, #000001 one shade of blue away from perfection, let me increase weight of Lota

[boreal bough]

yeah no that's a success. anything at the final 1% can be counted as a success

[undone bloom]

do any of mine reach true black? just to see if zsnr can do it

[boreal bough]

tbf, true black can already be achieved by the base model - its only true white that's really hard

[urban fjord]

0.13% away as pure white is 765 values away from pure black.

[hearty ginkgo]

@boreal bough true white almsot base then refiner no lota or what ever the other things are called

[urban fjord]

I'll try pure white LoRA after seeing if I can reach #000000

[hearty ginkgo]

whats a lora?

[boreal bough]

getting closer

[shy kelp]

Close.....ish. 😄

[true wing]

I think your bot had an accident...

[shy kelp]

Still seems like the ghetto version though. 😄

Very Portal like in design

[paper phoenix]

think of it as dlc for a model. it adds certain expansions, like trained on a character or a style or object or place not already in the model.

[undone bloom]

it does seem that even for a zero snr model solid white is harder all else being equal. also love the logo it made lol

[hearty ginkgo]

can u use them inside comfy?

[paper phoenix]

yessiree

[livid basin]

So diffusers currently uses EulerDiscreteScheduler by default

[true wing]

yeah and one of the examples is how to use a lora

[livid basin]

What scheduler do people recommend for the diffusers library?

[paper phoenix]

https://comfyanonymous.github.io/ComfyUI_examples/lora/

[hearty ginkgo]

ok ill look at it

[autumn forum]

i DID IT

but the face takes up the whole image

[undone bloom]

inference sampler? i'd just use whatever is 2M SDE Karras, and override the scheduler to exponential if there's a way to do that. it will produce the most coherent results by far

[livid basin]

hmm

[shy kelp]

KITTAH!

[livid basin]

Thanks @undone bloom

Seems like that isn't an option in the diffusers library (sorry, my space bar is broken)

[paper phoenix]

theres a bit of a knack to using loras, as a rule, dont use them too strong (generally 1.0 is too strong) and dont use lots of loras at once as they tend to fight. not hard rules, but rules of thumb.

[livid basin]

KDPM2DiscreteScheduler?

[shy kelp]

I'll check out you guys later. Too tired. Peace out all!

[urban fjord]

Alright, training a white LoRA now and see if that works well.

[undone bloom]

I suspect it is DPMSolverSDEScheduler, not sure though

[boreal bough]

in sdxl all my loras now go brrrrr at 1.0 strength 😄

[livid basin]

This is the list

[boreal bough]

oh well

[shy kelp]

Much better!

[shy kelp]

Saving this one.

[urban fjord]

I wonder if training resolution matters on single coloured training data....

[undone bloom]

ugh. they have the SDE sampler but not the 2M one. the SDE one works fine but it is a lot slower

[livid basin]

One sec, looking at their documentation

[urban fjord]

LoRA training is done, so can I at least get #FEFEFE...

[hard fractal]

remains to be seen

a whole lotta papers claim a whole lotta things.

[boreal bough]

a robot holding a rifle, in the style of Aaron Beck, concept art, artstation, grey bulletgproof vest, batteries in body, kitbashing, nurgle

style prompt

in the style of Aaron Beck, concept art, artstation, large brushstrokes, grey bulletgproof vest, batteries in body, (kitbashing:1.3)

negative:

text, background, signature

with those prompts, you should be able to continue engineering your ideal result

[hard fractal]

and, certainly the applied team has found out that not every paper applies to XL

[boreal bough]

(dont question the nurgle - I know)

@hard fractal may I ask you something in private?

[shy kelp]

I'll give it a shot. 😄

[hearty ginkgo]

ill try it

[livid basin]

Fair

whelp, DPM 2 doesn't converge to the same result when I vary the steps

[urban fjord]

Strange result, I've not managed to get perfect white yet, but somehow a LoRA model (to the left) just trained on a completely white image for 20 steps managed to improve the image (in my view) by framing the cat better. (Base only)

[autumn forum]

have you used the word "true white" in the prompt?

[magic dawn]

Could just be the random seed

[urban fjord]

Same seed for both images.

[boreal bough]

...might be worth trying to prompt a 100% white image with Vit-H and see if there are any words biased towards it XD

"transparent" used to work well 1.5 - but in sdxl its been giving me mixed results

[urban fjord]

Getting close, 254,254,254 on the top part of the background. 99.6% white.

(Left Lora, right base)

[livid basin]

@undone bloom I got it working, thanks for your help

pipe.scheduler.config["use_karras_sigmas"]=True
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

Gives you the DPM++ 2M Karras behavior

[hearty ginkgo]

@boreal bough

[livid basin]

Exact same image after 50 steps as the default diffusers scheduler after 200 steps

[urban fjord]

The last bit of white seems to be the hardest.

[boreal bough]

dont forget that offset noise will mess up the last minimal part of the lora - unless you're training it with like 0.0000001

[urban fjord]

I think for most purposes 99.6% white should be good.

[boreal bough]

yep. its close enough that a "white" color picker, picks it up - so automated background removal works

also it shouldnt stand out on paper anymore

[hearty ginkgo]

the max batch size my pc will handle seems to be aroudn 10-12

[boreal bough]

for inference?

with how much vram?

[hearty ginkgo]

idk

all i know is it does all 12 at once like it does 1 it for the first image then 1 it for the second up to 12 images then goes and does 2 and so on

i think

[boreal bough]

#✨｜sdxl message

for the max batch info that no one needs XD

[hearty ginkgo]

whats that?

[boreal bough]

just a fun attempt to see how many images I can fit into a single batch, without running out of vram

[hearty ginkgo]

what gpu u have?

[boreal bough]

all infos are in the link

[hearty ginkgo]

so u got a 4090 which is 24gb

and the vram is faster so i uses less vram then for exmaple a 3090 which is also 24gb i have the 3070 tho which is 8

lol

its werid tho for example the 1080 ti is 12 gb of ram i think but my 3070 out numbers it with less vram

and i was thinking of a 4090 but idk i decided to just wait

[old thunder]

so a gpu with gddr6x use less vram than a gddr6 ?

[livid basin]

no

[hearty ginkgo]

in theory based of my testing yes becaase i have a 3070 and i also have another gpu which is a 1080 ti that i plug into my pc to use mutiple gpus at once and my 3070 excees in every catorgy using less vram then my 1080ti but also the 1080ti vram is G5X VRAM

[livid basin]

...

[old thunder]

cool , so what should one look for in a gpu , higher vram ,cuda , tensor ( for sd) ?

[hearty ginkgo]

Idfk I'm just talking from past tests I did I don't know why and can't tell you how I can only say from past test

[hearty ginkgo]

But I know one thing for a fact never get a k80 24gb

[old thunder]

difficult to buy a gpu , 840

840$ jut in import tax

[hearty ginkgo]

We're do u love?

Live*

[old thunder]

India

[hearty ginkgo]

Oh

[boreal bough]

for training, 3090,4090 can actually use less vram at higher batch sizes, if cudnn is properly implemented

but other than that, vram is vram

[hearty ginkgo]

Just not a k80

At batch 8 ur using 11gb of vram allrwady at 12 batches I do just barely 8 max

[queen ether]

Hey y'all ❤️ loading 0.9 into 1111 crashes my browser lol

[peak dove]

Is 1.0?

[peak dove]

I have vlad a1111 working but only using base sdxl - refiner sdxl crashes my setup

[quasi remnant]

auto1111 still does not have official SDXL support(yet), these models don't work out of the box, it's like the change to sd 2.0

[peak dove]

At startup with a1111 sdxl, use --medvram --backend diffusers

[queen ether]

I'll try that now

[spring fulcrum]

Check this out... New composition and lighting... a bit over processed by the upscaler but still pretty good... any advice or criticism is welcome.

[peak dove]

This by refiner or base sdxl?

[spring fulcrum]

it goes through the base then refiner then an upscaler

[peak dove]

OK, if it was me I'd soften that lightburst by her face and try for a more even backlit effect. Excellent piece of work all the same

[spring fulcrum]

I agree... but I have been having trouble with the wording. I'm still sort of new to this. Its really hard to think of how you want something to look in words. lol

[queen ether]

@peak dove running it with the backened argument made it so 1111 wouldn't load at all and with just medvram loaded it fails to load the XL model. I give up lol. I don't want to run an extension I'm just gonna wait for the full release and hope that works

[peak dove]

What version of a1111 are you running?

[queen ether]

I have the gitpull command so it's always up to date

[peak dove]

What version ... vladmandic?

[hearty ginkgo]

Do you upscale the image 2 timea?

[queen ether]

I don't know what that is 😀

[hearty ginkgo]

U upscale it two times?

[spring fulcrum]

no it seems to go through 2 separate versions of an upscaler with slightly different effects. I just pick the one I like the most of the two upscaled results... One is more smooth, one is less smooth... it works when I want different outcomes.

[hearty ginkgo]

Try the ultimate upscale node

[spring fulcrum]

where do I get that?

[peak dove]

OK, I have AUTOMATIC1111 and Vladmandic AUTOMATIC1111 - this vlad version works very well and sits alongside the other version

[boreal bough]

hence why I said "that no one needs". Unless you have a 24gb vram its a moot graph - since due to background optimization, more vram is used on low batch sizes to speed things up

[peak dove]

The vlad does excellent sdx

Sdxl

[queen ether]

I'll have to look in to it

[boreal bough]

was just a fun experiment

[peak dove]

https://youtu.be/xhbtNRnvwv0

[spring fulcrum]

If you post an image you have used with that ultimate upscaler node I can copy your setup..

[hearty ginkgo]

does that work

[peak dove]

https://youtu.be/eg54fIOHlBU

[hearty ginkgo]

@spring fulcrum

[peak dove]

D/load Vladmandic, then follow up installing sdxl in the Seb Kamph video

[spring fulcrum]

yes and no... I need to download that upscaler now... I forgot what folder to put it in. lol

[peak dove]

For what its worth, I upscale using Topaz GigaPixel ... yesterday I did about 500 files batch upscaled (from 1024x1024) to 6 x (6144x6144)

Now, if only Upscaler had a batch function ... !

[hearty ginkgo]

In the custom node folder I think

[spring fulcrum]

did that but now it looks like this

It cant find the upscaler

[hearty ginkgo]

Check the terminal is it loading successfully

[spring fulcrum]

Ok it looks like it was loading. I had to close out the terminal and relaunch

Can you screenshot your setup of nodes. I want to see if my upscaler node is different

[hearty ginkgo]

you might need to mess with the upscaler settings also change the steps down to 10 or 20 for the upscale if you want

[livid basin]

hey peeps, just to confirm, SDXL is still limited to 77 tokens?

[soft zealot]

FWIW Ive added a Random Prompts Node (from WAS suite) into my SDXL workflow.

One downside is that if you load an image generated by it into Comfy you will get a new random image.

Partly to compensate for that I've also added a section to right some basic info to a text file in a prompts directory(NB this currently generates a random prompt to text file even if Standard Prompting jumpered and Random prompt not used)

I think its creating some interesting imagery.

[autumn forum]

how dide you get those right angled wires

[soft zealot]

Yechnically I believe All SD models are limited its just that A1111 for example does some form of concatanation to make tokens from token blocks (I'm sure one of the mor eerudite members can explain better) allowing a virtual unlimited token length.

There is probably a way of doing something similar in ComfyUI

[livid basin]

Awesome, thanks. I'll probably just truncate, seems like if the prompt is that long the rest isn't worth keeping

I think that is what the bot is doing as well, since it warns you about prompts too long

[spring fulcrum]

Ok so I now get an upscaled image but I have no Idea if it is saving the images or where they get saved if it is saving them

[soft zealot]

its a scripyt I found.

Download to \ComfyUI\web\extensions and change from .txt to .js

0 is straight lines
1 is slightly curvy
2 is standard spaghetti

[autumn forum]

thanks!

[livid basin]

Most helpful dude of the hour

[spring fulcrum]

Is there a way to set a specific output path in ComfyUI?

[hearty ginkgo]

It saves in the output

Folder

[spring fulcrum]

ya but I want to make multiple folders for separate projects.

[visual glade]

--output-directory

[hearty ginkgo]

That's above my pay grade i got no idea

[soft zealot]

just prefix the output path witht he directory you want, If it doesnt exist it gets created

NB that does work in the STandard COmfy Image Save however I use the WAS Suite so I can create Time Dtaed Folders & filenames

[soft zealot]

Im also working on an overly complex method so that I can simply Add the output path as [Folder] or [File] .

Also wirting a text file alongside the image

Its close, still have a few bits to sort

[spring fulcrum]

I dont see anywhere on my setup that says output path though... How do I add that to the node

[soft zealot]

the default path is \ComfyUI\output

so in a standard Image save if you added the below it would create a folder in the default folder called MyProjectName and all files will be prefixed ComfyUI

[spring fulcrum]

oh so I can do that right in the filename_prefix area... Ok that is easy

[soft zealot]

aye its only seperated out in WAS Suite Image Save (which IMHO is much better as allows things like image type to save as amongst other things)

[spring fulcrum]

So is (WAS Suite Image Save) a separate github repo I would need to get if I want to use that setup?

[soft zealot]

yes, theres a lot of stuff in there but be warned, documentation is "sparse"

https://github.com/WASasquatch/was-node-suite-comfyui

[latent zodiac]

#✨｜sdxl i have recurated a Nightcafe list of SDXL artists to be easier to navigate, enjoy ! https://charming-strudel-a6a8b6.netlify.app/

[west breach]

has anyone else had weird outputs from inpainting with sdxl? It seems to be inpainting over the masked area, but the unmasked part of the image looks like it has noise added, when it shouldn't be touched at all

[soft zealot]

Ta

One to add to my bookmarks along with the one I already use

https://supagruen.github.io/StableDiffusion-CheatSheet/

[buoyant crow]

SDXL&WDXL same setting test.
The results may be better if the CFG is raised (the image's is 24) and the NP weighting is relatively stronger.
I think the color atmosphere is very good, although I think it is still a work in progress.

[spring fulcrum]

Ok well here is the first test with this setup, Thanks @hearty ginkgo ... And the Folder Structure is working great. Thanks @soft zealot

[livid basin]

Alright, I just launched a job to generate ~12000 SDXL 0.9 images, lets see how they turn out

[spring fulcrum]

I like this Ultimate Upscaler A lot more than what I was using before.

And a few quick Gundams just for kicks

[winter raptor]

holy long legs

[spring fulcrum]

Is that supposed to be John Wick?

[winter raptor]

Nah

Just a dude in a rainy cyberpunk street

[spring fulcrum]

All right folks... Have a good night... I'm out

[dusk mica]

Good night

[sweet wyvern]

i have a question :3

wait, i should fix my name

ok

they changed name from Stable Diffusion to Stable Foundation, shouldn't the XL model be called SFXL ? 😛

[spring fulcrum]

lol

[peak dove]

Isn't SFXL that arts and music festival held in Austin Texas every year? 😄

[serene ravine]

@spring fulcrum Hey Scorp I was going to ask, what is that visual/node scripting I see above? Love the visual stuff! Not in the manual AI game atm, need to design my next build first, but for me everything is visual (although a visual prompter doesn't yet exist) anyway was going to type before, but you said you're off now lol

[dense chasm]

ImageTalking Wav2Lip plus CodeFormer still can't compare with the D-ID render,it's a long way to catch up

[sharp robin]

Hey mods. I think bots down.

[soft zealot]

Is it because of @high skiff that SDXL seems to like dogs 🙂

[latent zodiac]

lol yes

[peak dove]

2.43it/s using RTX 2070 and vlad a1111 sdxl 0.9

Here's a treasure-trove of today's output

Anyone want a list of top quality prompts?

[pallid lotus]

barbei

that SDXL seems to like dogs

[tropic turret]

I’d be happy with a list of good positive & negative prompt keywords to use.

[peak dove]

https://github.com/roma-glushko/genart-prompts

[tropic turret]

Thank you. I’m not on my PC at the moment, but I’ll check it out when I am.

[bronze stump]

You used any optimizations? I got 8sec/it on vlad on 1080ti. Sure, should be slower than 2070,but still not this bad...

[peak dove]

I'm using --medvram and --backend diffuers

[bronze stump]

Did you tweak any other settings?

[peak dove]

Not sure which to tweak?

What would you recommend at all?

I get a warning GPU High Memory Utilization - mebbe that's what is slowing it?

[bronze stump]

i really have no idea lol 😅
i think i was missing xformers and the latest update - now i got 2s/it
Still cant load the refiner in vlad - works in comfy tho..

are you using the refiner?

[bronze stump]

the --medvram doesnt seem to do anything on linux..

[verbal frost]

Is it because of @high skiff that rev animated 1.2.2 draws porn when asked for a cat lol

[serene ravine]

Hey can I ask what sort of systems you run things on, what kind of speeds things are?

Been procrastinating my next build forever lol

[bronze stump]

ryzen7 5800x 1080ti, 1024x1024 2.5s/it

[serene ravine]

what kind of steps for better quality 100+? so 250s at 100it?

[bronze stump]

30it are plenty most of the time imo

[serene ravine]

And how big are these models? I've observed the AI game but not been part of it on my 970 lol

[naive goblet]

The poses and hands seem to be improving. With fine tuning it will be much better

[bronze stump]

pretty big lol

[serene ravine]

5.76gb i assume

[bronze stump]

yeah its only the refiner. i think with comfy i have 6gb of vram usage all the time

[serene ravine]

so bigger vram doesnt necessarily mean faster?

[true wing]

more VRAM should speed up things like larger resolutions and generating multiple images at once

[serene ravine]

ah yes see your point there. i suppose the question is 24gb 4090, or high end 48gb, or, as i keep doing, wait for the next level!

what about core memory, does that come into play with anything?

[soft zealot]

There are 2 elements.
Processing Power and VRAM

Processing power has more of a dierect impact in terms of speed of generation.

VRAM is the working space, the bigger the working space the better the processing power can work

[true wing]

I think a big consideration, especially if you're going to be doing more then just stablediffusion like local LLM things, is that you simply won't be able to fit models into your GPU at all if you don't have enough VRAM. Personally I think VRAM is king if you're just considering using AI models. By core memory I'm not sure if you mean the CPU cache or regular Ram but either way it should be a minor concern unless you're running an ML model off of the CPU, which is going to be slow no matter what.

[serene ravine]

well the system wont just be for AI, try to make an all rounder for programming, art, music. but AI is the future, I want to make something at least 5 years future proof!

[floral island]

all bots seem to have issues

[true wing]

Yeah the bots have been slow all day and now sometimes they don't seem to work at all. I think the influx of people looking for info on XL, and the push by Stability for more voting, has overwhelmed them. Just a guess though.

[true wing]

I'd compare the performance and just see if it's worth the trade off in VRAM. You definitely want as much VRAM as possible so if you're talking like a 10% hit for double the VRAM I'd do that.

Yeah I don't actually know, it's just a guess. They could also be messing with things on the back end, their primary concern is testing after all.

[floral island]

i bought a 3080 10gb a little more than a year back, and recently i upgraded to used 3090 just for vram

when you're doing a lot with AI stuff, having a lot of VRAM to spare is not even a luxury actually

[true wing]

Yeah I'm on a 3080 10GB and it can be a struggle and there's definitely models I just can't run at all.

[serene ravine]

wondering what the 5x nvidia series will be. it's a shame that you can't simply upgrade vram haha

[floral island]

when i started running LLM models it was OOM all the time, that's why i upgraded

[true wing]

Yeah I have to look for a lot of optimized and quantized stuff just to get it to run but you lose some quality there

[serene ravine]

im a bit of a perfectionist so highest quality is what matters to me

i been living in 4k for about almost 10 years now lol

[floral island]

the extra amount of parameters you'd be able to fit because of quantisation gigantically outweighs the loss in quality

[true wing]

I mean nobody really knows what's coming down the pipe so I don't know what to add. Personally my biggest concern would be not being able to run a model at all because I don't have enough VRAM. On the other hand most of the public models are going to be targetted at mid to high end users and 24GB could certainly be enough to run a lot of them.

[true wing]

Oh for sure, it's just a shame to have to run an inferior version at all.

[serene ravine]

yes i heard the 4090 is fantastic for most things (and cheaper than those 48gb cards!)

[floral island]

i think that 4060ti 16gb model will be pretty interesting for enthousiasts, but the pricing is just off...

[true wing]

I need to figure out how to get an AI to just print me money 🤣

[true wing]

I agree with that and I'd add I don't think I'd even bother upgrading to anything with less than 20GB VRAM with how things are advancing

[serene ravine]

i think things are advancing at a fast speed now, with AI AI is like CD->Napster

[floral island]

i'm one of those lucky SOB's that can use their company A100 if it's not in use, and when running a 65B parameter model, you're certainly not noticing "loss in quality"

[serene ravine]

well thats lucky hehe

[true wing]

Yeah that's what I've heard. But even quantized I just can't run some models which sucks. Oh well I'm sure by the time I'm ready to upgrade we'll be on like the 6xxx series and it will have insane VRAM

[floral island]

usually you're running out of memory for inference, i had the same issue on my 3080

i could load a 13b model, but around ~1200 tokens it just crapped out

you might want to try to do the reverse: use GGML (cpu) models, and offload to gpu, it's not as fast as GPTQ, but it won't crap out on you

[true wing]

Yep this is it. Like technically there's a some bigger models I can run.... if I only want them to have like two sentences of context

[true wing]

I've used CPU and it's slow as shit, haven't heard of offloading to the GPU though. I'll have to look into that, any good examples I can start with?

[floral island]

https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md

this is the guide to follow

[thick grove]

How fast is SDXL on cpu? (i don't have a compatible gpu)

[floral island]

you'll need to do an extra step for llama.ccp so that it can offload layers from cpu to gpu

[true wing]

Thanks, I didn't know llama.cpp utilized the GPU at all. When I first heard of it it was CPU only.

[floral island]

it IS cpu, but with the extra steps it can move layers to GPU

even when pushing all layers to gpu that way it's slower than GPTQ, but it's massively faster than GPU->CPU

[serene ravine]

is another case with VRAM, that (as i think i observe) people might use multiple models to create one output?

[true wing]

Oh yeah I understand that, it's just the reverse of when you offload layers to the CPU from the GPU, I just didn't realize the other way around was faster or that llama.cpp had it

[floral island]

i'm not sure you'd want to try to do that... i've tried a 1.5 model once on cpu... it's in the order of minutes for a single small image, not even decent quality

[floral island]

llama.ccp is optimized for cpu, so it doesn't take as big a hit because it's on cpu, gptq just craps the bed once it does

[true wing]

I don't see why not. Or you could run like an LLM for prompt ideas and then feed them into SDXL with them all in memory.

[floral island]

this is what i've been doing for a while now ^^

[serene ravine]

i know alot of users probably just 'make me an image with a model'. but technically we can do alot more by putting things together

[true wing]

For sure. Like if you got deep into comfy UI you could chain models together. Like start with a txt2img then a chain of img2img's and end with the refiner and/or an upscaler. Lots you can do if you have enough VRAM, unlike me.

[serene ravine]

and thats what i mean, creating something which is without limit

give it a year, i think 24gb will be below limit of things we want and ai is capable of

[true wing]

I mean like I said there are already models that exceed that. There's models that require 80GB VRAM or more, especially if you get into like training and stuff. That's why I was saying unless the price or performance difference is significant VRAM is king for AI

[winter raptor]

I’m on a 2060 max q 6gb vram

[serene ravine]

at this rate ill be waiting for the vtx 6090 with 128gb vram

[sharp robin]

What’s ur speeds?

[winter raptor]

I haven’t measured, but fairly sloth like

I’ll look next time and @ you

[sharp robin]

Ok thanks

[true wing]

Yeah so I just looked up the prices and it looks like 48gb is 4x the price of 24GB VRAM. In that case I'd just get the 24GB if it will do what you want for now and just plan to upgrade in a year or two.

[peak dove]

refiner crashes vlad

[serene ravine]

miss the days of cheap video cards!

[bronze stump]

i got it working with refiner:
before, it would always get out of memory warning with refiner, i enabled the following options:

• disable live preview
• enable attention slicing
• Move base model to CPU when using refiner

[true wing]

same. Honestly I bought a prebuilt (and customized it) for the first time in over a decade just because I couldn't even buy a 3080 for like a year. This was at the height of the crypto bubble, at least there's some availability now

[peak dove]

I just got sdxl refiner to work on vlad a1111

5.9 it/s

[bronze stump]

lol did you do the same things as me?

[true wing]

I literally bought the most barebones prebuilt that would give me a 3080 mostly for the 3080

[bronze stump]

wut?
1024?

[peak dove]

No, I just tried it - and this time it worked

[bronze stump]

2s/it is my best

[peak dove]

wow, getting 8.57 it/s

[bronze stump]

hoooowww?

[peak dove]

Dunno - it just "happened"

[bronze stump]

can you dm me your config.json?

[peak dove]

I am copying prompts from Issue 3 of MidJourney Magazine - these are base sdxl Prompt = rat salad days inspired by louise zhang mark briscoe arthur adams gerald brom michael cheval tony diterlizzi

[peak dove]

False Alarm - one picture output (rubbish quality) then crashed!

[bronze stump]

:(

[jolly creek]

I wonder when will A1111 release sdxl support. On developer branch there already is webui 1.5 version

[true wing]

My understanding is it's already out on the /dev branch and slated for release alongside XL 1.0

could be wrong though

[jolly creek]

Could be the case

[sharp robin]

There’s a sdxl branch that supports base and refiner

[peak dove]

can you provide a link at all?

[sharp robin]

cd stable-diffusion-webui
git switch sdxl
webui-user.bat``` install it in separate location just in case.

Or just git switch sdxl

[peak dove]

Thankyou

[jolly creek]

Yes saw that, but update officially has not been released yet

[sharp robin]

And switch out to that one

[sharp robin]

What do u mean?

[jolly creek]

webUI 1.5.0

But it does work, have not tried it yet tho

[rapid jasper]

They are waiting for the "go time", that not means it is not working now

[rustic surge]

where is sdxl1.0

[sharp robin]

This is the 1.5 Janis is talking about u can try that one tho I havnt tested it yet. The sdxl one does work. From my tests tho. https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/11862

[sharp robin]

Postponed for ab a week

[rustic surge]

No!!!

[rapid jasper]

Yes, u can try it in bots for now

[peak dove]

git switch sdxl - INVALID REFERENCE sdxl

[rustic surge]

Can SDXL be used for model training now? Does the webui support SDXL yet?

[sharp robin]

Idk how u got that error but u can try this one then #✨｜sdxl message

[rustic garnet]

kohya_ss supports SDXL, yes

[rustic surge]

I am still waiting for SDXL 1.0 to be open sourced.

[rapid jasper]

Like we all

[floral island]

https://tenor.com/view/tinder-my5n5-kermit-gif-24133413

[rustic surge]

Can we communicate in Chinese here?

[peak dove]

Hao bu hao?

[rustic surge]

Hao

[peak dove]

Due le

[rustic surge]

Ni ye shi zhong guo ren?

[peak dove]

Wo shi Ying Wen

[rustic surge]

ying wen?

[peak dove]

English

[rustic surge]

oh!

[peak dove]

Fa wen = French

etc

[rustic surge]

i know little english

[peak dove]

Ying Guo - England

[rustic surge]

Yes!!

[peak dove]

Ni shi Han?

[rustic surge]

han?

china

[floral island]

i'm pretty sure this channel/server is mainly english, so if the conversations grow long, i'd suggest taking it to dm

[rustic surge]

what is dm?

[peak dove]

Han - Han people. China has 7 ethnic groups, the Han People being the top of the tree so to speak.

[floral island]

direct-message, one on one chat

[peak dove]

Han - the box with the stroke top to bottom

[rustic surge]

yes ! i am han

you very know china

[peak dove]

Wo de mingze shi Jack. Ta de mingze?

[upbeat summit]

greetings

[rustic surge]

wo de ming zi shi

Chen Jie

[peak dove]

Xixxie

xiexie

[rustic surge]

shi ni de mingzi , No ta de ming zi

[peak dove]

Ni shi laoshi?

[rustic surge]

i am a chinese student

NO teacher

[peak dove]

Xuesheng

[rustic surge]

Yes Xue sheng

da xue sheng

big student

[peak dove]

My Chinese was learnt 40 years ago - taught to me by a Swedish woman who lived in Beijing for 30 years

[rustic surge]

Oh!

amazing

It's been so long!

[peak dove]

She used the Beijing 'dialect' - a lot of rrr rrr rrr sounds 🙂

[sharp robin]

Panda express Beijing beef sounds good

[rustic surge]

Beijing dialect , I not understand

i from GuangDong

[peak dove]

In Beijing (apparently) they speak with a pronounced rrr rrr rrr sound when they talk

[smoky patrol]

Nice to meet you, Senderai

[rustic surge]

Beijing is 2000 km away from here.

[peak dove]

And 10,000 km from here 🙂

[rustic surge]

Nice to meet you!

haha .I'm starting to get curious about rrr's accent, I haven't heard the Beijing dialect before.

[smoky patrol]

My Chinese name is Lin-Lin. 🙂

[rustic surge]

Hello Linlin

[rustic surge]

Very cute name.

[smoky patrol]

Haha, thank you. My students gave it to me

[sharp robin]

Why do Chinese have two names? Are both legally recognized or is one just a nickname?

[rustic surge]

Your student is china people?

cool

[smoky patrol]

I taught English in China

[smoky patrol]

Yes, many years ago

In Hubei provence

[rustic surge]

Chinese national ID cards can only have one name, but most Chinese people have two names. I don't understand why.

[rustic surge]

Hubei is 2000+ km away from here.

[peak dove]

I would love this name if I was Han - 走狗

[rustic surge]

what?

[peak dove]

Hunte Dog

[rustic surge]

Zou Gou is a very disrespectful name.

[peak dove]

Hunter Dog

[rustic surge]

What are you talking about?

[peak dove]

Yes, Mao usd it against supporters of the Imperial West

Capitalist Running Dogs

[smoky patrol]

I have been to Guangzhou, and to Beijing, and a little of Shanghai, but that is as much travel as I have gone

[rustic surge]

People who betray their country are indeed called "zougou", or "hanjian".

[rustic surge]

I go to university in Guangzhou.

[jolly creek]

https://tenor.com/view/brick-hill-off-topic-mopaiv-spacebuilder-alex-dunn-gif-19842686

[smoky patrol]

Guangzhou is beautiful; it was very fun to visit, and I had a very fun time

[rustic surge]

But I have never left Guangdong province.

[grizzled warren]

Would the Chinese also call an American traitor like that?

[smoky patrol]

There are a lot of beautiful places in China. But yes, let us return to topic

[floral island]

#🌶｜off-topic

[smoky patrol]

What has everyone been creating in SDXL today?

[rustic surge]

American traitor is ”ZouGou“ or” MaiGuozei“

[floral island]

things. not sure if good things yet. but things

[rustic surge]

”Han Jian“ is china traitor

[shy kelp]

Not much because the bots are failing all over the place, but I have to say the results from the different models do indeed feel significantly better than 0.9 so far

[floral island]

@smoky patrol the worst part is, that when you interact with the sdxl1.0 bots, sdxl0.9 just feels so... mediocre xD

[peak dove]

Today's SDXL Creations

[jolly creek]

Just myself waiting on sdxl 1.0

[smoky patrol]

I've been working on a list of prompts for you all

[rapid jasper]

yes, its definitely look better now

[boreal bough]

A man named StableD, about to retire, looking at an ai painting, a memory of the past

[rustic surge]

What improvements does 1.0 have compared to 0.9?

[smoky patrol]

AWWW, lovely rat babs.

[peak dove]

Wow - gorgeous!!!

[grizzled warren]

Zei? I'd think that would be zhe...

[rustic surge]

i like it！！

[west breach]

hi Sunny 🙂 I've also been working on prompts. Did you see the style prompts the bots are using were released today?

[rustic surge]

is zei/贼

[smoky patrol]

I do so hope you know how much this picture puts joy into my heart

[smoky patrol]

I do love the addition of the orange. Good contrast

[floral island]

double exposure is such a fun tag to work with

[smoky patrol]

You mean from last night?

[rustic surge]

Can I use ControlNet in SDXL?

[west breach]

For me it was this morning 😄

[smoky patrol]

N o p e

[peak dove]

Can I setup AUTOMATIC1111 to open at 1024x1024? It is a tad tiresome having to adjust 512x512 each time?!

[smoky patrol]

I've only been awake this past hour and a half

[peak dove]

Dropp ded gawjus!!! ❤️

[floral island]

this is a good prompt ^^ i like it

[smoky patrol]

What's your prompt?

[west breach]

They pinned them in this channel

[rustic surge]

How do I use SDXL on Discord?

[smoky patrol]

Have some unicorns on basketballs!

[boreal bough]

#1100170514670039070

[sharp robin]

Top right bottom left there is a button u click and it loads the last settings u used.

[halcyon tusk]

hey everyone! just a friendly reminder to please go to the bots and vote either A or B on the images, it'll help the final finetune of the model! your voting improves 😎

[sharp robin]

Like amgic

[floral island]

SDXL0.9:
g:a double exposure of a silhoutte made of elements, drawn by studio ghibli and carne griffiths l:abstract lineart

the workflow is inside the image (comfyui)

[grizzled warren]

There are bots on this server. Type "/dream" command in the bot channels and you'll see the options.

[rustic surge]

Ok,thank you

[smoky patrol]

I believe that that you speak of the second message, in which case, this be the one I meant!

[sharp robin]

Bots been having a tough time lately

[smoky patrol]

We're having a few hiccups at the moment

[peak dove]

Prompt = black and white photograph of an attractive sexy young woman styled by paul freeman. Euler A, 20 steps, cfg 10, seed -1, SDXL 0.9

[rustic surge]

it is slow

[smoky patrol]

Yeah, like I said, apologies, but we are having some issues right now

[grizzled warren]

Normally it takes about 30-40 seconds to get an image.

[floral island]

damn this prompt slays 😮 just press generate for lovely images 😮

[grizzled warren]

But it's a test, so the stability staff reconfigures the bot very often

[rustic surge]

Then I'll wait until SDXL 1.0 can run locally.

[floral island]

you take the time for all the issues you need! we'll all be here like drug addicts anyway 😄

[grizzled warren]

But that's okay, because the point is to give us the best SDXL 1.0 possible

[grizzled warren]

Nah, the bot is okay. Just try it 30 minutes later, chances are it will be fixed

[rustic surge]

hao

[smoky patrol]

Pppsh, ain't that the truth. Just with the ADHD, I got all the issues I could ever need, haha!

[grizzled warren]

It just wasn't made to compete with Midjourney, it was made to gather the community's feedback

[smoky patrol]

Those flowers are so good

[peak dove]

I am currently compiling a text file with many of the MJ prompts from Issue 3 of the MJ Magazine ...

[rustic surge]

In China, many people compare SDXL to Midjourney, but I personally don't think they are comparable.

[peak dove]

Those B&W women were one such prompt - the rat babies too

[floral island]

one is open and has a huge community, the other is closed, and also has a huge community

[smoky patrol]

In case you weren't hungry

[floral island]

we need those very tasty anime versions of those

[rustic surge]

i am hungry

i want to eat dinner

[smoky patrol]

You're not wrong

[peak dove]

I agree - I mix artists in MJ (and other generators), and MJ's results are so much more distinctive

[smoky patrol]

Well, so am I, haha!

[halcyon tusk]

is this your breakfast? 😋

[peak dove]

I take a $30 sub, make thousands of photos at MJ, then back off for 3 or 4 months ...

[floral island]

still the same prompt?! where does it even come up with this shit?! got a lot on my mind

[smoky patrol]

I am having a fruit cup at the moment, to be honest, haha! I shall probably have some yogurt, too

[sharp robin]

One is open and free the other is basically in a jail cell.

[rustic surge]

The only use Mj has for me is: generating training datasets for lora when I'm training it...

[smoky patrol]

Well, you see

[peak dove]

Using Eface's prompt

[rustic surge]

yes

[floral island]

damn 😮 this one is so good too

[smoky patrol]

Maybe I'll steal your prompt

[sharp robin]

The urge to ping @daring smelt

Ooops

[floral island]

our prompt! https://tenor.com/view/meme-our-now-gif-21036569

[sharp robin]

Ok goodnight

[smoky patrol]

I mean....

You don't wanna see what I was gonna make...

[sharp robin]

[west breach]

this is what i got for got a lot on my mind

[floral island]

actually, that's pretty damn cool too

[smoky patrol]

I love it

Let's see, hehehe....

[peak dove]

SDXL 0.9 #prompt4diversity

[smoky patrol]

Enjoy these words

[peak dove]

See my video how to use ChatGPT to generate prompt after prompt after prompt ... this one is for cars - you can change it out as you need (and please use the prompts you see here) ... https://youtu.be/VpTEdsspxbw

[peak dove]

Supercool

[smoky patrol]

Thanks!

[barren olive]

SDXL vs Anything V4

[boreal bough]

sheeeeeeesh. this one is hitting mad good levels

[barren olive]

Really like how SDXL improve on humanoid arts

Almost Midjourney level for free

[floral island]

but it's not even fine-tuned yet

[barren olive]

Has anyone created fine-tune/custom model sdxl yet

[floral island]

this is just 'the base model'

[soft zealot]

two versions. first one using the same prompt data as your original in Linguistic & Supporting and then inverted them (used the linquistic as supporting and supporting as lingustic plus kept the seed static) because I could 🙂

[floral island]

yeah, it's really cool switching stuff around at times 😄

[floral island]

woof!

[barren olive]

The detailed are amazing. Can't wait when someone make fine -tuned version of SDXL. Imagine it can even beat Midjourney

[floral island]

pretty sure it will

[barren olive]

What's your favorite Model in base SD?

[floral island]

my own model mix, darkalfa v1.1 🙂

[barren olive]

Do u plan making it to sdxl? Also is your model available to download?

[peak dove]

Deliberate V2,seekARTmega,DreamlikDiffuson 1.0

[floral island]

https://civitai.com/models/64843/dark-alfa

[smoky patrol]

Love the right

[floral island]

it's only a mix, so it's not trained by me. but the first model was a drunk accident, so i kept mixing with that knowledge and got a model which has pretty damn good comprehension

[barren olive]

Honestly can't wait for anything v6 Fine tuned for sdxl. If this is only standard SD imagine what could it do on XL

[floral island]

but SDXL just blows that straight out of the water

[peak dove]

Oh my!!!

[soft zealot]

a few from earlier

[floral island]

base 0.9?

[soft zealot]

yes

[barren olive]

@floral island do u remember this prompt?

[floral island]

oh yeah, i know that image ^^

[barren olive]

Do u know the prompt?

[floral island]

isn't it embedded in the image?

[barren olive]

I kinda want to recreate it on sdxl

[tribal lantern]

I use https://www.midlibrary.io/ for ideas about artists, only to figure out that what is shown there often looks nothing like the artists work, I thought SDXL did those artists badly, but it's at times actually closer to an artist style than the extra pretty detailed colorful and saturated stuff shown there

[floral island]

lemme boot up A1111, i should be able to find it

[boreal bough]

instead, use Interrogator with Vit-H loaded, to query an image you wanna replicate

often there are random words, that are extremely biased towards the artstyle you wanna replicate

[grizzled warren]

They are both similar and fundamentally different.

Midjourney is a paid service based on SD, intended for casual users. This is why it does a lot of sneaky stuff under the hood with your prompt, and often defaults to "Midjourney style". It also appears to generate a lot more images, only showing 4 best rated by automatic aesthetic scorer. Out of 20, or maybe even 40 variants. And it also runs bigger models, this is why SD1.5 had a hard time competing with it.

Stable Diffusion, on the another hand, is an image generation platform. Which does precisely what you ask it to do, without any hidden prompt shenanigans or sneaky additions to your workflow.

SDXL appears to have very similar output quality in comparison with MJ and vastly superior prompt adherence. If you take your time and cherry-pick a single best of five image, you might actually beat Midjourney in quality more often than not. But most importantly, it puts you in the driver's seat with stuff like ControlNet (in the future), latent space tricks, inpainting tricks... So if you know what you are doing and take your time, you'll always beat Midjourney. It's even possible with 1.5, but it takes a lot of skill, since the base model is inferior. SDXL is on par, if not superior to MJ's model.

[boreal bough]

or even for clothing articles, where the wrong word was learned to represent a clothing article XD

[peak dove]

Gosh!

❤️

[peak dove]

But I (for one) am not in it for the artists's look - I want a completely new look - a voice or style I can call my own!

[floral island]

dang bro, leave some for us!

[tribal lantern]

I'll do that, often times I just want to mix and match styles

[smoky patrol]

Hahahaha

[sharp robin]

When will we get clipg interrogator?

[floral island]

oof, it's a lyrics prompt, those don't do well on SDXL yet

`As the (fairy :1.3) flies
Through the (battlefield :1.3) below
Horrors unfold,

Blood stains the ground,
Cries of grief fill the air,
Madness and destruction

(war:1.1)
(sunset:1.1)

Sorrow in her heart,
As she witnesses the scene,
(The horror of war,:1.1),
(world war 2:1.2), (1945:1.2), agony, panic
(best quality:1.3), (highres:1), (detailed:1.3), (incredible:1.3), (perfect:1.3), (perfection:1.3), (illustration:1.2), medium shot, floating, tears`

[shy kelp]

I made this with SDXL 0.9 and 1.5 inpainting🤍

[barren olive]

This is good enough. Gonna try it with Anime style @floral island

[peak dove]

If I mix remedios varo and rob gonsalves in MJ - u get this incredible geometirc effect (almost like Escher) - whereas in Wombo or Blue Willow or Imagaine or Stablecog, or Fusionbrain or Clipdrop etc, there is nvever that same dizzyingly vertigenous geometric effect

[barren olive]

Each ai has different Prompt language. You can still get same results if u know how to prompt it correctly

[smoky patrol]

Did you say...up the ante, maybe? Yeah, I think that's what you said! ❤️

[floral island]

hmm... doesn't want to give me fairies XD

[barren olive]

Anime has... Weird effect

[floral island]

that 1st one tho 😮

[barren olive]

Ayy

I bet this two would be great friend like geralt and jeskier

[smoky patrol]

Good, good 👏

[rustic surge]

what is Clipdrop

[barren olive]

Post apocalyptic vibes

[floral island]

i REALLY like how the base model can be so diverse already 🙂

[smoky patrol]

https://clipdrop.co/

[barren olive]

Imagine sdxl 2.0 in next few months

[barren olive]

Honestly before i dont really care for sd when it first came out but after the fine-tuned models came out i cancelled my mj subscription and fall in love with it

[floral island]

uh... task failed succesfully?

[smoky patrol]

What was your prompt?

[barren olive]

I've also made this with very complex prompt. Like how it turned out

[floral island]

An ethereal fairy-like creature floating amidst a sea of flowers, with vibrant colors surrounding them. The background appears blurred, as if the focus is solely on the subject. The mood seems peaceful and serene, conveying a message of hope and love.

[barren olive]

Not to mention SDXL can spell

[peak dove]

Some SDXL 0.9 country scenes

[smoky patrol]

I see

[floral island]

well... no, now that i think about it... that's about right

it's a fairy-like creature. and it's not wrong

[smoky patrol]

Love

[smoky patrol]

It isn't

you got what you prompted

❤️

[floral island]

😄

[smoky patrol]

It is fairy like

[floral island]

uhoh... i found older prompts of thiccccccccc

lets see how sdxl handles that

[barren olive]

@floral island Bot doesn't work for me

[floral island]

first, last non-thicc

[floral island]

i'm generating locally

[barren olive]

Whats your specs?

I doubt my pc can handle SDXL

[floral island]

beast-pc

[sharp robin]

What’s urs?

[floral island]

amd 7900x, 3090, 64gb ram, enough disk space XD

[barren olive]

Rx 580, i3 6100

[smoky patrol]

Perfection

[smoky patrol]

Also perfection

[barren olive]

Fk

[sharp robin]

Damn. Amd.

[floral island]

it also functions as the heater for my living room

[smoky patrol]

purrfection

[barren olive]

Whats wrong with amd?

[jolly creek]

Cozy winter evenings guaranteed

[smoky patrol]

As I said earlier, there are some technical hiccups right now; sorry about that!

[floral island]

i've not had my heating on since february: my home is a+ certified isolated, so i'm REALLY having trouble venting excess heat

[boreal bough]

thanks for sharing ❤️

[barren olive]

Can someone make this? Clipdrop has like 1500 people

[floral island]

so pretty!

[barren olive]

Prompt: prompt:a humanoid fox with leaf dress and fairy wings Using a flute on top of a tree on sunset, digital art, photorealistic illustration, cinematic, trending on artstation, dramatic,cinematic

Also with anime style

[boreal bough]

also this one turned out well

[barren olive]

SdXL is so demand. I wonder if Midjourney would make their ai free again

[floral island]

that prompt just pumps out amazing shit like no-one's business!

[boreal bough]

lemme try something dumb XD

[boreal bough]

works well enough XD

added , (lithograph, risograph:1.2) to it

[peak dove]

❤️

[boreal bough]

changes it from drawing, to lithograph style print

[barren olive]

Whats that prompt?

[floral island]

wait, comfyui supports (weights:1.2)?

[boreal bough]

yep! to make sure you dont get it wrong, select the words, then do Ctrl+Up

[smoky patrol]

Mmm, hair

[barren olive]

Really like how sdxl can make dramatic now now

[floral island]

TIL!

[smoky patrol]

keep posting, keep posting

[barren olive]

Also made some demon adts

[floral island]

i used ctrl+up so many times in a1111/vlad so it's second nature

[floral island]

weird when i'm not in those ui's and it doewn't work xD

this guy be stealing ur women

[boreal bough]

btw, you can use notes to save your prompts for future use in comfy

[smoky patrol]

He does have pockets

[peak dove]

Afrofuturist cowgirls SDXL 0.9

[peak dove]

❤️

[barren olive]

Tried making the most ridiculous thing ever

[floral island]

comfy-xl?

[boreal bough]

all versions of comfy

but yes - xl workflow

cat jumping on a box, lithograph, risograph
N: hyperrealism, text, signature

[floral island]

and this is just as cute ❤️

[smoky patrol]

I'll take more of these...with curly hair

[boreal bough]

and for the 90s version of it XD

[boreal bough]

cat jumping on a box, 90s aesthetic, Retro illustration by Alex Varanese
N: hyperrealism, text, signature

[barren olive]

I dont remember the prompt. This is very old and I can't make new art since bots and clipdrop is super slow

[barren olive]

Btw whats the prompt for this?

[floral island]

if it's made in comfy-ui you can simpl run comfy ui and drag it

[boreal bough]

depends on in if it was pasted, or uploaded
paste removes metadata

[smoky patrol]

It's all good

[floral island]

pasted/uploaded?

[barren olive]

Whats the prompt for the fox?

[boreal bough]

if you do "copy image" in chrome, then paste into discord - it wipes the metadata
if you take the saved image in the file browser, and drag it into discord, then everyone can use it get your layout + generation data

you can usually tell, by checking the filename of the images in discord. if they are called "view", then its guaranteed that there is no more metadata in them

(since that's the name they get when you copy/paste from comfyui)

[floral island]

i'm using firefox, so pretty sure it's a chrome issue then

[boreal bough]

ah, you might be right about that

[boreal bough]

just chrome and brave that wipe the data

[floral island]

they be stealing hearts!

[smoky patrol]

It has...male anthropomorphic fox in it...and furry art 😛

[smoky patrol]

THAT THEY DO ❤️

[smoky patrol]

Great shadows

[floral island]

i feel sdxl is also a lot more generous in lighting, being able to go much darker/lighter than default sd could

[boreal bough]

very much!

[peak dove]

I made a small prompt file at Google Drive - I shall add and add to it. Most these prompts are from Issue 3 MJ magazine (all prompts at MJ © MJ - but free to use). https://drive.google.com/file/d/1eJ3nvnwNWmoZ3Lwud73FqzbA0q8bRVlH/view?usp=drive_link

[barren olive]

Whats better? Deep-floyd vs stable diffusion XL?

[boreal bough]

A male lawyer waits underneath a bridge in london, 1920s, night scene, moonlit scene, film still
(horror:1.2) still, grunge, vignette, chromatic aberration, dark, lovecraftian

[smoky patrol]

so good

[floral island]

damn, i love this one!

[smoky patrol]

ahhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh

[floral island]

i think that's one of my llm random series: "characters doing something random of effort, in a serene environment"

[peak dove]

Spectacular

[smoky patrol]

nods quietly indeed

[floral island]

basically, i'm not even putting any creative effort -> "llm make me something pretty" -> passes that on to SDXL "here is something pretty"

i'm worried about humans as a species XD

[smoky patrol]

you shouldn't be 😛

[smoky patrol]

love the bottom right

[smoky patrol]

love the palette

[smoky patrol]

That truly is amazing

[floral island]

ooh, coming up on some of my favorite llm prompts 😮 lets see if sdxl can do them!

think i might need a bit more refiner for this... or sdxl sucks at one of my fav prompts 😦

[boreal bough]

share?
often its just one word messing things up - or a lack of specific negatives

[floral island]

it's basically the same as all the other images i've done earlier

they're not bad, but compared to the ones i got from my anime model, these are... lacking

[peak dove]

Anyone interested in dark, sexy, muscle car, city lights prompts?

[boreal bough]

XD

oddly specific

[boreal bough]

like this, but more anime?

[floral island]

that's pretty nice though

NOT SD XL OUTPUT was hoping more like this, but that's just the model, i suppose

[boreal bough]

ahhhh

[floral island]

sdxl nails the actual prompt a lot more, whereas this model straight up ignored the purple actually

[boreal bough]

makes more sense

I was working in a different direction XD

[floral island]

damn, this is so nice!

[boreal bough]

getting closer!

and done

[floral island]

damn, that's pretty good 😮

[boreal bough]

P:
(simple background:1.3), (orange haze:1.2), (anime:1.4), An ethereal figure stands amidst a backdrop of fiery reds and deep purples, with a single tree standing tall in the center of the painting. The figure appears to be made of light, with delicate wings spread wide and a gentle gaze fixed on the tree. The leaves of the tree glow with a soft inner light, casting a warm glow across the entire scene. There is a sense of movement and change in the air, as if something momentous is about to occur. The colors and textures of the painting come together to create a dreamlike atmosphere, full of mystery and wonder.

Style:
(simple background:1.3), (orange haze:1.2), dreamlike atmosphere, mystery, wonder, (anime:1.4)

N:
hyperrealism, text, signature

[floral island]

could be poster art just like that

[boreal bough]

simple background was the keyword to tone down what was going on, and anime pushed up like hell XD

also haze can be used to control color

[floral island]

you throw p in the G textbox and style in the L textbox?

[boreal bough]

yep

[boreal bough]

🤣

[boreal bough]

tghis prompt combo works pretty well for various simple anime like images

[floral island]

that's so cute!

[floral island]

ooohhh-kaaayy

[tribal gale]

nice!

love mgs3

[boreal bough]

can be adapted into everything 🤣

[chrome furnace]

Hey peepz, would running 2x 4060 ti 16gb make any sense with sdxl?

[tribal gale]

I don't think inference works with multiple gpus?

[floral island]

i don't think you can inference on multiple gpu's at the same time

[tribal gale]

but you could run it with one of them

[floral island]

you could game and inference at the same time though

[soft zealot]

however I believe you may be able to run 2 instances of your UI of choice with each working on a different GPU

[boreal bough]

(simple background:1.3), (grey haze:1.2), (anime:1.4), snake from metal gear solid, standing at the end of the world, retribution, the hope of mankind, nier automata

[chrome furnace]

Thx for the quick answers. Have not thought of that, but doubt i can focus on 2 things 😄

[upbeat summit]

that's very awesome!

[latent zodiac]

ok so I have fixed the UI a little more, now you can easily search via starting letter of an artist as well as copying artist name to clipboard to use in your prompts: https://charming-strudel-a6a8b6.netlify.app/

[chrome furnace]

thanks to u 2, i will look into this.

[latent zodiac]

This is a curated list of SDXL artists

[tribal gale]

yoji shinkawa works nicely

[floral island]

didn't the dude actually do drawings for mgs? XD

[tribal gale]

yep lol

he made all of them, continued on to make Death's stranding as well

[floral island]

got a 8400 artist count wildcard list, want to give it a go too? XD

[floral island]

i have such conflicted feelings about death stranding

it's a work of art, for sure

[tribal gale]

I loved it personally

but I'm a kojima-stan

[boreal bough]

worlds best postapocalyptic postal service worker simulator

[floral island]

and certainly a refreshing gameplay style, but... oh man... it tended to drag at times

[floral island]

i remember placing a specific bridge, it was SOOOOOOO well used

[true hazel]

the story is just too convoluted, bit harder to relate to than MGS

[floral island]

placed it, came back a bit later, used a gazillion times

[tribal gale]

I wonder how that game holds up today, one of the really cool stuff was playing it at launch with everyone building stuff

[floral island]

i think you only get a few from the 'global world'

i mean, else the snow mountains would be filled to the brim with those catapult thingies

dunno what they're called again

[tribal gale]

idk I remember people building highways lol

[peak dove]

Nice Prompt =there are flowers, birds, butterflies around the starfish, nordic illustration, flat style, high detail

[floral island]

i made it a personal vendetta against those mountains, you could go from that middle area to the other side of the mountains using those .. uhh... rides?

[tribal gale]

but I can totally understand people that didn't like it, most gameplay is just you planning routes and fighting the enviroment

[peak dove]

Another Nice Prompt = cocktail illustration, minimal, modern (add hibiscus and Davi Augusto if needed)

[floral island]

yeah, it certainly was a fresh take, and i didn't dislike it at all... but some of those cutscenes were a bit... long

[tribal gale]

I loved it personally, including the story

[upbeat summit]

the outlines are beautiful

[boreal bough]

👍

[peak dove]

So cool

[tribal gale]

that prompt but with yoji

[boreal bough]

lithographed!

[tribal gale]

seems like neon punk turned it to something else lol

[boreal bough]

2nd lithograph

[peak dove]

Here's a multicoloured flat vector look - change the city when u like Prompt = orthographic perspective, san francisco cityscape with floating financial iconography and symbols, flat design, simple and minimalist, business illustration, sharp and clean lines, infographic, bright colours, dribbble, freepik, pinterest, vector art

[sharp cloud]

What are your favorite simple things that even SDXL just doesn't seem to grok very well? Mine are facial piercings and watering cans. The can here at least is quite accurate, unlike 1.5 creations

[boreal bough]

I see your neon, and raise you with lovecraft

[peak dove]

This is so cool

[west breach]

collage

[tribal gale]

in origami

[boreal bough]

add this to positive and style, to add only a tint of horror: (works for all styles)
(horror:1.2) still, grunge, vignette, chromatic aberration, dark, lovecraftian

[smoky patrol]

oh, nice!

[floral island]

ART THEFT, I SAY!

oh wait... wrong community xD

[boreal bough]

XD

[peak dove]

Banksy-ish! Nice

[smoky patrol]

Love these

[floral island]

i've seldom seen you not like art in this server 😄

[smoky patrol]

love how the ink kind of drags and looks like it smudges...lmao..that's great

[smoky patrol]

Well. I love art. So whadya WANT!?

[tribal gale]

yeah that's pretty much yoji's style

[floral island]

nothing more, i love the arthusiasm!

[tribal gale]

espeically in the old mgs1 days

[smoky patrol]

s t a r f i s h

[floral island]

weird weed-growth, but ok

[smoky patrol]

stars blankly into the ether

[smoky patrol]

Yeah; I absolutely love it--anything to do with ink sign me up. Totally a fan.

[boreal bough]

oops. should have stopped while I was ahead

[boreal bough]

starfish is... less present XD

[smoky patrol]

oh. well then. CONTINUE

[smoky patrol]

It's almost gravitational art LMAO

[floral island]

this?

[smoky patrol]

look here. That's called. a wreath necklace.

[smoky patrol]

Don't make me get out my watercolor.

[smoky patrol]

ooooo. Great highlights! and skulls! ❤️

[boreal bough]

this is what happens when you feel the need to fill up the 77 tokens to the last one XD

[smoky patrol]

omg, are we in a game? I'm gonna bap it? Or...maybe it's supposed to be tamed.

[peak dove]

Here's a nice dragony prompt - a mesmerizing and immersive painting of a mythical griffin reading a book, created in the style of yayoi kusama, showcasing her iconic use of polka dots, vibrant colors, and hypnotic patterns that envelop the griffin in a fantastical world of literature and infinite possibilities

[floral island]

is 77 the hard limit for comfy?

[boreal bough]

oh god the colors! they burn XD

[boreal bough]

I think the rest are simply dropped - but would need confirmation from comfy for that

[floral island]

not the workaround a1111 did?

[peak dove]

A111 allows more than 75 'tis true

[boreal bough]

and they arent simply dropped in the background?

[floral island]

for A1111 they are grouped by 75 tokens, and then sent averaged

[peak dove]

You tried it? I love Yayoi Kusama, add her with Lisa Franks and ...

[sharp cloud]

This is probably the best I've gotten. A little small maybe but otherwise not bad.

[boreal bough]

honestly not sure for comfy now. but hey, I present you 200 tokens!

[safe spear]

Any news about the release? Like how much will it get postponed?

[boreal bough]

no more exact dates on sdxl, for the sake all mods & face of sai

[safe spear]

Makes sense, I just hope the model is released during this or the next month.

[boreal bough]

"It'll be like a week or so"
#✨｜sdxl message

[peak dove]

Colourful Colourful Prompt = paper art, xinjiang province, landscape, hami melon, grape, exquisite details, colourful, clean background, 8k

[floral island]

famous last words

[boreal bough]

I am filled with fear after the last one XD

[smoky patrol]

totally vibing how everything looks like it's on different layers and the colors are being pulled together

[boreal bough]

👍 eyes still work. all good!

[true hazel]

when you said some malarkey

[west breach]

trying out some 'graffiti' styles

[boreal bough]

XD changing theme turned into a ghost pokemon reveal

[peak dove]

I didn't see your dragony picture ? 😄

[peak dove]

Wowee

Kewl

I'm just doing some RLFH (or whatever they call it) ... 😄

[marble kindle]

so funny thats weebs are sad they cant make anime tiddies with sdxl

[strange mist]

hello guys, I'm having Error no file named model_index.json found in directory in vlad's fork 😭
can anyone please help me?

[true hazel]

what was the prompt?

[floral island]

one week tops, and titties will be back on the menu

[marble kindle]

the model is gonna be censored

[peak dove]

Just searching Vlad for model_index.json

[floral island]

and then people will train stuff back in

[true hazel]

I get boobas on locally 0.9 all the time, with nipples even

[strange mist]

wdym?

[peak dove]

Under Models/Diffusers you should have base diffuser and refiner diffuser - the index.json should be in there with them

[strange mist]

perhaps, I should redownload? 🤔

[peak dove]

Why not - or just rename VENV to something like oldVENV, and let venv be rebuilt?

[strange mist]

I'll try redl

[floral island]

living on an island lol

[smoky patrol]

loved the warped thing going on ❤️

[eternal fog]

I wouldn't even bother people just like to moan about it. They never listen when you tell them that. Or they think that it's going to let them make hardcore stuff by default lol.

[eternal fog]

It's so easy to train there will be nsfw models try soon after release I can almost guarantee it lol

[peak dove]

I just allowed VLAD to reconstruct a new VENV file and all was well

[urban fjord]

Does anyone know why Kohya-ss uses a lot more VRAM now than before when training SDXL?

[rustic garnet]

how do you read prompts from a file (line per line) using ComfyUI?

I haven't found any node for that

[boreal bough]

is there?
I only know about drag and drop

[boreal bough]

sure you disabled clip?

[rustic garnet]

that's strange... I mean, that should be a very common use case, shouldn't it?

[boreal bough]

witht he full unet only command

[boreal bough]

its mostly backend directly to frontend interface

[west breach]

my custom node can read line by line for each batch

[rustic garnet]

can you share code?

[west breach]

https://github.com/bash-j/mikey_nodes

if you use the __wild_card__ syntax and have your txt files in comfyui/wildcards, then set control after generate to increment, it should go line by line

[eternal fog]

Do you have the most recent version. Because it uses a lot less for me with the newest update

[peak dove]

SDXL 0.9 Vlad AUTO1111

[urban fjord]

I'm using bmaltais/kohya_ss, so I guess that might be why.

[eternal fog]

Yeah that's a little behind

The main koyha scripts got fixed last night and now I can train with 10gb vram

[rustic garnet]

hm, should the text file appear in the "styles" list?

[dense chasm]

when will sdxl 1.0 release ETA?

[urban fjord]

There is a SDXL 1.0 stage the 26th.

So I guess it will either be released then or release will be announced then

[rustic garnet]

ah, okay, wild__card has to be filename

[urban fjord]

I think there is a couple of pinned posts about it.

[west breach]

yeah, like this

[urban fjord]

Has anyone had any weird problems in ComfyUI when trying to generate the same image with and without a LoRA in the same generation?
Something feels off but problem doesn't disappear when I remove LoRA.

[strange mist]

redownloading seems to have fixed the issue. thanks 😄

[peak dove]

Prompt = a drawing of a woman in a suit and shirt, in the style of retro feel, charming character illustrations, dark white and pink, editorial illustrations, digitally enhanced, I can't believe how beautiful this is, eye-catching Vlad AUTO1111 SDXL 0.9

[rustic surge]

Does SDXL support nsfw?

[vocal stream]

only kind of, but probably will with finetuning

[eternal fog]

Why's everyone so horny