✨｜sdxl

[trail bay]

ty

[trail bay]

ahhh I'm not using CLIPTextEncodeSDXL

oops

probably have to update

[trail bay]

nvm have it. just need to figure out how to use g vs l and all that haha

[livid basin]

If this isn't modern art idk what is

[winter raptor]

also recommend installing this plugin, it allows you to install missing custom nodes/models https://github.com/ltdrdata/ComfyUI-Manager

[spring fulcrum]

Im not sure how to get creepier images. Its just seems to be too tame here

[hearty ginkgo]

so im using the workflow for umm sytan how do i make ultrawide photos?

[glad fulcrum]

Do the base and refiner models share the VAE?

[spring fulcrum]

these are looking pretty sinister... yay

[trim orbit]

edit those to a different aspect ratio

[vast narwhal]

Have you tested Dreamshaper?

[high skiff]

I don't want to get two ahead of myself at the moment, or be too optimistic, but it looks like there could potentially be hope for the 3090

It is still refusing to boot in my new computer, however I just got it to boot no problem with the old computer (with a brand spanking new windows install)

[jolly void]

N

[high skiff]

Alright, I can confirm now with no uncertainty now that the GPU is the problem

[peak dove]

ComfyUI - I want an Image Browser like A1111 where prompt and settings can be recalled ... is there such a thing in comfyUI?

[vast narwhal]

Dreamshaper XL, quite different results comparing with the base

[peak dove]

Delete VENV and then restart ComfyUI?

[sharp robin]

im 100% sure seller knew ab it.

have u contacted theM?

[high skiff]

They have been working with me for multiple days now to try and find solutions. I have tried it everything feasible under the sun to try and give them the benefit of the doubt, but I can't do it anymore

I have now officially escalated my original claim with PayPal to their money recovery team, and they will be severing our connection and taking an information from both sides

It is undeniable that the GPU is faulty now, and I should not be held responsible for a graphics card that started showing problems within less than 1 hour of appearing in the mail.

I have spent over 24 hours cumulative in the last 4 days working on this problem, and I found absolutely zero solutions, and nothing but a long list of reasons why all of these issues are pointing directly at the graphics card being dead on arrival

[sharp robin]

damn welp that blows, deal was too good-ish to be true in the end. hopefully u get ur money back

[lusty raptor]

😦

[peak dove]

Carpe Diem. For what it's worth - bend over backwards to buy a new (not used) GPU. Bite the bullet - and the expense - you'll live longer 🙂 Do ou have patreon/PaypalMe at all?

[rich orchid]

For anyone else getting errors when launching SDXL using the Vlad Automatic installation, try launching the app using this command line ".\webui.bat --backend diffusers". I have a i9 13900k 4090 system, but had to use that to make it work.

[azure oxide]

you can buy him a really big cup of coffee, link is on his github

[high skiff]

I have buy me a coffee, but there is no chance in hell I'll be able to afford a new 3090, not would I want to, cause the GPU is too expensive for what it's worth

[rich orchid]

Can SDXL save images in .png format retaining the metadata like SD normally does? Default seems to be jpg with no metadata.

[peak dove]

I am using VLAD A1111 - I start from a command line --medvram --backend diffusers

[rich orchid]

What does --medvram do?

[peak dove]

It recntly stopped working due to Xforners conflicts - but since I am not using --xformers, I am not sure what to do - except jump to the excellent (but featureless) ComfyUI

Medium VRAM usage - I use an 8Gb RTX2070

--lowvram - even lower vram usage

[rich orchid]

I think I just read that using tiles low ram gpu can work with Vlad A1111... ComfyUI seems like a headache.

[hard fractal]

Literally nothing has changed ❤️❤️😘😍🥰♥️♥️💙💜

[spring fulcrum]

--medvram does not seem to be a working command line argument for ComfyUI in case anyone was wondering.

[peak dove]

I was surprised at how easy to instal and setup ComfyUI really is - and don't worry about its "noodle nodey interface" - its a cinch!

[peak dove]

I only use --medvram in VLAD A1111

[high skiff]

I'm gonna miss the entire launch of SDXL cause some jackass decided to sell me a broken GPU

[peak dove]

Share a desktop with a friend?

[spring fulcrum]

Did you buy it used?

[high skiff]

It doesn't just work like that, I don't have any friends willing to do that, nor would I want to put that burden on them

[high skiff]

Yes

I'll get my money back eventually, but it will be long after SDXL is interesting anymore

[peak dove]

We'll be on 4kSDXL before you know it - or 8kSDXL!!!!! 😄

[spring fulcrum]

GPUs are pretty much the only thing I don't trust used due to Digital coin mining. They use the absolute crap out of a GPU and sometimes completely mess up the bios of the card with custom crap

[high skiff]

Used is my only option, so I'll just have to take this risk again I guess

[peak dove]

As I push towards an A1000 - I might have a very usable RTX2070 available ....

[spring fulcrum]

What card and budget are you looking for?

[high skiff]

It's not that I don't have a graphics card, it's that I don't have a graphics card capable of doing the research that I'm trying to publish alongside my workflow

[peak dove]

Oh no - there's me thinking out loud again! Ignore me 🙂

[high skiff]

It is not all me trying to sound ungrateful, I immensely appreciate your guys' support, but I don't expect anything from the community

[tall lion]

I keep getting this error when I try to get to the part of local tunnel - Earlier two parts go smoothly.

/tools/node/bin/lt -> /tools/node/lib/node_modules/localtunnel/bin/lt.js

• localtunnel@2.0.2
  added 22 packages from 22 contributors in 2.14s
  Traceback (most recent call last):
  File "/content/drive/MyDrive/ComfyUI/main.py", line 64, in <module>
  import cuda_malloc
  ModuleNotFoundError: No module named 'cuda_malloc'

[high skiff]

I do have a GPU that will run SDXL, but I don't have a GPU that will do what I need to do with SDXL, which is the whole point why I scrape together every penny I had (and even every penny I didn't) in order to buy this GPU

[high skiff]

I need minimum a 3090. It has to be 24GB VRAM

So I'm gonna have to wait for this guy to drag out this claim as long as possible before he has to fork over the money, and then go through this whole process again with a new seller

[spring fulcrum]

How much was that one?

[high skiff]

Immensely frustrating is the most understated way to put this

[peak dove]

CUDAS Malloc is CUDA Memory Allocation

[high skiff]

$600

Hell, I don't even have the money for it. My mom paid it for me to give me some time to sell my 3080 and 3060ti to pay it off

So I am just stuck, and so is my mom until this jackass hands my money back over

[peak dove]

ModuleNotFoundError: No module named 'cuda_malloc - you probably have a PATH error - and installed this to your appdata file and not the ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\python_embeded\Lib\site-packages

[spring fulcrum]

Are you in the US?

[tame raven]

which workflow are you using for this? I've tried creating 9:16 images but the perception is always weird

[high skiff]

Yes

I'm going to sleep. I'm so done with today and all of this.

[visual glade]

is the file cuda_maloc.py in that ComfyUI folder?

because it's in the repo; https://github.com/comfyanonymous/ComfyUI/blob/master/cuda_malloc.py

[rich orchid]

If I choose two images in batch in A1111 each will be different because it increments the seed by 1 for each image. How do I get SDXL to do the same?

[sharp robin]

on comfy?

SDXL is just another model

its not a UI

[rich orchid]

on Vlad A1111.

[peak dove]

SDXL also increments 1 by 1 - set Seed to -1

SD.NEXT/Vlad - set Seed to -1

[rich orchid]

If I set seed to 500 in A1111 and set batch to 4 images, it will increment each image by 1 seed. Right now in Vlad A1111 it makes all images exactly the same, apparently using the same seed for each one.

[peak dove]

No, if you set the seed at 500 - it stays at 500 - setting it to -1 it will increment

the only other way is to setup Agent Schedular, and increment 500, 501, 502, 503 etc for each new picture

[rich orchid]

Thanks. I'll look into it more.

[peak dove]

I agree in "ordinary" A1111 that 500 witl self-increment by 1 step at a time

[rich orchid]

Yeah, thought so. Can't get Vlad A1111 to do it. When I change seed to -1 I get error... cannot reshape tensor of 0 elements into shape

[peak dove]

Sounds like a bug - I have an Xformers bug - so I cannot get into Vlad today ,,, have posted at Github and await his advice

[winter raptor]

This keeps happening and I can't seem to correct it

the girl is too smoothed out

[tall lion]

yea but it says cuda_malloc not found while trying to install localtunnel as part of comfyUI on google colab. Thanks

[peak dove]

Sounds like an instal PATH error ...

[tall lion]

I will redownload the notebook and try from scratch again on colab

[spring fulcrum]

Does anyone have a good workflow with SDXL I could copy?

[peak dove]

If you want prompts (A1111 and Comfy) I got prompts

[spring fulcrum]

Prompts would be good

I could use some new ideas

I like some of the galaxy in a bottle things... or would like to try

[peak dove]

Often when d/loading, stuff can end up in %appdadata% folder instead of in python ... if u forget to select PATH when installing Python

[dense chasm]

the new LlaMA2 has reward and safe model,just like the nsfw combined with SD,but still a long way to catch up with GPT4

[tall lion]

This is awesome. Is there a prompt or workflow for this?

Seems to use a test.ckpt and says not found. Is the ckpt in hugginface? Tks

[peak dove]

Simple - Galaxy in a Bottle, style of Breathtaking night landscape with syellecafen

[spring fulcrum]

well its not in a bottle but still looks amazing

second seed

[soft zealot]

I Uses @high skiff s as a staring point

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[spring fulcrum]

3rd attempt and we are getting close

[tall lion]

Are you using a custom upscaler for htis. When I load this into comfyUI by dragging, I get an error about a SDUPscaler node

[spring fulcrum]

yes you would need to get this and put it in your custom nodes folder in comfyui

https://github.com/ssitu/ComfyUI_UltimateSDUpscale

[tall lion]

Awesome. Can't seem to get the workflowt o work though on Colab ComfyUI.

[tall lion]

Great let me give it a shot.

[spring fulcrum]

Ill give it a shot

[soft zealot]

ahhh trying to replicate/improve on the default COmfy Prompt

[spring fulcrum]

Is that what the default one is?

[soft zealot]

yarp

[spring fulcrum]

so the , with the empty space and , is not a mistake?

These are starting to look really good. BUT... what I want is for the bottle to be glowing and illuminate a pitch black room with the light emanating from the glowing galaxy in the bottle

[royal fern]

that's not a mistake. I understand it as "kicker"

[spring fulcrum]

what does that mean?

[royal fern]

something to steer the conditioning. you can actually add random words and see the composition change

[royal fern]

the white space is a conditioning without a meaning

try with and without and check the difference

(some models more than others)

[spring fulcrum]

So I can use that in any prompt and it fills it in with something random?

[royal fern]

no

[tall lion]

Excellent that worked in Google Colab also. Had to adda s snipped out of code to cd into custom_node and do a git pull. Tks

[royal fern]

it just steers the conditioning, the same as like changing the resolution by 10pixels might end up with a totally different result

[spring fulcrum]

no problem... always happy to help

[spring fulcrum]

that is interesting... I'm going to give it a try

[royal fern]

to break the monotony (eg: generating always the same face) I often add a completely random unrelated word at the end of the prompt

it can't be something with a strong weight (like say... "banana") but it often works to get that little unexpected variation

[peak dove]

nightime intergalactic landscape in a bottle syllecefan SDXL 0.9 ComfyUI

[spring fulcrum]

what is this syllecefan?

is that a style or artist? or ...

[peak dove]

SYLLECEFAN "is knowing the unknowable!" 😄

[spring fulcrum]

nice

[peak dove]

Actually, Syllecefan is probably the name of the originator of the prompt I modified 😉

Cowgirl in white preparing pomegranates in a glass bottle

[uncut steeple]

[peak dove]

Dope

[peak dove]

When I made my prompt - they came out very blue without me asking to - yours are very gray/green - you asked for those colors?

[spring fulcrum]

nope not at all

[peak dove]

Amazing how your bottle universes are muted grays

[peak dove]

Graphite gray

Moody

[ionic dragon]

damn, did you buy it from the link I shared with you?

[peak dove]

Ain't she gawjus?

[spring fulcrum]

those pomegranates nice

This one came out really cool

[peak dove]

Kitsune

[peak dove]

^..^<

[tall lion]

The workflow by @spring fulcrum is freaking phenomenal

Tks

[spring fulcrum]

Thanks 🙂

[soft zealot]

mmmm, nice melons

[spring fulcrum]

lmao nice mellons

[royal fern]

melons

[peak dove]

Her name must be Mellonie? 😄

[peak dove]

Mama! I'm scaahed!!!! 😦

[peak dove]

G.O.A.T. 😄

[spring fulcrum]

lol

[peak dove]

Pomegranate Cowgirl In White, White Horse

[spring fulcrum]

that is an interesting style

[peak dove]

FULL PROMPT - Painting of white clad cowgirl pomegranate juice white horse blue sky, preparing and slicing pomegranates by Diego Rivera VICTO NGAI Zdzisław Beksiński georgia okeeffe

[tall lion]

It is possible to train using Kohya Trainer for SDXL. Check this out https://github.com/kohya-ss/sd-scripts/discussions/662

@timid sonnet are you running your own chkpoint?

[spring fulcrum]

This is getting pretty good here

[peak dove]

Cool - the right side especially

[peak dove]

Very Route 66 - nice!!!

[silent oxide]

this one is awesome

[peak dove]

Do you have a powerful nVidia GPU?

[ionic dragon]

not really

https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/tree/main
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-0.9/tree/main

[peak dove]

Here is the ComfyUI d/load - but only if u have an nVidia GPU.

[ionic dragon]

you can even run on AMD; but its slower

https://github.com/comfyanonymous/ComfyUI#installing

[peak dove]

U can use the dream command here in Stability AI, or use SDXL 0.9 on NightCafé and/or Clipdrop.co

[serene flare]

scarlett johanson

deam scarlett johanson

[peak dove]

Go to bot cahnnel to use / dream

bot channel

Anybody know how to find prompt used in ComfyUI - is the info stored somewhere?

[ionic dragon]

#1100170312106127410

[peak dove]

If you feel that I am not communicating effectively and you didn't know what I meant ...

... but you dID kNOW wHAT I Meant 🙂

[ionic dragon]

its make the generated image better

[royal fern]

the pedant police has arrived

[peak dove]

Nep - SDXL is a two-stage process. The SDXL Base model makes an image. This image is txt2img format. This image in the second stage is fed into SDXL Refiner in a kind of (but not exactly) img2img fashion. Some of the txt2img data is also sent from the Base model - so the SDXL Refiner is used as the 2nd stage of the SDXL process

... kind of ... 🙂

[royal fern]

it is an img2img, just at latent level

[soft zealot]

you ar edescribing SDXL0.9 which is a pre release . This behaviour may or may not remain the same on full release of V1

[royal fern]

latent2latent 🙂

[eternal fog]

I'm using my own, but it should be fine unless you are pushing the resolution too far. Also make sure the height and width and target height and width on the SDXL Clip encoder match aspect ratios with the empty latent you supply to the Ksampler.

[full parrot]

Students face, looking determined and focus.

[peak dove]

Go to the Bot Channels and use / dream

Because the Bot Channel will produce SDXL 0.9 images for free

[royal fern]

SD devs said to just use 4096

at least for the 0.9

[peak dove]

But you don't have a powerful GPU?

So, you want me to ask "Do u have 1?" 🙂

[eternal fog]

If you try use a 1:1 aspect ratio for clip and then a 9:16 for the image it will stretch out the perspective. It's why people keep getting long people in their generations

[peak dove]

Are we flirting? 🙂

[soft zealot]

what do you consider the bare minimum "powerful GPU " to generaate images because I quite happily use a 3 generation old 1080ti without any issues

[royal fern]

width/height and target_width/height need to be the same (4096)

but it doesn't need to match the latent

[peak dove]

But you don't have a powerful GPU - or - But you don't have a powerful GPU?

[royal fern]

I'm just reporting what they said on the official server

[eternal fog]

I've done loads of testing. If it doesn't match the latents aspect ratio it messes up the perspective, usually more noticeable with people.

[peak dove]

My SDXL 0.9 in ComfyUI is often 768x1024 - to make poster size

One was a statement - the other was a question

[indigo carbon]

got SDXL working on A1111 without silly extensions, this is amazing!

[royal fern]

I can't say I experience the same but haven't tried much

[soft zealot]

I stick with the numbers @high skiff used in his workflow and as long as you dont go far beyond the bounds of 1024x1024 I havent noticed any distortion issues.

Ie the same behaviour as SD1.5 or 2.5 , as long as you dont try to generate stupidly oversized images then you get little disrortion

I tend to (mostly)stick to these max generated sizes

2.39:1 1280 x 536, 1024 x 432, 768 x 322, 512 x 216 (Cinema/anamorphic)
3:2 1280 x 856, 1024 x 680, 768 x 512, 512 x 341 (Professional Photography)
4:3 1280 x 968, 1024 x 768, 768 x 576, 512 x 384 (General Pictures)
5:4 1280 x 1024, 1024 x 816, 768 x 516, 512 x 408
16:9 1280 x 720, 1024 x 576, 768 x 432, 512 x 288 (Wide Screen)
16:10 1280 x 800, 1024 x 640, 768 x 480, 512 x 320

[eternal fog]

Yeah these are fine. I'm talking about the sizes you can provide the SDXL CLIP encoder

You can mismatch sizes, which in my experience can cause strange issues when the aspect ratios don't match.

[royal fern]

left aspect ration match, right all 4096

[eternal fog]

Well the smaller ones aren't

[soft zealot]

I leave them hard coded at 4096x4096 and dont touch them

[eternal fog]

It's usually portraits where it's more noticeable. But it doesn't always happen. I was just answering the guy what my workflow was for the portrait image and why I'd changed it.

[soft zealot]

and providing i stick to the sizes I mentioned I see very little if any stretching

[eternal fog]

Changing those clip values can change the images drastically. It's quite interesting.

[royal fern]

not saying that you are wrong, I'm sure you did more experimenting than myself

I just reported what I was suggested

[eternal fog]

Yeah I saw that suggestion too. It's fine most of the time. I just like messing around with the values and I noticed that when I was getting stretched out people that helped.

[jolly creek]

ayy man you finally made it work 😄 congratulations

[spring fulcrum]

cause its awesome

and science

maybe

[royal fern]

and reasons

[spring fulcrum]

do you even lift

the case is so dirty but its a beast... ill take it to the car wash later and hose it down

[jolly creek]

I saw carpet and almost had heart attack. But then noticed that pc is lifted up.

[spring fulcrum]

omg you run it on your phone

lol

jk

[dusk mica]

i run sdxl on my fridge /s

[jolly creek]

I run it with pen and paper /s

[royal fern]

the bruh intensifies

[spring fulcrum]

lmao

must play battleshits

i run it on the apollo 11

[dusk mica]

wat

[spring fulcrum]

BRUH

potato power FTW

[dusk mica]

im glad you dont have to type these very long negative prompts anymore

(most cases=

dumb question, dues easyn work on sdxl in A111 or does it need to be retrained?

does easynegative work on sdxl

[tall lion]

soap water always works 🙂

[dusk mica]

i cant seem to be able to get them to show when using sdxl

[ionic dragon]

@eternal fog is there a way to link a text node to these nodes

so I can just type in the prompt once

[spring fulcrum]

These came out pretty good... yay space

I wish I knew how to train a model or embeddings or a LoRA this could be pretty cool to just load up when I want

[eternal fog]

Those are text nodes. They go into the SDXL clip encoders. If you always want the same text in both just connect 1 of them to both text inputs.

[halcyon tusk]

you can think of it as something close to img 2 img, the output from the base model gets sent there for detail and refining

[spring fulcrum]

Did you do these in SDXL???

I mean in ComfyUI with SDXL???

[craggy agate]

Just because I feel like I couldn't get this exact being right in any other ai so I'm excited sdxl can do this

[smoky patrol]

Oh, that's really cool

[indigo carbon]

nope I'm actually a genius. I managed to get it to wwork with A1111

[craggy agate]

Spoopy new creepypasta?

[spring fulcrum]

What did you do to get this to work in A1111?

[craggy agate]

There's a branch you clone

[halcyon tusk]

there is a branch that let's you load it

[indigo carbon]

okay, this isn't simple so bear with me- first step is to download this :https://huggingface.co/stabilityai/sdxl-vae

then change commandline_args to --opt-sdp-attention --xformers --theme dark --no-half-vae, well, you don't have to do all that, but you do need --no-half-vae

and that's pretty much it, it works WAY better than comfyui

It's not the nodes, it's the inference. A1111 has it's own magic tricks

[spring fulcrum]

I really like your images but I have to say I really like ComfyUI I have been able to get some pretty amazing results so far

I'm going to run your A1111 setup though and compare. I like good surprises

[halcyon tusk]

depends, you have a lot more control and margin for experimentation with the node system

[spring fulcrum]

I also like that it is so easy to share different workflows..... that part is miles ahead of A1111

just drag an image or load the .json file

[halcyon tusk]

yeah the image metadata stuff rocks

[indigo carbon]

Idk, I personally have a lot of experience with nodes. and SDXL works insanely better on A1111 for me. It's also much faster

[spring fulcrum]

I think I will still set it up and give it a try

[spring fulcrum]

For the SDXL-VAE what exactly am I downloading.... Just the sdxl_vae.safetensors file???

[indigo carbon]

yeah, add SD_VAE to quicksettingslist then select it.

[spring fulcrum]

Do I still need the base model and the refiner?

[indigo carbon]

just the base. it doesn't support the refiner just yet, but man, this is great

I recommend using hiresfix for now, it kinda does the job

[spring fulcrum]

Ok so just this file??? and where do I need to put this?

[indigo carbon]

yes

models/VAE

[spring fulcrum]

Ok and where is the quicksettings list you were talking about... Can you screenshot that?

[indigo carbon]

user interface/quicksettings

[spring fulcrum]

I did that and it still says failed to load model then it reverts back to another model

[indigo carbon]

disable all extensions except for crucial ones, happened to me as well.

[spring fulcrum]

are any of these crucial?

I disabled all of them and still no dice

[peak dove]

I use Vlad AUTO1111 for SDXL, and ordinary AUTOMATIC1111 for SD

[indigo carbon]

oh, what? this is weird

works for me

[peak dove]

AUTOMATIC1111 uses diffusers and not safetensors

Set the diffusers path in settings

[spring fulcrum]

Do I need to get the model without safetensors?

[peak dove]

U can d/load both diffuser and safetensor - diffuser goes in the stable-diffusion-diffusers folder - and then appears under Models drop down

Start VLAD/AUTO1111 using webui --medvram --backend diffusers

[indigo carbon]

bro are you blind?? the fans are on the back of the case

NO IT'S NOT. what kind of PCs do you know??? the air should go out of the radiator

plus, no one cares what PC you have. leave him alone

[spring fulcrum]

where do I put the command line arguments?

[peak dove]

Open a CMD Prompt, then open the stable-diffusion directory where A1111 is - then write in the Command Prompt webui --medvram --backend diffusers

[indigo carbon]

okay, I can immediately tell you are genZ

of course you don't

well, usually they do. people learn things. but no one here judges you before you start judging them. you kinda violated the server's rules.

[jolly creek]

cough cough

[indigo carbon]

okay, good luck I guess. I NEVER talked to someone that doesn't know what year they were born

[jolly creek]

Off topic

[spring fulcrum]

https://tenor.com/view/awkward-lol-gif-20709392

Well that escallated quickly

https://tenor.com/view/tgif-good-morning-gif-25334088

https://tenor.com/view/kitten-rawr-cute-punch-gif-14558168

I take no sides.... except the dark side of the force

What do you all think of these???

[spring fulcrum]

Do i still need the --medvram with a 4090?

hmm

--opt-sdp-attention doesn't work in vlad1111

[spring fulcrum]

Ok it runs with the SDXL model... Do I need any other files besides the Base Model or is it all good on just that one in Vlad?

[spring fulcrum]

That looks fantastic... Did you run the prompts I had?

[peak dove]

If its running - don't touch anything! It'll be OK ...

[indigo carbon]

yep

[spring fulcrum]

no VAEs or anything needed?

[peak dove]

but if it does crash, just delete VENV, and wait the 20 minutes it'll take to rebuild it 🙂

Like I say - if its up and running - leave it alone 🙂

Just rerunning some old MJ prompts thru ComfyUI

[strange mist]

did any of you get img2img to run on vlad1111? 👀

[peak dove]

I don't think that img2img is supported by SDXL - it is a whole different ballgame

[uncut steeple]

you're hilarious

[strange mist]

I think it can be done on comfy 🤔

[dusk mica]

sd next

[spring fulcrum]

Someone made a different version of A1111 it has some different features and supports SDXL

First few good results from Vlad A1111

I haven't used this interface before and haven't used A1111 in about 6 months... Can anyone tell me how I can upscale these or get these to be upscaled automatically as they are generated?

[high skiff]

Not the exact one, but one very closely linked ._.

[ionic dragon]

oh ok, did you talk with him so you can return it back to him?

[high skiff]

Likely gonna need to force it on him, but we'll see

I've been in contact the whole time

[ionic dragon]

ok, i hope you have done the transaction through paypal g&s

[high skiff]

I did, already escalated it to a proper claim

What?

[spring fulcrum]

he is not talking about you

[high skiff]

What are you talking about, you have nothing to do with this conversation lol

[strange mist]

he's mistaking you for someone else, or just trolling 🤷‍♂️

[high skiff]

Nep is always a troll, or at least I hope he's not serious lmao

[peak dove]

I think Nep is just spoiling for a fight!

Just ignore him ...

Yummy, this rum and raisin ice-cream is way cool - let me write up a prompt so all y'all can have some! 🙂

Troll prompt

[strange mist]

looks more ghost-y 👻

[peak dove]

She 8 all the eyes scream 😦

[strange mist]

eyes scream
👀

[strange mist]

tbf, if she suddenly showed up on my house, I'll give her all of my ice cream 😨

[peak dove]

I'd give ice-cream to these two ...

[spring fulcrum]

Does upscaling work in Vlad A1111 or did I set it up wrong???

[white hatch]

Just playing around

[peak dove]

Much less grief using Topaz GigaPixel imho - but that's just me!!! 😄

[inner kelp]

Trying out multiple colors , behind a textured glass, looks good till upscale.

[spring fulcrum]

I only have access to an old license for GigaPixel so I'm not sure if it will be able to do the upscaling I need

[peak dove]

Some freaky pix

SDXL allows me a one pass Upscale of 6x - so my 1024x1024 becomes 6144x6144. Without SDXL, it used to be a 6x followed by a 2.75x etc

But one day, and just for fun - I'd like to try an AUTOMATIC1111 Upscale/InPaint/OutPaint/ControlNet etc etc

[inner kelp]

More Glass experimentation.

One more

[peak dove]

Some Neon Flamingoes

Here is a wildcard Prompt = Stock Image ... it has amazing results

[boreal bough]

dont know how I misread your prompt. but here's "Neon Flamingo Evangelion"

[spring fulcrum]

What prompts are you using for these?

[indigo carbon]

a crazy old scientist with a long white beard holding a blue light, still from a fantasy movie, cg artist, the electric boy, by Adam Paquette, inspired by Rube Goldberg, brandishing cosmic weapon, 3 - d 8 k, juno promotional image, sad wizard, asian old skinny scientist with a big beard and beard, by Peter Mohrbacher

[hearty ginkgo]

Does sdxl support image to image

[west breach]

local yes

[spring fulcrum]

How would I use image to image in ComfyUI?

[indigo carbon]

A1111

[west breach]

add a load image node, then encode vae, then feed into the first sampler, set denoise to .25 or start step to above 0

[spring fulcrum]

do you have a workflow setup with this already?

[indigo carbon]

A1111 doesn't work with the refiner yet, but it's definitely the best way to use SDXL

for some reason it doesn't work for everyone, but when it will, that's going to be the meta

[west breach]

i'll make a quick one

[soft zealot]

google to the rescue 🙂

https://www.parents.com/parenting/better-parenting/style/generation-names-and-years-a-cheat-sheet-for-parents/

[sharp robin]

@west breach after testing the prompt node two things, when selecting no style the addition of a l clip box and output would indeed be handy. Also I guess this may be unfixable but when doing the styles since it is a wildcard it makes comfy ui start at the begging. Removing built in optimization. For ex before if you wanted to increase a samplers step you start at sampler step w the node tho it restarts the entire gen as it thinks it a whole completely different input. Otherwise everything working amazing thank you.

[west breach]

if you aren't using wildcards, set the control after generate to 'fixed' to stop it triggering the whole workflow

might be better to create another node without styles? then you can enter your own style prompts

[sharp robin]

Ahhhh. Wow. Much intelligence you possess.

[spring fulcrum]

I understood some of those words

[sharp robin]

Yeah that’s how it’s rigged up currently. It’s just a mild inconvenience to switch out the pipes

[west breach]

haha, dUo an I are talking about a custom node I made

[sharp robin]

It’s a badass node.

[spring fulcrum]

so that doesnt have anything to do with the JSON you provided?

[west breach]

no, that shouldn't require any custom nodes, I hope

[chrome furnace]

Hey peepz, can somebody give me a hint how to make an image perfectly round? I have an image of a mushroom from beneath and it is not round. Thx guys!

[soft zealot]

@west breach FYI this is not quite 16:9 as it equals 1.75

The correct measurements are 1368x768

[sharp robin]

Oh apart from an inconvenience it would unclutter two nodes

[west breach]

resolutions have been set to the closest multiple of 64

[soft zealot]

what was wrong with using 8 ?

[spring fulcrum]

Can you send a link to the upscaler model you use?

[soft zealot]

eg if I use your AR of 16:9 for creating wallpapers there are gaps each side of the desktop on my 16:9 monitor

[west breach]

do you have any? you can select a different one

[west breach]

https://huggingface.co/Shandypur/ESRGAN-4x-UltraSharp/tree/main

[spring fulcrum]

Thank you

[sharp robin]

U can blame SDXL. They trained it at these ratios.

[west breach]

Yeah it's using the resolutions listed in SDXL documentation. You can add your own by adding a user_ratios.json file to the root dir of comfyui

[spring fulcrum]

holy crap folks... I can run ComfyUI and VLAD A1111 at the same time and it doesn't crash... IDK how that works but its awesome

[soft zealot]

its like running excel and a web browswer at the same time.

They are 2 differenty programs

Its call multi tasking ;o)

[west breach]

just can't run batches at the same time

[soft zealot]

unless you have 2 GPUs and have one allocated to each program ;o)

[spring fulcrum]

ya but my graphics card still only has the same amount of vram to pull from

[peak dove]

Prompt = Stock Image

[tribal gale]

Diffusion models are trained in 64 increments which means there's no real 16:9, you can generate and crop to get your desired ratio

[soft zealot]

Ok , just I know that @wicked frigate I think it was thats done a A1111 ratio plugin thingy used multiples of 8 for that whcih can give you a true 16:0 (I havent checked how far out any othe others are)

[tribal gale]

Yeah there are workarounds, but you'll get the best quality when generating at the trained resolutions

[paper phoenix]

link me up im interested in that bud

[soft zealot]

https://github.com/thomasasfk/sd-webui-aspect-ratio-helper

[west breach]

@sharp robin new Prompt With SDXL node if you git pull.

[paper phoenix]

cheers mate

[sharp robin]

Ill do it in ab an hr and report back. Or a little later.

That looks like sex tho.

[west breach]

I have to go to bed anyway 😄

[sharp robin]

Goodnight

[west breach]

also added a new HaldCLUT node for applying film emulation to the final image @upbeat summit

[spring fulcrum]

before you go can you upload the image you have in that screenshot. that looks great

[west breach]

goodnight 🌜

[narrow seal]

Hello, does anyone know how to generate 16:9 images ? Without uncut?

[soft zealot]

this is either a very subtle troll or just coincdence lol

[narrow seal]

I'm not trolling ?

[soft zealot]

anyways I just bgenerate at 1368x768 for 16:9

some people may disagree on that ;o)

[trim orbit]

why would that be a troll? you got bad juju.

@narrow seal different aspect ratios aren't always possible on web services. try using dreamstudio, they've got aspect ratio settings.

[narrow seal]

So it's not possible on web

[trim orbit]

the website you use doesn't appear to allow aspect ratio changes. dreamstudio does

[narrow seal]

Shame i have just bought pro

[soft zealot]

hey I did say "or just coincidence 🙂

It was just mildly amusing that only 10 minutes earlier we had this

#✨｜sdxl message

[trim orbit]

is that clipdrop?

[narrow seal]

Yes it is

[pure moss]

/dreamdog

[narrow seal]

I thought that channel was for clip drop stable diffusion

[trim orbit]

it's weird that they keep the image generation limited to squares without negative prompts, but they're still pretty good. they've got all the other tools you can use with pro too.

[trim orbit]

naw this server is for all things stablity and this channel is for all things sdxl. clipdrop is a site stability owns

[narrow seal]

You can turn your square photo to 16/9 but it regenerates sides, and it doesn't give good results

[high skiff]

the SDXL bots are looking pretty solid ATM

Very excited to see how much more quality I will be able to squeeze out on launch

[trim orbit]

i dont think uncrop uses sdxl yet

[narrow seal]

Guess it's sdxl 0.9

Are you guys using 1.0 somewhere there

Well i have no idea what I'm talking about too. Just saw clip drop, images were good and bought the pro plan

[vale eagle]

You could try the bots which are 1.0 candidates

[indigo carbon]

yeah, the bot uses SDXLa SDXLb and SDXLc

[high skiff]

you can use the 1.0 candidates for free in this server in one of the bot channels

[trim orbit]

uncrop is the tool they have to extend edges and stuff. it doesn't use sdxl 0.9. it uses 2.1 i think

[narrow seal]

Unlimited generation? Or does it have limit

[high skiff]

unlimited

[indigo carbon]

I think one of them is better than the other two

[narrow seal]

Damn i have paid for nothing then

[high skiff]

for sure, or one is really bad of the 3, cause I keep seeing some ugly ass gens

[trim orbit]

you wouldn't be able to tell. results are randomized. you'd have no way to know if a result comes from a certain model.

[high skiff]

one of the models is like super HDR look, and it looks crusty and gross lol

[trim orbit]

that way people can't cheer on their biases

[high skiff]

all I know is some of the images look consistently as bad almost all of the time

[indigo carbon]

Idk man, I think I can slightly tell, even with random settings

[trim orbit]

that's just confirmation bias

we all believe that our beliefs are true and accurate

[wicked frigate]

x8 is required due to how the code works, x64 is sometimes preferable just because the model was trained on x64 inputs so it tends to be kinder to those

[dense chasm]

anyone has a clue how to generate the Hollywood director Christopher Nolan profile looks like himself

i just confused sdxl can't work well on a lot of celebs image

realistic people is unreal

[high skiff]

Like a sort of celeb and protection thing

[trim orbit]

love these tanks it making. this lora is working well. i think i'll publish on civit on sunday, my next day off. Want to try a few more things with it. Will take these findings into the 1.0 version

[spring fulcrum]

a new master of Monopoly.... Dare to play... try passing go with this guy

[high skiff]

one of the models, or settings combinations that happens pretty frequently looks like shit lmao

what is thissss

[high skiff]

I am not sure what is up with whatever is happening with the hyper overly tonal contrasted images lol

the tank looks dope, the rest looks crusty

somethign feels very weird with the SDXL bots at times haha

[trim orbit]

the lora uses starcraft unit images to train. i cut the background out of 90% of the images. so if i don't prompt a backgorund it geneartes against grey. yeah the background is just "grassy meadow" and it do look jank. that aint my lora, that's just sdxl being bad at grassy meadows

[high skiff]

In that case I'd assume its just bad prompting

genuinely what is happening with the tonal contrast lol

I hope they release info on why this looks so shit so we can steer clear

[steady grove]

That's my assumption too. I still gotta up my game with prompts. I don't have a bunch of saved clips for xl yet. Might cobble a cheat sheet together on sunday

Prompting on xl is sooooo weird and different

Anyone got tips?

[high skiff]

I already have some documentation on it, tho it looks like 1.0 is gonna make all of the community info useless tho

[trim orbit]

same issues for me. i just bang out prompts to get a test out. no real thought to them other than how do i draw out the lora best?

[trim orbit]

naw it still uses the same two clip layers. the prompting methodology should still have a ton of relevance

[high skiff]

we will have to see

[gentle mirage]

Oh shi any mod here? I made a prompt to bot-1 that doesn't comply with the rule #4. Can anyone delete?

[dense chasm]

any suggestion about the adjustment?have u tried with modify workflow or different prompts?

[gentle mirage]

Sorry didn't realize there was a gore restrictions, though it was for nudity

[upbeat summit]

greetings

[high skiff]

At the moment I am extremely discouraged from all of my SDXL work, so I am on hiatus right now. Trying not to be too down about the whole thing

[trim orbit]

you got a bit of a sky is falling view on 1.0. maybe you've been listening to pseudo a bit too much

[trim orbit]

that guy seems like a bit of an anchor

[high skiff]

Pseudo and I know a lot more about the way SDXL works than a lot of people do, and it has a lot of core issues, but they shouldn't be things we couldn't fix honestly. Its monumentally better than any previous SD version, even with its issues

And I don't have a sky is falling view on SDXL, I am actually very excited, which is why I am upset because I will be missing the start wave of interest in SDXL

[peak dove]

Magritte and Varo - The Multiverse, Snow Crash, in the style of rene magritte and remedios varo

[trim orbit]

thats a vague statement. sdxl is so niche that of course anyone with interest will know more than most people.

being discouraged based on conjecture is just a personal choice. i'm sorry to hear you've gone down that road. hope ot see you on the other side

[high skiff]

Did you even read my message?

like, genuinely

cause it seems like you just brushed it off lol

[trim orbit]

oh no. "do you even read?" is not a reply that bodes well. ttyl sytan. enjoy your friday afternoon

or morning whatever it is where you at

[high skiff]

Bro, I just said I am not upset at all with SDXL and you are using this condascending tone of "that's a shame to see you be so stupid"

Like, I legit just said I am only discouaged cause I don't have the means to do the research I had planned anymore lmao

If you think I have some negative view towards SDXL after pouring almost every waking moment of my time into it, IDK what to say

Like yeah, SDXL has some issues, but everything has issues, and I lookforward to working with people to try and fix/lessen the effects of them

[trim orbit]

yeah. thats what i said. that you hate sdxl... 🙄

ngl. these conjecture spirals are dumb

[wicked frigate]

my meowmeows! i love them 😻

[high skiff]

"You being discouraged based on conjecture is just a personal choice. I'm sorry to hear you've gone down that road"

Anyways, I haven't, there was legit no conjecture at all in my statement, but ok

[wicked frigate]

I gotta convince the team to make SDXL 2.0 an exclusive catpic generation megamodel. It's what's best for society,,

[visual forge]

hmm, probably having a common issue, but I'm trying SDXL in A1111 and getting either black image gens with the vae on HF or off-color images with an older vae

[indigo carbon]

actually, that is without the refiner. and it's still might be the best way to run SDXL(A1111)

[trim orbit]

industrial revolution began when we started sharing cat paintings. coincidence?

[high skiff]

Reason I am discouraged = I was sold a faulty GPU and will now miss all of the research I had planned for the release of SDXL

reason I am NOT discourged = some weird conjecture thing you are assuming from me @trim orbit

[hearty ginkgo]

He never actually said he hates sdxl

[high skiff]

SDXL rocks, even just 0.9. I have been able to do some amazing things with it already, which is WHY I am upset, cause I lost the means to do so by being scammed by somebody

[trim orbit]

I was speaking towards this context. you never brought your gpu up this morning, and you should really not assume that's the primary focus of a topic in #✨｜sdxl

[wicked frigate]

use the VAE that's in the model (in auto that's, uh, turning VAE off, since the auto VAE option is actually a VAE replacement option). Also make sure you launch with --no-half-vae

[visual forge]

will give it a shot thanks

[indigo carbon]

also, when does A1111 implement the refiner?

[peak dove]

^..^<

[high skiff]

I mean, its not conjecture if its something I have live tested

Prompting in 1.0 will be different for sure, but I am not sure just how different

It looks like it could be different in a positive way

[peak dove]

SD.NEXT (VLAD) AUTOMATIC1111 works with refiner ...

[high skiff]

but comparing identical gens from just a week ago, images don't look even remotely the same now, which means you need to prompt a different way for the same output, which is fundementally a change

but I think it could be one for the better honestly

it seems to listen a little more (to its own detriment often honestly)

[indigo carbon]

SD.NEXT is a bad UI. I tried making identical images using a 1.5 model with both, the A1111 version was WAY better.

[trim orbit]

thanks i know that. hyperboles are fun for dramatic effect. the flippant tone of that post is conveyed well, at least i thought so, with the eye rolling emoji. maybe i could add a /me rolls eyes extraordinarily command to get that across better? next time i'll try. thanks for chiming in on this "doesn't matter at all" topic

[high skiff]

alright, so you aren't being an ass to exclusively, dope. Good to know for the block note

[trim orbit]

it's not really a contest. they're both just gradio ui's that have beta support for sdxl. automatic's doesn't prompt sdxl right either.

[peak dove]

I use Vlad for SDXL (and ComfyUI); and I use GRadio AUTOMATIC1111 just for SD

[high skiff]

I kinda wanna try A1111 for SDXL, just to run some direct comparisons sometimes against mixed diffusion

[indigo carbon]

I was talking about normal models. A1111 is by far the best

[visual glade]

I would disagree very much with that most people can't even figure out how to install it

[trim orbit]

available on the release_candidate branch today. git switch release_candidate

[peak dove]

I have used ComfyUI and Vlad A1111 - image quality is equal ...

[visual glade]

people keep forgetting about all the other user friendly uis

[peak dove]

... in SDXL

[high skiff]

its not very hard at all, I have had people who have like no program or SD experience install it in like 10 minutes

[high skiff]

yeah, but what about compared to partial diffusion?

[trim orbit]

that's just selection bias then. both do 1.5 "normal models" the same and if they aren't, you don't have the settings tuned to match each other. likely different vaes

[indigo carbon]

idk, I tried both- A1111 was insanely faster and better, and it didn't even use the refiner.

[peak dove]

A1111 has a CFG up to 30 - ComfyUI up to 100 - so CFG 10 in A1111 is actually CFG 30 in ComfyUI?

[trim orbit]

they're both just gradio ui's and arguing about what does 1.5 "normal models" in #✨｜sdxl makes zero sense

[visual forge]

That did it, thanks. I set the VAE setting to "None" and not "Automatic" in this case (just noting for clarity and for anyone looking later)

[visual glade]

a1111 is not faster, at best it's going to be the same speed as comfyui

[high skiff]

no, they are the same, just limitedin range as over 30 is basically never used

[high skiff]

and that is pretty optimistic to expect as well

[indigo carbon]

man, it was 20% faster. I tested both

[trim orbit]

configuration layer 8 issue

[high skiff]

I run A1111 and comfy together now (I use A1111 for inpaint cause it works so much better for me over comfy), and A1111 is always at least 10% slower for 1.5

[indigo carbon]

then your inference isn't optimized

[high skiff]

what do you mean?

[trim orbit]

compared to comfy? if comfy is slow for you, then you're doing somethign wrong

[high skiff]

comfy is a much lighter install, it makes sense that auto would have some bloat and slow it down a bit

[peak dove]

See my models list ...

[high skiff]

comfy is actually much faster than auto now with that BF16 VAE. That thing is black magic

[trim orbit]

among teh popular UI's that people can install easily, comfyui is unequivocally the fastest out there

[indigo carbon]

Idk man. I tested all settings with both, A1111 was much better and faster

[high skiff]

that seems like a physical error, there is no reason comfy would be slower otherwise

[trim orbit]

configuration issue then. telling you man

[high skiff]

Comfy has a much more efficient way of diffusing

[trim orbit]

theres no reason comfy should be slow

[peak dove]

ComfyUI was up and running 30 seconds after it landed on my desktop. A1111 I think arrived on a Wednesday, (or was it a saturday)?

[visual glade]

20% sounds like the difference between SD1.5 vs SDXL at 1024x1024

[trim orbit]

your testing is flawed

[high skiff]

I am gonna have to agree with flowwolf here right now

[dense chasm]

this is a good profile,but not the realistic Christopher Nolan i wanted

[trim orbit]

i think people are using default values in automatic. the 512x512 width and height

[peak dove]

SDXL is 4 x (512x512) - so on my 8Gb VRAM machine, it can take about 4 times as long as SD

[visual glade]

or people comparing samplers that call the unet twice on comfy vs euler on a1111

[high skiff]

yeah, or maybe an inefficient node setup to achieve the same results

[peak dove]

Yes, some models make a hi res pass

[indigo carbon]

idk. I'm making images like this with SDXL on A1111 with 6.7it/s =\

[high skiff]

on what GPU?

[indigo carbon]

4070TI

[high skiff]

that seems weird

[dense chasm]

actually it will be fast on automaic than comfyUI

[trim orbit]

i love that eular is popularized as a word in the lexicon again. 1) because it sounds like Bueller, and 2) because Euler discs are frigging cool. i won't pretend to know anything about the mathematician himself though. maybe he was cool. probably just some nerd though.

[peak dove]

(Someone anonymously post a ComfyUI picture, and an A1111 picture ... we can then have a guessing game over 10 items ... winner gets an nVidai A6000

...an A6000 which has done a year of data mining 😄

[trim orbit]

on that not, flowwolf is out for the day. it's friday! you know i gotta get down on friday

[high skiff]

sonds like Bueller? They are pronounced way different (which I learned to my own dismay lmao)

Euler = Oiler

[trim orbit]

thats just the correct way of saying it . like that matters

[lusty raptor]

Euler was pretty cool though, gave us lots of handy math

[indigo carbon]

idk, whatever my A1111 inference is it's way better and faster than comfy is for me =.

[high skiff]

fair enough

[peak dove]

As no-one said to his girlfriend "eulerv me?"

[high skiff]

I know its Oiler now, and I still say you'll-er lol

[trim orbit]

everyone says it "eular" like it's spelled. eular. eular. eular. see.

[peak dove]

Euler, Euler, Ferris Buehler!!!!!

[peak dove]

Is this A1111 SDXL? But not the Vlad version?

[indigo carbon]

Euler isn't that good. In my opinion DPM++ 2M Karras is best

[visual forge]

took too many math classes where you'd get chalk thrown at you to call it "you'll-er" 😛

[indigo carbon]

yeah

[high skiff]

yeah haha

I only found out recently 😅

[peak dove]

My favourite models are most of them 'cept DDIM

[indigo carbon]

SDXL doesn't work with DDIM

[high skiff]

I find it interesting how many people are against DDIM, when its won my blind vote for quality in this server basically every time for SDXL-

[trim orbit]

i use ddim cause i'ts hella fast

[visual glade]

sdxl works fine with ddim

[peak dove]

Heun, Uni-PC, Euler and Euler A, Karras ++2M,

[visual glade]

there's no reason it shouldn't work

[trim orbit]

then once i find a prompt i switch to something that cooks longer

okay now i'm really out. ☮️

[high skiff]

although, I will say that Pseudo did just have a big discovery with an issue with DDIM, and I think he fixed it, but don'tquote me on that

[peak dove]

I got a very faint kind of sketchy outline using DDIM ... but no, it's not good

In A1111 at least

[indigo carbon]

I don't have any problems with DDIM. but DPM++ 2M Karras is just as good, and it's slightly better with detail

[high skiff]

my whole workflow was made specifically to work with DDIM, tho I have a big feeling the sampler preferences will change with 1.0 cause of how different it is

[agile quarry]

is there a ddim_eta setting in comfy? its hidden away in a1111 but i like it allot adjusting between .7-.9

[dusk mica]

comfy feels like magnitudes faster than a1111 for me

[high skiff]

this has to be one of the most fucked SDXL images I have ever seen lmao

I am at a loss for words lol

[visual glade]

you can change ddim_eta in the code but it's not exposed in the ui

[peak dove]

Blondie SDXL 0.9 ComfyUI

The Weirdz

[agile quarry]

so i can try to make myself a node ?

[high skiff]

I heard that Kohya optimized his SDXL training scripts, so maybe I am not completely out of hope for my research

[peak dove]

I got myself a beef and curry node at Lidl

[high skiff]

although my tests are likely to be way slower now unfortunately

[peak dove]

Node?! Noodle 😄

[high skiff]

Nodette

[high skiff]

yup

I have heard its very easy

[hearty ginkgo]

How much optimized?

[high skiff]

no idea, hopefully enough for me to not be completely screwed on my research

[peak dove]

I'm going to wait until 1.0 before I try that Lora - First Ever SDXL Training With Kohya LoRA

[indigo carbon]

Isn't SAI going to release specialized finetuning tools with SDXL1.0?

[high skiff]

no idea

[spring fulcrum]

@indigo carbon Check these out

[high skiff]

they said they are, but I wouldn't hold my breath

they could release tools, but that doesn't mean they are gonna work/be good

[hearty ginkgo]

We're you not able to use the 3090 or 4090 I forgot what u had?

[high skiff]

its ruined, completely dead

[hearty ginkgo]

Wait when did that happen?

[high skiff]

they sold me a scam GPU, and now I am fucked for like the next month

[hearty ginkgo]

We're you overclocking?

[high skiff]

no, the GPU itself was already dead

he sold me a faulty card

either that or it miraculously died in transit, in the original box, with an antistatic bag, wrapped in bubble wrap inside another bigger box with no cosmetic damage

[peak dove]

Carz

[lusty raptor]

wait, what did you do with your old gpu?

[high skiff]

I still have it, but it can't do what I had planned

[hearty ginkgo]

What did you get it off?*

[high skiff]

I dropped all the money I had (and a lot of money I didn't) on that 3090, and now I am completely tied up until this jackass gives me my money back in paypal claims

[lusty raptor]

ah, sucks, but at least you're not totally hosed

[high skiff]

for my research I am

but oh well, that ship has sailed

[hearty ginkgo]

Ebay paypal?

[high skiff]

no, paypal goods and services

they are on my side ATM

but I need him to not fight it forever so I can get a new GPU ASAP, otherwise all he is doing is hurting me for no reason

[hearty ginkgo]

Since it's goods and service you should not have that hard of a time it's not like you did friends and family

[high skiff]

its not really the hard time thats the problem, its that its gonna take forever and I am gonna get a GPU WAY after the launch of SDXL when my research will be meaningless

[hearty ginkgo]

Could always be worse

[spring fulcrum]

it could be snowing

[dawn cipher]

I'm hoping the zero terminal SNR model wins, so that offset noise isn't needed. It's difficult to train with offset noise.

[high skiff]

that's kinda a toxic mindset

Discounting all issues as "it could be worse" is a good way to get complacent

[hearty ginkgo]

Didn't mean it in a toxic comment just ingeneral

[peak dove]

Teddy Bar Samurai

[autumn forum]

I had a car I was gunna sell to a friend for 13,000$ and it burnt into flames as he was driving it home before he had paid me cuz I was being nice and I was in a different state and afterwords he wouldn’t pay be because he said “it was the cars fault” even though his friend test drove it and did fine. Now I’m out 13k. Lawyers said it’s to hard of a case to handle and unfortunately would probably go more into his favor.

[high skiff]

I'm not trying to say it can't be worse, cause of course it can, but that doesn't discount that I have been royally taken advantage of while trying to do good for myself and others.

I am not even out $600 right now, my mom is, and thats not fair to her either

[spring fulcrum]

Are you a wizard?

[autumn forum]

Let’s hope it gets resolved soon!

[high skiff]

I am strong arming the fuck out of the seller, cause I am done

[visual forge]

holy crap, I really like the way SDXL generates so far.

[high skiff]

I am giving him the option to just fold early, or I will burry him in proof and evidence in that paypal claims court

[autumn forum]

It’s very good!

[visual forge]

I'm just playing around with old prompts in 1.5 style, is there any best practices for XL? I feel like with SD 1.5 I gradually worked my CFG down all the way to like low 4s to get anything good out of them

[high skiff]

I don't recommend using prompts from 1.5 for SDXL

you will likely be hindering results that way

[eternal fog]

@high skiff what was your old GPU?

[indigo carbon]

yeah, SDXL prompting is way easier than 1.5

[visual forge]

How so? (I just started using this 20 mins ago so steer me correctly )

[high skiff]

MSI 3080 Trio X

[eternal fog]

You can train LoRAs at least then

[high skiff]

in 1.5 you use tags, in SDXL, it plays a lot better with linguistic prompting

1.5: Dog, grass, sunset, field, detailed, photograph

SDXL: A photograph of a dog in a grassy field at sunset

it also allows you to do more specific prompting a lot easier

[visual forge]

oh, I always wrote slightly more narrative tags anyway (rather than the booru style)

[high skiff]

tho its not as consistent as it should be at the moment

[visual forge]

and then a lot of attention pushing

[high skiff]

there is a little hope for my research not being completely DOA, but not having access to a proper GPU to test my theories is gonna hurt my findings for sure

[high skiff]

my biggest limit now is time, as training a LoRA on the 3080 will be way slower than on the 3090

[high skiff]

weezord

[indigo carbon]

oh yeah, lora:add_detail would be insane with SDXL

[dusk mica]

wonder if one day text2img will be combined with text generation

[high skiff]

my fractional offset kinda already does that built into SDXL's pipeline

[eternal fog]

It's not too bad, depends what you are wanting to do. Something like 1200 steps takes about an hour with a decent sized dataset.

[high skiff]

yeah, hopefully I can find better settings that I was using

[eternal fog]

And it seems like LoRAs overcook really quickly anyway

[high skiff]

cause even on the 3090, that Na'vi LORA took 2 hours, so likely closer to like 3 on the 3080

[eternal fog]

So you only need like 600 steps or less

[high skiff]

compared to 1.5, no way

you can be very brutal with SDXL compared to 1.5

let me do some math, just a sec

[eternal fog]

Interesting, maybe that's because of the blue skin. Everything I've tested it was done within 30 minutes.

Well "done"

[indigo carbon]

oh yeah, finetuning SDXL should be way easier than 1.5, SAI said they will release tools capable of that.

[high skiff]

I trained my Na'vi LoRA for 6900 steps

[high skiff]

only question is, will we be able to use them

[eternal fog]

I feel like if I did that long with my stuff it would majorly overcook

[indigo carbon]

likely, yes

[eternal fog]

I've seen people recommending wildly different training settings though

I've been using mcmonkeys

[high skiff]

mine actually did end up being pretty overcooked, but I found better results from under-driving the overcooked LoRA than using the properly saturated one

[nimble heart]

so it's like those precooked chicken strips you thaw out and microwave

[high skiff]

the results came out pretty damn good by using the overcooked LoRA at .5

left is raw SDXL, right is with the same LoRA trained into 1.5 as a refiner pass

[steady grove]

I get great results from half powered Loras I train too

[eternal fog]

How does SDXL do avatar stuff without a LoRA?

[high skiff]

terribly lol

it just doesn't lol

no LoRA

[steady grove]

It tries if you add tons of emphasis and contextual tokens

[eternal fog]

I've found it very difficult, even impossible, to get "alien" skin colours

[high skiff]

and that's the least bad one lol

[steady grove]

But generally sucks

[eternal fog]

I tried to make a red skinned demon yesterday

And it kept either putting them in red skin tight clothes, or would cover them in blood

[high skiff]

even with my LoRA, you have to strengthen it with the term "cyan skin", which helps immensely

[steady grove]

Was going to say that's not half bad attempt in my experience. Of course it's your best case finding lol

[nimble heart]

specifying man or woman immediately tries to make it human. I found "humanoid" is a decent alternative that lets me make lizard people somewhat consistently

[steady grove]

Loras work really well to nudge existing knowledge into the front more. Like, big boobs. Holy I get humongous boobs if I train 5 pics of big boobed ladies.

[high skiff]

[eternal fog]

Interesting, I'll give that a go.

[high skiff]

so yeah, the LoRA is a big success lol

[steady grove]

Looks like the guy from Jack Reacher's avatar

[high skiff]

oh, cool side effect

Cause it still didn't pick up the blue skin so well, you can actually do different color skins

I got purple one time, but I deleted it to save some space for a grid

[eternal fog]

Can you make a blue skinned alien dog with it?

[high skiff]

but yeah, overall it works really good

[steady grove]

Next hardware acquisition, a NAS for grids

[high skiff]

not in the capacity to be running SDXL on my PC at teh moment unfortunately

[eternal fog]

😦

[high skiff]

got to get everything back together from the hell that was fucking around with this broken 3090

my 3080 is in, with no drivers ATM lol

[nimble heart]

Looks like it's tonemapped like the first movie. Wonder if a new dataset consisting only of stills from the 2md movie would improve quality some.

[high skiff]

anyways, I have chores to do, I will come back and test some more with the LoRA when I can

[high skiff]

I tone mapped the images to fix that, these images were just early and bad prompting

[vale vapor]

gaddamn this is good....how has img2img been? sadly i'm not getting any great mind blowing results. it looks like images sd1.5 models can reproduce, i'm seeing no leap in image quality sadly

[high skiff]

you can see I improved the tones a lot more later on

[steady grove]

Jacked

[high skiff]

I have a type 😅

without the cyan skin reinforcer lol

unique effect lol

but like I said, it should be able to do any color skin cause of that

even was able to pick up some of the headware

[nimble heart]

and the neck radios on the earlier image

[high skiff]

it can also blend with specific people if prompted properly

[steady grove]

I've notice that Loras get more of the smaller details brought into xl

[high skiff]

got it to blend with Taylor Launter

[steady grove]

Very well too

[high skiff]

I ahve theories on how to improve results, just scared to try and train a LoRA again and get burned

[crystal jungle]

I got stable diffusion earlier but for some reason everything I generate is trash

What're the best settings?

[vale vapor]

anyone getting these high quality results using img2img in sdxl?

[quasi remnant]

define trash and what model are you using

[crystal jungle]

Model?

Where do I see that

I'm new to it

[quasi remnant]

what checkpoint are you using?

[crystal jungle]

oh

sdv14

[quasi remnant]

that's really old and doesn't gen good stuff typically

[crystal jungle]

ohh

so will it take a really long time to update it?

[quasi remnant]

for sd 1.4-1.5 models you wanna use more finetuned models

depending on what you're trying to make

but the one that gens the best stuff right now is the SDXL model

[crystal jungle]

i can send a reference image of what im doing to make quality wise

[crystal jungle]

Could I send you it in DMs, it's a little NSFW but I don't wanna make NSFW but the quality of it is really good

[quasi remnant]

sd 1.4 isn't gonna produce good results

you're gonna want XL model or a finetuned model

[crystal jungle]

what would you say is the best?

for a medium end pc

[quasi remnant]

define medium end

[crystal jungle]

1650 ti

[quasi remnant]

that's low end

[crystal jungle]

oh

[quasi remnant]

you'd wanna look up finetuned 1.4-1.5 models and probs stick to like 640x640 if you're even able to handle that

you barely have any VRAM iirc

[indigo carbon]

WTF

i prompted "mcdonald's"

wtf

this is SDXL's weak point

[visual forge]

fast food?

[quasi remnant]

try more than one word? i don't know any model that does good with just one word

like "A mcdonald's restaurant" or something

[hidden lily]

how can i merge sdxl model?

[quasi remnant]

find another model finetuned on SDXL and merge it with that? something like Waifu Diffusion XL for instance?

[wicked frigate]

... how many imgur dead links have leaked into the training set, oof

[quasi remnant]

in the same way you can't merge 2.1 models with 1.4-1.5 i think you need to merge sdxl models with other sdxl models

[eternal fog]

I wonder, was your training set collected before or after Imgur did that big purge

[azure oxide]

Ooo that's spooky

[hidden lily]

I have many sdxl models and I want to merge them

[eternal fog]

Many? There's about 3 public ones lol

[hidden lily]

sdxl
sdxl refiner
wdxl
not waifu diffusion
art diffusion
dreamshaper xl

[eternal fog]

Hmm so I tried this and it works ok, but as soon as I try to make them a particular gender, it just completely ruins it and then they are just human with red clothes again 😦

[nimble heart]

add gender specific features at a very low weight

[eternal fog]

You wouldn't want to merge the refiner, I'm not sure it's even possible.

[quasi remnant]

don't merge the refiner

[hidden lily]

okay. I just want to merge other models. is it possible?

[indigo carbon]

does that exist?

[quasi remnant]

only a kind of like alpha version of it. https://huggingface.co/hakurei/waifu-diffusion-xl

[indigo carbon]

oh, cool

I didn't know we started finetuning SDXL already

what other SDXL finetunes are there?

[eternal fog]

Some people got it early. That waifu one is just a test.

[eternal fog]

Yeah it doesn't want to do it. Adding female even at weight 0.1 instantly just makes it a normal woman

[quasi remnant]

i don't see why it shouldn't be, but it might be more intensive than merging other models since XL models are huge

[azure oxide]

Is there any reason you want to merge dreamshaperxl with sdxl when it was trained on it?

It would be equivalent to merging dreamshaper7 with the base 1.5 model lol

[hidden lily]

with wdxl

[azure oxide]

I guess you could do it for the ollobrains approach where you merge literally everything for fun lmaoo

[barren olive]

wait there is already dreamsharperxl??

wow that was fast. is it official or fanmade?

[azure oxide]

I can already see him merging mega model with sdxl 1.0 when it releases

[eternal fog]

oh no, not that guy

[azure oxide]

Hahahahah

[nimble heart]

I meant like features that implies female without saying it. On 1.5 tunes "feminine" worked okay. Think "midriff" worked as well since it had booru data.
Something similar probably works on SDXL. I made a lizard woman once but I didn't save the image so i forgor

as in proper lizard fully scaly no ears etc. like the government agent kinda lizard

[livid cradle]

i have a technical question if anyone is willing to help me out:

So i got this product mockups. that i want placed on various backgounds.. how do i tell SDXL the type of product is in the image? im using image to image masking function to generate the background for it.

i know the image is placed there, but can it know to make the background fit? how do i achieve this magic? 😄

[indigo carbon]

holy shit, it does: https://tensor.art/models/617046080397449350

that's dreamshaper SDXL

[livid cradle]

sometimess i get lucky depending on the prompt, but sometimes, i wish it could of generated more content aware background with the image im feeding it.

is he also considering the type of product and sizes that is beneth the mask?

im building a product photography application based on SDXL API and can't figure some stuff out.

besides the fact that sometimes it changes the proiduct under the mask for no particular reason... even though mask is full black.

[azure oxide]

mabey inpainting is the route u want to go? alternatively you could try only generating the background and then edit the product in the old fashioned way with an image editor

[eternal fog]

I've managed to get 1 out of 100 tries and I can't post it because they are naked lmao

[nimble heart]

If you'rre trying to force a specific color instead of just species mixing that might be harder

[eternal fog]

I just want red skin lol

[nimble heart]

d e m o n

[eternal fog]

Yeah that doesn't work

[nimble heart]

fuck

devil

[eternal fog]

As soon as you mention anything to do with a woman

It just stops even trying

Even just clothes

[nimble heart]

sdxl needs to touch grass confirmed

My demon women have blue skin one my bench prompt though

think I had "body paint" or something in it

[eternal fog]

One thing it's doing which I understand why, but it's annoying. Is making normal colour skin, but then with very very sharp red lighting

[nimble heart]

which it just kinda ignored and made the skin itself colored

[nimble heart]

lighting so sharp it turns the skin red ig

[eternal fog]

Yeah but you can tell it's from the light

And the skin is still white or black

[sudden cliff]

I can get fantasy skin colors but then the whole scene is that color

[nimble heart]

I'm updating one of my libraries but after I'll test those non-human prompts a bit. I had it working super well in 1.5 tunes like realisitc vision and dreamshaper

[eternal fog]

Yeah it's doing that too

[nimble heart]

if comfy has BREAK syntax support could try that

[eternal fog]

Don't believe it does

[nimble heart]

hm

[eternal fog]

Every single time it gets it close, they have no top on lol

[nimble heart]

maybe skip the color on CLIP L but that might make it harder to get at all

[eternal fog]

colour is only on CLIP G

It sort of does dudes, but it looks like a guy in a mask lol

This is the closest I've got when trying to make them female

But there's always normal skin on it somewhere

[high skiff]

I am back

alright, my PC is setup with the 3080 installed and working drivers

3090 has been banished to the shadow realm, as I can't even stand to look at it ._.

[green python]

no ai other than parti has benn able to do this, wow

[spring fulcrum]

SDXL in ComfyUI no specific model

[high skiff]

alright, time to test my na'vi LoRA more

[upbeat summit]

have fun 🙂

[upbeat summit]

right is slightly sharper?

[high skiff]

right is with a very minor 1.5 pass

0.05 denoise with the same LoRA trained into it

[upbeat summit]

coherence is great in both

[high skiff]

yeah, SDXL Learns a lot for sure, especially since this is not with a text encoder, just the U-Net

[upbeat summit]

and according to a new reddit post by stability, training SDXL 1.0 will be a lot better / easier

[high skiff]

Its tsill trying to learn the ears, so I will be adding more images into the dataset for later

[spark bear]

looks good, but too close to the actual movie character so likely heavily overfitted and spitting out stuff close to the dataset

[high skiff]

she looks fairly different to the actual movie, and I am prompting simply

thats what she looks like in the movie

so there is already a lot of facial structure variation

[spark bear]

looks similar, with a bit of that "handsome squidward" thing sdxl likes to do to faces, with a stong jawline and high cheek bones

[agile quarry]

Can I merge sdxl with my Lora? Right now i have to choose LoRa (lyco actually I think) or refiner. With both I crash . Last time I tried at least.
I tried in merging them in comfy but I get an error and idk if I'm doing it right. I set sdxl as both models and attach the Lora to the second checkpoint loader. I set the merge model node to 1. To just copy the second checkpoint but it errors

[high skiff]

@spark bear

there you go, prompting for some different ethnicities

I will likely be able to generate more racial combinations, and then train these back into the LoRA

[paper phoenix]

youre better waiting for 1.0 and training again.

[high skiff]

for sure

I am just dicking around right now, learning some stuff

might change with 1.0

[eternal fog]

There was a reddit post earlier from one of the SAI Devs that implied LoRA Training was easier on 1.0

But they'd also trained for an insane amount of steps so

¯_(ツ)_/¯

[upbeat summit]

https://old.reddit.com/r/StableDiffusion/comments/1557jsx/sdxl_10_loras_are_way_easier_to_train_than_09/

[eternal fog]

Yeah that's the one

50 epochs on 400 - 1200 images lol

Seems massively excessive

[high skiff]

depends on how many repeats

[upbeat summit]

I hope I can still make TI embeddings with 10-15 images 🙂

[high skiff]

if you train like how Caith does, thats not much at all

[eternal fog]

What do you mean ?

[high skiff]

30 epochs on 50 images how caith does it is like 20 minutes

[eternal fog]

On what GPU?

[high skiff]

@boreal bough Care to elaborate? I don't wanna give away your info

[eternal fog]

I mean you just don't need that many

I've done 6 epochs on a image set of about 150 and it worked fine

50 epochs on 1200 images seems silly

[high skiff]

if you knew what his method was, you would understand why he does that, tho i don't wanna give away his info on it

its very interesting the way he does it

[eternal fog]

And it would take hours, even on a decent GPU

[high skiff]

it produces good results, just in a very different way

[high skiff]

again, not the way he does it

like I said, 30 epochs on 50 images in SDXL would take like 20 minutes on a 3090

[eternal fog]

Saying, not the way he does it isn't helpful lol

[high skiff]

Thats cause I don't wanna share his info if he doesn't wanna share it himself. Just know that it doesn't work how you would assume, and there is really no speed hit to it

so if SAI is using his method for training them, then it could take way less time than you would assume

I know I was blown away when I tried it lol

It just feels illegal haha

[eternal fog]

Anyway my point was, that them saying it's "Easy" to train but then providing an amount that would take ages, especially if you don't have some server grade or top end GPU isn't what I'd call "Easy"

[high skiff]

and my response is, if they train it like Caith does, then it would run fine on a consumer GPU

likely an 8GB GPU could do it

[eternal fog]

I mean I've got good results on datasets trained for 15 minutes

Although that only seems to work on people/characters