✨｜sdxl

[eternal fog]

IT doesn't seem to for styles

[high skiff]

as have I, so I totally get what you mean

[eternal fog]

^That isn't what I wanted lol. But I thought it looked cool.

[high skiff]

@spark bearsome moar

just messingw ith other races and such

the blue is a little strong, but its a pretty small dataset and I have to reinforce the skin color

[spark bear]

how does it look with other artstyles , outside of photorealism?

[eternal fog]

How are you captioning your datasets @high skiff You doing it manually?

[pallid path]

is there some refiner upscaling workflow for comfy

[high skiff]

I have yet to try that, and likely won't for a bit

[pallid path]

in order to make higher res image without duplication

[high skiff]

yeah, i always manually caption

[spark bear]

its still doing the high cheekbones/strong jawline thing for every face..I really hope 1.0 doesn't have that tendency, its really annoying

[eternal fog]

Just upscale it with a pixel upscaler and then run it through img2img. It washes out detail though.

[high skiff]

or well, i run it through a tagger and then hand curate it and add tags I see as well

[pallid path]

I see...

I was just thinking of making 1080p wallpapers with SDXL

I did make some decent ones

[eternal fog]

Yeah it's annoying. It will do stuff like fix eyes, mouths, hands (a bit), but it removes the texture off the skin and stuff in the background

You can try a latent upscale, but you have to use a lot of denoise

[high skiff]

finetunes will likely allow for higher res upscaling, just like it did for 1.5

[pallid path]

how do I quickly copy an image to clipboard in Comfy

[high skiff]

trying to use high res fix with 1.5 base is pathetic lol

[pallid path]

I can only copy to "clipspace"

[high skiff]

click open image, then right click there

[pallid path]

oh so it was copying the node then

[high skiff]

you're welcome

just know that people can take that image and have all your settings and workflow

[pallid path]

does it contain private info

or only just the workflow

[agile quarry]

Prompt too

[pallid path]

that doesn't matter for me

I'm just using a workflow off of reddit lol

and a bunch of generic prompts

eh for safety reasons I'll remove

if anyone wants the workflow I can just link them the reddit page

[high skiff]

prompt, nodes, settings, model names, everything you can see on yoru screen, you share

[pallid path]

well that's nothing private for me, but out of being paranoid I'll keep that removed

[eternal fog]

If your using a chromium based browser though and you copy and paste from the browser it strips all that data

[pallid path]

Wait so if I copy paste from the web browser then it wont have the metadata

does it only have it when I post the image directly?

[eternal fog]

If it's Chromium based

[pallid path]

I'm using firefox

so I suppose it doesn't strip the metadata

[eternal fog]

Then I believe it still carries the data

[pallid path]

I wonder what SD 3.0 will be about

since that will have the opted out artists and stuff

maybe it will be some clever "architectural" upgrade

[eternal fog]

Scuffed Rock

[boreal bough]

I preach my settings twice daily XD
I have no secrets

[pallid path]

yeah joe biden looked a little odd too

[eternal fog]

What are these magic settings then

[high skiff]

1 repeat lmao

1 repeat, a shit ton of epochs

it makes some very interesting results

[eternal fog]

I mean that's what I've been doing anyway

Mainly because I can't be bothered waiting

[high skiff]

cause if you think about it, running a high BS with 1 repeat would have several different images being gradient acumulated together, rather than many versions of the same image

its an interesting concept for sure

but yeah, if SAI is doing 50 epochs of like 600 images at only 1 repeat, its not nearly as long as you would assume

compared to how I do it

[agile quarry]

My usual "style" dataset is 289 images typically do one repeat every epoch is better and I usually wish it went longer. (Only tried sdxl training once so far though)

[boreal bough]

first of all training speed
Unet Training: 1e-3 <- ideal setting for 95% of situations (From dataset size of 10~1000)
Dimension/Alpha = 8/1
Unet training only! (--network_train_unet_only)
Resolution = 1024,1024
Bucket size = 512,2048
repeats = 1
epochs = around 20 should be where the model is 'perfect'. Train for 40 to be sure. (scales very slowly with dataset size - but not nearly as much as in 1.5)

[high skiff]

@boreal boughhey man, any reason why we uses such different alpha and dim compared to 1.5 with SDXL?

I was curious about that

[eternal fog]

That's basically what I've been doing, except DIM 16

[boreal bough]

offset noise = 0. (you want to use the original sdxl offset noise which is hardcoded into kohya if you leave it on 0)

[high skiff]

so far I am very happy with the diversity I can get out of this LoRA with a dataset this small

[urban fjord]

How many training images Sytan?

[high skiff]

23 :>

I could easily get a dataset of several 100

[eternal fog]

my problem with large datasets is my lack of enthusiasm to caption them all lol

[high skiff]

also curious about training a multi Layer LoRA, with different concepts

like some Avatar Creatures, as well as some of their nature/wildlife flora

[high skiff]

yeaahhh, thats same here haha

[urban fjord]

Thats why I do single digit data sets Arron

[high skiff]

tho, I am able to see what tags to remove from my auto tagger, as it supports removing some tags

[agile quarry]

Even though you're not training the text encoder captions still matter right?

[eternal fog]

i did a single image LoRA, it worked better than expected, but it bakes certain things in quite hard.

[barren olive]

guys btw what is lora?

[urban fjord]

I forgot, I was supposed to try to train a new hybrid LoRA so I can get centaurs.

[barren olive]

someone said they can blend image with lora

how does it work?

[eternal fog]

It's a way to train/finetune the model to add new stuff to it

[high skiff]

I would assume so? I am not sure

I know it makes clip do nothing

but like, I can still trigger single concepts this way, so IDK

[eternal fog]

yes they still matter because your training the UNET

And that uses the captions

[high skiff]

yeah

[barren olive]

is it better or the same as Stable diffusion training on google collab?

[eternal fog]

I wish highres fix worked properly 😢

[high skiff]

a person in our research server has text encoder training working

[eternal fog]

One of these is a pixel upscale, the other is trying to make it "better" via img2img

You can tell the img2img one because there's no texture on the ground

[boreal bough]

very very much.

[barren olive]

is that gigapixel?

[high skiff]

gigapixel is so antient by now

[boreal bough]

same as how prompting has changed for sdxl, the same should apply to captioning.
good captions are more important than they used to be - especially that the most important words that describe your thing are well chosen

[barren olive]

because gigapixel tends to make fake detail to make theyre image more detailed

[barren olive]

antient?

[eternal fog]

No it's 4x-Ultrasharp and it's not adding any detail, it's just the img2img takes it away

[high skiff]

thats kinda how all pixel upscalers work lol

[eternal fog]

It's not too bad on images like this just using Ultrasharp

Just for faces, especially ones that have dodgy eyes it's awful

[boreal bough]

if you want to teach sdxl a block of text for example, using "text" will give you only garbage, while using "word" gives immediate results.
a bit of an extreme example, but important to know

[eternal fog]

Img2img fixes the face, but it smooths out everything else.

SD 1.5 didn't do that, so it's annoying.

[barren olive]

ya but gigapixel has acually good AI to make fake detail seems real

[high skiff]

gigapixel is pretty outclassed by a lot of new upscalers

[barren olive]

and some ai just make it more sharper

[barren olive]

like what?

[boreal bough]

also, dont try to redefine 'key' words unless you're brave enough. like 'boy', 'girl', 'portrait' <- it will work, eventually. but just using any other word that describes the same thing saves you soooo much trouble

[barren olive]

dude i tried tons of ai upscaler and enchancer but nothing come close to Gigapixel and they always update theyre dateset/ai

[high skiff]

I mean, if you like it, go for it

trying different body types/ages/locations/hair

[spring fulcrum]

I had a gigapixel license from a few years ago... What are some new ones because I may not update my license for it if there are better ones. What are some of the newer better ones???

[agile quarry]

Tried using "word" with the bot
Prompt: a man holding a sign with the word "smile" in bubble text

[boreal bough]

wasn't gigapixel corporate focused?
so it works in real life applications, but doesn't shine in anything that deviates from practical use cases?

[high skiff]

basically

[boreal bough]

not sure if I'm confusing it with another company right now x_x

[high skiff]

They are locally run gans, like 4x ultrasharp, which is what I use all the time

its very good for realism. It adds more detail than a lot of others, but without the crunch of most of the high detail ones

[boreal bough]

just checked. yeah that was them. topaz labs are great for work situations.

[spring fulcrum]

I like the 4x ultrasharp... thats what Ive been using.. I just thought you meant there were better ones than that hiding somewhere

[high skiff]

that one already beats gigapixel pretty handily for realism

[boreal bough]

cant be that long for photoshop to add ai upscaling though. its a low hanging fruit - unless they plan to do it at a level where it doesn't compare, but beats locally run upscalers

[high skiff]

they

they have it

they have had it for like

2 years lmao

its called detail enhancer I think

I used it a while ago, along with gigapixel

[barren olive]

@high skiff ow shit youre right https://www.reddit.com/r/StableDiffusion/comments/12vppkk/gigapixel_ai_for_second_stage_upscaling_seems/

[barren olive]

i see more detail with ultrasharp when zoomed in

is there any ultrasharp on google collab? i wanna try it

[high skiff]

yeah, its a great little ESRGAN, and it runs super fast on local hardware

[boreal bough]

the one for camera raw? yeah. but that one only works on real camera photos.
I mean a general style upscaler that works on any image like the typical esrgan

[high skiff]

oh, you can convert any image into a dng and use it that way 😅

[boreal bough]

uffff xD

[high skiff]

thats how I used it on non raw images lol

[barren olive]

@high skiff is there any link for ultrasharp on google collab?

[high skiff]

its just a single 60mb file

just a sec

there you go

[indigo carbon]

WHY IS THERE A FACE

[barren olive]

thats ultrasharp. i also notice ultrasharp has blocky artifact

[indigo carbon]

no, there is a face wtf

[barren olive]

probably has Face restoration and ultrasharp thought thats a face lol

[upbeat summit]

yeah that's normal for face restoration or diffusing more than you should

[barren olive]

what do i do with this? i dont run any ai locally on my pc

[high skiff]

no idea then

[barren olive]

@high skiff https://colab.research.google.com/github/tensorflow/hub/blob/master/examples/colab/image_enhancing.ipynb

is this correct ai?

[peak dove]

Gigapixel can learn what detail is missing and can add some in a very intuitive way, detail that was not there in the original, but does not look out of place when added!

[tall lion]

On your drive in Google Colab, you can go into Drive/ComfyUI/Upscaler and then uploadit there. I did it this morning and it works. ComfyUI on Colab and SDXL

[high skiff]

IDK, I know 0 about non local AI

[barren olive]

ow shoot google colab can run sdxl normally?

[spring fulcrum]

I keep getting weird markings and scratches and other things with the Upscaler in ComfyUI any advice would be appreciated.

I made the boxes on the face to show where the weird markings are

[agile quarry]

https://github.com/comfyanonymous/ComfyUI#colab-notebook I'm colab pro and add RAM but idk if you need to. Maybe free wont be able to use the refiner?

[barren olive]

thanks alot. can i also fine tuned using this notebook?

[barren olive]

ah nvm. i just check guess not. but thanks alot for this link

[tall lion]

Look at how it is setup. Once you do run the environment setup code on Colab (assuming you loaded the comfyUI notebook), you should see the drive on the left side. Chk out my setup.

[barren olive]

do u have gigapixel?

[agile quarry]

https://github.com/Linaqruf/kohya-trainer it's not in the readme but if you look in the files there's a colab for sdxl training

[spring fulcrum]

yes but its a really old version

[tall lion]

Yes Google Colab ComfyUI can use SDXL pretty easily. Just drag and drop any image from this forum (sdxl form) into your comfyUI page and voila, the workflow is set!

[barren olive]

what is comfyUI? never heard of it before. is it like automatic1111

?

im still new to SDXL stuff

[autumn forum]

It’s a node based workflow ui

[agile quarry]

But like Factorio version

[tall lion]

ComfyUI is the UI to use to create imges using different stable diffusion models (SDXL being one of them). Easy to use on Colab, since you get free GPU as well (upto a limit)

[barren olive]

https://www.youtube.com/watch?v=KTPLOqAMR0s

comfyui looks like im editing or developing some games lol

[indigo carbon]

new minecraft version hits different bro

[barren olive]

lol this ui seems like im coding or something

[tall lion]

Looks daunting at first but slowly you get the hang of it! 🙂

[barren olive]

can i use like different sdxl model like dreamsharper xl?

[tall lion]

You could , but I would recommend that you stick to SDXL Base and Refiner models for now.

[violet zinc]

Does anyone know if SDXL will be able to generate text consistently?

[barren olive]

kinda. its a mixed bag

sometimes it can generate text flawlessly sometimes its not

[tall lion]

So, I came to know of Colab like 3 days ago, setup my first note book 2 days ago and created my first SDXL images two days ago. I am not a coder by any means and I do strategy work for a healthcare company. So , some of the videos on youtube realy helped me get upto speed 🙂

[barren olive]

Like this

[barren olive]

really excited to train my own images. sd is already good at training artstyle i cant imagine when i train An artstyle in sdxl. thank you so much again

[violet zinc]

Cool that it can do this now 🙂 I wonder what the success rate is and where it might struggle.

[barren olive]

in my experience it has 25% success rate. if u generate 4 images atleast one can have complete text

[spring fulcrum]

A nice to have feature in ComfyUI would be a stop rendering button.

[soft bone]

optimizer? lr scheduler?

[soft bone]

view queue > cancel

[spring fulcrum]

Thank you... That is exactly what I needed

[eternal fog]

I wish text further away was as good as it can be on signs now

[boreal bough]

adamw8bit - constant with warmup 5%

[indigo carbon]

dude, SDXL is such a good model

[quasi remnant]

yea; nothing really compares to sdxl right out outside of like the private models like parti

[indigo carbon]

parti?

[fresh path]

results with controlnet and lora are yet to be seen

[quasi remnant]

google's imagen, microsoft also has NuWa

neither are available to the public afaik

[fresh path]

i thought they had something better than imagen

[paper phoenix]

wasnt aware of either

[fresh path]

imagen > muse?

[paper phoenix]

do spill the tea

[indigo carbon]

nah, google doesn't even get close

[quasi remnant]

o i meant parti was google's big image generation thing

i forgot they also literally named one of them imagen

xD

[quasi remnant]

nah, gotta give credit where it's due, parti 20b and nuwa-xl surpass anything available publicly rn

[indigo carbon]

even SDXL?

[quasi remnant]

microsoft is a bit more open though, nuwa might eventually see the light of day

[fresh path]

parti isnt high res

[quasi remnant]

1024x1024 is what we're at and that's what parti 20b is at

in terms of coherence it's better afaik

[fresh path]

oh, do you have the source for that?

[agile quarry]

Imagen is finally available to some Google cloud customers need to request and be approved. Existing Vertex customers only I think

[indigo carbon]

what about detail?

[quasi remnant]

if you check the images on the page they're all 1024x1024

like these are all 1024s

[fresh path]

oh true lol

it didnt look like it

some kinda blur

[sudden cliff]

I have my doubts that either exist as such. Model architecture, but since nobody can touch them, it's likely hype. Just like Bard before its release.

[quasi remnant]

MS has invested so much in AI + the demos they have r pretty extensive for nuwa/nuwa-xl

i see no reason to doubt it

nuwa would probably get integrated in Bing AI or something down the road

[indigo carbon]

Idk, I refuse to believe after all SAI's efforts SDXL is still not the best, when I use A1111- it gets to an insane level of detailing

[quasi remnant]

sdxl is amazing

but googles & ms resources r like infinite

lol

ms put $10bil down for AI stuff

not too long ago

[indigo carbon]

nah, it's not as good with details as SDXL

[sudden cliff]

But I'd also argue that Midjourney V5 is probably mostly superior in the ways that SDXL is good. DeepFloyd IF is definitely the most accurate text to image model that we can demonstrably show.

Moreso than SDXL

[indigo carbon]

that's not true. even 1.5 models are way better than MJ. let alone SDXL..

[agile quarry]

Meta has a new model too https://ai.meta.com/blog/generative-ai-text-images-cm3leon/

[quasi remnant]

i wouldn't say 1.5 models are but for SDXL many people voted the SDXL pictures better than midjourney iirc

[sudden cliff]

you have not seen Midjourney 5.x based models

[quasi remnant]

so midjourney's next model might be SDXL based

[high skiff]

for a 20B model, thats pretty damn rough

[quasi remnant]

yea; the parti paper is older

nuwa-xl is showing better progress rn

[high skiff]

understandable in that case

[indigo carbon]

oh, I did, people here can vouch for that

[sudden cliff]

As far as unreleased products, in the world of stuff like that, there is so much fakery about products that maybe 'could be'

I fully expect Imagen was never finished

[high skiff]

MJ 5.x does look incredible, I will say

But while its images have a lot of detail, I find its "realism" images never look that real

like, they have a ton of detail, but it looks overwhelmingly artificial

[quasi remnant]

i wouldn't doubt the capabilities of google/ms, even if the products never get released

[indigo carbon]

SDXL is far better than MJ. Idk about Imagen

[sudden cliff]

I saw one really silly thing with MJ, you ask for different nationalities in buildings and it always goes so over the top

[agile quarry]

Custom 1.5 models can definitely be better then MJ

[high skiff]

they can be, yeah

[sudden cliff]

The capabilities of a company must at least in some degree be tied to something that eventually gets released 😄

[high skiff]

especially for more stylized things

[sudden cliff]

I expect Google has amazing research

[tender ocean]

what model/prompt did you use for that, they look gorgeous

[high skiff]

https://tenor.com/view/flashbang-gif-22572243

[eternal fog]

lmao

[quasi remnant]

yes, custom finetuned models can surpass

[eternal fog]

Me in the middle talking about nsfw in SDXL

XD

[quasi remnant]

the images have the metadata on them still, you can drag directly into comfyui if you have it

was testing a random prompt

[indigo carbon]

yep, I remember myself actually making a 1.5 model that surpassed MJ

[quasi remnant]

it also output these

[high skiff]

if MJ is the mind of Holz, I would hate to see what the mind of Emad looks like lol

probably V2.1 honestly

[indigo carbon]

not even close

[high skiff]

messy, incoherent, all over the place, and full of fake information

[tender ocean]

I'm still kinda new with sd, is comfyui in the files or in the webui?

[quasi remnant]

comfyui is a different UI to the web UI, i don't know if you'd be able to recreate that with auto1111 rn or if it even has sdxl support yet on the non dev branches?

if you have comfyui you'd just take the picture and put it into the ui itself

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

because that also lets you have finer control with the prompt

i didn't use clip vit-l

only vit-g

and also a refiner model

a1111 would send it both into vit-g and vit-l which would be a 100% diff result afaik

[eternal fog]

yes it would

[delicate grotto]

oi, thanks for the help a few days ago in setting up comfy

[tender ocean]

Half of that was gibberish to me

[indigo carbon]

Idk, A1111 does SDXL perfectly, except it isn't implemented with the refiner

[quasi remnant]

Basically, if you install comfyui, and the 2 SDXL 0.9 models, you just put the picture into the web browser with the UI loaded up and it'll copy all the settings immediately

and you'll be able to see the full workflow that made that

[indigo carbon]

like, does Emad really think MJ can outdo these?

[quasi remnant]

does it have a separate text box for vit-g and vit-l?

[paper phoenix]

if its any consolation we didnt all get qualifications in this, we started where you are and kind of fumbled around together.

[quasi remnant]

because i rarely use vit-l

[high skiff]

So... then its not perfectly 😅

It has LIETRALLY everything, but half of it

Then its not ALL OF IT lmao

[delicate grotto]

mmm well, a1111 is great, but the vram usage is huge
i can barely do 768x768 on a1111
and easily 2048(non sdxl, sdxl falls at ~1900x1900 on 6gbvram) on any model

[quasi remnant]

using vit-l decimates the quality unless it's used for a very specific reason for the prompts i use

like i'll sometimes add a few supporting terms in vit-l

but i don't want vit-g copied to vit-l

[indigo carbon]

idk what to tell you, what ever it is it's better than any ComfyUI workflow I used.

[quasi remnant]

and yea, refiner is pretty big part of the process for image clarity

[tender ocean]

I used the github link and am downloading it, I guess I'm just unpacking the .zip in my C:/AI/ folder?

[high skiff]

I would assume thats user error in that case

[quasi remnant]

nooooooooooooo

[quasi remnant]

git clone the repository

so you're able to update it

[soft bone]

any news on training the refiner yet?

[indigo carbon]

like, dude, whatever A1111 did it damn works

[delicate grotto]

after you do that, you can link your SD venv in here:

[pallid path]

why is the chat active

[sudden cliff]

It's funny I've found that without vit-l often key attributes are completely missing from my prompts. I'm thinking that by skipping this you're missing some of the knowledge in -l

[pallid path]

I see no announcements

[tender ocean]

ok, how do I launch git, I have like 3 executables in my search bar

[high skiff]

I mean, it looks fine

[delicate grotto]

and only change this path, keep the rest

[tender ocean]

Bash, Gui, or CMD

[quasi remnant]

you can do it through windows terminal/powershell

this is what the workflow looked like btw

[agile quarry]

Does using the basic condition node prompt both? Or just vit-g

[quasi remnant]

but if you just used green text in a1111 it wouldn't do same result cuz it'd be feeding into vit-l too

[eternal fog]

They'll probably be putting both text encoders into the model, which it fine, it just gives you a bit less control

[quasi remnant]

also no refiner

having it be denoised by refiner gives it a lot more clarity

[pallid path]

does the fp16 vae fix help

in comfyui

[sudden cliff]

I bet there are specific prompts where either g or l have destructive knowledge, and times where g or l have specific knowledge that you'd want

[high skiff]

the BF16 one does

[quasi remnant]

yea

that's why

[high skiff]

it speeds up generations massively at the end

[quasi remnant]

you put specific terms in one or the other

[pallid path]

my GPU supports BF16, it's ampere

is it on by default

[sudden cliff]

But the trick is SUPER long term (I guess) learning which terms should go where

[delicate grotto]

cat diffusion when?

[sudden cliff]

like maybe l has a great idea of what an elf looks like but g doesn't

[quasi remnant]

i more or less have an idea through testing what words i wanna put where

but i never wanna directly copy paste

or like

use same terms for both

[indigo carbon]

*SDXL now

[delicate grotto]

they didn't go far enough
i want a cat with wings, a devil cat, a cat with a fork, a cat playing poker

[quasi remnant]

i even sparsely use negative unless i really need to

[sudden cliff]

some kind of interface to be able to mark works to 'stay off' of one prompt or another might be interesting. Or literally... someone should do one of those charts, trying each prompt term in each combination of each clip model

none, l, g, both * each word

[soft bone]

u using caith's params?

[high skiff]

no, I am using very different ones

[pallid path]

Sytan is there some tutorial for BF16 vae

and in comfy

[paper phoenix]

you can type in cmd in the address bar of any folder to open a command prompt there. git has a nice windows gui for keeping things up to date but unless you're actively coding its not v useful. think of git as like the update service for projects, you can largely just have it there.

[agile quarry]

Make a Sumo wrestling na'vi

[high skiff]

you would need to install the nightly build of torch, which I do not remember how to do

[quasi remnant]

i remember

[high skiff]

ok, now that is a lot haha

[quasi remnant]

pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

[paper phoenix]

ok i want this prompt now mate.

[pallid path]

so it's some experimental feature for now 😔

[quasi remnant]

this is how you install nightly torch

for whoever needed it

[pallid path]

oh that's it, thanks

[soft bone]

care to share? these settings arent working for faces

[tender ocean]

okay, so how do I get started, Idk wtf to do

[pallid path]

the generation times are good enough for me so I won't experiment with BF16 for now

[delicate grotto]

my heart can't take it anymore

[eternal fog]

bubble twins

[high skiff]

I am not gonna share my settings at the moment as they are EXTREMELY wasteful

Like, 70% of the LoRA was wasted on over training, and I need to tweak it a ton

[paper phoenix]

we should maybe move to the tech channel, this moves pretty fast

[tender ocean]

#🤝｜tech-support

[high skiff]

The one thing that is pissing me off about this LoRA is it just like... randomly decides to not generate a Na'vi, and instead generates a black guy

it keeps happening, and I have no idea why

like, WTF

[pallid path]

lmfao

[soft bone]

gotcha. I'm trying to do the same, just need a coherent starting point for the process, and so far i'm getting major distortions. might be my captions

[high skiff]

its so stupid, and I have no idea how to stop it from doing that

[soft bone]

im sure training the text encoder will help once we figure that out

[urban fjord]

Sytan: I can't remember having issues like that with different skintone LoRA with 1.5. But things are a bit different with SDXL if you're not training TE.

[pallid path]

does prompt weighting help?

[high skiff]

good point

[pallid path]

and that too

[high skiff]

there is a guy in our research server that has TE training working

its just hacky

[pallid path]

yeah we should wait for SDXL 1.0 anyway

[high skiff]

I don't understand it well enough to explain, but he did get it working

[pallid path]

I heard SDXL 1.0 is delayed because SA wanted to help with 0.9 Loras and stuff

[soft bone]

is it hard because it's a guessing game or is it hard because it's demanding?

[high skiff]

it requires you to train 2 LoRA's

[soft bone]

no biggie tbh

once we know the params to use

[pallid path]

I also saw some post that suggests that we can use T2I-Adapter with SDXL

[urban fjord]

That seems to be a misunderstanding and it isnt' really delayed due to 0.9 capabilities.

[high skiff]

I need to find a way for SDXL to stop racially profiling Na'vi lmao

[soft bone]

It would suck if we end up having to train 3-4 Loras for each subject (unet, TE, and refiner-and/or refiner TE)

[eternal fog]

If you train a new one for it yes.

That's what they did for that scribble thing

[high skiff]

cause it just keeps generating black people

[pallid path]

aah

[eternal fog]

I think it's something to do with how it captures general skin tone. Because earlier when I was trying to do a red skinned demon, it kept generating people with darker skin as well

[high skiff]

the second I put african in the negative, all it generates is white guys

[eternal fog]

He comin

[tender ocean]

What happens if I put multiple models in the models folder? do I get like a character selection screen or an error that kills everything? 😂

[high skiff]

oh wow, that one reminds me of Mao Mao haha

[high skiff]

Just change the eye color and its basically Mao Mao haha

[eternal fog]

lol

I've got another

[high skiff]

Cheshire Mao Mao lol

[soft bone]

how does one use xy plot node in comfy?

[eternal fog]

It moved away from full cat person lol

[delicate grotto]

anime style sometimes does that

[eternal fog]

we back

[delicate grotto]

for a reason the prompt "red" is very dominant

even at 0.1 for me

[ionic dragon]

@eternal fog I am using the same prompt for the prompts, is there any way to connect these 2 nodes using another node, so that i can just input the prompt once?

[eternal fog]

Just use the one node and plug it into both

[delicate grotto]

the cat on the chrismas tree

[ionic dragon]

to which nodes is it connected, its too difficult to find the path when the noodles are straight

[eternal fog]

They plug into the Text_g and text_L in the CLIPTextEncodeSDXL Nodes

[ionic dragon]

ok

[paper phoenix]

this doesnt even look like a kitten.

[indigo carbon]

I moved on in life

[soft bone]

wtf is pooled output

[nimble heart]

is it not just adding --bf16-vae when launching comfy?

[pallid path]

might try thx

[ionic dragon]

the smoke is realistic af

[errant yarrow]

a fully tattooed young pretty and attractive Brazilian tall female fitness supermodel wearing biking and super muscular with six pack in the bedroom with face and chest towards camera full front ultra realistic 8k image hyperrealistic. Her each and every body part is fully covered with tattoos.

[nimble heart]

also in my case the patched fp16 vae on HF is actually faster than the standard cast to bf16 by quite a bit for decoding. Encoding seems broken though.

[barren olive]

Is that Midjourney?

[barren olive]

I like this artstyle

What did u use?

[indigo carbon]

SDXL0.9

[barren olive]

@delicate grotto what prompt/artstyle did u usw for this? I loved the Soft shading/anime style like league of legend

[jolly creek]

soo good man

[indigo carbon]

yeah, the power of A1111

[barren olive]

@indigo carbon do u have any idea for this artstyle?

[barren olive]

Or this?

Seems close enough for that artstyle

[delicate grotto]

mmm can't recall
i think
a catgirl in a purple yukata, sakura trees in the background, pink background

5 steps with the style: Enhance, Cinematic
15 steps with: Enhance, anime, digital art, Cinematic

[barren olive]

Do u use ComfyUI and custom Model? @delicate grotto

[delicate grotto]

sdxl 0.9 pruned - comfy

[noble shoal]

Also archived a very colourful splashy style with "dreamy, (style of Hannah Yata:1.1), colorful Psychedelic Art, neon colors" in my prompt. sdxl 0.9 pruned - comfy

[delicate grotto]

how do you get a prompt from a photo using comfy? simply dragging it into the web browser ui doesn't do anything

well, i can't replicate the original but it does look alike i think

[noble shoal]

You probably will not get a workflow from the drag and dropped image since Discord is probably messing with the EXIF data from the uploaded images.

[delicate grotto]

anyway, for anyone that was interested

this is as best as i could replicate it

[azure oxide]

workflow should work, but you gotta copy it from file browser tho iirc, not from the comfyui

[delicate grotto]

well, here are 4 pics from the directory itself, not sure how to extract info from comfy

[azure oxide]

just drag it into discord

[noble shoal]

Thanks, could replicate it.

[delicate grotto]

ah wait, it makes the entire workflow with the pic

damn, comfy is good

[noble shoal]

Yeah, just drag and drop. Comfy is awesome.

[delicate grotto]

i thought i was going to get a prompt and steps only

[noble shoal]

Wait until you have missing nodes 😄

[delicate grotto]

at least its not a bsod

[noble shoal]

Why is there no refiner in the workflow?

Or did i miss something?

[delicate grotto]

can't load a refiner

[noble shoal]

Let me check that out real quick

[delicate grotto]

i need to do it in two workflows...

sadly

one to save the latent image, and one to pass through the refiner after deloading the refiner

[shy kelp]

how can I remove kids from sdXL

I don't want any kids generated by this

%0

[delicate grotto]

goddamit joe

stop it ahahahhaha

[shy kelp]

on god I won't sniff

just want to remove

all any way no posssible

[azure oxide]

cant really do that easily as of yet

[shy kelp]

how to do it

like lets say

I will train this xl

with thousands of 18+ images

how I will stop

this shit

genrates

kids and arrest me

whats the best way?

[delicate grotto]

its time to stop...

[shy kelp]

its real question I ma pay

like 2k usd

who can do it

[azure oxide]

joe penna himself said that if u can delete trained things from models that he'll hire you on the spot lmao

[shy kelp]

I don't want to get my ass arrested for this sht

yea like umm

maybe we can train like

hmm

we can collect all keywords ok these pedos search

and we can train for those keywords

a fbi logo

no it wont work

if it will generate a %0.1 percent change

a kid

I go arrested

tf is this how to

tell stable owner

don't use any kid shit

[azure oxide]

youre not going to get arrested for accidentally creating that lol

[shy kelp]

but I will make a website

for +18 images

lets say

than what

[delicate grotto]

he's trolling...

[shy kelp]

nope

its fact

[azure oxide]

i think hes genuine

[noble shoal]

Jesus christ, i was 2 seconds away

[azure oxide]

yeah it gets tricky if youre hosting a service

[noble shoal]

Here with refiner

[shy kelp]

like you are not a developer or like u are user?

[shy kelp]

yea

[near pulsar]

hey people already asked on #🤝｜tech-support but you seem more active, but im getting something weird, basically with A1111, i tried the default SDXL0.9, the Dreamshaper one, and the WaifuDiffusion one, and WaifuDiffusion runs flawlessy, like, it's perfect, but the base 0.9 and the dreamshaper one just, they take forever to load, and when they're loaded they seem to generate the image "fine" but then when it reach 99% it throws out a cuda memory error, which is weird because it does that with every resolution i try, even 512x512, while WDXL works till 1024x1536

[shy kelp]

so there must be a way

either you will create your own model

that has %0 kids in it

or like how we can

[delicate grotto]

mmm add 3d to the positive

[shy kelp]

fix XL because its better is it

I thinked to add a

age detector

[noble shoal]

K, let me see

[shy kelp]

before um releasing the image

[delicate grotto]

or pixar, the refiner kinda ruined it and flattened it

[shy kelp]

it will check for age

[shy kelp]

but it costs

and not always works

[delicate grotto]

not sure how the refiner works, but try to use it on the last 2 3 steps?

[azure oxide]

i saw another host post how he handles it: he replaces prompts of what he thinks will create that into another word. so child becomes ape or something. obviously that means noone can ever make pics of children ever with his service but that's a small price to pay

[shy kelp]

yea me saw it to but

its not %100

its like %99.9 and

[azure oxide]

right

[shy kelp]

if a guy is a fully pedo

he will find a way

I am sure

and you get ur ass whoped

like trained model

must have %0 child image I guess

all adult

[delicate grotto]

sdxl was trained specifically to work on 1024 right?

[azure oxide]

from what ive seen, most github repos will have a license stating that the dev isnt responsible for anything that a user does with their tool, im sure there's something similar for a service provider

[shy kelp]

yea yea but umm

you need to make sure before giving it away

they will take down ur server domain and shit

[azure oxide]

i can see that happening, yeah

[shy kelp]

there must be a way like to not generate any %100 no chance u know

I can add a age detector and reverse prompts

[noble shoal]

Don't know if that any better. I am more the realistic guy

[shy kelp]

its 11 yo 💀

[delicate grotto]

yep, looks better, didn't get flattened

[near pulsar]

you using comfy or auto1111?

[shy kelp]

yo bro are you good with stable dif

[noble shoal]

Comfy

[shy kelp]

I was gonna buy a mercedes I invested in this nude stuff need help

[near pulsar]

mhh damn

comfy seems better optimized than auto1111 rn

[noble shoal]

At the moment yes. I always worked with Auto, but i completly changed my workflow.

[near pulsar]

i prefer the gradio UI too much to change it tbh

[delicate grotto]

i tried your workflow

[noble shoal]

Yeah the learning curve is pretty steep. But once you get a hang of it, you will aprechiate Comfy. I am no fanboy or something, and i loved Auto, but now i stick to comfy

[noble shoal]

How did you do that?

[shy kelp]

IMPRESSIVE!

[delicate grotto]

ah yeah, its the nodes on the right

after deleting its all good

[shy kelp]

Is the sdxl 10 times slower than other models??

[delicate grotto]

where did you get the pruned refiner?

[noble shoal]

Not realy, it generates just in higher resolution.

[indigo carbon]

on high resolutions it's faster

[shy kelp]

So you think its better than any other is it

[near pulsar]

tbh i've found that WHEN it works it's faster

[shy kelp]

Can I use it on my 16gb ram gpu thing

train it

[noble shoal]

Some guy oploaded it one day after the leak on huggingface

[shy kelp]

Can you send 💀

[weary kiln]

What to the DM buttons under the bot image results do? Who gets DMed?

[near pulsar]

you

so you have them saved if you want them for later

[noble shoal]

Hell no.

[shy kelp]

haram

[delicate grotto]

too bad, its not there anymore

oh well

[near pulsar]

also, is the refiner really needed, sometime i see it improves images, sometime it's worse

[shy kelp]

refiner is good

are you using comfi gui

whats the best gui

[indigo carbon]

good question. A1111 isn't using the refiner and it runs it the best

[noble shoal]

I have read that you can prune them by yourselfe with automatic. Don't know if thats a thing

[shy kelp]

How to you use it on a1111

it needs api key

[indigo carbon]

dev branch

[near pulsar]

yes you can

[weary kiln]

I don’t see anything in my DMs; maybe it’s because of restrictions in my account settings?

[near pulsar]

aye could be

[weary kiln]

Alrighty. No big deal.

[delicate grotto]

if anyone knows where there is a pruned no-ema sdxl/refiner please send it here

[near pulsar]

love this shit

it even does that with 64x64 images wth

[noble shoal]

Sorry, i have searched without luck. I am sorry, but i can't upload it. You might have to convert it by yourselve.

Here is a thread about it on reddit https://www.reddit.com/r/StableDiffusion/comments/14uaoca/how_convert_sd_xl_base_09safetensors_to_pruned/

[near pulsar]

you can sort of prune the main model from 13gb to like 6.5 with auto1111 but it didnt seem to work with the refiner, the size was the same

[indigo carbon]

WHY DID IT MAKE TILE IMAGE

[noble shoal]

Base: 6.775.435 kb / refiner 5.933.544 kb. Can confirm that they are both fitting on my RTX 4070 12 GB in Comfy

Generation Time: Approx. 20 sec 1208x1024 with my shitty workflow inc. refiner

[noble shoal]

I think the better question is: How did you make a tile image?

[near pulsar]

weird firey revolver

[autumn forum]

i get tile images a lot on certain seeds with short prompts. was the prompt short? hopefully 1.0 fixes that. happens a lot to me too

[noble shoal]

What did you actually wanted to achive with that prompt?

[delicate grotto]

Well, i closed some stuff
A 3060 mobile 6gb can fit it too

~30 sec render time

[noble shoal]

Uh, thats nice. Lets not look at the ghostly thing top left

[soft zealot]

some people may say that what you focus on speaks to your state of mind ;o)

[noble shoal]

30 sec is ok for that setup i think. Was it now with the refiner?

[delicate grotto]

Yep

[noble shoal]

Nice. Oh btw, yesterday i found this here somewhere:

[delicate grotto]

Im still not sure how the refiner works, but i didnt notice any difference

[noble shoal]

Style: Enhance
Positive: breathtaking {prompt} . award-winning, professional, highly detailed
Negative: ugly, deformed, noisy, blurry, distorted, grainy

Style: Anime
Positive: anime artwork {prompt} . anime style, key visual, vibrant, studio anime, highly detailed
Negative: photo, deformed, black and white, realism, disfigured, low contrast

Style: Photographic
Positive: cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

Style: Digital art
Positive: concept art {prompt} . digital artwork, illustrative, painterly, matte painting, highly detailed
Negative: photo, photorealistic, realism, ugly

[delicate grotto]

Yeah, i got it on a note, extremely useful for the style

[noble shoal]

tyle: Comic book
Positive: comic {prompt} . graphic illustration, comic art, graphic novel art, vibrant, highly detailed
Negative: photograph, deformed, glitch, noisy, realistic, stock photo

Style: Fantasy art
Positive: ethereal fantasy concept art of {prompt} . magnificent, celestial, ethereal, painterly, epic, majestic, magical, fantasy art, cover art, dreamy
Negative: photographic, realistic, realism, 35mm film, dslr, cropped, frame, text, deformed, glitch, noise, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, sloppy, duplicate, mutated, black and white

Style: Analog film
Positive: analog film photo {prompt} . faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage
Negative: painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured

Style: Neonpunk
Positive: neonpunk style {prompt} . cyberpunk, vaporwave, neon, vibes, vibrant, stunningly beautiful, crisp, detailed, sleek, ultramodern, magenta highlights, dark purple shadows, high contrast, cinematic, ultra detailed, intricate, professional
Negative: painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured

Style: Isometric
Positive: isometric style {prompt} . vibrant, beautiful, crisp, detailed, ultra detailed, intricate
Negative: deformed, mutated, ugly, disfigured, blur, blurry, noise, noisy, realistic, photographic

Style: Lowpoly
Positive: low-poly style {prompt} . low-poly game art, polygon mesh, jagged, blocky, wireframe edges, centered composition
Negative: noisy, sloppy, messy, grainy, highly detailed, ultra textured, photo

Style: Origami
Positive: origami style {prompt} . paper art, pleated paper, folded, origami art, pleats, cut and fold, centered composition
Negative: noisy, sloppy, messy, grainy, highly detailed, ultra textured, photo

Style: Line art
Positive: line art drawing {prompt} . professional, sleek, modern, minimalist, graphic, line art, vector graphics
Negative: anime, photorealistic, 35mm film, deformed, glitch, blurry, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, mutated, realism, realistic, impressionism, expressionism, oil, acrylic

Style: Craft clay
Positive: play-doh style {prompt} . sculpture, clay art, centered composition, Claymation
Negative: sloppy, messy, grainy, highly detailed, ultra textured, photo

Style: Cinematic
Positive: cinematic film still {prompt} . shallow depth of field, vignette, highly detailed, high budget Hollywood movie, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy
Negative: anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured

[delicate grotto]

Some of them btw mess up with anime

[noble shoal]

Style: 3d-model
Positive: professional 3d model {prompt} . octane render, highly detailed, volumetric, dramatic lighting
Negative: ugly, deformed, noisy, low poly, blurry, painting

Style: pixel art
Postive: pixel-art {prompt} . low-res, blocky, pixel art style, 8-bit graphics
Negative: sloppy, messy, blurry, noisy, highly detailed, ultra textured, photo, realistic

Style: Texture
Positive: texture {prompt} top down close-up
Negative: ugly, deformed, noisy, blurry

Sorry for the spam

@delicate grotto 😛

[indigo carbon]

weird, it got better when I prompted "beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful beautiful "

[paper phoenix]

wonder how reptition actually plays out in clip.

[blazing hemlock]

I got latent walks working
https://twitter.com/Birchlabs/status/1682531741100015617

Birchlabs (@Birchlabs)

#SDXL conditioned on interpolated CLIP hidden states
thanks @cafeai_labs for the compute, and @RiversHaveWings for answering my questions

▶ Play video

Twitter · Jul 21, 2023, 11:23 PM

pedantically it's interpolating hidden states rather than latent embeddings but that's as close as is possible

[jaunty adder]

My biggest hope for improvement between 0.9 and 1.0:

The refiner model: All of the faces turn out more Caucasian and generally more plain looking than the intended expression. Like if a person is another ethnicity they come out looking at least half white. Also if the original face has a more stylized rendering, it almost always turns out more photoreal in the refiner version.

[near pulsar]

i really dont understand why they went with the refiner model to begin with, im no expert, i dont understand the process, but it seems to just to make the process more complicated

[crystal jungle]

bruh

what the fuck

everytime my generation gets close to done it disappears

[fresh path]

what is joe byron talking about\

[crystal jungle]

does anyone know how to fix this

[blazing hemlock]

it implements eDiff-I's "ensemble of expert denoisers" idea
https://arxiv.org/abs/2211.01324

[fresh path]

it reminds me of gpt 4

what if they had like 8 different refiners trained on different aspects or styles

and combined them

I guess it wouldnt be local at that point

[crystal jungle]

it's happened like 7 times in a row bro what is going on 💀

[blazing hemlock]

if the conditioning is good enough, model has enough params, and the things you want are represented in the data: one refiner is enough

[fresh path]

okay, thanks luluco

[blazing hemlock]

the refiner already is an expert of an "aspect" anyway

it's a high-frequency details expert

[fresh path]

ooh thats right

[blazing hemlock]

and I guess it's biased towards being responsible for aesthetics

because it's conditioned on LAION's CLIP

[fresh path]

do you know if diffusion models in the future could get the text right? Like if you ask it for a newspaper

[blazing hemlock]

DeepFloyd IF already does well on text

[fresh path]

true

[blazing hemlock]

but I think the people who've done best, conditioned their UNets on text encoders which tokenized the prompt into characters rather than words

[fresh path]

does that also remove the issue of two word objects being split up? for example dragon ball

[blazing hemlock]

hmm is that even an issue?

[fresh path]

it sometimes puts a dragon

[blazing hemlock]

hmmm

[fresh path]

its not a good example but

[blazing hemlock]

it's the most normal thing in the world for the tokenizer to tokenize your prompt badly

but attention still learns which tokens cause each other

the token embedding it learns for that token, may end up a bit tangled with multiple concepts

[fresh path]

yeah it does sometimes

[blazing hemlock]

I think if you're not getting a dragonball, the more likely thing is it wasn't sufficiently represented in the training set of the text encoder, and/or of the Unet

[near pulsar]

does some of y'all got a good ComfyUI workflow?

[blazing hemlock]

even tokenizing it really well can't make up for the text encoder or Unet not having seen it during training

[fresh path]

solution is to train a lora and put a unique token in the training data?

like dbz

[fresh path]

in pinned

[blazing hemlock]

hmmm loads of solutions

textual inversion can search the text encoder's embedding space for coordinates that would already do the trick, even though they have no name in tokens

[fresh path]

ooh

[blazing hemlock]

LoRA can't change the shape of model weights, so if you want to extend the text encoder's vocab: you'd need to resize the input embedding and output projection

(there are ways to do both of those without a full finetune)

then you'd finetune text encoder (you could use LoRA)

[fresh path]

I think thats beyond me, its not really an issue though

[blazing hemlock]

then you'd finetune Unet (you could use LoRA)

but I think you can get away with the text encoder not understanding what you're talking about

lots of people just overfit when they finetune the Unet

just make it always draw dragon girls

[fresh path]

dragon girl fumo

[blazing hemlock]

that's a good idea

I do need to do a fumo finetune

[fresh path]

I have fumo irl so I can just stick it to my monitor

[blazing hemlock]

I made a fumo textual inversion embedding
https://birchlabs.co.uk/machine-learning#textual-inversion

I was amazed how close textual inversion got; it doesn't involve finetuning the text encoder or the Unet

[fresh path]

theyre so cute

[blazing hemlock]

the thing that keeps stopping me from doing a proper finetune is I don't have a decent fumo dataset

[fresh path]

theres a bunch of really good 3d fumo scans

[blazing hemlock]

there's a lot of data here, if a way can be found to de-watermark them
https://fumo.website/fumo_checklist.html

[fresh path]

like https://sketchfab.com/3d-models/project-cirno-fumo-3d-scan-efd2a7f4dbf048c1a18438db7f86c4b9

true

most of them dont have watermarks?

[blazing hemlock]

I'll take 10

[fresh path]

oh the touhou logo

[blazing hemlock]

or whatever that icon says

it's all solvable, but more work than I really want to do

[sudden cliff]

I just realized that the formatting like %date:yyyy-MM-dd-hh-mm-ss% works in comfy for filename outputs

[near pulsar]

aint gonna lie this doesnt seem optimal

[upbeat summit]

very cute 🙂

[hearty ginkgo]

heylo

[trail bay]

something is wrong with this. what is your workflow

[spring fulcrum]

Got a couple new LoFi girl style images... What do you folks think?

[fresh path]

perfect except for the eyes

it really struggles on them

[trim orbit]

really looks like somethings going wrong with that one mikey

[winter wind]

They look pretty good. Reminds me a bit of cyberpunk for some reason.

[trim orbit]

i'm not sure what lo-fi means in this context. i usually see it as "low fidelity" nice images though. maybe a little underrefined. no refiner?

[fresh path]

HiFi girl

[winter raptor]

I absolutely love this, would you mind sharing the workflow and prompt?

[sharp robin]

a terrifying creature lurks in the shadows of an abyss, another creature is spooked + fantasy art, imaginative themes, surreal elements, mythological creatures, dreamlike scenery

[high skiff]

So

I have news

I wish not to get my own hopes up

but my 3090 is currently working without issues

and I am not sure if I have found the permanent fix, or if this is just a temporary postponement, but whatever it is, its the first time since I got it that I have not immediately bluescrened after basic high intensity compute

[trail bay]

was it on the new computer? or working on the old one?

[shy kelp]

It's not the 3090 correct?

[high skiff]

If this is the fix, it definitely was the 3090

[high skiff]

Both of them

[visual glade]

what's the fix?

[trail bay]

interesting

[high skiff]

The seller and I exhausted practically every option, and I considered the GPU completely dead, so we did some more reckless things

[shy kelp]

Yes what fix, and what's wrong with the 3090?

Did you get it used?

[visual glade]

did it involve an oven?

[high skiff]

I disassembled the entire GPU and started running it without the backplate, to diagnose any potential heating problems on the VRMs and the pads that he included, it was still crashing at that point

So then I wanted to test with the backplate reinstalled just out of curiosity, so I disassembled the whole graphics card again, reassembled it with the front and back on, and it hasn't had any problems since

I don't know if maybe the pads weren't making proper contact, or potentially there was too much mounting pressure on the GPU, which I know can cause problems with CPUs

[trail bay]

ahh

[shy kelp]

Sounds to me like the seller installed some pads incorrectly

[high skiff]

I'd likely assume that it is the mounting pressure situation, as the seller has nearly two months of squeaky clean windows event logs, and temperature logs and everything that show absolutely no out of the ordinary characteristics for the GPU

So our current guest is that maybe the box is staying some form of high impact that messed up the mounting pressure asymmetrically, so one side of the die was experiencing more pressure than the other, which can cause some pretty big issues with CPUs, and I would assume GPUs as well

The thermals look identical, the performance looks identical, but it's not crashing anymore

I genuinely have no idea, and I don't want to get too over hopeful that the issue has been completely solved, but this is the first thing out of dozens of different things I've tried that has yielded any form of improvement, let alone a complete eradication of the problem as of now

[hearty ginkgo]

What are you using for thermals? Does your app you use show the the vram temp ect?

[shy kelp]

I just got my 3090 from eBay myself. The seller used it for light gaming only. It's working fine so far for me.

[high skiff]

But just for a comparison, my GPU has sustained over 60 BSODs, every single time being identical without fail

The criteria for the BSOD was doing some form of thing that required the GPU to hit its maximum clock, waiting for it to fall back down to idle, and then having the GPU try and reach its maximum clock again

Every single time that led to a crash

[shy kelp]

Are you saying it might have gotten bounced around in shipping?

[high skiff]

As of right now, I have run through that exact same system almost 70 times without a single hiccup or BSOD

So the single thing that failed every single time on the graphics card has now worked consecutively over 70 times

[high skiff]

That's the guess, because the seller 100% knows his shit about GPUs.

I'm pretty educated on the topic, but this guy was running circles around me in terms of specifications, tests that we could run, potential hardware fixes, all sorts of shit

He's got a squeaky clean record of like 300 and some sales on eBay, and a phenomenal rating as well

The entire time he was optimistic, helpful, and very polite, and overall just seemed confident that there was some weird or stupid issue at play

He stopped me from giving up several times, and it seems like it may have been for good reason

He was immensely confident that he did nothing wrong on his part, as he was literally running live tests for me with logs and everything as I was contemplating buying the GPU

And I know for a fact that I did nothing wrong, is I handled the GPU extremely gently, and I've installed graphics cards dozens of times

It really does seem like some form of fluke mounting pressure problem

[lusty raptor]

congrats, and that's damn weird

i might even take it back apart and look for a short

[shy kelp]

but you solved the problem @high skiff ?

[visual glade]

hopefully it keeps working

[high skiff]

As of right now, yes

[shy kelp]

good

[high skiff]

I have no idea if it's a temporary, or a permanent, or whatever

[shy kelp]

I think i got lucky with my 3090. Had it for a week and a half and so far it runs very nicely. image iterations which took 30 mins on a 4GB 1050 Ti is now taking 30 seconds. lol

although, with more complicated images with multiple controlnets it does tend to sound very loud with that founders edition turbo fan. lol

[spring fulcrum]

LoFi girl is a sub genre of music... its a youtube channel and its all about easy to listen to music with no lyrics that helps some folks study or just relax. The LoFi Girl is the face of that channel. Lookup LoFi Girl on youtube. Even if you don't find that you like the music you will see what I mean... And yes its meant to be a combination of the Neon CyberPunk with The LoFi Girl in a photo format instead of an illustration. Sorry for the wall of text. Just figured it could use the explanation.

[trim orbit]

don't appologize! that's totally the sauce i wanted. see, i'm old. i don't get some lingo sometimes. wanted to know where that good old oldy, lofi, was coming from in a modern context. thanks for laying it all out for me.

[spring fulcrum]

Thank you 🙂

[lusty raptor]

funnily enough, lo-fi tries to go for the older analog warm tones, what would have been called hi-fi back in the day

[trim orbit]

heh, yeah but hifi was analog stuff that was just purrre with zero noise. it was pretty impressive gear. i was never alive while it was the hot market, but there's a level of quality to old hifi audio systems that was just.. mmm..mmmmmmm!

[spring fulcrum]

I think in the context they use it in its less about the quality of the music and more to signify that the music is low energy and relaxing

[livid phoenix]

I keep getting this error message here everytime I try to generate any image: NotImplementedError: No operator found for memory_efficient_attention_forward with inputs: query : shape=(1, 4096, 1, 512) (torch.float32) key : shape=(1, 4096, 1, 512) (torch.float32) value : shape=(1, 4096, 1, 512) (torch.float32) attn_bias : <class 'NoneType'> p : 0.0 cutlassF is not supported because: device=cpu (supported: {'cuda'}) flshattF is not supported because: device=cpu (supported: {'cuda'}) dtype=torch.float32 (supported: {torch.float16, torch.bfloat16}) max(query.shape[-1] != value.shape[-1]) > 128 tritonflashattF is not supported because: device=cpu (supported: {'cuda'}) dtype=torch.float32 (supported: {torch.float16, torch.bfloat16}) max(query.shape[-1] != value.shape[-1]) > 128 Operator wasn't built - see python -m xformers.info for more info triton is not available smallkF is not supported because: max(query.shape[-1] != value.shape[-1]) > 32 unsupported embed per head: 512

[lusty raptor]

could be, there's only so much to go around

[livid phoenix]

Do you guys know how to fix this? Bc I can't load any image prompts in Stable Diffusion

[spring fulcrum]

I might be able to help

[livid phoenix]

Thank you

[spring fulcrum]

Have you edited the .bat file you use to run ComfyUI?

[trim orbit]

lofi hiphop might've been what she started with

[livid phoenix]

Uh... kinda? It's the WebUI-user.bat file

[trim orbit]

triphop beats i used to listen to had lowfi subgenres

i'm over it

[spring fulcrum]

Are you using ComfyUI or VLAD A1111?

[livid phoenix]

I open it up in Notepad and added in the prompts to COMMANDLINE_ARGS: --xformers --autolaunch --skip-torch-cuda-test --no-half

It's A1111

[spring fulcrum]

ok let me check and see what I have in that one.

What graphics card do you have?

[spring fulcrum]

It might be

[trim orbit]

sdxl only works in the release_candidate branch of auto111

[livid phoenix]

Where do I find the graphics card?

Look: can we take this to DMs? Bc I don't want this chat getting crowded or drowned out

[trim orbit]

git switch release_candidate
git pull

[livid phoenix]

Where do I imput that?

[trim orbit]

hmm. in the terminal. you might want to wait for the model to be released fully instead of just throwing it at a random install that you don't understand. I could explain how git works to you, but theres the next question, then the next.

I mean, they're git commands. where do you figure? in the root of the folder. someone should tell you this. beta software requires some knowledge about how the software works, generally.

learn your tool chains. read the manuals.

i believe in you

[livid phoenix]

Look: perhaps we should take this to DMs, and you can explain the step-by-step process in installing Stable Diffusion and guiding me through it

[trim orbit]

do i seem like a person who feeds people step by step?

[livid phoenix]

Bc I've clearly been thrown into the deep end

[trim orbit]

sink or swim

[livid phoenix]

How am I supposed to know how to use the program?

Look: do you know somebody who can help me with the installation process?

[spring fulcrum]

Hold on there are videos that can walk you through the whole process

[trim orbit]

/points at the entirety of the internet and it's vast annals of knowledge, with a wide grand gesture

[brisk vigil]

I would like to see images created with sdxl as an example of its power, where can I find them?

[livid phoenix]

I've followed each and every one of those videos - step by step. Nothing seems to work

[brisk vigil]

it hasn't been released yet right?

[trim orbit]

the #1100170312106127410 and other rooms let you make images

[spring fulcrum]

That is what brought me to my question... What graphics card to you have... If you don't have one it changes the entire process... If you do have one we need to know if it is an AMD or NVIDIA graphics card... These are questions I can't answer for you.

[livid phoenix]

I'm trying to find my graphics card

I'm on my device manager. Where do I find my graphics card?

[spring fulcrum]

It would probably be something you would know about because they are really big and pretty expensive... It's not really something you would miss.

[visual glade]

if you want the simple way of running SDXL get comfyui and the workflow pinned in this channel

[livid phoenix]

No, I don't have a graphics card. It seems to be inside my computer screen, which also doubles as a monitor

It's basically a hybrid of sorts, but I don't think it contains a graphics card

[spring fulcrum]

Then you would need to follow the installation process to use the CPU and RAM instead of a graphics card... The next question is are you on a MAC or a Windows PC?

[trim orbit]

its an all in one. it's ancient

[livid phoenix]

Windows PC

[spring fulcrum]

now we are getting somewhere

So, honestly unless this is very new and very powerful it is probably going to take a very long time to generate an image with your PC if it can even handle it... However you need to set command line arguments that force it to use the CPU and RAM and not to look for a graphics card.

[livid phoenix]

Alright, so I would need to have Stable Diffusion installed into a CPU and RAM

[spring fulcrum]

not exactly

[livid phoenix]

Yes, I got my PC a year or so ago. So it's brand-new

What command lines am I looking for?

[spring fulcrum]

The program gets installed on a hard drive. The program needs to know that it needs to use your CPU and RAM the CPU is the brain of the computer and the RAM is what stores temporary Memory and helps the CPU

I will need to look up the command line arguments for A1111

I need a few minutes unless someone else knows what they are

[livid phoenix]

Well, now I need to reinstall everything again

Great

This is so confusing

[urban fjord]

Depending on your computer you will be best off using ClipDrop or the bot here to use SDXL as running on CPU and not GPU is really slow.

Alternatively you can use Google Colab.

[spring fulcrum]

I'm sorry I'm not trying to be confusing... There is a big learning curve though if you aren't familiar with your systems components

[livid phoenix]

It's alright

[spring fulcrum]

Just give me 5 minutes and I will try to get those command line arguments and then we can go through those steps

[livid phoenix]

Alright I'm re-installing Python, Git, and the Stable Diffusion 1.5 and checkpoint

[spring fulcrum]

You probably don't need to do that

Try to put this for the command line argument

--use-cpu all --precision full --no-half --skip-torch-cuda-test

[livid phoenix]

I'll try

[spring fulcrum]

It should look exactly like this spaces and all

The one on the left is the one you use the LoRA on?

[trim orbit]

left is off. discord reversed the order i put them in.

both suck though it's just my prompts

[spring fulcrum]

So you want them to look like the one on the right? The blue one with 4 tracks?

[trim orbit]

towards the asteroid goal

[spring fulcrum]

It looks very futuristic

[spring fulcrum]

are you having any luck?

[trim orbit]

i threw 200 in game graphics from the terran race from starcraft and a bunch that marketing art at a lora. but didn't tag any of it with starcraft words so it's more generalized.

[spring fulcrum]

I like it.

[livid phoenix]

I'm waiting for Stable Diffusion and the checkpoint to finish downloading

[spring fulcrum]

I might try to learn how to train my own model or LoRA at some point. I would really like to see some High End Gundams.... If you know what those are you might just want to switch over to those from tanks... They are pretty badass

[livid phoenix]

Alright, managed to clone everything on Stable Diffusion into a new folder

[spring fulcrum]

Awesome... Let me know when you edited your webui-user.bat file

[livid phoenix]

I did

Copied everything into the commandline args

[spring fulcrum]

alright give it a try

make sure you save it

[livid phoenix]

Now I just gotta wait until I can copy my CetusMix checkpoint into the models file

I did

[trim orbit]

i got mechs in the dataset. you've heard of gundam but not starcraft?

[spring fulcrum]

Let me know if you have any success with it.

[livid phoenix]

Alright

[spring fulcrum]

I heard of it sure... I used to see it when I used to log into my World of Warcraft account. I never had any friends who played it so I never got into it.

[spring fulcrum]

If it still fails with the CetusMix checkpoint you should try it with the basic original Stable Diffusion 1.5 checkpoint... and if that doesn't work then it might just be that your computer cannot handle the image generation process.

[livid phoenix]

Alright, I'll try. Even if it might not generate the results I'm looking for

[spring fulcrum]

right but at that point its really about finding out if your computer is even capable to do this or not and less about the specific results

[livid phoenix]

Well, it's a pretty new computer that has plenty of space

[spring fulcrum]

Its more complicated than that

[livid phoenix]

So it's just compatibility

[spring fulcrum]

Its about the parts in the computer... think about it like this.. Would you use a toyota prius to race a porsche? or would you use your everyday car to enter in a nascar race? The parts matter... It doesn't matter if its a brand new toyota prius if it doesn't have the parts to win the race... That is the best way I can think of it.

[livid phoenix]

I understand

It’s starting to freeze. I think it’s working

[spring fulcrum]

It might take a really long time to generate one image but give it a chance.

[livid phoenix]

I know

Let’s hope it doesn’t give me some kind of error message or whatever

And it looks like I might have to manually open Stable Diffusion once all of this has finished installing

Okay, it’s given me a couple of notices

That a new release of pip is available - 23.2
And if I need to update, I need to run: venv/scripts/python.exe -m pip install —upgrade pip

[trim orbit]

it feels messy. hmm.

[trim orbit]

oh that guy? you don't need to worry about him

[livid phoenix]

Alright

[trim orbit]

https://www.youtube.com/watch?v=-u9zr6oF4uA thank you for allowing me an opprotunity for a solid gold reference

[livid phoenix]

lmao

Okay, it works

[spring fulcrum]

Awesome..

Can you upload the image?

[livid phoenix]

Warning: caught exception 'Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx', memory monitor disabled

Uh... is this a problem? I got it in the cmd.exe file I'm running for the WebUI User program

[spring fulcrum]

well if you have no graphics card there would be no NVIDIA driver so at least the error makes sense

[livid phoenix]

I'm generating an image

Okay, I managed to successfully generate an image

[spring fulcrum]

https://tenor.com/view/dancing-happy-dance-praise-yas-excited-gif-18656424

[livid phoenix]

Added in "Dimentio from Super Paper Mario, masterpiece" into the prompt

Thanks for helping me out, @spring fulcrum

[spring fulcrum]

It takes a while to get a prompt that works perfect but at least now you are able to keep trying with the prompts

Your welcome... Happy to help.

Good luck with your images

[livid phoenix]

Now I'm gonna figure out how to add in some styles to better improve my AI preformance

[spring fulcrum]

https://tenor.com/view/good-luck-you-got-this-you-can-do-it-you-can-do-this-suerte-gif-18340019

All right I'm going to bed. See you all tomorrow.

[livid phoenix]

Alright, any idea where I can find all kinds of styles I can use for Stable Diffusion?

[agile quarry]

prompt them. you just describe what you want. here's the style prompts the bot uses though

[livid phoenix]

okay thank you

[agile quarry]

you can see images with prompts here. kinda see how different people prompt https://lexica.art/

[trim orbit]

while yeah, good example of a good prompt database, it's all skewed to their highly finetuned version now a days. another problem is it's all towards the clip from sd1. One of SDXL's encoder is clip and one is openclip

[agile quarry]

I think they're using 1.5 for the first time

[trim orbit]

mb

[agile quarry]

https://civitai.com/ @livid phoenix here's a popular host of custom models. Different models or adding LoRa's can impact style allot

[livid phoenix]

Wait... was I supposed to put the character checkpoint in the models folder? Or does that not really matter?

[agile quarry]

I'm not sure what you have. If you got a character LoRa they go in a specific folder, textual inversions go in a different folder

[livid phoenix]

I have cetusMix

Apparently it's a style checkpoint

[agile quarry]

models/Lora #Lora
Models/stable-diffusion #full model

[livid phoenix]

ok thx

[agile quarry]

It's a full model . Even LoRa's are checkpoints it's pretty much the same as saying model

[livid phoenix]

Okay, where do I find these

[agile quarry]

the folders to place your models? mine looks like this sd\stable-diffusion-webui\models\Stable-diffusion where ever your sd folder is are you using A1111

[livid phoenix]

Yeah, I'm using A1111, and yes: I did figure that out. I put in the cetusMix style in there

[livid phoenix]

No, I'm talking about where I find the Lora models I need

models/Lora #Lora
Models/stable-diffusion #full model

Where can I find these so I can put it in the folder?

I'm on Civit.ai

[agile quarry]

adjust filters

[livid phoenix]

I did

[deft turret]

If you're looking for a specific character or style, just try to search it directly on civitai. The results will have little text on corner telling you whether it is a checkpoint or lora or textual inversion (TI). Then you can adjust the filter accordingly if you're getting too many results

[livid phoenix]

Okay

[deft turret]

For characters and styles, Lora, TI and LyCoris will work for you

[livid phoenix]

Got it

Another thing: how exactly do I make fusions of two or more characters in Stable Diffusion?

[deft turret]

TI will go in embeddings folder

[deft turret]

You can use two loras in your prompts. Just adjust their weights acc to your needs

[livid phoenix]

Alright

How do I make it so thatI get male models?

more specifically: how do I edit my prompts such that I get male characters?

[agile quarry]

you can just prompt things together with the base model too bobby hill as spider-man words describe what you want its text to image