✨｜sdxl

[latent river]

Someone gotta train sdxl on loads of backrooms and liminal space images

[dapper current]

[autumn forum]

im sure a lora will come out with it. i do like the look of pool rooms as i have done a few in blender so maybe ill train a lora for it 😛

[covert viper]

Sdxl

[hard fractal]

I'd love someone to train a deep focus LoRA.

[latent river]

1.5 never gets the poolrooms right

[ionic dragon]

use laion dataset, so if they are there then they had, or no

[hard fractal]

@wicked frigate – speaking of LoRAs...

how about we release one along with SDXL?

[dapper current]

[ionic dragon]

do you want any specific lora?

[autumn forum]

im sure sdxl could make some pretty awesome ones. ive already generated a few and it did surprisingly well

[spark bear]

split depth lora for trippy faux 3d animations..hmm i might try that

[hard fractal]

a generalist LoRA that's trained on deep focus photographs.

[covert viper]

Yo guitarman, everyones beating around the bush, sdxl coming out today/tomorrow?

[ionic dragon]

like landscape photography?

[tribal magnet]

i'm still waiting for SD to learn what a razor blade is. not a disposable razor, not a straight razor, just s regular old iconic blade. For art. And also science. ;_;

[autumn forum]

the joke is to beat around the bush lol

[nimble heart]

it'll come out when Emad finishes playing Tears of the Kingdom

[covert viper]

Next game of civ

[boreal bough]

can we get a guide on meaningful use of TE during LoRA training? cause the results of that are just chaotic unless you do TE first and then freeze it

[latent river]

How did they turn out?

[civic sigil]

Pretty sure the stage tomorrow is a good estimate lol

[ionic dragon]

||<t:1690394400:R>||

[civic sigil]

The mystery is controlnet when

[wicked frigate]

Great idea, we should ask @tiny walrus to help make one

[hard fractal]

could be! or, oo, I wonder how well SDXL knows split diopter

this kinda look

spielberg loves 'em

[nimble heart]

when you stare at it too long it looks like the foremost guy is photoshopped in lol

[spark bear]

hot damn its the soggy bottom boys

[latent river]

Chilling in Chile eating chilli while it’s chilly

[covert viper]

Tbh they both look photoshopped in or greenscreened

[ionic dragon]

i think a model would be better

[hard fractal]

oh, it sure don't know split diopters

from SDXL there.

1.0

[ionic dragon]

if can suggest the settings, i would love to collect the dataset and training on my machine

[hard fractal]

nah.

[tribal magnet]

can sd generate "magic eye" pictures, like the ones that used to be all over the mall in the 90s? because thats something the world needs.

[ionic dragon]

you just want to add depth effect?

[hard fractal]

Some of the LoRA types have more params than the original SD 1.5 model 🤯

[knotty lotus]

[autumn forum]

i got this with a film still from a move of 2 men split diopter style

[hard fractal]

deep focus images, yeah.

[hard fractal]

ooooo

nice

[hard fractal]

🤷‍♂️

[boreal bough]

I'm just looking forward to the LoRAs that achieve finetune levels of changes, with under a 100mb

[civic sigil]

I dont mind a chonky lora as long as it can be trained on a 24gb card

[latent river]

I’m the future we will laugh at sdxl

[hard fractal]

I mean, you can train the full model on a 24GB card.

[boreal bough]

tbf, they dont need to be chonky.
I managed to fit over 100 concepts, fully learned, nothing overfit, into 43mb

[knotty lotus]

just gunna one-up you here...

[hard fractal]

lol, stop the release

fix cat eyes.

[wicked frigate]

wow, kaj, yknow what would be even better than one clean UI over comfy dropping soon?

[autumn forum]

agreed, sdxl should make perfect cats all the time..

[civic sigil]

I doubt I can on windows tbh

[boreal bough]

but wouldn't that do more damage than good?
considering batch size of 1. Unless something changed about finetuning itself, which I'm not aware of

[latent river]

Almost 2 years ago…

[hard fractal]

why no grad accum?

[tribal magnet]

no no no we need BEYOND perfect cats, so those little nerds can get body image issues like the rest of us

[knotty lotus]

[hearty ginkgo]

That's sdxl 1?

[eternal fog]

If it's from 2 years ago, then no it's not lol

[latent river]

It’s from mindalle anyone remember that?

[tiny walrus]

Can do--I was doing an exploratory one the other night. Totally feasible

[hard fractal]

aww

[hard fractal]

Kali, got any cool 1.0 pics?

[autumn forum]

so yeah it doesnt know split field diopter very well no matter how i prompt haha

[knotty lotus]

make a miku lora 🙏

[latent river]

Anyone got nice car drifting pics

[hard fractal]

Love this one

[boreal bough]

because I completely forgot that that exists XD
good point 👍

[tiny walrus]

yessir, comin' up

[dapper current]

GA and Batch Size is like Peanut Butter and Jelly

[civic sigil]

How do you enable this with kohya

[hard fractal]

oh, that I'm not sure.

[tiny walrus]

I am also not sure fwiw

[civic sigil]

Idk if there are other trainers, psuedo said he was working on one but he implied you can only do batch size 1 on a 3090...

Waifu uses the official SAI code?

Or just their own trainer

[latent river]

Thoughts on a1111 becoming obsolete?

[dapper current]

I highly doubt it will become obsolete

[knotty lotus]

for Waifu Diffusion XL they used a modified version of kohya's trainer afiak

[dapper current]

Overshadowed? Perhaps.

[spark bear]

everdream has grad acum

everydream

[urban fjord]

I thought Kohya-ss had grad accumulation. There is at least a setting for it.

[wicked frigate]

re training and batch sizes and all: use lora lol

can set batch higher with lora

and like... why would you not use lora when lora is an option? Excluding the case of full heavy model retunes like waifu

[boreal bough]

not to be confused with --gradient_checkpointing

[wicked frigate]

Anything where you'd expect to be able to do it at home anyway, is better down with LoRA

lora was already a better option usually on SDv1, it's much more so on SDXL

[wicked frigate]

(both because the base model is a harder target to train, and because loras trained on it are more powerful)

[boreal bough]

tbf, I still wouldn't finetune on my 4090. just my high end lora will run a total of 30 hours

and that's a mere 5k dataset

[civic sigil]

I wouldnt mind if it takes a week lol I have two computers now anyway

[boreal bough]

at that point move to runpod and get that 8stack of A100s

[civic sigil]

I like full models, idk exactly how loras work

[boreal bough]

don't forget that 1.5 lora =/= sdxl lora

[civic sigil]

With full models I feel more comfortable doing merges and stuff, not sure how that works with lora

[urban fjord]

We do not need 100s of identical merges.

[civic sigil]

How is it different

[autumn forum]

i thought they were trying to say that sdxl loras are much more powerful and would basically eleminate the need for whole finetuned models unless its a really big change?

[visual glade]

long term it's likely, short term probably not

[civic sigil]

I mean I have my own datasets with art styles that are not super well understood by the base model imo

Or community finetunes

[half ivy]

would it be possible for each node to show a percentage count thingy

[civic sigil]

But I would still like to merge with other community finetunes

[dapper current]

I think there are going to be a lot of great UIs for SD and they'll all have their places with their respective pros and cons.

The biggest thing holding back Auto1111 is the fact it's built on Gradio

[urban fjord]

LoRAs are perfect for this.

[rustic temple]

the SD userbase should migrate to Comfy and InvokeAI. that is all

[urban fjord]

I've yet to see anything better than Gradio

[tiny walrus]

seconded

[civic sigil]

How do loras work with supermerger?

[knotty lotus]

[civic sigil]

Id rather just use full models which I vaguely understand how they work

[agile quarry]

What do you mean? You can use old LoRas with sdxl?

[urban fjord]

You don't merge them.

[civic sigil]

Yeah but I want layer specific merging

[visual glade]

it's not gradio but the code quality that's going to make adding new features more and more difficult

[urban fjord]

And Gradio's inbuilt --share is really nice.

[nimble heart]

output folder is just full of mangled kitties after a debugging session

[urban fjord]

Yeah, the code quality is a mess but it looks like Auto1111 has become a better developer and has long stopped pushing unstable comits to master.

[dapper current]

Oh I know that one first hand! Even with a rewrite though Gradio is tougher to extend as people are going to want more functionality.

[autumn forum]

poor kitties

[nimble heart]

a1111 didn't even have a repo license till somewhat recently. I don't think he ever intended it to blow up, but since his was the first "extensible" UI it just kinda did

[upbeat summit]

you need to diffuse them to save em

[boreal bough]

the theory of how a lora works, didn't change at all.
but sdxl model has way superior clip models + base weights and more parameters.

therefore lora can:
A.) make use of those extra params
B.) concepts already exist, you're just improving & bringing out existing hard to prompt for information

lora can't:
• Turn sdxl into waifudiffusion xl

[trim orbit]

there's an obvious solution to everything. rewrite in rust

[dapper current]

Auto1111 is a fork in and of itself from hlky's original webui

iirc

[urban fjord]

Problem with rewriting it in rust is that the people making extensions doesn't know rust.

[knotty lotus]

on it..

[civic sigil]

I guess the main feature I want a full finetune for is layer specific merging with other finetunes

[urban fjord]

Yeah you don't do that with LoRA.

No need to merge just to merge.

[nimble heart]

my rocm 5.6 benchmarking kitties are worse lol

[urban fjord]

Sure you can merge a LoRA into the base model but that defeats the point.

[knotty lotus]

oh really? I didn't know that

[autumn forum]

big ol floof mess of kitties

[urban fjord]

Take Auto1111 frontend and put it on ComfyUI backend and it's perfect.

[tribal magnet]

oh, btw... um. how BIG is sdxl? im sure i wont need a new drive, but... will i need a new drive? lol

[trim orbit]

til too

[urban fjord]

About 18.5 GB with base + refiner.

[visual glade]

the a1111 frontend isn't very good in my opinion

[nimble heart]

you probably could tbh. All the rust UIs have wasm/web support

[urban fjord]

The auto1111 frontend is still the best out there.

[boreal bough]

but previously, we needed finetunes because the model genuinely couldn't produce certain images/artstyles/photos due to a lack of information & parameters
sdxl doesn't lack this anymore. sure it has its issues, but not on a level where you need a finetune for 95% of situations

[dapper current]

This was done recently kinda, but you still have all the same overhead from Auto1111 which doesn't really give you the true benefit of Comfy

[tribal magnet]

ahh, thanks. not too bad, i have plenty of room for it

[dapper current]

Not a true Front only with Comfy back

[knotty lotus]

a1111 frontend isn't maintainable and actually doesn't lend itself to a good iterative workflow believe it or not

[nimble heart]

and there's an egui extension that adds node graphs. could rewrite comfy in rust

[civic sigil]

I havent tested a full finetune with it yet but the lora with my favorite dataset was kind of underwhelming

[trim orbit]

i got comfybox working last night. i might try to put together an sdxl ui in it for funsies

[urban fjord]

I've tried doing iterative workflows with ComfyUI and it is terrible for that.

[civic sigil]

Maybe because it is only rank 8

[boreal bough]

and you're sure it wasn't a lack of lora training settings?

[knotty lotus]

can I drop this tweet? I think I can drop this tweet, why not... https://twitter.com/kaj718/status/1683985659172892672

kaj (@kaj718)

https://t.co/JoEC29mptN

local inference is sweet!

▶ Play video

Twitter · Jul 25, 2023, 11:40 PM

[civic sigil]

It could be

[boreal bough]

How big was your dataset, and can you post the full captions of a single image as a sample?

[civic sigil]

Is this the project mcmonkey has been talking about?

[knotty lotus]

he's probably talking about a different thing actually

[nimble heart]

that a native UI or is it a web framework

[knotty lotus]

yes

[civic sigil]

Maybe but I dont want to share the artist I'm using lol

[urban fjord]

I'm optimistic to whatever mcmonkey is hinting towards.

[autumn forum]

will definitely be using that ui heck yeah. hope it has inpainting implemented too at some point

[boreal bough]

replace artist with <artist> or something. just want to get a feel for your captioning style

[ionic dragon]

is 1.0 good at generating faces when in distance?

[boreal bough]

better than 0.9. that's for sure

[autumn forum]

this is awesome

[nimble heart]

okay 99% its a web framework.
or you just really like material design

[knotty lotus]

it's react, running on a native webview using tauri

its open source, you can take a peek

[nimble heart]

damn you actually did rewrite it in rust

[wicked frigate]

that's right, two new clean UIs over comfy lmao

[boreal bough]

I think I cry. fiiiiinnnnaaaalllllllyyyyyy a ui that doesn't remove the results after queuing more
@wicked frigate ❤️

oh wait. is this the work of kaj? O:

[wicked frigate]

i was waiting to infodump about my one til tomorrow's stage event

[knotty lotus]

yeah there's still some more cool stuff to come

[wicked frigate]

kaj is building the super duper easy/beginner friendly StableStudio integration for comfy, I'm building a more poweruser oriented thing

[autumn forum]

that is the best ui ever and cant wait to try it out on 1.0 . i like comfy;s nodes, i like invokes style inpaint, and ease of use like a1111. BEST DAY EVER

[nimble heart]

are you using libtorch through ffi or is the backend still python

[boreal bough]

in that case. I take my love away from mcmonkey and give it all to @knotty lotus ❤️

[wicked frigate]

no rust but we're all moving away from "python everywhere all the time"

[dapper current]

[knotty lotus]

no it's a forked version of comfyUI, same backend as everyone else, means you can use your same custom nodes

[wicked frigate]

alright fine you don't get my one then

[boreal bough]

😭

[nimble heart]

Tauri is rust I thought

[urban fjord]

mcmonkey: Does yours support prompt-editing and x/y/z/...?

[dapper current]

Oh does it ever support x/y/z

@wicked frigate is the author of Infinity Grid

[wicked frigate]

Tauri has rust stuff but coding is usually typescript

[knotty lotus]

tauri just serves static files in a native webview and manages IPC/events/etc

[hard fractal]

you are

[wicked frigate]

XYZ and all the other letters too yes lol

[visual glade]

just make sure you sync your fork with the latest code

[wicked frigate]

and kaj and the other engineers!

[ionic dragon]

if its bad at distant images, then it would make no sense to train a lora on conematic photography

[lusty raptor]

ooh!! ooh!!

[visual glade]

there's some fixes that are important if you want it to work on older GPUs

[hard fractal]

I'm either sticking with Python forever, or moving back to assembly.

[upbeat summit]

this took a couple of seeds without controlnet or img2img heh

[visual glade]

if I had time I would rewrite comfyui in C

[dapper current]

Let's just say @wicked frigate 's can do a/b/c/d/e/f/g/h/i/j/k/l/m/n/o/p/q/r/s/t/u/v/w/x/y/z and depending on your setup VERY fast too

[wicked frigate]

i'll teach you C# myself

and you'll ignore me and use ChatGPT to write it for you

[hard fractal]

plz no.

[wicked frigate]

lol

[hard fractal]

i'm a mac boi.

[dapper current]

[boreal bough]

we tortured the bots a bit, to test distant faces. results were significantly better than expected. I'd give it the highest score of all the existing AI Image Generators that currently exist

[civic sigil]

Plot twist: They specifically trained a bunch of images of text that says SDXL

[wicked frigate]

AND I NEED A MAC DEV TO ENSURE COMPAT IS PERFECT thank you for volunteering

[upbeat summit]

are you sure?

[tiny walrus]

btw, saw your reddit post! Try ✨[Prompt/style]✨ instead of ~~[Prompt/style]~*~ *

and lmk what you think

[wicked frigate]

you don't read your DMs apparently

[visual glade]

should use stableLM instead

[dapper current]

😅

[boreal bough]

50% chance of very rich programmer / 50% chance of did his fair share of game/app hacking XD

[hard fractal]

https://tenor.com/view/freewilly-yeah-freedom-wet-perfect-gif-9911674

[civic sigil]

That's their next release, they snuck in SDXXL images in the training data to build up hype

[nimble heart]

that'd make custom nodes super annoying unless you included a whole scripting language interpreter in it

[ionic dragon]

@hard fractal i have the image links of entire 1x.com, so if someone's interested I can share them, do you want them?

[upbeat summit]

Cool, that you saw it 🙂 Thank you again for sharing your findings. I guess discord somehow changed the syntax. I will try it out for sure!

[tiny walrus]

nah, they both work

I just think the emojis work better

[visual glade]

at least I wouldn't have the memory management issues of pytorch + python

[ionic dragon]

SDXXL when?

[upbeat summit]

yeah, I've done some emoji experiments back when SD 1.4 was in beta heh

[dapper current]

What about SDXS?

[hearty ginkgo]

Nah

Sdxl tate

[dapper current]

[hearty ginkgo]

I'm jp

[nimble heart]

a1111 changed the default allocator to tcmalloc a while ago which supposedly helps a bit

[tiny walrus]

they even worked with VQGAN. an under-explored universe in my opinion 🙂

[ionic dragon]

no dont think of it,
why to release a mini ver, SDXL can run on 6gb cards(probably) so is there a point to get SDXS

[visual glade]

I don't understand how that would help

[nimble heart]

supposedly fixed their memory leaking problem

[visual glade]

I think it just delayed it a bit

[boreal bough]

July 18th

[civic sigil]

If there is a user friendly front end maybe I will release my crazy animation workflow

[trim orbit]

SDXXXVIII is the one with janetjackson's nip slip

[boreal bough]

the band aid treatment is starting to become impressive XD

[upbeat summit]

I really like the detail in this one. great images!

[nimble heart]

I don't even think it fully fixed it, I think it just delayed it long enough that it's not a massive issue anymore

[boreal bough]

starting to look like a final fantasy series game

[visual glade]

I doubt regular malloc has memory leak bugs so switching to another one won't fix anything

it's probably one of the most tested functions

[trim orbit]

maybe the romans did have a real oracle that was a bunch of guys abacussing really hard

[visual glade]

and malloc doesn't have anything to do with python or pytorch memory management

[trim orbit]

1it/month

[boreal bough]

with enough abasusses, you can simulate the whole world

sounds like a monty python skit

[knotty lotus]

I don't think an abacus is turing complete is it?

[boreal bough]

it is with a human attached

[urban fjord]

If you can create a nand gate with an abacus it is turing complete

[hard fractal]

Funny enough? We do have one of those, haha

[upbeat summit]

SD2049

[hard fractal]

lil baby model

[urban fjord]

Are you going to release it SDXS?

[hard fractal]

hell no

It sucks

[civic sigil]

Is that the famous 30 FPS model

[knotty lotus]

fair

[hard fractal]

That's the famous distilled model

[west breach]

Is the refiner a 512 model?

[autumn forum]

is the hype train over? i needa go to bed lmao

[dapper current]

[hard fractal]

We'd have to give you all a couple days warning before putting it up on the bot

People are like "he STILL HASNT RELEASED IT"

Like... ya. You think distilled 1.5 base was any good?

[boreal bough]

go sleep! XD

[nimble heart]

that's just what I read in this issue which caused them to set LD_PRELOAD to the system's tcmalloc dylib
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/9323

[hard fractal]

Now, distilled SDXL 👀👀👀👀

[upbeat summit]

are you really sure (sure)?

[trim orbit]

🚂 🚃 🚃 🚃

[tiny walrus]

😅

[dapper current]

[hard fractal]

Distilled SDXL 30fps 3D hologram video will be out July 18

[upbeat summit]

SDXL distilled

[spark bear]

hey, thats when controlnet drops

[hard fractal]

SdXL

[autumn forum]

i think i am, ill see yall tomorrow when sdxl, maybe, possibly, probably, releases. even though ill be at work im definitely tuning into the super stage..

[trim orbit]

holograms are actually a really cool physical phenomenon and 3d projections aren't holograms

[urban fjord]

Isn't it about 14 hours till release?

[hard fractal]

Release of what?

[urban fjord]

So you can sleep with good conscience.

[trim orbit]

🤓

[ionic dragon]

any estimate when control net would release?

[hard fractal]

July 18

[urban fjord]

Release of SD 3.0 of course, SDXL has just been a distraction from the real truth

[dapper current]

In the year 20xx

[visual glade]

it probably helps if tcmalloc is more efficient with memory with lots of small allocs and frees but it's not a real fix

[urban fjord]

We all expect SDXL but instead we get something even better.

[hard fractal]

but in reality, no - we'd rather reinvent ControlNets

[ionic dragon]

https://tenor.com/view/nervous-sweating-sweating-bullets-sweating-gif-17507069

[urban fjord]

Why reinvent it when it already works well?

[hard fractal]

@visual glade had an INSANELY BRILLIANT idea

[trim orbit]

because thats how tech improves

[hard fractal]

I posted about why on Reddit

[urban fjord]

Sure reinvent it eventually if you can do it better, but the original is still good.

[dapper current]

SDXS might be unncessary...

but...

[civic sigil]

That sounds like it will take a while

[ionic dragon]

@hard fractal a simple question, how do control nets gets trained, sd is trained through images and captions, how is it trained?

[hard fractal]

original would make a 1.3 billion param model for every single ControlNet

[trim orbit]

SDXXX is the one with VinDeisel an IceCube

[hard fractal]

That's bigger than SD1.5

[visual glade]

controlnets are not very efficient, they can be more efficient

[civic sigil]

I wouldn't mind

[urban fjord]

There is too many controlnets, most of them are not really needed.

[civic sigil]

I hope someone trains controlnets other than SAI then

[trim orbit]

the various models i don't think are the issue at hand

[hard fractal]

Then you're free to train one -- id imagine some people will from the community, the code is out there

[dapper current]

Most CN are trained with the preprocessor image along with the original image and caption to teach the model what it should do

[boreal bough]

<t:1690394400:R>

[trim orbit]

anyone can. the OG were all made on the author's 3090

[hard fractal]

But, we care about normal folks running it.

[nimble heart]

I tried looking into how a1111 is able to eat 60 gigs of ram but their backend is split into so many weird semi-abstraction things that I have no idea where to begin with it. Think it'd actually be faster to just write a new gradio UI than trying to fix it.

[ionic dragon]

https://www.reddit.com/r/StableDiffusion/comments/158n1td/controlnet_sdxl_10/

[urban fjord]

How do you get Auto1111 to eat 60 GB RAM?

[deep dagger]

Feeding XL outputs directly into img2img animate diff in comfy.

[nimble heart]

XY grid with different models

[urban fjord]

You probably have a lot of caching turned on in Auto1111

[nimble heart]

nope

[trim orbit]

fun!

[nimble heart]

I'm set for 2 models and 2 vaes

[hard fractal]

Time for C0NTR0LN3XT

[dapper current]

StableControl

[civic sigil]

My whole workflow is gonna be on hold since I hardcoded in the 1024x image size and there's no controlnets for SDXL 😢

[hard fractal]

ComptrollNet

[knotty lotus]

StableNets

[ionic dragon]

Joe, can we get few more preview pics of the control net?
any pose pictures?

[hard fractal]

No pose adapter trained yet

Changing arch first

[eternal fog]

Are you guys looking at something like this for controlnet?
https://canqin001.github.io/UniControl-Page/

[hard fractal]

No.

[urban fjord]

You just have the line Controlnet?

[hard fractal]

Kinda.

Kiiiinda but no.

[civic sigil]

I hope their will be a good model for facial expressions

[urban fjord]

Not a fan of it anyway so I don't care too much that you're not releasing it.

[ionic dragon]

ahhh, you are still figuring it out.
so it wouldn't be released anywhere under a month

[hard fractal]

Doubtful

Very doubtful. Expect community ControlNets before ours

[civic sigil]

Dang even 2.0 had a depth model on release

That's a shame

[dapper current]

We really want to innovate and provide new amazing stuff for everyone

[nimble heart]

How does SDXL scale with controlnets? 1.5 every layer of controlnet had a pretty substantial performance hit.

[trim orbit]

a shame might be harsh language. 2.0 certainly didn't have controlnet anything at release

[civic sigil]

It had a depth img2img model

[boreal bough]

@hard fractal will you release like a 100 image dataset sample with your training captions for both clips attached?

[trim orbit]

2.0's depth awareness is 256x256 pixels iirc

[civic sigil]

Which was fine, it was a good model

[trim orbit]

it's neat but experimental and kind of, meh

[civic sigil]

Really? I thought it was solid

[urban fjord]

It was attached to 2.0 which meant I never used it.

[trim orbit]

seems a weird thing to shame them over

[civic sigil]

I mean it's just a step backwards. And I'm not shaming anyone, just saying it's a shame there won't be controlnets ready on release when Emad implied there would be a while back

[dapper current]

It's because we took a different step forward

[ionic dragon]

they trained a new clip model?

[dapper current]

We had them

[nimble heart]

apparently you can disarm the a1111 memory bomb with some cursed super user commands

[urban fjord]

How do you know that you're actually taking a step forward without an original version to compare against?

[civic sigil]

What's the harm in releasing if there's going to be community weights anyway

[dapper current]

They don't work with these ones anymore

No decision was made lightly

[civic sigil]

Ok, well I'll trust the community then lol. Hope the SAI version is good too

[trim orbit]

if i knew how to make useable controlnet datasets, i'd do some training myself

[dapper current]

The info will be out there and you definitely can!

[half ivy]

what is cr?

[azure oxide]

iirc either the controlnet github or the a1111 controlnet extension github has instructions on training your own controlnet model

[civic sigil]

There's a guide on how to do it, obviously generating the dataset could be difficult depending on what you're training

[nimble heart]

custom node looks like

[trim orbit]

i imagine it's a lot of image pairs. photo and a corresponding openpose read on it

[civic sigil]

There's a docker for training cnets, really easy to use

Idk about for xl obviously

[trim orbit]

yeah i've looked into it. doesn't seem like it'll take much to work out

[trim orbit]

can't wait! well i guess we'll have to, but i don't wanna!

[upbeat summit]

https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[boreal bough]

oh shit, how did I not see this already being released
https://github.com/Stability-AI/StableStudio/tree/tauri

[shy kelp]

where can I find that custom node?

[boreal bough]

thought it was just a preview

[upbeat summit]

if you want to know what's up in the world of ComfyUI just bookmark this url - everything mentioning comfyui on github sorted chronologically: https://github.com/search?o=desc&p=1&q=ComfyUI&s=updated&type=Repositories

[trim orbit]

its not what i thought it was. it's a front end for stability's online services. open sourced

[dapper current]

It's just been kinda sitting there... being quiet 🙃

[nimble heart]

My repo: https://github.com/Beinsezii/bsz-cui-extras

[dapper current]

Take a look at that specific branch though

[boreal bough]

you can connect to localhost

[spring fulcrum]

Can you post an image with that you made with your setup... then I can just drag it into the UI

[nimble heart]

image embed meta only has the workflow, not any custom nodes

if you already have the node ig this would work then

[trim orbit]

looks exactly the same to me. it generates locally now?

[dapper current]

Using Comfy

[spring fulcrum]

I just need to put your two python files into the custom nodes folder right?

[nimble heart]

yea

[dapper current]

😏

[trim orbit]

you guys have been working towards convergence this whole time i bet

https://tenor.com/view/shia-la-beouf-shia-la-beouf-clapping-brando-gif-4502586

[boreal bough]

damnit XD lemme use my existing installlll

[dapper current]

@knotty lotus ^

[trim orbit]

guess you gotta build that tree yourself? the actual release page takes me back to main branch. i'll figure it out

[nimble heart]

symlink everything

[shy kelp]

Yup Just tried it out, wonderful app
told it to install to my H drive and it just downloaded 5+ gigs to my C drive -_-

[boreal bough]

I worked too long in IT to ever use symlinks

[shy kelp]

why even ask me where to put a 22mb exe file if you are just gunna fill up my C anyways?

[nimble heart]

what does that even mean

[dapper current]

hey symlinks can be great lol

especially in the world of AI checkpoints

[trim orbit]

because thats where appdata lives. cant you move your user folder to a drive you prefer if you want control over c like that?

symlinks only hurt if you're trying to sort out someone else's

[shy kelp]

yes, there is a folder called appdata on my C, which a lot of programs use for small files. If that was where the entire FREAKING PROGRAM was meant to go, then program files wouldnt exist 😛

[boreal bough]

no longer an issues - but in the past many apps (and multiple version of windows even) would never end some routines due to a never ending browser agent

also other reasons

[trim orbit]

appdata, yea. you seem pretty upset about it but it's being used as designed

[boreal bough]

I took the rich alternative, and got a 4TB SSD just for AI

[nimble heart]

anything that walks directories without a maximum depth safety is straight up insane should be locked away

[shy kelp]

that is in fact, not the point of appdata. its for stuff like settings, configs, small tools a program used. not the program

if 95% of your program is in appdata, you failed the point of appdata lol

[trim orbit]

minecraft?

[spring fulcrum]

Where do I get the SD vae.safetensors?

[shy kelp]

when did I ever talk about minecraft

[trim orbit]

one of the most succesful games in history and is 100% installed in appdata

[civic sigil]

I did this too but python and huggingface cache, appdata, etc bloats up my boot drive anyway

[visual glade]

that's why I like portable standalones

[shy kelp]

and obilvion on the xbox restarted the whole fucking system sometimes when loading... doesnt mean its right lol

[wicked frigate]

that took weirdly long but yes yes it does work in colab

[trim orbit]

yup. you find work arounds instead of blaming the developers. layer 8 problems

[nimble heart]

official SDXL vae is on hugginface. Alternatively you can just hook the vae decoder straight into the Base Checkpoint vae

I just like having my VAEs separate

[boreal bough]

will it let us use existing comfy installs? please

[wicked frigate]

yes yes it does

[boreal bough]

❤️

[trim orbit]

classic over heating issue. people would keep their xboxs in an unventalated cubby hole or in the sunshine

[wicked frigate]

i've been working on it half the time with a hacked internal version of comfy so it kinda had to work with existing installs from the start lol

[boreal bough]

now I wonder... am i redownloading 0.9 weights? 🤣

[trim orbit]

not the overheating you're used to either. this was the psu tripping in the xbox

[shy kelp]

no it wasnt, it was intentional because of memory use

[visual glade]

just make sure it also works with the public version

[boreal bough]

cause this is taking a lot longer than it should

[wicked frigate]

it does, i got it fully running out of the box on an empty windows 11 VM by autodownloading the official public comfy ref

[nimble heart]

appdata on windows is where you're supposed to install software that isn't system-wide. Not sure what the issue is. It's like linux' ~/.local/share/

[visual glade]

nice, should be good then

[civic sigil]

Yeah probably to some mystery hidden folder on your boot drive that only power users know where to find lol

So annoying

[trim orbit]

there's only a couple games on the xbox that caused memory leaks. and only a couple versions of those games. oblivion wasn't one of them

[dapper current]

I mean, Xbox 360 would overheat even if you kept it completely clean and in open air. They all needed a reflow too because of the red ring of death

[wicked frigate]

did i mention stablestudio is best for beginners and my ui is better for the experienced powerusers lol

[shy kelp]

https://www.gamesradar.com/the-elder-scrolls-3-morrowind-restarts-your-xbox-and-microsoft-showed-bethesda-how-to-do-it/ phil spencer lied huh?

[boreal bough]

if you ever planned on having a leak, now would be a good time XD

[trim orbit]

hacking and repairing og xboxs was one of my side hustles and i cleaned about $30-50k from fixing them up over the years

[dapper current]

Did you do PS optical drives too?

[shy kelp]

and yes I realize this is morrowind and not oblivion, I dont play bethesda games so I got the names mixed up, but point still stands

[trim orbit]

bro that wasn't oblvion and it was a literal software call to a reset function.

[dapper current]

They were another one haha

Especially PS2

[trim orbit]

less so.

[shy kelp]

" if you're running low on memory you can reboot the original Xbox and the user can't tell."

read the article, use your eyes 😛

[trim orbit]

only because i knew a lot less people with ps2s in my region

[visual glade]

don't tell SD UI devs about this trick....

[dapper current]

"SD UI devs HATE this one trick..."

[wicked frigate]

ssssh, stage event literally like 13 hours away

[nimble heart]

too bad they already know it. webui 1.4 literally added a "restart" button to the settings page.

[dapper current]

Yeah then maybe you'll get more info

[boreal bough]

but I'll be DMing a DND session 😭 I wont be there until the event is over

[nimble heart]

specifically intended to clear the memory leak

[wicked frigate]

lmao

[trim orbit]

did you read the article? it wasn't ever while loading. it was while playing the game. geeze.

[wicked frigate]

auto's working on an actual fix for mem issues

[shy kelp]

"sometimes you get a very long load, that's us rebooting the Xbox"

[trim orbit]

talk about misrepresenting the problem

[shy kelp]

hmmm 😛

[deep dagger]

@visual glade possible to have a "x" "y" tiled latent node so we could create seamless images with XL in comfyUI? Like the Tiled box in A1111

[visual glade]

tiled never worked well for me which is why I didn't implement it

[trim orbit]

i just still don't understand the original point. that your c drive is small?

[visual glade]

what is a1111 going to do to try to fix the mem issues?

[wicked frigate]

tiled is super useful for certain niches cases, eg generating textures to put on 3d models

[shy kelp]

I said it shouldnt be dumping the program into the appdata as that is not the point of appdata, you said but minecraft does it so thats ok, and I pointed out other programs in the past do fucky things. Doesnt mean its the right way to do it xD

[visual glade]

I have some ideas to lower comfyui memory usage by a bit more but I would have to have it implemented a week ago to have it properly tested and ready for 1.0

[boreal bough]

@knotty lotus
it ignores the location I gave it for the install :/
this is a real issue, as that is a lot of space for the C drive. Especially for non-power users

[shy kelp]

Ayyy, just like I was saying

[wicked frigate]

oh i misremembered what RAM issue there was a fix about: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11958

the memleak it had on linux was solved by integrating the tcmalloc fix thingy

[floral island]

treesize is ur friend, finding big files 'somewhere'

[wicked frigate]

other memleaks idk about

[floral island]

goodmorning my fellow excited enthusiasts 😄

[shy kelp]

Also for anyone else reading, its not that my C is small, its a 1TB drive
Its that my H is 16TB and I would like big things that dont need to be fast, there 😛

[nimble heart]

sdxl leeks now in imax 3d

[civic sigil]

Ok Ill try that

[dapper current]

For Windows I like using Windirstat

[boreal bough]

ufff. it even downloads the whole SD2.1 checkpoint :/
(hope that changes tomorrow!)

[floral island]

remember after install, run as admin -> else you can't find files that are behind admin folders

[floral island]

i used too, treesize is newer and multithreaded, waaay faster

[shy kelp]

I would reccomend looking at wiztree, if you have not. Instead of going folder by folder bruteforce, wiztree uses the file table to do it really fast

[deep dagger]

You should try running XL and AnimateDiff in the same node tree while watching youtube videos 😅

[nimble heart]

why test when you can push straight to main

[dapper current]

oh no way!

I will check it out 😄

[shy kelp]

yeah wiztree can do my 1tb C drive with 4 million items in about 5-10 seconds lol

[floral island]

i'll be checking out this one too, always good having more options

[shy kelp]

has the same fancy box view at the bottom and everything

[floral island]

winget windows mvp 😄

[hard fractal]

I'm glad these issues are coming to a head.

[sour obsidian]

Yea its a pretty fun one. Could put it on the list after the next couple things

[spring fulcrum]

@visual glade @high skiff Do either of you know what this is or if it might be related to why my gens are slow?

[hard fractal]

Just need some training data.

and maybe a QR-loss.

haha

[sour obsidian]

Hahah

[nimble heart]

think tcmalloc was only a partial fix as well. More just delayed the bomb

[hard fractal]

I mean, I'm sure there are python QR decoders.

[visual glade]

yeah knowing pytorch I give that a 50% chance of not fixing anything

[hard fractal]

comfy bringing the snark, I love it.

[wicked frigate]

i hear tomorrow there's gonna be a UI that gives you a choice of what, if anything, you want to autodownload while it's installing, yknow something like

(made up image, not real, mockup, fake news 2023)

[floral island]

tested it, i think i like wiztree more than treesize 😄

[visual glade]

no that's just a harmless torch warning

[hard fractal]

Applied Team: "Comfy is better."
A1111 Stans: "Nuh uh."

SDXL 0.9 is, uh, released.

A1111 Stans: "But... halp?"

[shy kelp]

hops back into bot 10
sees someone asking for 'a little yellow boy'
Damn we gettin racist up in here huh

[hard fractal]

let me see

[shy kelp]

wiztree and Everything are 2 must haves if you do a lot of files lol

[civic sigil]

Yeah this tool is basically telling me my whole C drive is completely filled by appdata, .conda, miniconda, and hugginface cache folders. Getting into ML was a mistake

[boreal bough]

the 10 hours after webui 1.5 released were... interesting

[shy kelp]

#1101178553900478464 message

[boreal bough]

sdxl turned into tech support for a while XD

[nimble heart]

tf you have installed

[shy kelp]

now I dont know if that user meant 'a little boy wearing yellow' and their english is bad but xD

[floral island]

i'm a madman, i don't need everything: i know where my stuff is lol xD

[boreal bough]

and this is why I've reduced my life to venv

[floral island]

i'm still a madman xD

[hard fractal]

hm. plausible that they meant blond

will keep an eye on future prompts.

[shy kelp]

yeah thats what im hoping

I could see if english wasnt your first language making that oopsie so

thats why I didnt instantly go full report mode lol

[hard fractal]

Doubt benefitting.

[boreal bough]

the simpsons are sad with your assumptions!

[hard fractal]

tomorrow?

[nimble heart]

oof. each version of torch with rocm is 1.6 gigs and ig it adds up

[hard fractal]

what's tomorrow?

[boreal bough]

July 18th is tomorrow

[hard fractal]

The applied stage, right?

[wicked frigate]

ye

[sour obsidian]

Who are they?

[civic sigil]

Yeah its annoying

I was strapped for cash when I bought my hard drive too

[wicked frigate]

^ did you see that stability dev that said they were gonna talk about a new ui during the stage?

[civic sigil]

Someone told me to get a small SSD for the boot drive and a big HDD, big mistake lol

[wicked frigate]

whoever that 'mcmonkey' person is, they seem really cool and handsome

[nimble heart]

that was true maybe like 8 years ago

ssds are so cheap now.

[dapper current]

Sounds like they might be hanging around that @hard fractal guy too much

[boreal bough]

sadly he's not allowed to talk about it for another <t:1690394400:R>

[floral island]

thta's what i've been doing -> it just got worse progressively

[visual glade]

I would still recommend that but 2TB SSD and 20TB HDD

[spring fulcrum]

https://tenor.com/view/coronel-gif-18840785

I do believe it is the day after today my good sir.

[floral island]

sdxl 1.0 is overpowered 😮 no man should have this power!

[dapper current]

hey nice that's a bot gen?

[floral island]

yeah

[boreal bough]

I've got 20TB SSD total, and 10TB HDD. I'm ready for SDXXL

[sour obsidian]

Ok youre right, we will delete it

[wicked frigate]

i got a fast ssd for my boot drive and a slow ssd for my data drive

[floral island]

https://tenor.com/view/404soob-pepe-pp-gif-21552696

[dapper current]

oh and rest assured everyone, the bot is pure 1.0 at the moment and will stay that way as I am already in bed

[wicked frigate]

4TiB C drive and 8 TiB E drive

(QVO drives are awesome, affordable super high cap ssd)

[nimble heart]

double check what's actually using the space. 99% of my hugginface folder is just DeepFloyd models which I haven't used in like 4 months.

[wicked frigate]

brb messing with twoperc's bot files while he's asleep

[spring fulcrum]

All the bots channels 1-10?????

[dapper current]

plz no

[boreal bough]

but what about your Dataset drive?
does it go clonk clonk, or are you using a slow SSD for long term storage of images you'll use once every blood moon as well? 🤣

[dapper current]

though honestly I bet if anyone actually opened up what I'm using right now code wise they'd throw bleach on their eyes, set their house on fire and run into traffic

[boreal bough]

80% of programmers be like

[spring fulcrum]

I would love to burn my eyes.... send it to me

[dapper current]

In all seriousness, maybe one day 😉

[spring fulcrum]

lol

[nimble heart]

just run the IDE autoformatter on it it'll be fine

[spring fulcrum]

lmao

[dapper current]

Yeah it'll nicely format all the gazillion things I have commented out from all the crazy and wacky things we've done over the last 4 months

[floral island]

go on vacation for a week and then re-open your code. you'd do the same, probably!

[spring fulcrum]

just ask chatgpt to rewrite it so it allllllll works with just one line of code 😉

[nimble heart]

also make sure you set your formatter line break to 80 characters for tty compatibility as well

[dapper current]

I just need to go to sleep and wake back up for it to happen

I jest lol

[wicked frigate]

I store everything on SSD lol

HDDs are so last decade

[hard fractal]

Do you all remember when you were REALLY MEAN TO US WHEN WE NEEDED A FEW EXTRA DAYS TO FINISH UP SDXL

[nimble heart]

nope

[dapper current]

... or today....

[nimble heart]

No idea what you're talking about

[dapper current]

when we just needed to compare some rougher stuff...

and I almost cried

[sour obsidian]

He did, I was there

[hard fractal]

y'all almost made 2% cry?!

Okay, that's it.

Someone add 2% hours to the clock

[spring fulcrum]

Did that really happen.... I only joined this discord server like a week ago.

[dapper current]

hahaha

Me crying, no. People getting upset, yes

[hard fractal]

Some people were menaces.

[spring fulcrum]

ummm when is he not a troll... hehe

[hard fractal]

True.

[boreal bough]

perfect. that's just enough time for me to finish DND XD

[dapper current]

Lets just say when you have the settings all nice and people are getting great images for awhile and you suddenly start randomizing things like CFG, Samplers and Steps in wide variety... People get a bit shell shocked

[sour obsidian]

https://tenor.com/view/andre-giant-no-wait-holdon-gif-9412107

[floral island]

we're counting from absolute beginning of time, right? that should give you enough time to finish up nicely

[sour obsidian]

We're counting back to the beginning of unix time starting now

[lusty raptor]

always weirded me out to see regulars not quite getting it when that happened

[civic sigil]

Which will come first? SAI's controlnet weights or the heat death of the universe?

[sour obsidian]

Tough to say

[boreal bough]

the randomized settings ptsd was real. also the influx of 'sdxl bot is trash' users was... painful

[hard fractal]

CompTroll-N3T

[dapper current]

You have no idea how much I appreciate you and the others who whole heartedly have been in this and understand the flux! I'm seriously going to talk to the community leaders here about getting a special role for those who contributed like you

[spring fulcrum]

Okay everyone I would like to introduce you all to my wives....

[deep dagger]

Community Archive Net?

[dapper current]

Part of the reason I want to restore full stats

[lusty raptor]

aww shucks. i really don't want anything extra, it's all for the science

[floral island]

i mean, it was stated that the sdxl bot was rlhf, so bad images were bound to happen...

[autumn forum]

ill take the role.. xD

oh shoot im supposed to be in bed.

[dapper current]

So am I. Something about a stage and needing to look alive tomorrow...

[daring smelt]

early sdxl badges wen

[hard fractal]

"omg I knew it, they're censoring SDXL at the last minute, gimping it before ||the stage event|| on July 26"

[hearty ginkgo]

Good morning

[hard fractal]

Where's Sytan? I need to change his roles

[boreal bough]

391 days be like

[daring smelt]

we need a pre-ban role

[lusty raptor]

saw someone on reddit (of all places) asking for a nsfw fix for sdxl, because clearly it's only broken/censored

[wicked frigate]

we finally got rid of all the non-cat generations :)
SDXL 1.0 will finally be the catpic generator we all wanted

[dapper current]

I mean, SDXL 1.0 is absolutely amazing at cat pictures

[hard fractal]

SDXL only makes one image now, but you can Drag it like a GAN

[dapper current]

Even when it's all jangled it always manages to pull off good cats...

[wicked frigate]

fun fact: I actually have that in another discord

[hearty ginkgo]

Isn't the true black and true white much better in sdxl 1?

[lusty raptor]

it's better at posing cats in action poses than it is with people!

[dapper current]

Cat eyes too

Like, perfect every time almost

[hearty ginkgo]

Send a photo

[daring smelt]

this is too good

[sour obsidian]

It only makes pictures of food if youre hungry

[boreal bough]

'frienemies' role confirmed?

[autumn forum]

the refiner sometimes messes up cat eyes..

[daring smelt]

SDXPelled

[dapper current]

almost

[hard fractal]

Parkat

[autumn forum]

one of my fav's

[daring smelt]

i'm not censoring you, i'm customizing you

[dapper current]

What's nice is honestly 1.0 is much better at detail, so even though the refiner still brings it to another level and helps in pictures with lots of small faces, most of the time stock 1.0 is actually great or sometimes even better

[shy kelp]

noice.

[sour obsidian]

Whats sdxl?

[dapper current]

Something about July 18th?

[daring smelt]

i'm sure she's a nice girl

[wicked frigate]

SDXL is a sad-laughing emoji with something on its chin and an eyebrow perked up

[nimble heart]

cats will sit on anything

[chrome flicker]

hi, today xl release or delay?

[thin nova]

I'm so excited for July 18th

[boreal bough]

<t:1690394400:R>

[hearty ginkgo]

Delayed by 2 months

[chrome flicker]

nice

[spring fulcrum]

@hard fractal @dapper current I am genuinly asking... What is it about hands that makes it super hard to generate well? Is it that there aren't enough images of hands to model from? Is there a way as a community we could actually help to get hands to improve more?

[dapper current]

https://tenor.com/6SBX.gif

[boreal bough]

I wish we had a July 18th every week

[autumn forum]

me when i get up every morning lmao

[nimble heart]

hands don't have a standard "shape" they're in so it treats them like scales or fur and just pattern-fills fingers into the image

[deep dagger]

All the different angles and positions hands can be in from any angle in my best guess.

[dapper current]

The problem with hands is due to the number of positions and angles they can be in combined with fingers. It's hard for AI to come up with a rigid "pattern" that is correct every time.

Also, go look at LAION one day.

[ionic dragon]

sadly i would have to celebrate my birthday every week

[spring fulcrum]

so having more pictures of peoples hands wouldn't really help much to improve the model?

[autumn forum]

probably not it might just give you more misshaped hands lmao

[dapper current]

Fine tunes do better with hands because they can ensure the pics are clear, high quality, and have similar poses for them

[floral island]

yes, sdxl does cats 😄 nice calico 😄

[dapper current]

More does not necessarily mean better

[ionic dragon]

@dapper current do you know any large fintuner?
i am willing to share my huge dataset with them, around 2million images

[dapper current]

SDXL does much better with hands mainly due to it's sheer size

[spring fulcrum]

Well I do appreciate the info... I was always curious about how I could somehow help with that.

[ionic dragon]

but i wouldn't agree upon that on hands and to generate realistic image

[shy kelp]

now that 1.0 is finalized and pushed to the bots 1-10, can we know which of the 3 models won? 😄

[spark bear]

also many community finetunes are overfitted so i guess that helps with good hands since its just spitting out stuff close to the dataset

[dapper current]

What's SDXL 1.0?

That, is the question 🙂

[wicked frigate]

what's a bot

[nimble heart]

think my GPU just hanged.
all my kitten pics are gone now...

[shy kelp]

who said anything about sdxl? I just said 1.0 😉

[deep dagger]

Is that the new name of Twitter?

[dapper current]

oops darn I've been figured out

[boreal bough]

did you ever do art?
When you draw hands '100% realistically', you'll first notice the weird thing, of when you look at your own hand in a normal non-posed position, it looks anatomically incorrect
that's issue No.1
next, fingers look similar, just having different lengths - but the lengths depend on view, bending and many other things
that's issue No.2
lastly, while you have 5 fingers, it's not like you can always see 5 fingers. often less are visible. so as an ai, it would under normal circumstances learn "5 or less fingers" (bad) or "there's a high chance of a finger being next to a finger" (still bad)
that's issue No.3

So how do we solve this using TI, LoRA or Finetuning?
The Answer: we don't
What we do is reduce the sample size to only 'posed' pictures where all 5 fingers are visible, or popular poses such as 'fist' or 'thumbs up' or 'pointing' - which essentially introduces extreme bias - but it's bias we want, at the cost of of flexibility

[wicked frigate]

probably, fits musks naming rules (always put an X in it)

[shy kelp]

#1100170312106127410 message someone is making ai art of 9/11 in bot 1

[dapper current]

Paypal used to be called X iirc

[sour obsidian]

We are officially renaming sdxl as SDTwitterL

[wicked frigate]

Twitter did take an L, that's true

[sour obsidian]

Hahah

[shy kelp]

is clip guidance not possible to do locally?

[sour obsidian]

Totally possible, I think diffusers has a decent implementation

[spark bear]

doesn't musk own x.com? kind of a waste, should be a site dedicated to UFO defense

[shy kelp]

seemed to me having that clip guidance in the api via auto made some truly killer images

[ionic dragon]

just imagine people searching twitter video downloader, now they have to search x video downloader 💀 💀

[vast flare]

Guys where I can find the notebook file setup in civitai / github pages?

[dapper current]

That... is an interesting thought

[vast flare]

To run them in Google colab

[boreal bough]

until musk makes a bad pun about buying SD. Then is forced to buy it for inhuman amounts. Then you all move to Fiji, and SD gets renamed to SD𝕏

[deep dagger]

I would be careful, he might try to charge you for using that character.

[boreal bough]

I'm safe as long as I stay in middle Europe XD

[ionic dragon]

anyways, twitter doesn't restrict porn so its legit to change its name to x

[spring fulcrum]

I'm sure someday it will get there... I realize its not an easy issue to solve... But at what point does the Intelligence part of AI get to where it actually will think about all the points you just made and put it all together on its own... I also realize it learns the way we program it to learn. Which is why it will take time. I'm sure it will get there at some point in time. I am excited for that day. But yes I agree with everything you said.

[wicked frigate]

to run what now

[knotty lotus]

comfyUI install location is located in settings.json in the roaming directory it makes for itself

[knotty lotus]

😉

[boreal bough]

there are definitely ways to train hands correctly. essentially do what SAI did for faces this time round.
its not an issue of the machine learning anymore, nor model limitations. Just a matter of painfully making that dataset where you finetune the full anatomical knowledge of a hand

[dapper current]

Why would it change for a stage tho

[knotty lotus]

plan is to allow changing install location for comfy & allow for downloading whatever weights you want

[wicked frigate]

well that's when we announce the change twoperc

[boreal bough]

much effort for little return, when overfitting does perceptively similar results for 80% of the usecases

[vast flare]

Models

[wicked frigate]

oh i just realized you posted in every channel lol

[spring fulcrum]

If someone would be willing to help me learn how to do that I would be happy to try and train something that can finetune it.

[lusty raptor]

speaking of hands, i remember sdxl really struggling with power tools in the past, so i tried some just now

[agile quarry]

https://github.com/comfyanonymous/ComfyUI#colab-notebook https://github.com/TheLastBen/fast-stable-diffusion

[autumn forum]

wow a few of those are actually close to real tools

[dapper current]

Tools, airplanes, pianos (still not perfect)... but much better in SDXL

[boreal bough]

will still request an 'auto' or 'manual' button during install, avoid polluting the poor C drive

[knotty lotus]

yes

[boreal bough]

its not even the finetuning.
just the act of gathering roughly 200 images per anatomical part of a hand. hand captioning all those images
it doesn't sound bad, but this is the act of turning 1 concept "hand" into roughly 300 connected concepts. And if you still think its doable, remember that you also need to obtain a fairy equal sample size separated by age, ethnicity, gender & accessories such as rings. Dataset also need to be high quality enough, so you can have a closeup of the hand, and one uncropped to show the context of the image

its a monumental task that would take years.
It's a good business idea if nothing else - to then sell well averaged datasets for machine learning, to reduce possible bias

[timid sonnet]

3

83FA

FAST

[spring fulcrum]

ummm... listen... I am tooo fat to climb mount everest... ok. I dont mind going for a walk to the kitchen but dayum... I see why it hasn't been done

https://tenor.com/view/now-i-understand-neil-degrasse-tyson-startalk-now-i-get-it-i-get-it-now-gif-21992239

That really does sound like a lot to do.

[boreal bough]

there's a fun conversation to be had here, about how challenging it would be to get a good nsfw finetune XD but wrong discord server to have that conversation

[spring fulcrum]

I bet if my wife wouldn't kill me that I could probably get that done in like 10 minutes compared to hands

[lusty raptor]

I'm happy that that nasty dotty texture is gone from images in the bot now.

[boreal bough]

wasn't that the watermark getting out of hand?

didn't check bots enough

[lusty raptor]

i could dig through some saved images and probably find an example, it was everywhere, but i don't think it was a watermark.

[spring fulcrum]

I am having trouble getting more in my images than like shoulders and up.... Maybe its the aspect ratio.. What is the closest resolution to 1920x1080 that is under the 1024 pixel mark?

[boreal bough]

I mean the 'invisible' SAI watermark, so that sdxl gen images that find their way online won't be used for future model training

[civic sigil]

720p

Which is 1280x720

[nimble heart]

well I wanted a pikachu being photographed in the wild but I guess that works too

[upbeat summit]

the nikon must be so heavy

[spring fulcrum]

If i use the 1280 is my card going to shit itself when I upscale it to 4k?

[boreal bough]

its a bit of a biased aspect ratio. You'll have to get your prompts and negative prompts right to make that aspect ratio work consistently with people

[civic sigil]

How do you plan on upscaling

[lusty raptor]

here's the "worst" one i could find. You can see it in the asphalt kinda. Many times in places of grainy-ish detail, you'd see these too-large bumps

[nimble heart]

probably why he looks so grumpy

[spring fulcrum]

a lieka dis

[civic sigil]

Hmm Idk

[boreal bough]

ah I see it. also isn't that from prompts?

it's the thing you get when you scan high end artwork

[civic sigil]

I think I've seen those dots before yeah, is this with refiner or no refiner out of curiosity?

[nimble heart]

their cheeks always look like seedless strawberries

[trim orbit]

if you get down to a defining a set of rules that are specific to making hands accurately, you have an atlas of a rule book

[boreal bough]

so prob an artwork tag got mixed in

[lusty raptor]

from what i could tell, it was one of the contenter checkpoints, but maybe from settings? hard to know from this end

[shy kelp]

does anyone know that comfyui plugin/node that allows inline lora calling? like lora:name:1 in the prompt itself and not having to load it seperately

[boreal bough]

also, adding ", descreened" at the end of your positive prompt removes it

[lusty raptor]

bot, pretty sure it wasn't the refiner, i saw a lot of it during the dark times of testing

[boreal bough]

though it shouldn't show up in first place

[spring fulcrum]

I feel like life is great right now... I'm going to wake up like a little kid and it will be Christmas morning with SDXL 1.0 nestled safely under the huggingface Christmas tree... What a time to be alive. 😉

[west breach]

14 hours to go?

[heady vale]

12

[boreal bough]

<t:1690394400:R>

[west breach]

That's the event yeah? Not necessarily the release time

[rustic surge]

国人？

[spring fulcrum]

heyy, i just want to think happy thoughts before i go to sleep tonight

[autumn forum]

I’m pretty confident they will release it. The devs in here earlier were talking it up pretty hard. ‘‘Twas a good time lol

[spring fulcrum]

I was here... and I'm hoping they do

[rustic surge]

It seems that I don’t need to sleep tonight,

[boreal bough]

Deutsch
but it auto translates into everyones native language 😄

[rustic surge]

oh

[autumn forum]

Oh yeah hehe sorry I’m so tired

[rustic surge]

Is there a live broadcast when SDXL is released?

[spring fulcrum]

thats what I got... not sure if that is the translation

[boreal bough]

wait a minute. why you not sleeping? xD go sleep @autumn forum you can play after work!

[trim orbit]

https://discord.com/events/1002292111942635562/1130919343941758978

[autumn forum]

Can’t sleep to excited lmao but yeah okay 💤

[rustic surge]

Go to bed early and get up early to be healthy, don't stay up late, my friend

[rustic surge]

thank you

[trim orbit]

be ready for any announcement

[hearty ginkgo]

Ye

[shy kelp]

for comfyui should I be setting my Clip Set Last Layer node before or after loading all my Loras?

[visual glade]

doesn't matter

[shy kelp]

ty 😛

[nimble heart]

I'll probably be asleep during the broadcast it looks

bloody europeans

yes, tovarich

[shy kelp]

nice kitty cat LMAO

i couldn't say the p word its disallowed

[rustic surge]

sdxl can nsfw？

[shy kelp]

sure can

[rustic surge]

wow

[shy kelp]

afaik the base model isn't trained with those kinds of images but it doesn't have strict filters like sd2.1 did

[boreal bough]

sdxl has not been trained on explicit nsfw data. That's as far as conversation is this server should go.

[rustic surge]

Shouldn't there be a situation where it cannot be popularized like sd2.1?

[kindred plinth]

yeah they should really train it on NSFW, as far as ive seen SDXL is just worse than SD1.5 in every way because they didnt train it on NSFW /s

[rustic surge]

SDXL is worse than sd1.5?

[kindred plinth]

yeah people dont know it yet but SD1.5 is and will always be the only good txt2img model

[boreal bough]

/s translates badly for non native speakers

[rustic surge]

I don't agree, I feel that SDXL has made great progress compared to sd1.5

[vale eagle]

he is joking.

[civic sigil]

Why can't we have the conversation, what's the big deal?

[rustic surge]

小智？

[boreal bough]

• cause for many people this is a server that is accessed at work
• there is no age limit to join here
• while completely valid to talk about, it's just in bad taste to do so on the official sai server, in front of sai employees who are negatively impacted by media coverage of nsfw in relation to SD

[civic sigil]

I mean I wouldn't want to have a graphic discussion, and I don't think it's in bad taste at all considering this is a community server about an open source model, no one should feel afraid to be critical of SAI in here of all places

[soft zealot]

why ?

With "portable" install you can save the file whereever you want, extract the contents, move the folder onto whatever drive you want and then run it.

Similarly when doing a git install, just start in a location that isn't (for example) your C drive

[boreal bough]

oh definitely, when it comes to being critical. I'm only referring to specifically "make nude photos", "train sdxl to make nude photos" & "can current sdxl be circumvented to make nude photos" nsfw conversations.
anatomy topics should be fine - but that's not what people want to talk about. especially since sdxl gets 10/10 when it comes to anatomy understanding

[vital wolf]

When will sdxl release?

[civic sigil]

10/10? Reproductive organs are anatomy too

[nimble heart]

when Emad finishes playing Tears of the Kingdom

[boreal bough]

?
yes. I agree.
but that's not what happens
it disregards your current settings, and first installs it to C appdrive folder regardless of what you do

[vale eagle]

SDXL is bad at drawing chair compared to some fine tuned 1.5 model

[nimble heart]

like a finetune specifically made for chairs or a general purpose finetune?

[boreal bough]

lemme rephrase that. contextual understanding of body proportions

which was a huge issue in 1.5 & 2.1

[vale eagle]

general

[boreal bough]

<t:1690394400:R>

[civic sigil]

I would still disagree with 10/10 but yeah it is a lot better

[vital wolf]

Thanks 👍

[civic sigil]

Looking through my output folder I really feel like the refiner model is not doing that much good for my gens

[soft zealot]

ok so what the flip is in this that the bot has moaned about ??

[boreal bough]

XD

I didn't even know there's a protection

[soft zealot]

I'm rereading it multiple times and I cant see any words used that havent been used elsewhere

[boreal bough]

I think it was the 'dragged over coals' part 🤣

[civic sigil]

Yeah I agree with you... I feel like people are conflating nudity with hardcore p*rn

[nimble heart]

4 random seeds dreamshaper vs sdxl for product photo of a wooden chair, studio lighting. SDXL followed the prompt a lot better but in turn it has much more non-euclidean geometry so pick your poison

[boreal bough]

totally. I mean I lived in Greece for 10 years - so that should tell you enough about what I consider to be nsfw and what not XD
but yeah - stuff like bikinis sdxl does real well. so that shouldn't be an issue, right?

[civic sigil]

Yeah breasts are mostly what people want to gen so it will not be a huge problem for the model gaining adoptance

[soft zealot]

my issue is that SDXL is a bit like many of the 2.1 models in that it has problems with nipples

[boreal bough]

this man does science 🤣

[spring fulcrum]

lmao

I figured it out

[soft zealot]

nope just the one word

[spring fulcrum]

not lingerie

[soft zealot]

how about lingerie?

[nimble heart]

more steps seems to help. could also be rng

[civic sigil]

Some nice ones here actually

[nimble heart]

so you can't say 3rotic but you can say PORN okay

[civic sigil]

But that sort of geometry SD will always struggle with

[boreal bough]

to keep it short. loras are real easy to train.
like I said many times before, finetune levels of improvement are now doable with a simple 43mb lora.
I don't say that in theory, I say that because I've made a fair share of loras for 0.9 to test the limits. so just give it 2 weeks or less - and then see how happy you are with sdxl

[soft zealot]

so nipples is ok , sexy is ok, but 3r0t1c isn't?

[civic sigil]

Where it looks wildly different depending on the viewing angle

[soft zealot]

cos that makes complete sense lol

[vale eagle]

Might be I was biased on bad SDXL photo vs good 1.5 photo

[kindred plinth]

bruh

[soft zealot]

the word snowflake itself is more likely to trigger IMHO lol

[spring fulcrum]

lol

[civic sigil]

My fear is they are all from roughly the same viewing angle. Diffusion models have no real concept of 3D, if you asked for the camera to be between the legs it would have no clue what it's doing, even asking for a top-down view would probably be too much for it. It's the same problem SD has with hands

[boreal bough]

this was solved though? at least with my lora of 2B Cosplay, I made gens from all angles, and could control them consistently

[kindred plinth]

unrelated but i really really cant wait for a good text to 3d large scale world

[boreal bough]

obviously cant say the same for chairs XD

[boreal bough]

same

[civic sigil]

For human figures maybe, enough training data can overcome the problem

[nimble heart]

train a lora for upskirt photos of chairs

[boreal bough]

I feel called out 🤣

nsfw chair images

[kindred plinth]

damm i was really looking forward to making chair porn with SDXL

[vale eagle]

nsfw chair

[shy kelp]

going to sleep now, see you guys tomorrow for the big event 🥳

[boreal bough]

I mean its only an hour of work. tempted to make a "upskirt chair" lora for 1.0 release 🤣

[nimble heart]

comfyui just added this as a feature. Using the conditioning timestep you could blend chair and pornography prompts to give the chair a massive horsecock or something

[boreal bough]

make a point that any dumb concept can be trained

[kindred plinth]

lol i didnt expect a serious response

[nimble heart]

I'm the most serious person you'll ever meet

[civic sigil]

Ok this inspired me to see if SDXL is any good at guns

Yeah it seems pretty good 👍

Of course the fingers holding the gun are bad

[nimble heart]

irl people can't decide how guns should look so the ai 100% can't. gonna get 2 magazines 3 charging handles and 17 selector knobs

[kindred plinth]

2 barrels

[nimble heart]

maybe if you specify something specific with a really strong presence like the 1911 it'll work better

[nimble heart]

I think it just assumes the gas block is a 2nd barrel

[kindred plinth]

add a rear facing barrel to kill people behind you

[boreal bough]

concept art + rifle makes for fun guns

[nimble heart]

concept art of a super soaker M60 machine gun

[boreal bough]

WW2 scene of a battlefield in germany, with everyone holding super soakers

[nimble heart]

artist's rendition of a Nerf guillotine during the french revolution

[boreal bough]

what in gods name 🤣

[nimble heart]

im asking the bot rn

[civic sigil]

Nothing too crazy

[nimble heart]

I don't think it can even make a regular guillotine nevermind a Nerf one

https://media.discordapp.net/attachments/1100170312106127410/1133663989189062767/artists_rendition_of_a_Nerf_guillotine_during_the_french_revolution_seed-0ts-1690356953_idx-0.png?width=1387&height=1387

[nimble heart]

ew handshake grip

[kind pendant]

with full auto trigger

[vital wolf]

I wonder if there will be a model like niji for sdxl in the future

[kindred plinth]

there will be like 50 of them

[nimble heart]

tf is 'niji'

[vital wolf]

I hope so, cause base and refiner is still not in par with niji.

[uncut steeple]

anime model I think

[nimble heart]

Ah

is it another anythingv3 mix

or is it actually trained

[kindred plinth]

no

[vital wolf]

It's a collaborative model between spellbrush and midjourney which mainly focuses on anime

[kindred plinth]

nijijourney

[nimble heart]

interesting

i imagine anime and porn will be the first things sdxl is tuned for

[vital wolf]

Trained, and nothing comes closer to it when it comes to anime as of now

[civic sigil]

Hmm from testing it seems like Comfyui's conditioning-average and similar nodes only work with the old CLIP

[ionic dragon]

does LAION have NSFW images?

[west breach]

I used conditioning average with SDXL 0.9

[civic sigil]

It only works with CLIP-L though

[civic sigil]

Yes but they don't train on those

In fact it seems to totally disregard your CLIP-G prompt if you use it

[peak dove]

1.0 day is finally here!!!

[quartz sequoia]

i have awakened... where is it?

WHERE

[peak dove]

Message 19.00 UK time?

[kindred plinth]

delayed to 2024

[peak dove]

20.24 ... 24 minutes after 8 p.m.? 🙂

[quartz sequoia]

that already passed

gang gang , gang gang

[chrome flicker]

any new release?

[shy kelp]

is today the day we get jebaited?

[cunning geyser]

im hyped

also yeah cant wait to see if its moved.

if it is, good, more time for me

[indigo carbon]

B

[quartz sequoia]

hmmm

do i want SDXL controlnet before SDXL 1.0?

[indigo carbon]

and I want TensorRT SDXL, that ain't happening either

[quartz sequoia]

who said its not

[indigo carbon]

I have been trying altering with the code to make it work. TensorRT only works on 1.5

[quartz sequoia]

wait do i need an extension to run a TensorRT model on Auto?

[indigo carbon]

yes

it's worth it though

[quartz sequoia]

id assume it works on a 3090

[indigo carbon]

obviously

[quartz sequoia]

what im i looking at in terms of speed improvements?

[indigo carbon]

Its an optimization made by NVIDIA. almost double the speed if not even more without precision decrease

[quartz sequoia]

so its faster than the speed of light

[indigo carbon]

pretty much, it's black magic

[quartz sequoia]

well i shall install it now

and watch my new pc die

on the day of SDXL release

[indigo carbon]

I bet after this community figures out how to use this with SDXL it might become a default optimization

[indigo carbon]

it's not relevant for you. 1.5 is nothing compared to SDXL. I advise you to wait until it supports SDXL

[quartz sequoia]

i can use the main AutoBranch right... not the old Dev branch?

[indigo carbon]

.

[quartz sequoia]

maybe i want super speed 1.5

[indigo carbon]

fair enough.

[indigo carbon]

wait but you have a 3090.. it's not already super speed with 1.5?

[quartz sequoia]

yeah... but why would i say no to more extra super speed

[indigo carbon]

it doesn't work with high resolutions.. the A1111 implementation isn't coded correctly.

[quartz sequoia]

ugh... fine

[indigo carbon]

however, I bet TensorRT for SDXL will show up eventually =]

[thick mist]

Will building Loras work the same with SDXL and do you need that "refiner" thingy when it releases today?

[indigo carbon]

LoRAs should be easier to make, the refiner is not as crucial, but it will be used by most people here I bet

oh will they work? no. SDXL LoRAs for SDXL

[thick mist]

Ty for the answer

[quartz sequoia]

what are you using to make these?

[indigo carbon]

SDXL0.9

[quartz sequoia]

are these SDXL 1.0

[ornate elk]

Has anyone noticed these red dots all over darker parts of images coming out of SDXL?

[uncut steeple]

First time seeing that

[nimble heart]

watermark

using Diffusers?

[ornate elk]

huh? stable diffusion is a diffuser

[ornate elk]

wdym?

[nimble heart]

no Diffusers the python library

or maybe you're using Vlad's ui?

[ornate elk]

yeah, a1111

[nimble heart]

a1111 im not sure. Vlad's uses Diffusers which has a watermark