✨｜sdxl

[indigo carbon]

Then try to write "sd-refiner" there

That should work

Tag me if you get into any trouble

[paper phoenix]

on a1111 should be same place

[indigo carbon]

Except a1111 can't run SDXL with refiner

[peak dove]

I found it - checked is sd_model_interface

[thorn lion]

(EDIT: This was dumb but we're dumb before we're smart and my direction was realigned by pseudo. I leave this here for anybody else to benefit similarly

what if you reinitialzed the weights and trained a model on ONLY hands with overly verbose and descriptive captions and then used, like, an adetailer or some other form of segmentation->inpainting pipeline? Although, that may defeat the purpose of "foundation model" but there is data within the foundation that conflicts with how we want to be able to control hands. So, just, take it out, no? (Obviously easier said than done but it needs to be said before it can be done)

[peak dove]

Or sd_model_checkpoint I mean

[indigo carbon]

Add refiner then apply

[uneven dove]

man i need that scene from billy madison, "Not a single word of your answer, was correct"

[indigo carbon]

SD_refiner or whatever it's called

[thorn lion]

I love being wrong, thank you!

[thorn lion]

i am going to watch it cause I also love that scene.

[peak dove]

SD_Model_Refiner?

[indigo carbon]

Exactly

[peak dove]

Add?Or Change?

[indigo carbon]

Add that, apply, restart UI, then enjoy

Add

Obviously

[peak dove]

OK, done, I am restarting 😄

[uneven dove]

i just watched it lmao

[paper phoenix]

we're talking about how to add things to quicksettings.

[uneven dove]

his explanation of the puppy who lost his way as "the industrial revolution" is also hilarious

[thorn lion]

been a good minute since I have. isn't the answer to the question, "the printing press"?

the puppy was a dog but the industry was a revolution.....ok....

[uneven dove]

honestly it's one of the best rants in cinematic history

[thorn lion]

in 10 words or less, why was i wrong?

[peak dove]

SDXL 0.9 Prompt - nocturnal animals with eyes shining in the night, jungle, detailed bushes, anatomical earth structure, dine, soft lighting

[thick mist]

So the full 1.0 version of SDXL comes out tomorrow?

[indigo carbon]

The refiner works?

[peak dove]

No

Its stuck - frozen

[delicate grotto]

Are you in comfy or a1111?

[thorn lion]

not asking to be taught. asking what direction to independently search for my answer with something more precise than "lolreddituser23 says so" or other forms of influencer rhetoric. I am not referring to you as lolreddituser23. You are clearly well educated and well informed. I want to do the hard work to get there. It's just hard when the rhetoric is so hype-oriented

[peak dove]

A1111

[indigo carbon]

Damn, did you select it in the sd_model_refiner drop-down menu

[indigo carbon]

I thought you said Vlad's

[peak dove]

I selected in in the Stable Diffusion checkpoint menu

Yes Vlad ... but these new manus I've yet to experience

Where is the sd_refiner_drop_down_menu?

[indigo carbon]

In the SD checkpoint menu, select base SDXL, in refiner menu select the refiner.

[indigo carbon]

Did you apply the settings?

[uneven dove]

it's not necessarily incorrect, it's just not simple like that either

[peak dove]

I have base selected - but no further option

[uneven dove]

there, 10 words

[peak dove]

WHERE IS THE REFINER MENU? (Sorry for shouting!!!) 🙂

[uneven dove]

no one knows what toolkit you're using unless you tell us

there's like 80 billion ways to run SDXL

i don't read either so if you already told us, well

[indigo carbon]

Uh idk then. @uneven dove can you help him? I gotta go

[uneven dove]

i gotta go do work.

[peak dove]

Thanks for all your help 🙂

Venice Carnival Grand Guignol

[thorn lion]

Thank you. As per my personal protocol, I'll keep rereading the papers about training methods and multimodal image/text understanding and evaluating/testing various resources available until my ability to rearticulate the complexities are second nature and then, with that validity on my back, try to acquire the specialized information behind the "it's not that simple" barrier through social communication with living beings. Since that is my labour to perform before I can perform the empathy necessary for me to ask for help while considering the other real human being across from me, I shall continue to look up all I can and read documentation before asking further questions on the topic.

[uneven dove]

what you're thinking of might be some kind of LoRA/LyCON or something

those are randomly initialised networks that are trained and merged into the base unet at runtime

i'm built differently

[molten breach]

release still planned for tomorrow?

[broken hare]

Are you guys using the Sytan setup, and also using an easy image to image option?

how do i get there?

[thorn lion]

The best words I have, when observing my thought, are that, in the [THIS IS A...] box, we have "specialized inpainting model" and, in [THIS IS NOT A...] box was "LoRA/LyCON"

[thorn lion]

base model for just hands then extracting a lora might be an overlap between what i'm thinking and, according the my best understanding, your communications

[uneven dove]

an inpainting model would need more knowledge than just hands in order to be able to compose hands into the scene though

[broken hare]

I havent heard anything to agree or disagree with that 18th target. But how cool is it to get a great new base, and then another one a week later

[thorn lion]

ahh....i see what I missed now

[thorn lion]

which would mean this. i see i see

thank you

[thorn lion]

to my understanding: going back down a level loses the contextual understanding, making my idea revolutionary in, like, 2015 but it's 2023 now soooooooo we can do better

[primal nimbus]

in how many hours SD XL releases? please

[eternal fog]

Target date has been given as tomorrow

July 18th UTC

[peak dove]

SDXL 0.9 Vlad AUTO1111

[ionic dragon]

Did you compare with comfyui?

[peak dove]

Not yet

SDXL 0.9

[hoary saddle]

what's the preference between A1111 vs Vladmandic? mostly concerned with the great API features and use them daily via python scripts, does Vladmandic still utilize the same API structure?

[visual glade]

both codebases have a bunch of technical debt but vlad seems to be making some good decisions like switching to diffusers

if you want good api features there are better solutions

[trim orbit]

what's good about diffusers? why haven't they been used yet?

[broken hare]

i havent made any workflows myself yet, just used others......is there a way to drag an image from my batch and easily use it as inspiration on next batch? using SDXL in Sytan's setup

[hoary saddle]

FastAPI seems pretty solid, was easy to understand for an idiot like myself lol

[trim orbit]

i woke up today thinking it was the 18th. i am too excited for this.

[autumn forum]

Same. Lol

[uncut zephyr]

Have you heard any updates? Are they still planning for tomorrow??

[marble kindle]

so will a 3060 12gb vram 16gb dram be enough?

[visual glade]

on comfyui yes

[marble kindle]

for 1.0?

[visual glade]

a lot of people use diffusers it's a library not a UI

[broken hare]

im using a quadro RTX 4000 8 gig and it works great in comfy

[uncut zephyr]

Yes, for the release of 1.0

[broken hare]

RTX 4000 and 3060 are pretty close in performance i believe

[marble kindle]

has 1.0 been leaked?

[fresh path]

no

but it comes out like, tomorrow..

[marble kindle]

oh

[visual glade]

1.0 has the same performance requirements as 0.9

[marble kindle]

what are estimated generation times for 3060

and does all generations need to be equal to 1024x1024

[trim orbit]

yeah it's a core library for inferrence if i understand it right. i'm curious why theres buzz around them now with sdxl coming out. they've been around for a while but it has always seemed they're a hastle. like models need converting

[broken hare]

nope, im changing resolutions to all kinds of things, above and below

[marble kindle]

will 1.0 work in a1111

[trim orbit]

i loaded up d8ahazard's new ui that uses diffusers and had to do some recompiling process to make any model i wanted into diffusers

[broken hare]

even if it does, just switch to comfy for now

[sage basin]

Depends on what samplers you use, but using ddim, I'm getting 85 seconds for a batch of four 1024x1024 images on my 3060

[uncut zephyr]

So far a1111 only has it as an extension. ComfyUI or Vlad (SDNext) can run it natively

[trim orbit]

i got it working but that whole process seemed like a whole un needed hastle for no benefit

[visual glade]

I don't think a1111 implemented the ensemble base + refiner pipeline yet

[broken hare]

yep, that sounds about right.... similar to me

[hoary saddle]

will probably switch to comfy when an API comes out for it, would be cool to just call in json files from python

[visual glade]

so for now it's comfy, vlad, maybe others

[marble kindle]

comfyui looks nicer but im so used to a1111 lmao

[fresh path]

comfyUI works with amd too

[visual glade]

there is an api

https://github.com/comfyanonymous/ComfyUI/discussions/909

[hoary saddle]

yeah i looked thru the links you sent last week, not wrapping my head around it as easily as i did with a1111

[uncut zephyr]

Vlad is a branch of a1111 and sdxl runs on it

[trim orbit]

i saw automatic1111 ui working with 0.9 this morning. my neighbor has 0.9 and we got it working in auto

[broken hare]

im still blown away that i can make batches of 8 images, and large ones without any error

[vast narwhal]

@visual glade There's a switch node for comfy ui? A reroute node that receive two sources and you choose which one by clicking like a light switch?

[visual glade]

not in the base but there's probably one somewhere

[uneven dove]

SD-2.1 partial diffusion with SDXL refiner ❤️

[broken hare]

i havent even tried using 2.1 models with refiner. nice

[hoary saddle]

lol what's this? invisible watermark? hah

[trim orbit]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/dev this branch allows sdxl

[uneven dove]

it's working pretty well @broken hare

[visual glade]

I don't think the SDXL pipeline is properly implemented, I think he just implemented the models

[void skiff]

I see folk use links to images in their prompts, when I look at them they do not look as the result, is this a function in SDXL that folk just use in a faulty way or is it just a prompt trend that do nothing?

[trim orbit]

it's not some conspiracy to track and control ai images. the invisible watermark has been around for a while now. we've been labling 1.x generations with watermarks.

The idea is that future scraping of the web will encounter junk from old models and won't want to train with those images potentially.

[uneven dove]

those two jesters are from my ptx0/pseudo-real model, and these two are from ptx0/pseudo-flex-base, through SDXL refiner (50% refiner strength)

[trim orbit]

the watermark identifies them as a legacy generation

[broken hare]

i watched a video comparing results to comfy with same prompt/steps and everything that he could make the same, and they just look blotchy and weird out of A1111

[hoary saddle]

appreciate the explanation Flowwolf

[trim orbit]

it seems pretty sus up front haha

[hoary saddle]

lol in this world we live in right now, yeah hah

[trim orbit]

if there was some cabal that wanted to track things they wouldn't use a library called invisible watermark.dll

[broken hare]

do you think we will get some other software to accompany release of 1.0?

[uneven dove]

trying to make it mechatronic

[void skiff]

All are so fixed with AI fakes now and want them to have hidden markings and so to spot faked images, but the best faked images are still made in Photoshop by skilled artists.

[broken hare]

i want tickets to this show

[uneven dove]

i'm getting way better widescreen results using my flex-base model into the refiner, than 0.9 base. that thing performs poorly outside of its base square res for me

[high skiff]

https://tenor.com/view/kermit-the-frog-meme-memes-gif-24277198

[vast narwhal]

comfy UI and Auto1111 can't make exactly the same image because they generate noise in different ways

[trim orbit]

love that final tracking refresh mania fun

[broken hare]

really? Im actually liking my 1280 width shots so far, but i think i will try it out with some 2.1 based models for fun. Artius for example

[uneven dove]

artius has really effective bucketing in the models it merged, it does pretty well too but not in portrait res

[broken hare]

is that a new computer on the way?

[trim orbit]

people get really hung up on platform parity in this realm. but why?

I get it for like minecraft. you want a seed number to always reproduce the same world for player consistency. But why are we wanting to recreate the exact same image again? can't we just save the image?

just feels like platform parity doesn't matter very much outside a few fringe scenarios

[high skiff]

my 3090 for my new PC

My new PC got here 3 days ago heh

its the final part

[rustic garnet]

it is a hastle. But the design of diffusers is the better one. In auto111 people tend to do stuff like model pruning and so on to keep file size small. In contrast, diffusers always just download the part of the model you need. That's why the format is a bit different.

Nevertheless, I agree with comfy that diffusers is a popular and widely used library. Just not by people who wants an UI and, therefore, use auto111

[trim orbit]

if the end user case is a convoluted process requiring conversion to diffusers befor a model can even be loaded, i don't think it's designed better imo

i've never heard any reasons why diffusers are better. just told they are. i'd like to know. i've looked it up but all i find are lots of toxic comment threads of people yelling at each other

[visual glade]

diffusers can load single files now

[trim orbit]

i only care about use cases. i'm not a ui developer.

[visual glade]

they are better because their code is decent

and their performance is better than a1111

[rustic garnet]

that's a strange argument. There are just two different formats out there. If you have your model in auto111 format you have to convert it to use it in diffusers

[visual glade]

I took some of their performance tricks for comfyui which is one of the reasons people say it's faster

[rustic garnet]

other way around: if your model is in diffuser format you would have to convert it to auto111 to use it there

[trim orbit]

https://huggingface.co/docs/diffusers/conceptual/philosophy i've found this but leaves me head scratching. saying they're not making it easy, they're not making it performant. i dn't care about it lacking abstraction. that's a developer preference. i see them hyped all the time, but i just don't get why end users should care.

[trim orbit]

safetensors wasn't invented by auto1111

[rustic garnet]

that has nothing to do with safetensors

[visual glade]

comfyui natively supports both formats by the way

[rustic garnet]

you can load safetensors in diffusers, too. The question of safetensors vs. checkpoints is a completely different thing

[trim orbit]

diffusers philosophy page says they're going for usability and not performance

[vagrant plover]

hype hype hype

[rustic garnet]

its about the naming of the weight matrices IN the safetensor or checkpoint file

[high skiff]

INDEED

[visual glade]

they still are more performant than a1111

[high skiff]

WOOOOOOOOO

[rustic garnet]

they are more academic. They want to implement stuff as it is described in the paper

[trim orbit]

when i used them on d8ahazard's ui, it was much slower. probably just their code then

[vagrant plover]

I KNOW HOW THAT FEELS LOL

[rustic garnet]

and don't get me wrong. Not everything in diffusers is great

and many things in auto111 are better BECAUSE they did not follow the research papers but experimented what was better for artists

[trim orbit]

i just don't understand diffusers at all. they're often heralded as something so great and bringing accuracy and benefits, but as an end user, i see nothing. the only thing diffusers has ever presented me with is a giant hastle

[rustic garnet]

it's just: auto111 is not a proper API. Its a gradio interface around the stable diffusion code which is growing over time and get less and less maintainable

[visual glade]

it's about development mainly

[rustic garnet]

when I started stable diffusion I looked into the auto111 code, run away, looked into diffusers code and finally could start programing with it

its just so much cleaner and easier to understand

[trim orbit]

so it's like, "we're using microsoft visual studio run time v1.284 instead of v1.187" kind of hype.

[inner ruin]

Curious if 1.0 will be better than 0.9!

[visual glade]

something like a1111 for example has to implement pretty much everything from scratch

[trim orbit]

totally meaningless to end user

[visual glade]

while people using diffusers can just update the library and have everything that's new

[trim orbit]

use the bots and see for yourself!

[rustic garnet]

you should never mix a backend with a frontend. But unfortunately, that is what is happening more or less with auto111 in my opinion

[peak dove]

Cannot see that diffusers/safetensors are a big deal - both produce stunning results!!!

[rustic garnet]

there should be a good and maintainable API which can be extended by developers and which should be independent from the UI the enduser is using

[trim orbit]

auto won't last. it has an expiry date. feels like duct tape engineering to get'r done. i appreciate that kind of workmanship, but at some point you've gotta come back and get it done right

[rustic garnet]

diffusers is not for endusers. Its for tools like ComfyUI that build upon it

[peak dove]

My Vlad AUTO1111 startup line = webui --medvram --backend diffusers

[trim orbit]

Vladmantic enthusiasts have been preaching that diffusers are the jesus of diffusion tech it feels like. i've been trying to figure out wtf is the deal with them

[peak dove]

Vlad AUTO1111 SDXL 0.9

[rustic garnet]

what is the problem? That people hype something 🤷 Thats the internet

[trim orbit]

feels more like the "Blast Processing" of diffusion tech to me. just some marketing buzz word that end users don't see at all

[rustic garnet]

but if that is your question: no, Jesus and diffusers are two different things

[silver flax]

hey guys i have a macbook and really want to get into stable diffusion, would you reckon runing it through google colab?

[rustic garnet]

for developers it is a great thing (and definitely not a marketing buzz word) because we have one stable code base we can all work upon

[trim orbit]

hmmm. no. that was not my question. if you read it more closely, you'll see i represented that particular point as a simile in a hyperbolic fashion. Analogy.

Blast processing was great for developers too, but when it showed up on sonic magazine ads it was just hilarious and meaningless

[visual glade]

picking diffusers just shows that he's doing the right decisions from a software dev perspective

[uneven dove]

look inward

[trim orbit]

you've got me blocked. consider that before pinging me in the future

[rustic garnet]

in the end there are, to the best of my knowledge, two code bases for stable diffusion

[peak dove]

As a complete non-techy/mostly artistic user of SDXL - I feel that diffusers and safetensors BOTH are doing a great job!

[rustic garnet]

the original one from compvis and SAI

and diffusers

[visual glade]

actually 3 because comfyui only uses the models from ldm and throws out everything else

[trim orbit]

humble flex

[ionic dragon]

can you please share the seed and the prompt including the -ve prompt

[boreal bough]

I lack the context. Are we comparing divine creation to diffusers library? 🤣

[uneven dove]

i told ya why, it's because you seem proud to not know stuff, or seek out wisdom. i think everyone in here understands your 'end-user perspective' and it is not anyone else's job but your own to educate yourself on how things work, or stop questioning others who work on it

[rustic garnet]

the original code base is not maintained. Its SAIs own thing and they use it for developing and improving their models. But they don't throw out any new features like controlnets, loras, sag and so on

thus, we are left with diffusers as the only really maintained and extended SD codebase. or we do what auto111 is doing and every tool reinwents the wheel and hacks all its features again and again in the same old codebase

[peak dove]

Seed = -1 Model = Euler A Iterations = 20 CFG = 14 Prompt = lowbrow, 8k, matte painting, in the style of zentangle fauvism fibonacci Lilia Alvarado Salvador Dali Rembrandt Kandinsky Mark Ryden Anthony Ausgang Robert Williams Amy Sol Kenny Scharf Camille Rose Garcia, rauschenberg, diana ejaita, basquiat, remedios varo, rob gonsalves, keith haring, zentangle, blek le rat, ntombephi ntobela, richard burlet

[trim orbit]

this is getting really immature as people ignore context. i'm not attacking diffusers. just trying to understand their benefits. asking questions isn't a bad thing in my world view.

[peak dove]

Using base SDXL 0.9 no negative prompt

[trim orbit]

i'll drop it now

[ionic dragon]

seed -1 randomizes the seed, so after the generation, you can view the seed

[novel birch]

is SDXL still on track for tomorrow?

1.0

[peak dove]

OK, but you require the exact seed?

[delicate grotto]

hopefully we get both of the SDXL's

there are a few versions, the large ones are "aesthetic" and the stock 1.0

[rustic garnet]

haha, to be honest, I asked the same questions about ComfyUI and this was also in no way meant as offense against this tool

[delicate grotto]

not sure which we will even get in the end, might be just sdxl 1.0

[boreal bough]

possible vram optimization?
I mean, even if we ignore all it can do for standard inference, finetuning or lora training would be hell without it

[peak dove]

Seed 3696284803

[fresh path]

do they have any theories on how diffusers can overcome the current pitfalls it has always faced?

[trim orbit]

dont ask me man. asking questions about them got a mod favorite regular all heated at me. didnt yuou just see me get ripped a new one because i admitted to not knowing ?

drop the topic please. or at least stop pinging me into it

[rustic garnet]

honestly, I don't know who said anything offending to you. Maybe you overreact a little bit?
Anyways 🤷

[trim orbit]

naw. drop it please.

[fresh path]

strange man

[silver flax]

hey guys i am working on something, basically they're mockups for clothing brands where you can add your own design on the t-shirt through a photoshop, for the mockups i want to create generate realistic model pictures like these 2 below, would stable diffusion be something i should look in to? 🙂

[trim orbit]

stable diffusion isn't an instant solution to your drop ship website. not without a lot of hhand crafting to get it right. That'll be r&d you need to invest into to get a jump on the coming AI generated drop ship sites. no one has made tools public for that yet.

[silver flax]

it's not for a dropship website, just want to provide realistic assets for startup brands with no capital to do photoshoots, i would edit it in photoshop so they could mock up their designs and use it to display on their website etc

i've been able to create these very basic ai generated models, but never really looked into stable diffusion so i figured this might be the place to ask

[rustic garnet]

if its just about making artificial model shoots, I would say that's definitely possible

[winter raptor]

So I'm trying to achieve an image of a man standing on a cliff/rock outcrop over looking a cyberpunk city, with a shattered/crumbling looking planet in the sky. Anyone have any recommendations on how to prompt for that? No matter what I do, I can't seem to get it

[boreal bough]

while you could probably get the relevant parts working (heads & hands, where you just overlay clothes in photoshop)
you wouldn't actually avoid the fact that its commercial usage
Not saying you cant use it, just that you need to properly read the license agreements (from Stability AI), and stay on top of new laws in regards to AI gen image usage

Odds are higher you'll save money by getting a one month subscription to envato elements - download like 1000+ images of models wearing various clothes, to which you then permanently own commericial rights

(also legal aspects change, depending on which country you deal with - which doesn't make using AI gen images any easier)

[silver flax]

yes exactly that, they need to be very basic models, like the ones i've sent, no designs, nothing special, because i would be adding the design onto it through photoshop

Just like this i will make it possible to mockups your own designs and change color etc, just wondered if stable diffusion would be a good option to generate the base models

[boreal bough]

depends.
1.5 -> get an anime checkpoint
sdxl -> use an interrogator running Vit-H on an existing similar anime picture, then use the prompt it gives you and work from there. It won't make much sense why it works - but it works XD

[pallid path]

it's that good now?

you can make it very specific and it gets that close?

cause using an interrigator in older models just kind of sukced

[winter raptor]

I've been trying to get an interrogator to work with SDXL, similar to clip interrogator. Are there any easy methods to get that up and running?

[lusty raptor]

https://github.com/pharmapsychotic/clip-interrogator-ext

[silver flax]

Are you using colab?

[boreal bough]

#✨｜sdxl message
yes. it can give you the images you seek.
but seriously, don't ignore legal issues. As an ex web dev, I've seen more than enough customers suffer real bad from images with shady licensing - and the costs for messing up there run between 1k~15k, so please take it serious

[silver flax]

thank you! i will definitaly have to do some research about this because i wasn't really aware of it

[boreal bough]

they still do - but Vit-H works like magic on sdxl

[pallid path]

thanks

[winter raptor]

How do you do that? Is that functionality built into ComfyUI?

[silver flax]

i have a macbook so i can't run stable diffusion locally so i googled how to run it through web browser and they told me to use google Colab to run it

[autumn forum]

Just download a model off civitai and put it in the same folder you put the 1.5 model. Switch to the new model. Type what you want. Enjoy

[lusty raptor]

this is the interrogator that caith mentioned, you feed it an image, and it gives you a prompt that should make something like that image. like anime ladies or whatever

[random zenith]

what time will be the SDXL launch tomorrow? I want to get it as soon as it's released lol

[boreal bough]

also check out this site
https://elements.envato.com/
most new artists/web designer don't know about it - but its a lifesaver when you're starting fresh

[winter raptor]

What prompt did you use for that?

Wonder if it is just struggling for me because of the aspect ratio im trying to use

[eternal fog]

Cinematic Shot of a man standing on a cliff overlooking a city, cracked planet in the night sky

And then Cyberpunk in the 2nd text box for style

[distant panther]

hello can someone say me why with the 3060 ti i get very low it/s

[eternal fog]

What aspect ratio?

[winter raptor]

9:19

[silver flax]

thanks! i have an envato elements subscription also,, just wanted to see what i could do with AI, i have used the basic Ai's everyone can use but just recently started to learn about stable diffusion

[eternal fog]

I'll try that, works at 16:9 though

[delicate grotto]

are you on mobile? and what res?

im 6gb 3060 mobile(non ti) 125W at ~ 1.25it/s and at times when it is hot at 1.3s/it - 1024x1024

[dense chasm]

Hydra jordan with sd 1.5 realisticVisionV40 and LyCORIS ,painful

[boreal bough]

aspect ratios are heavily biased 🫂
but try to like move it up and down a bit, often things get a lot better by making a small compromise

[distant panther]

im on pc, res its 512x768

[winter raptor]

Got some interesting results this run through

[silver flax]

if i want to run stable diffusion via my browser, do i have restrictions to when running it locally? i have a macbook so i'm afraid i can't run it

[delicate grotto]

yeah, seems kinda low for a reason
is your cpu old get? also what is your xformers version?

[distant panther]

how i can se xformers version

[boreal bough]

if you want to use AI commercially, you can use Adobe Firefly - on the legal side they've cleared everything for their users

[distant panther]

cpu its ryzen 5 5600x

[eternal fog]

The cracked/broken planet is hard. Because do you mean the planet you are standing on, or the one in the sky. I don't think it can differentiate

[delicate grotto]

it looks really really good on your end

[uneven dove]

lol @ people coming in here to ask if there are restrictions. this only means so many things

[distant panther]

how i can se xformers version
cpu its ryzen 5 5600x

[hoary saddle]

anyone have a good example python script and workflow.json for img2img using comfy

[distant panther]

models/stable diffusion

[winter raptor]

I meant the planet in the sky. Yeah, it's super tricky, been trying in sd 1.5 as well, and I couldn't even get the cracked planet anywhere there lol

[delicate grotto]

you are using comfy right? should show you when you start it

[delicate grotto]

at the very top

[high skiff]

RIP this channel

[high skiff]

long gone are the fun times, now its all basic tech support ._.

[eternal fog]

You know there is a #🤝｜tech-support channel right

This is specifically for SDXL discussion

@winter raptor

[boreal bough]

I'm gonna quote this next time we devolve again

[uneven dove]

@boreal bough don't make me call you out

[winter raptor]

I'm liking the composition you are getting

[delicate grotto]

doesn't show here, anyway go to tech-support and ask around, i can't remember the command to update xformers

[distant panther]

oke

[eternal fog]

Exploding planet works quite well, but maybe not what you have in mind

[autumn forum]

Same place you put the 1.5 model

[uneven dove]

this channel feels like K-Mart now for some reason

[winter raptor]

Who are we kidding, the channel always devolves

[uneven dove]

real K-Mart vibes.

[autumn forum]

It’s gunna be even worse when 1.0 comes out

[eternal fog]

#🤝｜tech-support
#🤝｜tech-support
#🤝｜tech-support

[silver flax]

What??? with restrictions i meant like can i run everything like i would if i ran it locally, i am only researching before i want to learn

[uneven dove]

can we bring it up to the level of Sears, if not Macy's, and ideally Radioshack but i know we'll never get there

[winter raptor]

I mean I'd prefer for it to not shut down like radioshack, so maybe we don't aim for that specifically lol

[boreal bough]

we'll just spam it back on topic with images & stuff XD

[uneven dove]

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

[silver flax]

thank you!

[uneven dove]

go nuts, kiddo

[winter raptor]

I've noticed it can do the exploding planet well as well, struggles with the shattered planet much more. Been trying to find a way around it using the split prompting, mixed results. Maybe a lora is required.

[eternal fog]

Maybe get a planet and then inpaint cracked planet

[uneven dove]

christian slater approves this message

[delicate grotto]

not sure what mac you got, but you almost for sure can run it on your CPU

[uneven dove]

PowerMac 5500

[delicate grotto]

why not do it in steps? first make the planet and then add a few steps for the cracks

[eternal fog]

I mean that's sort of what I meant

[autumn forum]

I heard someone say that there will be multiple sdxl 1.0 models. Y’all think that could be true? Cuz I don’t believe it

[lusty raptor]

eh, i was a bit annoyed with sytan's username/hardware saga

[delicate grotto]

btw, found a way to not let it fully denoise a pic in the amount of steps the K-sampler outputs?

[silver flax]

macbook pro 2020
2GHz quad-intel core i5
intel iris plus graphics 1536 MB
16 GB 3733 MHz

[eternal fog]

What do you mean?

[delicate grotto]

yeah, seems like CPU mode is the only possibility for you
it takes much longer to output pics like that but it is possible

[uneven dove]

same

[silver flax]

using googles clouds bases CPU servers would solve this issue right?

[delicate grotto]

when you lets say give the k-sampler a task and ask it for 5 steps - it would usually fully filter the noise into a picture
if you give it 20 it would still fully filter the noise

[eternal fog]

Yeah, just set "Return with leftover noise" to enable

[delicate grotto]

mmm right that is an option

[eternal fog]

it sounds like your talking about what we do with the refiner

[delicate grotto]

nah, just prompt injection mid steps like in au1111

[eternal fog]

Yeah you can do that as well

But it doesn't always work very well

[delicate grotto]

is there a specific command?

[eternal fog]

No, just 2 KSamplers with different prompts fed into them

[delicate grotto]

mmm yeah, so what i've been doing so far
so it does transfer the noise... in the preview i couldn't see any noise so i though it fully filtered it after x steps

[visual glade]

this is a complex example but this is how it can be done: https://comfyanonymous.github.io/ComfyUI_examples/noisy_latent_composition/

[eternal fog]

If you tell it to yes

[visual glade]

the previews don't show noise because they show the denoised image not the Xt

I thought it was more useful than showing the image gradually denoising itself

[delicate grotto]

any way to show the noise too?

[eternal fog]

why would you want to

[delicate grotto]

just to know how much noise is left to be filtered, more noise means that more can still be changed without adding more noise

i usually use that to know when to inject the prompts in au1111

[delicate grotto]

you did great tbh, it works better than au1111 and i gotta give ya props for that, got a donation page?

[eternal fog]

It sort of works, put the prompt "Cracked Planet" in half way through a 30 step render

at 15 steps

at 30 steps

[patent badger]

what's the minimum amount of images you would recommend for training lora on SDXL?

[boreal bough]

while extremely cool, I also doubt it. But a man can dream? right? xD

[civic sigil]

It's pretty believable to me, I would like a photorealism model and an anime model, maybe a traditional art model too

[boreal bough]

1~30 <- hard mode. would not recommend
30~50 <- lowest that works without too much effort. make sure to tag backgrounds as well!
50~100 <- good amount that avoids 90% of issues
100~500 <- getting harder again (to keep up good tagging), but you can achieve really cool things!
4000~10000 <- while training for 2 eternities, you can achieve finetune levels of results on sdxl with just a small 43mb lora

[delicate grotto]

in pickApic there are 2 SDXL models
SDXL and SDXL aesthetic

[boreal bough]

(if it's just your own, or anothers face you wanna train, divide my numbers by 3)

[visual glade]

no the project doesn't have any expenses

[winter raptor]

That’s actually really good, would you mind sharing the workflow for that?

[eternal fog]

Just a 2nd ksampler with a 2nd text encoder.

Feed the first ksampler into the 2nd with if ending the steps before it's finshed and return with left over noise.

Then have the 2nd text encoder have something different in that's going into the 2nd ksampler.

[delicate grotto]

mmm well, then im glad that your project went so well

[autumn forum]

Hm that’s interesting.

Also interesting is that if they release information on styles or if those were just injected prompts

[winter raptor]

I’ll give it a try, I haven’t played around with changing the steps, noise output, etc, in the samplers yet.

[eternal fog]

It's the same sort of thing you do with the refiner

[amber fulcrum]

Hope it is not spam - my fav dragons

[delicate grotto]

good question, because the 2ndary prompt was a surprise and it might have been it

[high skiff]

not looking forward to it

[winter raptor]

These are really cool

[high skiff]

likely as it is already, I will probably be abandoning this channel

[high skiff]

These are so dope

[eternal fog]

Imagine standing on 4 legs 🙂

[boreal bough]

same. but thanks to the bots, they'd be reverse engineerable with a few hours of effort. so we get them either way 😄

cause I'm really liking the sticker style

[boreal bough]

those are chefs kiss pretty!

[trim orbit]

weird to see the resolution attention problems at 524x524 pixels while also, this is an sdxl channel not 1.5 RV channel

[amber fulcrum]

workflow should be in the images

[visual glade]

if anyone has slowness issues on the latest nvidia drivers on ComfyUI update it and try: --cuda-malloc

[eternal fog]

Does that fix the awful memory management?

[boreal bough]

"Churrarura" poor dragon got named after spare ribs :"D

trying it now ❤️

[visual glade]

that's what I want to find out

[eternal fog]

I could update my drivers to test

[winter raptor]

Idk what this is, but it looks cool lol

[uneven dove]

The #SDXL Merry-go-Round

[uneven dove]

have you tried microsoft mimalloc with 2MiB hugepages?

[boreal bough]

thanks for prompt ❤️
translates well into other styles c:

[clever verge]

@lilac wren Like your F40 so much that I had to do something similar.

[civic sigil]

Is it just me or changing the prompt halfway through the steps only seems to affect the background?

[boreal bough]

are you outputting noise as well?

[civic sigil]

For the last sampler? No

[boreal bough]

for the midchange

[civic sigil]

I do

[boreal bough]

odd 😦

[civic sigil]

For the first sampler

[boreal bough]

too many steps perhaps?

[eternal fog]

If your initial image is already established it will be hard for it to change major components

[delicate grotto]

Try lowering the denoise value on the first one

[amber fulcrum]

cool - I play around with subject prompt and the part with location - it seems working well. Also keyword location in context seems to work well, like "tundra location, ancient ruins location, old ruined pharmacy interior location" and similar

[delicate grotto]

The 2nd one should be at 1 i think?

[civic sigil]

It just seems like the background is more malleable than the foreground for some reason

[boreal bough]

in case you havent tried it yet, :: also works, but it changes how the weights work significantly

A detailed portrait of Amy Quesada :: City center dripping with black ink and black slime in the background, lights reflecting gasoline colors:: Bojan Jevtic + Ashley Wood :: maximalist intricate detailed :: ray tracing :: hyperdetailed, maximalist, psychedelic, post-apocalyptic, photorealistic, 64k resolution concept art, dynamic lighting, trending on Artstation

an extreme version using it, to show what its capable of

[amber fulcrum]

thx!

[eternal fog]

What does :: Do?

[boreal bough]

and how I used it for your prompt

[[fantasy monstrous compendium zoology illustration]], (epic colored painting of legendary animal fauna style "Churrarura" animal-hybrid, dynamic pose, Full body, (aquatic triceratops necromancer bone-dragon which body is made from bones, with visible skull and bones through decayed flesh, (skeleton dragon:1.6)) roaring:1.1), (by Norman Rockwell and Kazuki Takahashi and Frank Frazetta and Takeshi Obata:1.2), MTG illustration concept, [dnd character sheet], symmetrical, (detailed biology specie), centered subject, highly detailed, broad brushstrokes, fine art, (fantasy cursed dungeon with giant hall and stone obelisks location:1.4) :: lithograph, risograph

essentially gave like 0.5 to the first part, and 0.5 to lithograf & risograph - which is why they have such a strong impact int he image I showed

[uneven dove]

uh, :: isn't handled specially 😛

[clever verge]

New option --cuda-malloc in ComfyUI improved generation time by a second

[boreal bough]

it's similar (but not the same), to using this

(prompt part 1, with this style going on:0.5) , (prompt part 2, with this other style going on:0.5)

[boreal bough]

good for even merging of styles

rather than overweighting everything

[eternal fog]

@visual glade --cuda-malloc seems to mostly fix the fucky memory issues. But it still dipps into it a little bit and causes a small slowdown. It's not where near as bad as before where it would try fill up RAM and crash my PC

[soft bone]

so it ignores prompt order and makes a 50/50 split

[eternal fog]

Where's the documentation for it if you don't mind?

[boreal bough]

kinda

[boreal bough]

not sure about official documentation
came across this
https://aidailies.com/midjourney/what-are-double-colons-weights-iw-and-multiprompts-how-do-i-use-them

[eternal fog]

Isn't that something that would have to be implemented in whatever system you are using?

Is it implemented in comfy

[uneven dove]

it's not implemented in most places other than midjourney

[boreal bough]

@visual glade am I spreading snake oil?
cause I could swear its working :/

[visual glade]

thanks for testing if nobody reports issues I'll probably enable it by default on the standalone builds

[eternal fog]

I think it's just the fact that : is a token, I don't think it's doing what you think it's doing

[boreal bough]

maybe XD

[eternal fog]

I don't think it would really make much sense for what the UI is designed for

As you'd be changing the input of a node part way through.

[uneven dove]

even among comfyUI, InvokeAI, the diffusers LPW pipeline, and Automatic1111, they all have different attention emphasis syntax

[eternal fog]

@visual glade Ok did a secondary test, it's definately a fair bit faster on my upscaling workflow.

A lot actually wtf

[visual glade]

comfyui just has (test:1.2)

[boreal bough]

cries in false confirmations

👍

[uneven dove]

stunning portrait of Caith consuming a horse size placebo pill, epic, cinematic

oops

wrong window

[eternal fog]

@visual glade Yeah this is like 20 - 30 seconds faster over my entire upscale workflow

[visual glade]

that's on the latest nvidia drivers?

[eternal fog]

yeah

Unless there's some really new one

Yeah newest one

[visual glade]

how does the speed compare to vs 531?

[eternal fog]

I cba re-installing the previous one, but I was getting something like 100 seconds for my entire workflow when changing prompt, now it's 80

[trim orbit]

using sytan's workflow, adjusted to 30 steps, i generate in 18 seconds with new drivers and that the cuda malloc switch. feels faster. i'll relaunch without the switch to see

[visual glade]

ok thanks for testing

time to enable it by default then

but only on the standalone

[eternal fog]

@visual glade Could you put the switches on the main page along with the keyboard commands maybe?

So people know they exist.

Or even just put run main.py with --help to get a list of switches

I was looking for a wiki page that listed them and couldn't find any

[civic sigil]

Wow prompt scheduling is so powerful

[trim orbit]

hmmm. my generation speeds are all over the place. without it i'm getting 11 and 10

[lilac wren]

😄

[trim orbit]

i'm on a 4080.

[eternal fog]

If it's for memory allocation it probably shouldn't really effect cards with larger memory. I forget how much the 4080 has.

[trim orbit]

going back to use it again. i'll try 5 more

16

[eternal fog]

Yeah, I have 10GB

So before on these drivers, it was being stupid and going OOM when switching to the refiner, but it didn't report that to the app for it to stop, so then it just starts leaching System Memory until it's full and your PC locks up

[trim orbit]

looks like across a longer set of gens, it's the same speed on both counts for me. what you say makes sense.

[quasi remnant]

sdxl needs like minimum 11-12gb VRAM

afaik

[visual glade]

if anyone that has the standalone updates now you should get it as long as you didn't remove --windows-standalone-build from the .bat

[trim orbit]

probably eventually

[visual glade]

I have to do some tests to see if it breaks AMD before enabling it for everyone

[boreal bough]

8~10 can be made to work - but does need to cut corners

[quasi remnant]

should we add this to the target line if we build from git?

[boreal bough]

...I should really save one of the 8gb layouts so I can repost it

[visual glade]

from git add: --cuda-malloc

[quasi remnant]

alright, ty.

[visual glade]

assuming it doesn't break anything I'll probably enable it by default for everyone soon

[quasi remnant]

ooooo

using like 5GB less VRAM now on vae decode phase

and it seems faster

quite a bit faster lol. nvidia really messed up something

i can batch 3 with VRAM to spare now

instead of capping out at 2

xD

What exactly did NVIDIA do to break it? because this is a massssiiiive difference

went from 17 secs per image to 13 secs and able to batch 3 at a time, and still have vram to spare, not enough for a 4th batch though xD

[sweet abyss]

Does automatic1111 work eith sdxl by default ?

[trim orbit]

/dev branch does

[vernal cloak]

You have to "cut corners" with 10GB of Vram? I thought 8 was sufficient for all features?

[trim orbit]

should probably merge tonite or tommorrow

[quasi remnant]

no way... lol

8 is definitely not usable for sdxl from what i know unless you also use system ram/cpu and this

[high skiff]

I have seen SDXL run fine on 6GH VRAM

[quasi remnant]

will make it take.... so long

[quasi remnant]

hooow? Lol

[boreal bough]

tiled vae - though I think comfy made that default for low vram systems (so you dont see it happening)

[quasi remnant]

they're def using system RAM/CPU

right?

[high skiff]

I have seen plenty of people running it just fine on 6GB

[boreal bough]

not sure what else is being done in background

[quasi remnant]

in which case it'd be like 5+mins per image

unless it's something else?

[high skiff]

The base model only uses like 5.9GB when genning, and the refiner only like 5.1, it's just VAE decode that uses a lot of VRaM

[trim orbit]

sacrificing speed. switching models in and out. different optimization tricks. comfyui gives lots of power to that kind of cause

[vernal cloak]

Wait.. So if you have 8-10GB of Vram. You need to use tailing, and not a native res?

[high skiff]

In my experience, personally, 10GB VRAM doesn't have any compromises for SDCL

[civic sigil]

Usually only for vae decode

[boreal bough]

vae only - basically colors are a small bit off - but not in a way that's a real issue

[boreal bough]

load/unload - wasn't 12gb the magical amount where that stops happening?

[vernal cloak]

Ah. Ok. Thanks btw.

[high skiff]

oh, well maybe, but it takes what, like a single second?

I think the text encoders take more time than that

[boreal bough]

depends on a lot of things. 1~20 seconds, depending on your environment (Ram/ssd/hdd)

usually on the lower side though

[high skiff]

also, 12GB was the magical number where people were egtting OOM cause the new NVIDIA bug was gettinbg big enough balls to keep the models cached

on my old PC it was like 2 seconds, now its like 1

[ionic dragon]

at what time is SDXL releasing?

[high skiff]

new GPU shuld be here any time now

[ionic dragon]

its already 18th here

[high skiff]

nobody knows

[boreal bough]

git your pitchforks AND LIGHT THEM ON FIRE

we needs 1.0 XD

[high skiff]

My guess would be in like 26 hours ish, about 3PM, thats when I feel it would launh

[trim orbit]

Stability is a UK company. where the fuck is it ! 🔥 ⛏️

[high skiff]

assuming its even dropped

[ionic dragon]

26hrs?

thats would be 19th

[high skiff]

for you, yeah

not for us

[ionic dragon]

true

[high skiff]

oh really? Thats surprising

[trim orbit]

its 7pm on the 17th in london. they're fucking lucky. no wrath today

[boreal bough]

I hereby predict, in 19.5 hours!

[civic sigil]

I hope they took the time to launch it with controlnets ready

[high skiff]

SAI's recokless abandon and hype mongering felt very American to me lmao

[trim orbit]

UK has precedence for machine learning in their law i think

[boreal bough]

lets see how right/wrong I get it XD

[trim orbit]

the rebel spirit is much much older than america

[autumn forum]

If that’s the case it’ll be tomorrow morning for me. Ugh works gunna go so slowww

[high skiff]

I'm just patiently waiting for my 3090 ||impatiently||

[eternal fog]

Don't think we do.

We are just fucked after leaving the EU so wanting ways to make money. If the EU is going to start blocking stuff the UK is a much better location for an AI Business.

[high skiff]

the suspence is agonizing lol

[autumn forum]

I hope you get your 3090 before 1.0 drops so you have time to install it before it drops

[high skiff]

oh I will

its out for delivery right now

I just need it sooner rather than later, cause I need to sleeeeeep

[eternal fog]

late

https://tenor.com/view/champoy-el-risitas-kek-issou-etu-gif-17837830

[high skiff]

maybe for you lol

[trim orbit]

s/waiting/dying

[high skiff]

am so tired right now

Let's enjoy these final moments before this channel gets flooded with people asking the same question every few minutes

[shy kelp]

whats that question gonna be? anyone willing to bet money?

[autumn forum]

“How to use sdxl”

[high skiff]

So far it's "when does it drop?" "does it do NSFW?" and "can I make money off of it?"

[shy kelp]

"why does my img2img look like shit"

[high skiff]

Gonna be full of that

[boreal bough]

soon/soon/soon?

[eternal fog]

the best ones are the ones that come in and just try write prompts directly into random chats

[boreal bough]

I mean you didnt ask legally XD

[shy kelp]

absolutely based

[high skiff]

Those are the 3 I keep seeing in here nonstop

[eternal fog]

I guarentee most of the people that ask if they can make money off it are children

[placid edge]

you should expect those sorts of questions when involving such a large and diverse community 🙂

[boreal bough]

I wonder if they'll stick with the responsible AI license, or write a custom one

[eternal fog]

Thinking they can become millionaires from having a computer generate images

Free money, no effort

[shy kelp]

well its probably more likely than doing real art 😂

[eternal fog]

These are the sort of people that get scammed lol

[thorn lion]

yo sytan when is SDXL coming out again?

🙂

[eternal fog]

Depends what art it is and who you have to buy it.

[livid basin]

18th? 19th?

[boreal bough]

I'm soooo happy I left IT in time XD don't think I could bear working through all the AI scams 🥲

[eternal fog]

It's going to be funny tomorrow when they delay it

[livid basin]

I wonder when the SDXL 1.0 weights are coming out, and if they will release the Style "photography/cinematic" weights as well

[pallid path]

AI scam?

[pallid path]

literally tomorrow apaprently

[eternal fog]

I mean that's only if you work in service desk. I did that for a few years, but it was internal so a slightly less amount of morons

[livid basin]

We can hope

[boreal bough]

#🌶｜off-topic if you wanna talk about it - but yeah, pretty bad

[pallid path]

they promised 18th 🤷‍♂️

[eternal fog]

They didn't promise

They said target

One of the devs was in here last night pointing that out

[high skiff]

Genuinely already wanna throw up

[pallid path]

well they will delay it if they have to

[shy kelp]

well any sort of visual art really, otherwise you'd need years of training, and in most cases you just end up doing industrial streamlines artstyles with little creative freedom

[high skiff]

Idiots who don't know what they are doing plastering bad info to the masses

[eternal fog]

Just wait for the Youtube videos

[placid edge]

as someone who actually goes out and shoots photos with a lot of kit, i wholeheartedly understand why people seek realism and photography from AI, doing releases for every single person if you intend to make money or even just practice is just crazy depending where you live

[civic sigil]

Didnt mcmonkey just say yesterday that they were releasing in a couple days along with the comfyui presentation layer

Its definitely coming on the 18th

[eternal fog]

He said target was 18th

[civic sigil]

Bro he was clearly implying they were set for the 18th

[shy kelp]

theres gonna be tons of img2img tutorials that dont acknowledge it being broken at this point

[soft zealot]

is this the latest version then?

(asking for a friend)

[pallid path]

they also said that they might do a delay and have like another beta model

[boreal bough]

tbf, their prompt guide is correct

no overkill tags, less commas, few on point negatives

[high skiff]

I have a potential solution for the broken img2img, but I need somebody with coding experience to fix a node for me ._.

[high skiff]

Second is the negative

[eternal fog]

I try not to use commas at all now, it's just an extra token to add noise into the prompt, it's rarely helpful.

[civic sigil]

Okay, they are going to release on the 18th. What type of last minute decision is gonna push them back?

[placid edge]

wont be surprised if it delays, not beholden to release before its ready 🙂 will be disappointed but not surprised

[shy kelp]

would be nice if they could update us on that before it releases so we dont hype ourselves up like little children on christmas eve

[boreal bough]

yep - just corrected

[eternal fog]

What's broken about it?

[eternal fog]

I think that was the point in his message, to stop people being daft and getting over hyped and then getting mad tomorrow

[high skiff]

It's a lot to explain, and a lot of guesses on my part, but I have had a pretty damn good track record with my guesses towards SDXL

[civic sigil]

What node

[high skiff]

I wonder if @uneven dove could help me work with it

[shy kelp]

saying something on discord is a poor way to get your message across

[high skiff]

It's a third party node

[eternal fog]

When I've used it, it seems to work well enough

[eternal fog]

Have you not seen SAIs messaging before lol. It's not brilliant.

[civic sigil]

My whole workflow is img2img so I'm curious

[high skiff]

Oh no, this new thing I am talking about is leaps and bounds beyond anything I have done with SD up until this point

[civic sigil]

Is it actually bad for sdxl

[thorn lion]

how is saying something in a public forum for discussion a poor way to get your message across?

[high skiff]

Also, just wanted to say that I had a staff member tell me my workflow is actually the right/intended way to use SDXL

[eternal fog]

I mean it is in their paper.

I think you just interpreted it properly.

[high skiff]

Which makes me feel pretty good

[boreal bough]

no kidding, this is a pretty damn good guide for prompts
I am amused by the one interrogator example they included - but rest is good

[eternal fog]

Care to elaborate?

[shy kelp]

theres no point using it right now other than experimentally.

[boreal bough]

guessing they just lurked in the bot section?

[high skiff]

No it's notttttt

That is a terrible way to write that prompt lmfao

[thorn lion]

oh boi........

[boreal bough]

so how would you tell a complete newbie to use it?

[pallid path]

yes tell us

[boreal bough]

genuinely curious

[civic sigil]

Well Sytan fix it before 1.0 drops lol

[high skiff]

The way I have it documented in my workflow, I included a hole multiple paragraph explanation as to how I recommend prompting everything

[high skiff]

I don't think I can

[boreal bough]

but that's for photorealistic only

[shy kelp]

depends what forum. if you post something in discord chat and expect enough people to clip it and share it, it won't reach very far

[trim orbit]

i get tensor size doesn't match nan error, how do i fix?

[eternal fog]

The way I try to write prompts, is imagine what these images would be called online. Filenames, how they'd be described.

Now an image of a person for example, when would anyone put "Detailed Beautiful Eyes" into an image like that.

[pallid path]

#🤝｜tech-support

[high skiff]

I need to fix a third party node to selectively upscale tiles of the image, then run a very low denoise pass on top of them, then pass that denoise pass on to a refiner sampler that only utilizes the third fourth of the sampling time scale

[high skiff]

Did I not show you how to prompt in pseudos server?

[eternal fog]

Is that the ultimate SD Upscale node?

[trim orbit]

all channels will become tech support

[high skiff]

Nevermind, I showed diodotos

[trim orbit]

it is inevitable /echos

[high skiff]

My prompting guide works consistently across all styles, it's just a formula

[boreal bough]

nope XD

[civic sigil]

So by img2img you mean refiner only img2img?

Refiner is kind of a dumb model Id rather use base

[high skiff]

Yes, image with base, denoise value, refiner on top as an img2img pass

[pallid path]

god this refining thing just complicates everything

[visual glade]

yes you can see it says: cudaMallocAsync

[pallid path]

I hope this becomes just a few sliders in Webuis

[high skiff]

It does, but the results speak for themselves when they work

[thorn lion]

I'm just curious if it was the only place it was stated. If so, sure. But if the message has multiple instances, i don't mind. it makes sense to me. it's on me to read the news and search for info, read the blog posts, threads, documentation, etc. Just my thoughts but I'm not sure how people live without search bars

[pallid path]

refiner strenght, refiner prompt, refiner steps, etc

[high skiff]

In my workflow for comfy, it's not even a slider, it's all setup for you, you just need 2 positives and a negative

[eternal fog]

@high skiff Are you just wanting to essentially add a proper way to do the mixed diffusion on the Ultimate SD Upscale Node?

Because I was looking at that as well.

[shy kelp]

some guy was talking about a quickfix solution on reddit but im not fully certain he understand what he's talking about or just pretends to

https://www.reddit.com/r/StableDiffusion/comments/14yggse/comment/js43436/?utm_source=share&utm_medium=web2x&context=3

[shy kelp]

well i think they should make a statement before tomorrow regardless, if they decide to delay it.

[rustic garnet]

I might try...?

[autumn forum]

Low key kinda wanna make a funny YouTube video when sdxl 1.0 comes out on how to set up and use it.. prolly won’t tho but I have some funny ideas running through my head😂

[thorn lion]

I think they should do whatever they want in retribution of some individual forcing them to make statements before they were ready by leaking 0.9

[eternal fog]

If this was true and he knows all about the code that's causing it. Why doesn't he fix it?

[thorn lion]

but that's probably just me

[trim orbit]

yeah he seems like a dunning kreuger kind of situation going on.

[high skiff]

They are right here, right on the nose

you either have 100% noise or 0% noise

[amber fulcrum]

Not what I wanted but looks cool 😄

[thorn lion]

no no no do it but do it on a 4 GB VRAM card and then rage about it not working and then insert a short clip of coming to discord to ask why

[eternal fog]

They are saying that it doesn't do that in the normal KSampler though. So surely there should be an img2img difference in using them. But I get the same results out of both for the same equivalent denoising rates

[placid edge]

a carnivorous rabbit 😮 you are twisted lol

[shy kelp]

so could you control the noise being applied better with a custom node?

[high skiff]

the problem is, I am not sure how to describe what I need to do with the node. The only person i know who can code my workflow is pseudo, as he was my partner in implementing it into the official diffusers workflows

[thorn lion]

and then the end of the video will just be you zooming in on wherever the recommended hardware/software is posted

[trim orbit]

big problem is people start using the base or the refiner wrong as they're trying to figure it out, and then they don't take responsibility. they blame the model and refuse to believe they're doing it wrong. so they doubel down on their workflow and make it "correct" with weird convoluted work arounds. always happens with new technology platforms. technically inclined newbs get in there thinking they're experts in the new domain, and they haven't even read the manpages yet

[thorn lion]

Olivio let me down..... 😦

[high skiff]

potentially

I have an idea for a workaround for it that could use just stock nodes, but it would likely be messy and lossy

[shy kelp]

i like potentially

[trim orbit]

ego and Bravada all over the place

[high skiff]

and would still have very little control

[eternal fog]

Surely @visual glade Is the best person to ask about this anyway

[civic sigil]

@high skiff Are you saying we should only use refiner for img2img? Or your fix applies to the base model too?

[uneven dove]

idk, img2img works fine for me

[shy kelp]

well i asked on element if anyone was looking into it but no answers

[high skiff]

my fix uses the base and the refiner on one image without using img2img, thats what the fix is

[boreal bough]

I'll make a megapost on reddit & civitai a few days later (prob sunday) - after making sure nothing changed about lora training
Showing how to do training - different results based on different training setups
• how to get around refiner issues when adding loras
Linking to working comfyui setups
Adding a bunch of prompts that work out of the box

and will try to keep it updated as things change

[uneven dove]

when you do it correctly even using img2img can have details preserved, it's a non-issue

[high skiff]

I am for sure, I just don't have the raw python knowledge to get it working myself

it should be noted that I am more focused on high res fix than straight up img2img though

[thorn lion]

ok ok so i'm supposed to use the refiner first, right? THEN the base?

(i'm trying to come up with the most absurd questions responsible so i don't get irrationally angry on release)

[civic sigil]

I hope they release refiner for 1.0 but I also hope it goes out of style with the community lol

[boreal bough]

and oh god I hope mcmonkeys grid is release by then - cause I'm not excited to stich together 200~400 images x_x

[high skiff]

same here haha

[uneven dove]

refiner 1.0 is "finished" according to emad twitter

[eternal fog]

Finished as in complete or finished as in gone?

[high skiff]

I wanna play with mcmonkeys grid addon big time

[uneven dove]

yes

[thorn lion]

is this the good infinity grid or the bad OOM xyz grid?

[high skiff]

who knows, maybe the refiner is gonna be outrageously good, we have seen nothing of the 1.0 refiner

[trim orbit]

i wouldn't call that reddit ranter right when he says the only potential fix is some magical solution that hasn't appeared yet

[high skiff]

its amazing

[boreal bough]

the good shit - with probably more features based on his excitement

[thorn lion]

WAIT we're getting a NEW NEW infinity grid?

brrrruuuuhhhhhhh

[trim orbit]

i'd say he fundamentally misunderstands things if he thinks only magic can solve it

[high skiff]

you can have a single grid that can automate 10's of variables, and its all stored in its own GUI that allows you to pick and directly compare images

[shy kelp]

well high res fix is actually what im trying to achieve aswell, but thats basically also img2img at a different resolution than the native resolution

[thorn lion]

magic is tech you don't get so makes sense

[thorn lion]

bruh this is my jesus

i have an 8640 image grid that just finished on runpod

[trim orbit]

bullshit artists, especially technically inclined ones, tend to weave a little bit of truth into their tapestries. gives it that truthiness feel

[thorn lion]

i love me the infinity grid

[high skiff]

if my understanding of how I got SDXL to where it is right now holds true, I am pretty sure I could get it working, I just need some proper brain power time

[uneven dove]

the only problem i was able to experience with img2img was when i tried refining the finished result

[shy kelp]

i can channel some to you with my crystal ball

[trim orbit]

i think asimov said that first, and it was more along the lines of technology that is sufficiently complex and unable to be understood is indistinguishable from magic.

I don't think solving noise problems is quite there yet. feels magical but it's just well made technology that is easy to grasp if you try

[high skiff]

I will say, my workflow works very closely off of his, and also is limited for the exact reason he states, and my workflow was good enough to make it into the new img2img code for diffusers, so I would say my guess hes onto something is pretty accurate

[civic sigil]

Yeah I've been wondering if that's what people mean when they say img2img is broken

[uneven dove]

it likely is

i don't know why the refiner can't do img2img outputs, it just can't lol

[civic sigil]

The base model seems to behave pretty normally for img2img in my tests

[uneven dove]

same

and SAI can't reproduce their issue either

[random hollow]

If I have the refiner enabled in "From image" in SD.next I keep getting "'Tensor' object has no attribute 'astype'" error. Error goeas away if I disable the refiner. Also refiner works perfectly in "From Text" .. anyone had that issue ?

[boreal bough]

I was shilling for it in this server - until mcmonkey tagged me. then I posted HIM the link to my favorite addon - the infinity grid... and saw that url I was linking was his own damn domain XD

[civic sigil]

Refiner is probably fine by itself for img2img with low denoising strength too I guess

[thorn lion]

I would sacrifice controlnet for the infinity grid

[boreal bough]

same

[shy kelp]

yeah so you're using the base model, which you can get to preserve SOME detail with very low denoise. But you are basically not getting the full SDXL quality out of img2img no matter what

[uneven dove]

it's actually great, it can improve the quality but it doesn't change the img

[trim orbit]

he's parotting the same research that's been going around. He has some right to his words. but he says things are unsolveable currently.

Yet people are doing img2img gens very well. there's a ton of experimentation going on. Perhaps he's overstating the problem.

[uneven dove]

puh-leese. i can use up to like .8 strength img2img

and it looks great

[thorn lion]

"is quite there yet" = "is sufficiently complex to warrent 'magical' techniques"?

[trim orbit]

the post seems to illustrate things as unsolveable

[shy kelp]

well maybe you're not doing images that are supposed to have fine details like skin texture in the first place, in which case it could work. But you're still missing a lot of potential quality increases by not using the refiner (which dosent work for img2img)

[uneven dove]

yes i am

input -> output

@shy kelp instead of calling me a clown, quit proving to the world that you are one

[shy kelp]

woah you turned shit into more shit*

[uneven dove]

tired of your stuff, welcome to the block list, child

[shy kelp]

👏

[eternal fog]

Hmm, the Ksampler and Ksampler Advanced nodes do have different results for Img2img, even with the same supposed denoise values

[civic sigil]

It's kind of annoying you cant end at step 0 in comfy

[trim orbit]

end at the beginning?

[thorn lion]

beginning at the end

[boreal bough]

if you mute the tab, does it still pass it on?

[civic sigil]

Like just to pass through when you dont want to do double prompts

[civic sigil]

Ill try

[trim orbit]

end of time and space beginning of every end

[thorn lion]

ouroboros

[shy kelp]

have fun ruining all your good SDXL images, kid

[civic sigil]

Yeah no

[uneven dove]

the danny devito transformation above was after putting the input through img2img, eight times

it doesn't nuke details, some people just don't know what they're doing

[visual glade]

this is a bit wrong the amount of noise returned if you enabled return leftover noise is the exact amount of noise the next step expects

[eternal fog]

@high skiff Confirm I'm not being stupid here.

If I have 20 steps 0.600 denoise is the same as starting at step 8 isn't it.

It's the same number of "denoise"

[visual glade]

this is why if you chain samplers it actually works

[uneven dove]

denoise at .8 isn't the same as starting from a given step

[shy kelp]

bro of course it dosent remove detail at .8 😂 you're basically regening the same image over and over and probably didnt use that much refiner strength for your first pass to begin with 😂

[uneven dove]

denoise adds noise, starting from a given step simply continues with the residual noise

[amber fulcrum]

Is there any img2img workflow for comfyUI (I mean JSON or image i can "borrow")?

[eternal fog]

What does the denoise value do then? Does it just add a different amount of noise?

But the starting at step add's some set amount?

[civic sigil]

Yeah but if you have add noise enabled

[eternal fog]

If that's the case shouldn't the KSampler Advanced also have a denoise slider so you can control the amount of noise

[uneven dove]

well,guess i don't know what i'm talking about lol

[boreal bough]

sytans workflow -> https://github.com/SytanSD/Sytan-SDXL-ComfyUI
or the one I linked is from mimizukari

[shy kelp]

it does, start at step is basically that

[civic sigil]

Thats what "add noise" does in comfy Im pretty sure

Adds noise based on the steps you have left or whatever

[thorn lion]

I will always remember the person that drove back to the store, bought new computer parts, went back home, did the hardware install, and it still wasn't working. Then I came, looked at it, and pressed the power button. When I'm working out problems, I usually narrate. They asked me what I just did. They heard me JUST say, "hmm I wonder if the button was pressed" and when I asked for context about their response, they said they didn't know they had to. At which point, I got mad. Now, I just don't help anymore. Not else you can prove that you have basic literacy and the ability to hold dialectic opinions

[autumn forum]

[eternal fog]

But how much noise does it add. What pseudo is saying makes sense.

There's a difference between run through all 20 steps, but just add less noise and only start adding noise at step 8 but a whole bunch of it.

[boreal bough]

you can use either of those two as a start for whatever project you wanna pursue in comfy

[uneven dove]

@thorn lion the best is "how do i do that?" and you have to direct them toward the glowing button

[amber fulcrum]

thank you - appreciate your help

[thorn lion]

OMG accelerate config

.....

[uneven dove]

@eternal fog at 1.0 denoise it adds 100% noise

the image is basically ignored

[thorn lion]

how do you get confused on accelerate config?

[civic sigil]

I would assume it adds noise depending on the step / total steps ratio like you were saying earlier

For the first step

[uneven dove]

10 steps, .8 denoise = 8 steps of noise added

[eternal fog]

Yeah, but you don't have that silder in the advanced sampler in comfy, you only have the start on steps, so you have no control over the amount of noise it seems.

I'm going to try add the noise silder into the node and see what it does.

[uneven dove]

steps are just an abstraction layer on top of the sampler's internal concepts

the start on step will likely calculate the amount of noise by that

you have zero reason to add more noise than you intend on scheduling

[boreal bough]

tbf, when I started out I managed it wrong twice in a row 🤣

[civic sigil]

I thought it was hip to return slightly noisy latents into our decoders now lol

[thorn lion]

things get abstracted to get simplified but people want less abstraction with more granular control AND for it to be easy, simple, and pretty

[uneven dove]

@civic sigil a field of "active research"

[shy kelp]

I think, and its just a guess, that the refiner only works well on certain noise patterns leftover by an image generated from 100% noise. If you just img2img by adding x% noise instead of 100% noise, it drastically decreases in quality because that does not make the same noise patterns (leftover by the base model) that it was trained on.

[boreal bough]

only if you operate in the fractional space between dimensions.
the twilight zone

[uneven dove]

the timestep zone

[boreal bough]

but it sounds more fun!

[autumn forum]

That’s actually an interesting thought. I could see that being the case. But who knows🤷‍♂️

[visual glade]

--cuda-malloc should now be enabled by default if you have torch2.0 and up installed

[shy kelp]

this also holds up with, the more noise you add for img2img, the better it gets. if you do 0.8 noise it's way better at doing its thing than 0.2

[boreal bough]

also very very dependent on the subjects in the image, and the supplied prompt
for photoreal its a lot more proactive - than for some artstyles where it does only the bare minimum of making lines neater and adding contrast

[thorn lion]

also same. I wasn't confused but it didn't work the way I wanted and it was cause I didn't understand fp16/bf16/fp8 and why they exist. but then I googled the same question people came to places to ask and found that I wasn't the only one that misunderstood. Then I learned from someone else asking the question and someone else performing the labour of answering it because the internet allows information and discussions to be preserved. then I tried again and got it right.

TL;DR try, think, search, read, interpret, try again

[eternal fog]

@uneven dove I added the denoise silder to the Advanced Node along with the start as steps and it absolutely makes a difference.

@visual glade Does the Ksampler Advanced, just add a set amount of noise at the starting step when you start from a later step?

[thorn lion]

but the ones i'm talking about are the ones who don't ask a direct question. they say "it's not working. what do i do?"

[upbeat summit]

Does anybody know if there is a ComfyUI keyboard shortcut that cancels the current running prompt?

[visual glade]

if you have add noise enabled yes

[eternal fog]

So you don't have control over the amount of noise it adds then?

[visual glade]

that's controlled by the step

[thorn lion]

and, like, it tells you.....if you have a more developed question like "why pick bf16 over fp16?" that's a different story. it's a basic foundational knowledge type nugget but it's new enough that it could spawn questions.

[boreal bough]

under running, there should be an X for you to cancel it

you can enable that by clicking 'see queue'

[visual glade]

sigma is tied to the step which is tied to the amount of noise added

[shy kelp]

@visual glade is there anything in the plans for making the refiner better for img2img/hires fix?

[uneven dove]

you'd have to document an issue with img2img first

which, you haven't

[shy kelp]

🤡 💩

[uneven dove]

there's some parts of this core architecture where there's no "fixing" it, it's just something that you were never meant to do

[high skiff]

No, that's not how that works

[shy kelp]

bro just block me already omg

[high skiff]

The denoise is the amount of noise it's added on top of the image, so where 1.0 would be pure gaussian noise, 0.6 is 60% of that total gaussian noise on top of the base image

It still does all of the steps from the smallest sigma to the highest sigma, or the lowest frequency details to the highest frequency

[uneven dove]

i remember when they tried perlin noise in the thing for a day or two, i don't think i ever saw that brought up again lol

[eternal fog]

I think their might be benefit to allow users to change it seperately. This is an image of a Cat, starting at step 2 and ending at 20. If I run this normally I get this man.

But if I manually add in a denoise silder to it and reduce the amount of noise it supplies I can get a proper cat and I can't get that exact image otherwise.

[upbeat summit]

Thanks, Caith!
I meant a keyboard+shortcut. Because right now in my 1920x1024 workflow, which obviously delivers lots of bad ones, I go through a lot of images. I keep canceling the generation as soon as I see I will not like how it is converging.

It would be way faster with a keyboard shortcut instead of clicking 😉

[shy kelp]

literally made a reddit post with examples showcasing exactly how it dosent work https://www.reddit.com/r/StableDiffusion/comments/14yggse/sdxl_09_currently_does_not_work_particularly_well/

and your own examples only confirms it

[thorn lion]

yknow

i find it interesting

that someone makes so much noise about something NOT working

on 0.9

but we don't have 1.0

I wonder. I WONDER

if they fixed it and didn't tell you

[eternal fog]

This isn't to do with the model though @thorn lion It's to do with how the inference is being done I think.

[uneven dove]

they've also asked that once an issue is acknowledged that it's kind of, moved on from 😄 and not to rehash it repeatedly, because it is not useful

[civic sigil]

I dont think it would be an issue with the model weights themself but the inference code

[thorn lion]

yup but there are certain people accountable to SAI, right? it's not out of distribution to consider them employing NDAs after the leak

[civic sigil]

Oh Arron already said that lol

[thorn lion]

or just courtesy from the developers like comfy

[shy kelp]

time will tell... but at least i could potentially have made them aware of the issue if they weren't. or made other people aware of it in case there's no fix tomorrow, so they don't go get themselves frustrated trying to get it working as well as well as it does for SD1.5

[winter raptor]

Wow, this is amazing

[thorn lion]

i bet comfy knows SO MUCH that we don't even consider

would never press them for it though

[uneven dove]

i would

points to the meat grinder

get in, boy

we will extract that knowledge

[soft zealot]

ah so a single pass rather than a double pass as per your 0.5 release

[civic sigil]

Is there any way to simply add noise to a latent image in comfyui?

[thorn lion]

well, I see what you're saying and you make a good point there. But I'm gonna bet that one opinion out of (anybody have stats on people using gen-ai?), or even from the perspective of one specific SDXL beta user amongst many, given that many many people try to make use of the pipelines and workflows and techniques that you are trying to employ, it is highly likely that they have much more information in regards to the issue that you are currently able to provide

[high skiff]

no, otherwise I would do that and would have solved this issue by now haha

[eternal fog]

I'm going to experiment with a denoise silder on my KSampler (Advanced) you can get some different results with it

[civic sigil]

Seems like it would be simple to do

[shy kelp]

probably. comfy is a community project though, and the solution might come from some community member if comfy is preoccupied with other things.

[upbeat summit]

I've been fine tuning this prompt all day - it was fun.
portrait of a battered defeated humanoid robot made out of silver metal standing on a hill overlooking the ruins of a destroyed urban city, from behind, golden hour, dystopian retro futuristic, natural light photo, Canon 85L f4.8, ISO320, 5000K colour balance, (pulp art by Robert Mcginnis:0.9) and (pixar:0.7)

and SDXL interpreted it really well! And many thanks to @tender timber . Got me inspired with the prompt he kindly shared.

[thorn lion]

"preoccupied with other things" could be fixing the thing and not being allowed to tell you about it

[soft zealot]

I fdont know if this helps but WAS Suite has a Latent Noise Injection Node

[high skiff]

this si how my workflow works

Purple is the base, green is the refiner

My workflow has the base handle the low tm medium frequency details, and the refiner handle the upper medium to high frequency details

[winter raptor]

Wow that’s an awesome prompt. I’m still waiting with bated breath for you to release your node setup and workflow one day lol

[trim orbit]

the comfy project has winamp level energy to me. winamp was such a feature rich professional grade audio player when it dropped for free. breaking open the capabiliteis of audio playback for all PC users

[nimble heart]

PSA imagemagick can create nice grids of automatic or N size with one command. pretty sure they have windows downloads you can just add to PATH and use in powershell

[high skiff]

if you do img2img, it spreads all of the steps across the lowerst to highest frequency detail

[trim orbit]

comfy really whips the llama's ass

[high skiff]

it does help, though I can only use stock nodes for my workflow on the official comfy wiki

I had WAS on my old install

[shy kelp]

cool cool i hope you are right

[nimble heart]

sorry but what county did you learn that from? I've heard many a redneck saying in my life and that's a new one

[boreal bough]

painful, but a good alternate solution 👍

the county of winamp!

[trim orbit]

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

[upbeat summit]

You have been posting some really good stuff yourself! This is an amazing image #✨｜sdxl message

[shy kelp]

auto1111 released their official SDXL support a few days ago tho, with no hi-res fix... i wonder why......

[thorn lion]

and, if i remember correctly, community projects have a very well established protocol for being worked on, which include pull requests, which i've understood you think is clown doo doo.......

so start your own with your own rules and guidelines and shut up
OR
be more amicable in your dealings with the community that allows passionately cares about the community project. whether you wanna make a linkedin profile pic or roop your face into hentai, we are all here for the same essential reason

because major members of this community are working on something we want.

Why would you....be rude.....in that situation? or disagree disrespectfully? Or be a conceited ass?

Am I calling you any of these things? idk. I forgot I was talking about you and started venting general frustrations

[soft zealot]

technically didnt someone release anss ADXL plugin to use with A1111 rather than A1111 doing it directly

[thorn lion]

someone released a comfy wrapper to load comfy in a1111

that gave me a chuckle

[nimble heart]

send the result to img2img, that's basically what their high res fix does anyways

[shy kelp]

oh? if that's the case, my bad

[civic sigil]

No reason not to use A1111 now

[winter raptor]

Do you have any recommendations for achieving a specific art style in SDXL? Like say there is a picture and I really like it’s style (whether it be a specific drawing style, photography style, etc). How would you go about trying to achieve it?

[boreal bough]

loving it ❤️

[thorn lion]

half a story. I remember seeing the comfy release and SAI staff being like, "we talked to A1111 but" (in my words) they couldn't understand why he was loafting.

[nimble heart]

auto bloating to the point of containing all of comfyui inside it is something else

[thorn lion]

like

if we're gonna do THE THING, at least be comprehensive. don't selectively include details

[nimble heart]

becoming the Emacs of diffusion UIs

[high skiff]

Because auto doesn't even have my mixed diffusion, let alone proper noise injection for img2img or high res fix

[shy kelp]

i really like that you can load up comfy so fast and work on your prompt while the model is loading... i hate that i cant do that in auto1111 (not without changing the code at least)

[civic sigil]

True

[thorn lion]

that was when i knew i had to switch tbh

[nimble heart]

I still have it for inpainting and the openpose editors tbf

[civic sigil]

Comfyui is really forcing me to learn python better cause of having to make a million custom nodes lol

[eternal fog]

@high skiff What do you think about this. It's not perfect because I'm messing around. But I put a denoise silder on the Advanced KSampler so you can control the amount of noise it applies more.

Left is what it would normally do, right is more steps allowed, but I've restricted the amount of noise it can use

[nimble heart]

also Sytan did you even sleep

[eternal fog]

It's changed the look but I did that on purpose to make it pronounced

[high skiff]

Nope

[high skiff]

Might need to test this in dM's potentially

[thorn lion]

it was great that the extensions ecosystem was so active but it wasn't managed. now, i'm not a big fan of regulation or someone telling me what to do in general but it is clear that we need some modifications to systems that are centered around deregulation or self-correcting communities

[high skiff]

I am sick, I woke up at 8PM

[nimble heart]

you have the circadian rhythm of a vampire

[high skiff]

It's 12:42 pm

[eternal fog]

Can if you want. I need to get some food first though.

[high skiff]

It's been established lol

I have diagnosed nocturnal sleep alignment disorder lol

[thorn lion]

well, I've been working on the raaaailroad

[shy kelp]

looks like its really messing up anatomy, but the detail on the armor is more fine

[thorn lion]

sorry i've been working on trainings recently with minimal generations and i still have to learn comfy

[high skiff]

Sounds good, I'm waiting for my 3090, impatiently lol

[eternal fog]

I cranked it up really far to get an obvious change

[thorn lion]

so i don't know what comfy CAN'T do yet or what is more my style to do in A1111

[thorn lion]

still out for delivery?

[civic sigil]

Load images from a folder

[nimble heart]

I don't think the extensions are the issue I think the main repo adds features faster than it fixes broken ones

[winter raptor]

I’ve really gotta learn what these terms and techniques are, like noise injection. Any recommendations on where to start learning about that stuff?

[nimble heart]

some of the stuff in the main repo should just be its own official extension, like everything in the Extras tab

[thorn lion]

okay so just....don't update once you have a version that does what you want?

[thorn lion]

but also this. i did like that comfy gives you a big blank canvas versus 9287346578923645 settings in A1111 GUI

(i assume. never opened comfy but i am remembering a screenshot)

[soft zealot]

I like the far superior memory handling

the first two humps are 4 images in SD.15 in COmfy and note how the VRAM clears

Then 4 in A1111 and you can see the VRAM is kept in use even when finished

The 4 using COmfy & SDXL

First hump is higher due to the VRAM left in use by A1111 and then comfy released it

[winter raptor]

I’m loving the skin detail on the right

[upbeat summit]

Here's a totally unnecessary 7680x4096 (4x1920x1024) test image using 4x-UniScaleV2_Moderate.pth for a quick and dirty upscale (the eyes, background depth of field not that great).

I also tried to give the prompt more realism in this variation.

[nimble heart]

really the only 'settings' comfy has are cli flags for optimizations and setting input/output directories. everything else is just done with nodes

[thorn lion]

OH SO I'M NOT CRAZY okay thanks you

[high skiff]

My GitHub hosts my workflow that features mixed diffusion.

It's at workflow where instead of generating a full image with the base model and then doing an image to image pass for the refiner to fix the flaws, it instead only does part of the base image with the base model, and then passes the incomplete image to the refiner to continue from where it left off.

It takes less time to do, and produces better results on average

It ended up working well enough that diffusers has implemented it as their go to image to image workflow in their pipeline now

[thorn lion]

sorry. settings and parameters

so that includes things you can set on the nodes but comfy is just prettier

[nimble heart]

according to comfy, comfy's default is equivalent to auto's --med-vram

[thorn lion]

and I play Satisfactory so spaghetti doesn't scare me

[paper phoenix]

yeah not letting go of ram is a serious a1111 issue. especially when you can count your ram on one hand 😉

[winter raptor]

It’s always the eyes with SDXL but I am loving the detail in this image

[soft zealot]

which is fine as thats what I ran A1111 at however COmfy releases the VRAM after generating whereas A1111 doesnt

[winter raptor]

Awesome, and how about things like noise injection? Do you play around with the noise levels when prompting, steps, etc, or just focus on the prompt?

[thorn lion]

this ELI5 aligns with my intuition after reading the paper. thank you for your contributions

[soft zealot]

and just runs quicker generally 🙂

[paper phoenix]

for a while early on i thought wonky eyes was a deliberate overt watermark on AI images!

[nimble heart]

Back in the 1.4 days I used to have to restart the server every few minutes as my 1070 would gradually run out of VRAM lol

think even now they're even adding a button in the UI to restart itself for similar reasons

[thorn lion]

i still occasionally have to restart the server after switching models and resolutions a few times

[paper phoenix]

a flush handle, like a toilet. 😄

[high skiff]

I do have an experimental workflow that I call fractional offset, but utilizes the fractional misalignments between 21 and 20 steps in order to inject a little bit more detail

[soft zealot]

lol

DOwnside howver is that COmfy does like to use system RAM to chche data in between runs , which again is fine as I have 64Gb of that and at least trelativly cheap & easy to upgrade but could be an issue for people with less than 32Gb (potentially)

[nimble heart]

yea when I first got my 24 gig card I made a big grid with all the 1.5 tunes I had and ran out of VRAM...

[thorn lion]

A1111? it already exists. In settings -> actions. it would purge the model but it didn't fix the issue and i dunno why

[thorn lion]

lol mcmonkey infinity grid ftw

no oom on grids

[thorn lion]

specifically 20 and 21 as opposed to.......actually, there is stuff i need to educamate myself on that you've already typed out. i'll go look for that

[soft zealot]

some people may say it because the author "borrowed" a signicant portion of the code however I have no conclusive evidence or proff of that

[thorn lion]

pssst hey look at this code I "borrowed"

torch.cuda.empty_cache()

-.-

[nimble heart]

data caching is super nice though. On SDXL for example if you change something on just the refiner it won't re-run the base again.
And on my hi res fix if I want to save the low res I can just wire up a save image after its done and not have to re-gen the low res latents

[thorn lion]

if you're gonna give me a story, give me the whole story. all perspectives

[misty sequoia]

So, does A1111 just suck now?

[soft zealot]

orthogonal , isometric, forced

[nimble heart]

does that actually fix the auto issues or is that a troll lol

[misty sequoia]

Or am I completely missing the point of the convo?

[paper phoenix]

people are overly mean tbh.

its my preferred UI, but it lacks in some areas compared to others

[civic sigil]

They didnt give them early access to the weights and are now criticizing them for not supporting XL, there is a very obvious agenda here

[nimble heart]

okay they do actually call it sometimes lol

[soft zealot]

reminds me of th ebad old days of creating a RAM cache to run windows in lol

[thorn lion]

NOTE WELL THAT I AM GUESSING
like, just unload model, clear cache, close torch, open torch, load model SHOULD fix the issue because that SHOULD clear the vram but if the issue is due to something other than VRAM not being garbage collected properly (which was what my understanding was based on what I read)

[paper phoenix]

https://media.discordapp.net/attachments/1008814691079958548/1119632550302847069/cuda-kirk.gif

[boreal bough]

right now, for sdxl specifically - yes
in all other instances, no

also this statement will probably change sooner rather than later

[misty sequoia]

Did they give ComfyUI access to the weights?

[paper phoenix]

🙂

[civic sigil]

Yes very early on

[thorn lion]

but, like, if there is something deeper, well, i'm not sure. but there was something else I was doing and staring at my nvitop and the shape was the same as the memory bleed i got from A1111 so I did what I said above and it, at least, stopped being so serious

[civic sigil]

And money lol

[glad fulcrum]

that's why sdxl works faster for me in AUTO1111, in COFMFY UI, it has to load the model every generation

[thorn lion]

i'm not trolling. I honestly feel like some people are trolling when they don't know some things but no, i am being serious and I think I am right and I am prepared to be corrected

[nimble heart]

Not really old, I have my browser download directory set to a 48 gigabyte ramdisk

[misty sequoia]

Do you know why they supported Comfy but not A1111?

[civic sigil]

Idk honestly

[rustic temple]

comfy himself is stability staff

[thorn lion]

they were in conversation with A1111 and Comfy around the same time. Comfy did the thing. so everyone used sdxl with them. A1111 only just did. idk why. that's all i've observed.

[thorn lion]

was he staff before he was staff

like, was he staff before ComfyUI? or was ComfyUI the thing that got them to hired them?

[vale eagle]

First Image2Image Workflow Attempt.

[nimble heart]

interestingly the empty cache is only in unload_model_weights() which a cursory glance seems like xyz_grid.py doesn't even call

[misty sequoia]

So then will we see A1111 be as optimized with SDXL compared to Comfy?

[thorn lion]

i genuinely laughed. was just caught up on the train of thought

[misty sequoia]

Because from what I've heard, A1111 is really vram heavy compared to Comfy with sdxl

[thorn lion]

ask A

[high skiff]

The misalignment between 21 and 20 just happens to coincide with the amount of steps that I normally generate at, technically you could do all sorts of different misalignments, that's just the one that works for me

[rustic temple]

I'm pretty sure he was hired sometime after starting that project but I mean he's in here so we can just ask lol

[thorn lion]

i'm just regurgitating information that has been publically shared. my abilities end where my memory does

[visual glade]

it was implemented in comfyui before because that's what we use to test the model

[sudden cliff]

In this particular example, it looks like the noise you're adding is lower resolution than the diffusion, which would result in that big blotchy noise you're seeing.

But I'm kind of curious, are you saying the flow of base model complete denoise -> add noise -> refiner denoise always results in some defect similar to that, or only for upscaling?

I tried flows that did the ''''proper'''' step stop -> finish denoising in refiner and didn't like how it lost fine detail coherence (for example, resulting in kimonos made of tiny bathroom tiles etc)