✨｜sdxl

[static prawn]

mh i might try, i just gave up with sdxl for the moment but i can take a look for u soon

[dense chasm]

you should pay your attension to another profession

[high skiff]

[boreal bough]

XD

[high skiff]

FINAL FANTAIVASY

but squeeshed

[eternal fog]

I think it's tried to merge the N and T

[soft bone]

2.12s/it for 3090 sound right to yall? on sde gpu

[eternal fog]

Which SDE

2M or the other one?

And actually 2s/it that's slow if you are doing 1024x1024

I get 2it/s on a 3080 at that res

[dense chasm]

i have some charm in some extent actually,huh

[soft bone]

just sde gpu and karras, 872x1128

[eternal fog]

That seems slow to me

[high skiff]

about, yeah

[soft bone]

thats without refiner btw

[high skiff]

SDE GPU is slow

[eternal fog]

2M is faster

[high skiff]

let me check what I get

[dense chasm]

it's good to know,but the base model define the majority, i tested 1.5 and 2.0 plus either

[eternal fog]

Which is what I use, about double the speed

[soft bone]

86 seconds for a batch of 4 just hurts

[eternal fog]

Oh that speed is at batch 4?

[high skiff]

thats the one that failed my aesthetics tests in this server horrifically 😅

[eternal fog]

Been working well for me

[visual glade]

2.29it/s shoud be what dpmpp SDE gives on a 3090

[eternal fog]

I think they forgot to mention they are doing batch 4

Which would explain the lower speed

[soft bone]

indeed

[hidden cobalt]

is there a tutorial for installing it into the a1111 webui now that it's merged?

(I have it running on h100 server but I couldn't find the instructions for actually adding SDXL model)

[high skiff]

you are using a pretty slow sampler, and likely also excessive steps

[soft bone]

just using the same i've used on 2.1 for months

[dense chasm]

you can never change a human being's nature,otherwise Elon will not provoke and Muck will never respond for a cage fight, the only thing will change forever is being conquered or with love , which one will you choose

[high skiff]

ok, what the hell are you on about lmao

like it was funny at first, now I think you are just actively trolling

[soft bone]

#🌶｜off-topic

[dense chasm]

the bots or robots will not make a decison emotionally nowadays, so we have a chance now still

[boreal bough]

if this turns out to be an elaborate hoax about us interacting with an automated instance of gpt-4, I'll be mad XD

[eternal fog]

It's started doing this sort of stuff now lol

[boreal bough]

pixel go brrrrr

[eternal fog]

I put "16-bit Sega" as a style

[uneven dove]

that's 12-bit DOS

YET ANOTHER FAIL

[eternal fog]

Yeah some of the style words do things completely different lol

[dense chasm]

find your true love ,that's the reason human survive

[boreal bough]

while I can always get the style I want, it's rarely using the words I'd expect. Half the time I ask Vit-H, and get weird names or styles - which then work perfectly, but why not the ones that should work???

[eternal fog]

Yeah it can be a little annoying sometimes trying to get certain styles

I've noticed that you sometimes need to either crank up the CFG, or heavily weight the prompt

[autumn forum]

whats the easiest way to ask Vit-H?

[eternal fog]

Pixel Art works though

[uneven dove]

so without doing anything fancy to share VAEs or Text encoders between the models,

• no CPU offload = 25GB VRAM, OOM on a 4090
• model CPU offload = 16GB VRAM, works great, adds about 10 seconds to the image gen
• sequential CPU offload = 8GB VRAM, adds 20 seconds to the image gen time

this is on a 5800X3D, the numbers are notably worse for a Xeon Gold

[boreal bough]

I have interrogator running in my old 1.5 webui

[autumn forum]

gotcha

[uneven dove]

sequential CPU offload with attention slicing and VAE tiling and VAE slicing = 90 seconds to gen an image, 2GiB of VRAM used

that 5800X3D, man

what a beast

[eternal fog]

Accidentally made a crap quality Mario level

[boreal bough]

sounds like a fun but terrible idea
gen image -> automate into mario level -> suffer XD

[eternal fog]

I see Minecraft still works as a style

I always struggled to get space helmets to work properly

Seems to actually put the head inside the helmet now

[uneven dove]

so blurry

0.9 ?

[eternal fog]

Yeah

[uneven dove]

it'll still gladly screw up helmet visors

[eternal fog]

It's not that blurry

[uneven dove]

just seems to be luck of the draw

it's so blurry my eyes are like "what the hell" lmao

[autumn forum]

best ive got of space helmets, it doesnt do a good job at those at all lol

[lucid raft]

a young guy walking through the stromy road as a part of a life.

[uneven dove]

stromy?

[sonic void]

anyone konwn how to fix it?

[uneven dove]

i hate to nit-pick but everything this guy does is a part of life

[boreal bough]

so this is just in hindsight... but could it be that a lot of people cant get training to work, while some can on the same amount of vram - be due to their card not having bf16 support?
It would explain the huge different in vram between different people

[eternal fog]

You need to install those nodes

[uneven dove]

nope, mcmonkey ran the test on a 2070

[soft bone]

why tf does it switch between it/s and s/it?

[uneven dove]

3080 and 3090 are the first ones with bf16 hardware support for the RTX series.

[boreal bough]

mcmonkey did it in fp16 on a version of sdxl where that wasnt broken yet, so he doesnt count XD

[eternal fog]

Because sometimes you are doing more than 1 iteration per second and sometimes it takes multiple seconds to do 1 iteration

[sonic void]

do u have a node source link?

[dense chasm]

Interstellar is expotentially hardness for human beings

[soft bone]

so use 0.xx

[uneven dove]

yeah i love how he's always comparing to us on old old versions "bcuz its what's lying around"

[eternal fog]

No idea what those nodes are, try googling them

[lucid raft]

@uneven dove dusty i mean !

[uneven dove]

too late

it's stromy now

[sonic void]

[lucid raft]

Can i get a another one.

[uneven dove]

you have two images remaining

[eternal fog]

lmao this is justmaking weird stuff now

[boreal bough]

although fp16 training may be a thing again, as of today again

[uneven dove]

no lol

the bf16 was needed for OpenCLIP

you get 'numeric instability' in the text enc

[boreal bough]

oh. oh well then 🥲

[uneven dove]

you can load the vae in fp16 tho now

[lucid raft]

I mean this kind of dusty.@uneven dove

[uneven dove]

ask for "burning man", then

[dense chasm]

what is four-dimensional space,time plus geographic,human beings survive is for love priority,if love is no longer exist,it's the time robot will tak prominent

[lucid raft]

@uneven dove sorry for the misunderstanding. Will you ?

[uneven dove]

four-dimension, like an altruistic version of measuring from zero rather than one? when in four-dimensional space, the time is embedded in coordinates, they are inseparable, you know, this why it's called space-time continuum no beginning or end

[autumn forum]

i do like the movie interstellar, one of my all time favorites

[uneven dove]

you have one image remaining

[lucid raft]

Can you get a well dressed guy moving away like the previous one. @uneven dove

[dense chasm]

u should ask for the aliens if they are exists,idk,four dimensions is all we can understand,but how we can grasp the fundemantal of more dimensionally meanings the existing deep leanring strategy,how knows in future

[uneven dove]

dimension is just a way to move

Length. Width. Height. these are dimensions

time, another one

[eternal fog]

Hmm I do get a lot of blurry images sometimes. It's weird, you'll have a bunch that are blurry and then they aren't again

[lucid raft]

Why ?? @uneven dove

[uneven dove]

sometimes it works sometimes it doesn't

[dense chasm]

move forward,who can tell, That's one small step for man, one giant leap for mankind,To see a world or the galaxy in a grain of sand

[uneven dove]

you know pale blue dot ?

[lucid raft]

@dense chasm Go fucking watch rick and morty together guys.
Fuck off !!
////////////////_____________

[uneven dove]

you have no more image credits

[lucid raft]

Look at the image you send me.

[uneven dove]

Frank still has a ton

[dense chasm]

can u tell me what's the implication of Prometheus

[uneven dove]

@dense chasm that's a cartoon from 90s? Prometheus and Bob?

Claymation

[dense chasm]

u can't aware of i'm currenly chat with u thr proxy agent across half of the world,lol..

[autumn forum]

using " shot on iphone" seems to work well at getting some of the blur out, also use photographer names that dont have any blur.

[uneven dove]

shadowsocks?

[dense chasm]

universy exists for almost biliions of years, we are negligible, dinosaurs disppeare as well

[dense chasm]

nope, chinese exists for almost over 5 thousand years lol..

[eternal fog]

I tried to use Shrek in a prompt and it's done him dirty lmao

Big ass monster

I'm not sure if this is much better though

[stray mantle]

SDXL 0.9 ComfyUI

[dense chasm]

where is Caribbean pirate,captain,huh

[autumn forum]

sometimes i wonder why so many people want sdxl to work in a1111 so bad. like i get it, i started out in a1111 but at least give a thought to using comfy. 😅 ignore me tho i just wanted to get that out of my system

[azure oxide]

Just don't mind what others prefer or do and focus on yourself, applies to all aspects of life tbh

[civic sigil]

Why not, it has tons of extensions and people already have it built into their workflows

[dense chasm]

forget about it,u should come to China for a visit, the chinese is awesome almost like the majority of American ppl, and you will fell in love with chinese food huh

[uneven dove]

it's called beating a dead horse

[civic sigil]

Personally I had already done the work of writing custom code for A1111 so I would have liked to see support earlier

[uneven dove]

ah yes, the sunk cost fallacy

[boreal bough]

the trick is to tell them to use the bot XD 50% of the time that is actually what they wanted. an easy and intuitive, no hassle, way of using sdxl

[civic sigil]

Who cares about the codebase dude all that matters is the actual ML programs

[boreal bough]

the other 50% are booba in your basement

[eternal fog]

Yeah but then they can't make their weird nsfw stuff

[uneven dove]

A1111 isn't even a good experience for users

so that argument falls on its face, too

[civic sigil]

See my previous response

[eternal fog]

They don't know that though. They press button, pretty picture comes out.

[uneven dove]

"who cares about the codebase" you just explained why you do lmfao

[uneven dove]

this

[civic sigil]

?

[autumn forum]

i understand theres a workflow aspect there, i get it. i guess for my workflow it just is easier to use comfy. may not be for others.

[uneven dove]

InvokeAI and Vlad's fork have SDXL support. will you switch? no, because you care about the codebase

[civic sigil]

I moved my workflow over to comfy actually

[dense chasm]

agreed, A1111 connection errors sometimes will pop up

[civic sigil]

But it took days

That's not a fallacy

It's actual time I could have spent doing something else

[boreal bough]

time to find out if samples are finally working in kohya x_x

[uneven dove]

it's just that your own reasons don't align with what you accuse others of.

[boreal bough]

technically, they should be

[civic sigil]

How so

[uneven dove]

i was using Diffusers on day one, not a whole lot of thought or work went into it. but you're wanting A1111 to work, because you "care about the codebase". but then, when we point out that keeping A1111 alive is "beating a dead horse", you say "who cares about the codebase?" - even though, you do. you stated you do. switching to comfyUI i guess you can bring that up if you want to muddy the waters, but the discussion started out as one thing, and that's what i'm responding to

[soft bone]

base only

[civic sigil]

I don't care about the quality of the codebase I just care about having a shared platform for people to write code for and tie in their existing image manipulation programs into to create a unified workflow, and first mover advantage is a real thing

[uneven dove]

that's what Diffusers is, and that's not what A1111 is.

it's literally a single platform that reimplements everything in its own way

[eternal fog]

I tried to use "Red Faction" as a style prompt and it's taken the RED very literally

[uneven dove]

Diffusers is a shared library that's in use by InvokeAI, SD.next, my own bot, Huggingface spaces, yada yada

[civic sigil]

Everyone makes stuff for A1111 including non ML image manipulation stuff, diffusers is just an API for diffusion models specifically

[uneven dove]

naw, Diffusers is a subset of the Huggingface toolkit, they even have Transformers support

the Diffusers ecosystem is actually expanding more quickly than A1111's which is one reason Vlad is switching to it

[wicked frigate]

most SD tools are released for diffusers before auto - people just immediately port it over

[uneven dove]

tbh he copied SGM code this time, not Diffusers

[civic sigil]

Comfyui doesnt even have some of the most basic image manipulation nodes you could find in even Chainner which is incredibly minimalistic, comfy is overly focused on SD

[civic sigil]

I doubt that but ok

[wicked frigate]

(I mean, new advancements that start as arxiv papers, usually release for diffusers or for compvis codebase)

(but research code is usually trash anyway ngl)

i'm personally a fan of quality abstraction layers - lemme just use a single API that can translate to whichever backend supports my needs

[wicked frigate]

comfy has tons of community-made addon tools, they're just not really centralized very well atm

[civic sigil]

Like I said not everything needs to be a bleeding edge ML solution, SD should be pipelined with lots of other image transformations that have been out for decades

[boreal bough]

but now you're asking for the level of tools/support that companies usually charge for
at that point, might as well use photoshop beta, which has everything you'd need - but is paid

dont forget this is open source (to a certain degree)

[uneven dove]

and also just trying to reinvent Diffusers.

[boreal bough]

on the plus side, due to its open source nature, you can be the change you want to see

[dense chasm]

so cool,automatic1111 adds-on extensions always connections errors, you are a AI professional experts as well🤝

[azure oxide]

This was the first thing that I noticed. Someone other than me needs to get on it and making it popular like civitai does with models. Heck, even civitai could add a spot for it.

[civic sigil]

Youy guys are misinterpreting what I mean, I just wish I didnt have to create a dozen custom nodes to make my workflow work in Comfyui (13 actually)

[wicked frigate]

what

[civic sigil]

For extremely basic stuff

[boreal bough]

dont think about it. its a chatbot XD

[eternal fog]

Create Pull requests with them in on Github so it can get put into the main thing

[wicked frigate]

i heard there might be a way that'll get easier in the near future

[azure oxide]

Honestly just looks like a translation issue, language barriers are tough but you get used to it after a while in the professional world

[dense chasm]

u told me once before how to set the cfg values appropriately with sdxl 0.9 i remember still, appreciate for your remind dude

[boreal bough]

well yeah - but the thing is that comfyui hasn't been around as long as a1111
so its a genuine time difference.
(also, if you want your nodes to become permanent additions, send him a pull request - as long as its user friendly, and not counter-productive for future innovations, I doubt he'll reject)

[civic sigil]

I dont have a github and it would genuinely take me twice as long to make the code ready for a PR than it did to put it together lol

[autumn forum]

its not a chatbot, just a chinese man with poor english. sometimes hes hard to understand lol

[boreal bough]

fair enough

[civic sigil]

BTW I dont want to hear about A1111 having spaghetti code when my workflow ported over to Comfy looks like this

[wicked frigate]

that's not spaghetti code

that's just normal spaghetti

you can slurp it up

[patent badger]

seems casual

[uneven dove]

the code for that is way more elegant and maintainable

[urban fjord]

I'm pretty sure there are more spaghetti there than a normal spaghetti dish

[boreal bough]

sample generation during training... is... very sketchy - to say the least

already crashed for multiple reasons XD

[civic sigil]

Sure, now imagine I want to make a 3x3 grid instead of 2x2

[azure oxide]

You can easily clean that up tho. My blueprints in unreal engine were chefs kiss 👨‍🍳😘👌 compared to the spaghetti that I saw others have. Like jesus people, clean up ya work.

[uneven dove]

i don't know, sounds like a problem you're trying to solve using the wrong tool

[civic sigil]

I agree

[trim orbit]

i dont care what the code looks like. as long as i got an environment that i can do creative workflows in.

I don't feel like comfy is there yet. Amazing tool you can wire up for specific situations, but it lacks integration with other systems so well. I'll load up comfyui for specific tasks, but i'll load up auto to be creative and keep things flowing.

[uneven dove]

for tiling, i have python scripts i use, that GPT4 helped me write (because i honestly hate trying to figure out recursive loops with off-by-one errors)

[trim orbit]

simple things like adding a lora to a prompt in auto, but it requires rewiring the whole net in comfy

[boreal bough]

its missing a lot of nodes - such as variables, which is why it looks like madness
also a UE4/5 user. So I feel the pain when using comfy and not having a lot of nodes T.T
especially the limited reroute options

[dense chasm]

it's ture,btw,chatbot will be banned in China,it's tedious and ridiculous to do that in China, they should focus on their niche market tbh

[uneven dove]

i thought webo or whatever it's called, had one

[trim orbit]

china has a lot of laws that the chinese people just don't care about

[boreal bough]

that problem may or may not solve itself in a few days - as a voice in the wind has told me

[trim orbit]

i grew up being taught that the 1 child policy meant chinese people were all killing their daughters to get a son instead. turns out no, that never happened. those poeple just did the human thing and raised the daughter and declared the son.

[uneven dove]

frank i just hope we can show more of your countrymen, this kind of art

i feel it is harmless

[dense chasm]

i will not argue with you about it, but there are lots of tunnels,as i chat with u guys

[trim orbit]

people overstate how much enforcement the chinese government has over ALL of CHINA

[visual glade]

the issues with comfyui are frontend issues because there's not that big of an abstraction between what you see in the frontend and what is sent to the backend

[trim orbit]

you nailed it. thanks for chiming in. Yeah. There's no abstraction view layer. I'm sure you of all people know it

[uneven dove]

there's a time and a place for this discussion, somewhere / somewhen else

[dense chasm]

you have no idea who i am,btw,who u care about,Imao

[trim orbit]

that was the only comment about it because it was pertinent. you've blocked me. move on.

[uneven dove]

i care bout everyone

inspecting illegal tunnel operation

i think it got their uniforms wrong...

[dense chasm]

it's a true facts,cuz China economy growth,being a while,ppl will not focus only in China ,Japan is another example

[autumn forum]

omg i made my ramen too hot

[boreal bough]

jealousy intensifies

[visual glade]

and the issues with a1111/derivatives are core issues that are difficult to solve

[trim orbit]

coughgradiocough

[urban fjord]

Gradio has some great advantages.

[trim orbit]

it's been a good little gradio

[uneven dove]

sous chef shrek cooking gold flake ramen

[urban fjord]

The issue is how Auto1111 uses gradio

[boreal bough]

as an ex web dev/designer, I am deeply pained by this statement

[visual glade]

it's not only because of gradio but that doesn't help at all because of how closely it ties the backend and frontend

[winter raptor]

I'm loving the fact that ComfyUI doesn't melt my computer, unline A1111 rn

[uneven dove]

my first approach to a platform was that, but the second one was to decouple everything so that i don't need to, for example, use a Discord bot as a frontend to interact with my pipelines

it's just one of the frontends

[trim orbit]

Comfy , you seem to be building something much more purposed and I feel like you're using an MVC model with real engineering experience behind it. I like what i see you doing over there.

but i want that presentation layer

[dense chasm]

Freedom is like a air breath,huh,i though once before,it's a trade-off in the end

[uneven dove]

not much air in a tunnel underground

[urban fjord]

If you seperate the gradio code from the rest of the code there isn't much issues with Gradio. But yeah if you mix it in with everything else then it becomes a mess but this apply to everything.

[trim orbit]

automatic1111 feels like the kind of engineering fishermen do to get their boat out as low effort as possible

[uneven dove]

well gradio has API clients that make interaction with it easy

[trim orbit]

the goal is to bring in fish afterall

[uneven dove]

i still don't use it, but i think that's pretty slick.

[urban fjord]

Auto1111 isn't using api_names in the code and that makes using gradio as an api a bit harder, but that's the fault of Auto1111 and not gradio.

[uneven dove]

what you doin in the back there shrekky

[boreal bough]

EasyDiffusion was my ideal ui... it just never got enough love and support T.T

[dense chasm]

maybe,different circumstances,if as a global perspective,in a country side,there is win-win situation for good

[trim orbit]

i don't like the programmer ui of all these gradio derivatives. invoke looks nicer but i still think they miss the beat somehow. i often can't get it to install or stay up to date either. their dependency chain seems fragile

i can use them but i've never liked a straight up form field as a primary interface

[visual glade]

comfyui was primarily designed as a backend and the UI I put on top was the easiest one that exposes all the backend functionality because my strength is not frontend dev

[trim orbit]

you achieved that goal so well. it's a very cool node editor

[ionic dragon]

@uneven dove did you source the dataset to train?

[uneven dove]

shreking bad

[ionic dragon]

whats the subject?

[uneven dove]

no idea what this is in reference to

[ionic dragon]

@high skiff was saying that he was collaborating with you to train a model

[high skiff]

I said we are considering it, but we don't know what we would be doing yet

[civic sigil]

Any plans for AnimateDiff support?

[ionic dragon]

oh, i thought you guys were kickstarting it on 18th

[trim orbit]

i remember in Maya 3d, when i used it more back around 2004-2005, i would edit nodes in the graph, but that would wire up to the controls in the front end. and i could create custom controls that would link to different values in the graph. i'd love to see comfy get there one day but i understand that's a tall order.

[uneven dove]

yep that needs 1.0 to come out first so that we know what it even needs a fine-tune for

[high skiff]

I never said that lmao

[uneven dove]

that's wild

Jamie, pull that clip up, where Sytan said we're kick-starting training on the 18th

[high skiff]

All I said was he and I were planning on doing a finetune at some point

[urban fjord]

People can add what they want as custom nodes. AnimateDiff isn't really a core feature.

[high skiff]

not sure of what, not sure of when lol

[ionic dragon]

my bad, i dont remember

[high skiff]

@uneven doveOh also, hows the a100?

[uneven dove]

it crashed some more but idk, stopped worrying about it lmao

been playing with the 4090's memory issues in SDXL today because it runs out of memory when trying to generate, which pisses me off, it's a 24GB GPU

[civic sigil]

What is the distinction between controlnet and AnimateDiff in that regard?

[dense chasm]

u know ,when i am in a young age ,i recognized some fuxxing assholes from State of Michigan, and some good men from Denver,so it's hard to tell,huh

[uneven dove]

i'm too lazy to fix it in a precise way so i've gone ahead and simply enabled CPU model offloading on the 4090

[uneven dove]

the solution is to run away and hide, any time people you don't know, start walking toward you. that is what i do.

[dense chasm]

not towards me, the fucking asshole in Michigan screwed over Chinese girls in my age,but i agree with u, at that time, i can't do anything

[soft bone]

#🌶｜off-topic

[uneven dove]

good luck dal

[eternal fog]

That's got some pretty good detail on the helmet and stuff

Little scratches and stuff like that

[trim orbit]

small fine grained detail, it's so much better at then sd 1.5. Even where 1.5 suceeded with finer detail, it almost seemed like a tiled aesthetic. something similar to photoshop's content aware clone tool. I don't catch those vibes from the details made by sdxl

[eternal fog]

I still can't decide what the best way is to use the text encoders height and width settings as you can get massively different results depending on what it's set as.

[boreal bough]

why yes training sampler, that is a blonde girl sitting on a chair...

"working" is a very broad definition it seems

[eternal fog]

This is the same as the previous image, only thing that's changed is the text coders height and width values

[trim orbit]

i dont even understand why a text encoder has a width and height and why it's so yhuge

i just go with it

[eternal fog]

I mean setting it too high, or too low gets you bad results

but it changes stuff quite a lot

[boreal bough]

you can enforce the bias, thats associated with existing resolution size + aspect ratio

useful for portrait fotos, in non portrait format

[trim orbit]

reminds me of a young daniel day lewis as the butchar

[eternal fog]

I use it so it matches the aspect ratio, but these changes are just from multiplying the amount

I've found 2 times the latent image resolution seems to work best

But it can change generation to generation

[boreal bough]

oh wait. no. I mean the other node you can also plug it in

that one simple stretches it

[uneven dove]

you don't have to tweak every input 😛

[eternal fog]

But there's buttons to press

[uneven dove]

i just leave the target/crop values at defaults

haven't even played with them once

i have zero curiosity about those

[eternal fog]

They change stuff a lot though

And sometimes for the better

[uneven dove]

all the more reason to leave them alone

[trim orbit]

the crop to 9000

+1

[uneven dove]

you see, i don't know what i don't know. ergo, i don't know what i'm missing out on if i never see it. ergo, i never have to wonder, "what if i changed this setting?" because i simply ignore its existence.

[sudden cliff]

Ooh what's the style? Makes me think of an Atari 2600 box art

[sudden cliff]

Oh I see

[boreal bough]

pseudo's feelings about LoRA summed up in one sentence XD

[uneven dove]

imagine i change the Target value? and suddenly, now i don't know what the default output would have been.

that's a nightmare

[eternal fog]

For the one you replied to it's just "Pixel Art" with a fair bit of weighting

[trim orbit]

damn it worked well too. that really nailed the aesthetic

[sudden cliff]

Is this diffusers? Why is the VRAM utilization so high at no CPU offload?

[eternal fog]

I mean I guess really it's the same as just changing the seed. There's so many possible results, it's not like it matters too much what you change to get the result you want.

[uneven dove]

it's high because it's SDXL

[sudden cliff]

Huh interesting

[uneven dove]

lora gives rajiv the cold shoulder

[sudden cliff]

inference? I'm staying below 20GB at batch size 8 no CPU offloading...

[eternal fog]

And sometimes you get very similar outputs.

This is left to right, Same as image size, double image size, triple image size.

[uneven dove]

there's many permutations of "no CPU offloading"

[trim orbit]

I saw kohya-ss has support for full bf16 training. i don't know what it means but the patchnotes say it only works with libraries built for linux as there's no compatible version of bitsandbytes for windows yet.

Would it be worth setting up wsl2 on my 4080 for that? what does full bf16 support even mean

bf16 uses more precision bits, but hardware support could mean optimizations there?

[eternal fog]

The crop values I don't touch because I don't understand them and for example this is the same image as above but with a crop value lmao

[uneven dove]

lmao, see?

i never have to suffer this

i live in blissful ignorance

[trim orbit]

almost looks like a great header for a document though

purpose

[uneven dove]

you dared to dream, Arron

[eternal fog]

I think I sort of understand what the targets are doing, but I don't understand the crops

Like again, same image, different crop value lol

[dense chasm]

so sdxl is primary rivarly is midjourney, we are all in the Gotham city online🤣

[uneven dove]

honestly i don't even think the crop conditioning did a damn thing for the model to learn how they work

[trim orbit]

i think we'll see better understanding with "crop" values as the release and official documents pour out

[uneven dove]

how could it?

[trim orbit]

in the training, i saw that you can train one image with many different crops through that conditioning

refined models might benefit from the values more

[eternal fog]

The only thing I can think of how crop could be used is that you get a wide image, and what you want is too out of the way, so you use the crop. But it doesn't do what I'd expect

[trim orbit]

it all ties into the crop conditioning that is built into sdxl. something new they came up with

[uneven dove]

yeah i've noticed the Diffusers project just overrides the crop values to zero during training

even when it crops the images

[boreal bough]

tbf, training adds a lot of values simply because they exist - even if in practice they are bad or incredibly destructive

[uneven dove]

^

this is the "innovation" that SAI did, adding new conditioning values in the hopes that it'll Just Work

[eternal fog]

It's like you get this image and then you add crop 256 on width and

like what lmao

[uneven dove]

they almost added the aesthetic score to the base model but it made it stop following prompts very well

[eternal fog]

What is it doing

[boreal bough]

I'm greatful its there - cause niche cases will find use for it. But it's not something that should be a core focus for anyone not experiencing issues with a lack of that setting

[urban fjord]

Experimenting with conditing values is a good thing as long as you can check if it is doing what you want or not.

[trim orbit]

the goal of crop conditioning was to prevent the cropped face problem from 1.5 base. Now the model has data telling it the bottom half of someone's face during training is just a cropped off portion of it and there are 200 pixels above it still

[dense chasm]

git clone comfy with cloudyusage like colab or diffuse locally,u should learn a lot anyway

[trim orbit]

inferrence exposes the crop value, but i don't htink they're meaningful yet

[uneven dove]

the cropped face problem in 1.5 was, unsurprisingly, because they randomly cropped images when training 1.5

[trim orbit]

yeah

[uneven dove]

SDXL training uses data bucketing

[trim orbit]

you've blocked me. i don't think you're reading my posts.

[eternal fog]

Even tiny values change things, this is 10 on crop_w

[boreal bough]

thank god for bucketing x_x

[uneven dove]

the crop issue was solved by community fine-tunes for 1.5 and 2.1

[boreal bough]

I cropped around 2k images in total, before it became a thing...

[uneven dove]

dunno why SAI thought they need to add new conditioning values

[eternal fog]

More stuff to make it look more complex?

[uneven dove]

no it's just new research and new research almost never makes sense or does what you want

[dense chasm]

i can feel you temper huh

[eternal fog]

No temper, just interesting how little values can change so much

[sour obsidian]

its less about inference and more about training at scale for pre-training. Since it allows cropping of images to not effect context in images as the model learns to associate the cropped spaces with the context in relation to it. It does also help with inference but the mixed aspect training mostly fixed that. The conditions give a full map of image context for the model to work with and build associations around

[dense chasm]

img2video with continuous frames of pics, it will be fun to make a video with scripts

[trim orbit]

https://github.com/Stability-AI/generative-models/blob/main/assets/sdxl_report.pdf To fix this problem, we propose another simple yet effective conditioning method: During dataloading,
we uniformly sample crop coordinates ctop and cleft (integers specifying the amount of pixels cropped
from the top-left corner along the height and width axes, respectively) and feed them into the model
as conditioning parameters via Fourier feature embeddings, similar to the size conditioning described
above. The concatenated embedding ccrop is then used as an additional conditioning parameter.
We emphasize that this technique is not uniquely applicable to LDMs. Note that crop- and size-
conditioning can be readily combined. In such a case, we concatenate the feature embedding along
the channel dimension, before adding it to the timestep embedding in the UNet. Alg. 1 illustrates
how we sample ccrop and csize during training if such a combination is applied

very cool report. they did train with buckets, but also random crops

its worth reading for anyone interested in teh technical side of sdxl

[uneven dove]

i don't see how that does that, when it merely knows "i am a subset of an entire image", but those additional tiles aren't provided after with the appropriate crop values, are they?

eg. when an image is cropped, all cropped pieces get fed in?

[trim orbit]

the crop issue of 1.5 was never solved. it was worked around.

[uneven dove]

and if these conditioning inputs do affect training and help it so much, why doesn't it require fiddling with the values during inference to maintain the result?

[trim orbit]

fundamental flaws in the base model can only be solved with a new base model

[sour obsidian]

it doesnt need to see all of an entire image when training at foundational scale. If it sees many images that are missing the left side and mention a person and then see some on the right and top and bottom with a person it will build proper associations for context in relation to the training augmentation (i.e. crop/scale)

[urban fjord]

Same, some of my earlier datasets were cropped becasue that were what I were told, then I padded my images until bucketing and just leaving the images as is became the norm.

[uneven dove]

that sounds like wishful thinking considering SAI has shown SDXL so many images of celebrities that it still doesn't even get their faces correct?

you'd need a pretty even distribution of training data for that to work, right?

[trim orbit]

the report is useful to read

a lot of research is applied

i wouldn't call it "wishful"

[uneven dove]

we can call a lot of SAI's approaches wishful thinking, especially putting SD 2.1-v through 1.9 million steps of training

[sour obsidian]

thats not wishful thinking, thats how it works haha. Also its seen a ton of terrible images of celebs tagged that way too. We dont specifically map over celebs with a good set of them, one good fine tune would likely fix it but we arent trying to make the celeb model atm lol. Anyone is free to train them once its out

[trim orbit]

you can sure. but you haven't read the report is what that tells me. nothing personal.

i'll call it applied research towards a goal

[uneven dove]

well caption dropout should be helping those unseen prompt features be extracted, right?

[uneven dove]

this is why i blocked you lol. i did read the report. it is lacklustre. it doesn't describe how they arrived at many of their decisions or even get into the details of how it works / how well it works. it is a boring paper.

comfy agreed with some of my points, on what it is lacking

you know they tried T5 as a text encoder but that's not even mentioned in the paper?

i'm always hoping there's a second one coming, but they've not mentioned it yet.

[clever verge]

pytorch cross attention seems to be a second faster on avg on a RTX2060S than xformers

21s vs 22s

[uneven dove]

someone was saying that earlier but with about a 3 second gap, on a 3080.

[boreal bough]

newer gpu (more cuda cores), more difference it seems

I'll also try it on 4090 once my lora is done in a few hours

[uneven dove]

more CUDA cores, Tensor cores, RT-TFLOPS, etc 😄

[clever verge]

I've been offered a new RTX4090 by my employer but I'm not sure I want to have it.

Not sure if it's worth it.

[uneven dove]

i'd ask why they're giving you one? and why not an A100 instead? 😁

[boreal bough]

H100! you're outdated

[uneven dove]

are you trying to get them fired? that's expensive

[clever verge]

Learning, keep on top of this

[boreal bough]

XD

[clever verge]

H100 would be nice.

But then I need a server too.

There's also an L40

[eternal fog]

Got OOM trying to do a 2nd pass lo

Thankfully Ultrasharp-4x is pretty good at game style images without wrecking them

[ionic dragon]

you wanna gift us one?

[eternal fog]

Hmm you can get some unexpected stuff with img2img

[boreal bough]

even if you invest 1k, along with 40 other people for a "shared" H100, you'd only get 9 days to use it each year, in a fitted timeslot XD

[uneven dove]

i added the variant buttons to my discord bot so it's super easy to do img2img now. and i've been using a lot more as a result. it can be quite fun

[boreal bough]

god its gonna be expensive to rent...

[trim orbit]

so we're dismissing all of the pertinent information about crop conditioning in the report, because some of it was lacking ? oh okay. throwing out the baby with the bathwater seems apt.

[boreal bough]

throwing out the base ingredients, along with the finished chemical solution
(along with the research of all ways to not produce the chemical solution)

[trim orbit]

or something

[boreal bough]

just felt like an unfair comparison. no complaints

[viscid girder]

What’s better SDXL 1.0 Inpainting or Adobe firefly generative fill?

[trim orbit]

the report has a ton of information about why crop conditioning was used while building a foundational model. all i'm saying is that ignoring that was a problem. it created a huge misunderstanding here. When i said it was used for training, people came back to dunk on me by saying "nahuh, there's Bucketing!"

so i brought receipts and then now those receipts are being called inconclusive or lacking

it just seems very irrationally motivated to me

[ionic dragon]

its subjective

[eternal fog]

Yeah it's cool, there were some people saying it doesn't work. But it absolutely does, just depends what you prompt it.

[viscid girder]

Is SDXL 1.0 Inpainting good tho?

[boreal bough]

photoshop fill has worked amazingly for me so far.
If I need a solution that works quick, and works well - I always switch to photoshop generative fill (but its also a paid service, so a bit of an unfair comparison)

[viscid girder]

Can you generate celebrities? It doesn’t work on Adobe firefly website but that’s what generative fill is on photoshop so idk

[trim orbit]

adobe will keep updating theirs and it'll keep getting better. SDXL will get controlnet inpainting plus community refinements. The best is the one that works best for your creative workflow that looks the best to your artistic eye

[boreal bough]

sdxl is my hobby generator
but if I need anything for work or real life situations - Photoshop generative fill is just plain reliable (since it was developed for real life situations)

[trim orbit]

objectively, adobe's will be limited in a variety of ways for legal reasons, since they're a commercial product

[viscid girder]

True

[boreal bough]

not a hobby tool. as a hobby tool it's equally good to sdxl at its best. It only shines when you genuinely need fixes for issues in photos that are sfw

[viscid girder]

Gotcha 👍

[trim orbit]

i'm sure you could fix nsfw photos if just the non naughty bits

[boreal bough]

had to use it for a sfw cosplay photoshoot - even that was too much XD even if you're trying to remove a stone in the background

[viscid girder]

Oof fr?

[trim orbit]

i haven't played with it much yet, but i bet they'll come to fix that. Fashion marketing and cosmetics will need capabiliteis. sex sells always

[viscid girder]

If someone has used the SDXL Inpainting, could someone post an example? I’m just trying to see if it’s worth downloading

[boreal bough]

cosplay is often tagged nsfw, so its a bit biased there. normal fotos work just fine

or crop, open as new image -> fix there -> move back to main image

[trim orbit]

can be worked around so easily lol

could probably set an action up to expedite that

[viscid girder]

Cause I’m trying to get a character to hold a gun and it blocks me, is holding a gun that bad?

Is there a workaround for it?

[sudden cliff]

So Firefly blocks you from editing certain images (trying to follow)?

[boreal bough]

definitely read their compliance guide for it. You can get banned from the whole service.
Violence is specifically mentioned as a no go

[trim orbit]

probably marekting departments idea. they don't want the adobe brand associated with guns in children's hands photos

just a long shot guess. don't read into it too much

[viscid girder]

Does someone have an example of an SDXL Inpainting?

[trim orbit]

ahhh they got compliance guides. yeah. legal and marketing shenanigans

that'll erode away as they find better ways to cover their ass

[uneven dove]

the spirit of adventure, 1955

[viscid girder]

Does someone have an example of an SDXL Inpainting?

[uneven dove]

keep repeating it a bunch more times

someone is bound to do the test for you

[viscid girder]

Chill bro

[eternal fog]

yep img2img works lol

[trim orbit]

weird ping. i dont even have access to an sdxl model at my home.

[viscid girder]

I just don’t know if there’s a place where I can see inpaintings and thought u might know

[trim orbit]

¯_(ツ)_/¯

[uneven dove]

just download it and do it yourself lmfao

[eternal fog]

Balance this out with a dude

[viscid girder]

But I don’t wanna 😭

[uneven dove]

looks smooth because you're refining the output @eternal fog

[eternal fog]

I've not refined either of those, I'm just messing around and haven't got the settings down properly.

[uneven dove]

ah

[eternal fog]

I'm not even prompting properly for it

Just interesting to see what it does

[uneven dove]

i use 20 steps and 5.4 cfg on the img2img input on Base model.

those are hard-coded because it's a pretty narrow window where the i2i seems to work properly

decided not to let people mess with that parameter group

[eternal fog]

The denoise strength seems to have a pretty definate drop where it no longer looks like the input image at all as well

[uneven dove]

i'm at .8 on my examples

it can be rough

should get better results from this input (left side) with 0.73 strength

[viscid girder]

Just to clarify this is 1.0 correct?

[uneven dove]

0.9

we do not have weights for 1.0 yet

pixar's spirit of adventure, 2018, cinematic

it seems that even if you stick to "1 megapixel area" resolutions, if you go too wide, integrity across the scene is violated

it's consistent

[viscid girder]

That Low-key reminds me of 1.5 from 2 years ago…

[uneven dove]

2 years ago? lmao it was released last October

not even up on 1 year old yet

[viscid girder]

Sure feels like 2 years ago lmfao

[uneven dove]

maybe you meant Craiyon or Dall-E Mini

[viscid girder]

Prolly

[uneven dove]

craiyon has upgraded since then

you should try it lol it has negative prompts now

[eternal fog]

Converting stuff into game style is pretty good

[viscid girder]

I’ve seen it it’s improved a tun

[uneven dove]

i was able to get a couple photoreal images from Craiyon a cpl weeks back

using OpenCLIP prompt style

[trim orbit]

reminds me of metal gear ps2 era

maybe ps3

its legitimately good at it

[ionic dragon]

they scraped the entire web, thats obv that its way better

[uneven dove]

Adobe?

[ionic dragon]

yep

[uneven dove]

they say they only used licensed inputs

who knows

[ionic dragon]

they have adobe stock

they'll use the entire dataset

[amber fulcrum]

hey guys - where I can find more workflows? I finding a lot of workflows in comfyui using nodes I don't have. Is there any repo/huggingface with some?

[uneven dove]

unfortunately it's not very centralised right now

[trim orbit]

gigo applies there. they didn't use a generalized scrape for a dataset. they are also adobe and have been doing photography research for decades

they have data sets

[ionic dragon]

yeah

they are a massive company

[trim orbit]

leaders of digital editing research i'd say

[ionic dragon]

if they want, they can even launch a very good txt2vid generator

[trim orbit]

civitai.com has a lot if you can wrangle the search function to show them.

https://civitai.com/?query=comfyui&view=feed

oops wrong reply button. meant that for @amber fulcrum

[ionic dragon]

oh ok

[eternal fog]

i don't think it can decide what it wants to do here, sort of half way between photoreal and not

The rain/water effects are cool though

[amber fulcrum]

nice! like shot from tekken cinematic

[steady grove]

there are only sdxl 1.0 release canadidates so far. 2 more days though so the feedback is probably locked in and they're refining the selected final run now

you cna try the release candidates throug the bots on the server last i checked

[eternal fog]

I'm trying to do a dragon ball z style but more realistic, but if I put dragon ball z in the top text encoder, it just does weird cosplay lol

[steady grove]

put "cosplay" in the neg

[eternal fog]

Or whatever this is lmao

[steady grove]

gokine using kaio-ken?

[eternal fog]

The results are "interesting" at least lol

[steady grove]

thats how live action netflix adaptation should be done

[eternal fog]

Lmao, it's going a little heavy on the booba as well for some reason

[steady grove]

so much boobscle getting flexed

[eternal fog]

I must be right on the edge on the prompts, because 1 seed it will be realistic, another seed it will be 3D Animation type look and then another it will just be Goku

And then adding movie still makes it actually look like anime?

I think it's just doing that it wants at this point

[amber fulcrum]

This one is amazing ^^

[eternal fog]

I'll do 4 in a batch and just show how different it's making stuff

(watch them all be the same now)

lol

[amber fulcrum]

mabe add "game screenshot" in negative?

[rustic garnet]

cool style! what is the prompt

[eternal fog]

I tink it's because my CFG is low

[eternal fog]

The prompt for that one is "Movie Still of a woman on a wet Japanese Street during the day, debris energy ball energy blast from Dragon Ball Z"

But it's not reliable. I've only gotten that type of style once.

[eternal fog]

What is that lol

[rustic garnet]

when I tried Dragonball once I got extremely uggly results as if it was inspired by bad fan arts

[true hazel]

gym girl

[eternal fog]

Yeah it likes to try Goku everything

like this

[eternal fog]

sus

lol it keeps using the "Ball" from Dragon Ball Z on other stuff

Giant Stone balls keep appearing

[fresh path]

what about doing dragonballz

or dbz

[eternal fog]

dragonballz would be the same as I imagine ball is it's own token, but I can try dhz

[fresh path]

yeah I think it is ruined by bad fan art

[eternal fog]

dbz*

[fresh path]

try some stuff in negative prompts

and try keywords like masterpiece pro idk

[eternal fog]

ok dbz might work

Ok looks like at least on the surface it does

[fresh path]

got this

[eternal fog]

On it's own yeah I imagine it will do that

[fresh path]

thats really good, 0.9 couldnt do that I think

[eternal fog]

This is 0.9

[fresh path]

same with goku by himself

ooh

[fresh path]

it messes upthe eyes for me

you put the refiner on it?

[eternal fog]

Nah, I'm just seeing what comes out

[fresh path]

try some fight scenes with two characters

it has real trouble with that

[eternal fog]

Yeah that never works

I'll try though

[fresh path]

whenever theres two people it morphs their faces together

[true hazel]

it sort of knows bleach, but merged multiple characters

[fresh path]

is that ichizen

[trim orbit]

sdxl really understands linguistic context so much better. it could help to say "as characters from dragonball"

in some kind of full complete sentence

[eternal fog]

ball is still it's own token

So it does other stuff as well

[trim orbit]

the context of tokens matters especially more so in larger models

[trim orbit]

I've seen Sytan's comfy ui brings in two different text inputs. I've been wondering how that works, but it's anice approach.

[fresh path]

thats better

[trim orbit]

he's really trying to intimindate her bangs with that gaze

[eternal fog]

Getting htem to fight is difficult, strange things happen

[true hazel]

hollow ichizen lol

[fresh path]

BLEETH

for some reason it makes ichigo angry

[lament rune]

Civitai I guess. For new nodes, use Comfy Manager to download missing ones.

[lament rune]

stability is non violent

I wonder if they filtered that out with nsfw

hence the no guns, swords etc

[eternal fog]

nsfw isn't really filtered out

You can get it quite easily

I've even done it by accident

[fresh path]

the discord bot censors words but it doesnt tell you

[eternal fog]

The bot does, but you can run it locally and you don't have that issue

[fresh path]

yeah

[eternal fog]

I'm trying to get it to do Piccolo

It knows he's green because it's giving people green hair

But they just look "normal"

[true hazel]

bleach visiting walmart arc

[eternal fog]

lmao, that's not Piccolo

[fresh path]

pick up the reishibites, aizen

-kenpachi to aizen while seireitei is being attacked

in SOULMART

[eternal fog]

I think you cursed me or something, it's just generated naked titty for no reason

[ionic dragon]

is there a tool only to upscale?

[lament rune]

did it have a male nipple?

[normal cliff]

why is that so blurry though?

[lament rune]

because that seems to be the way it goes

[eternal fog]

No, it was a full naked breast, how they should be

[compact beacon]

vhs lora

[true hazel]

it's pretty good at drawing other people in bleach style, joe biden bankai

[fresh path]

lol thats awesome

some kind of donald trump monster at the top

[eternal fog]

This isn't doing what I expect, but it's doing some decent images

[fresh path]

sexy assassin vs 100 unidentified men in black scene

classic

[delicate grotto]

is this workflow ok for comfy with refiner?

[lament rune]

is this public!?

[compact beacon]

civital

[normal cliff]

try finding it on civitai

[eternal fog]

You don't use the refiner like that.

You use the Advanced KSampler, Start with the base and then do the last few steps with the refiner.

[lament rune]

all I see are 1.5 and 2.1 lora... unless your workflow redoes it with a pass in those

[delicate grotto]

ahhhh

[compact beacon]

https://media.discordapp.net/attachments/908912109142827082/1079351357737279498/cat-chomp-fireball.gif

[true hazel]

trump bankai lol

[fresh path]

Bankai... Kyodaina Kōtei

[eternal fog]

Hmm, it can sort of do Psycho-Pass

Although I'm not sure where this outfit comes into it lmao

[delicate grotto]

how do you make comfy use dynamic ram?

[eternal fog]

What do you mean

[delicate grotto]

i have 16gb ram, i need ~24 to 30 to use sdxl + refiner

[eternal fog]

You shouldnt at all

[paper phoenix]

that doesnt sound right at all TYP

[delicate grotto]

odd, im getting lots of out of ram, it asks for ~24gb

[lilac wren]

SDXL know Ferrari F40 ❤️

[eternal fog]

I have 16GB RAM and its fine

You've not restricted your page file size have you?

[civic sigil]

Sometimes it feels like the background and foregrounds are totally seperate in xl

[delicate grotto]

don't think so

[delicate grotto]

looks like this atm

maybe there is a command for unload after use?

[eternal fog]

why are your images so small for a start

[civic sigil]

Im curious whats the point of seperate vae in xl

[delicate grotto]

i think you are using dynamic ram to load it all, because combined it should be 12gb + 6gb

[trim orbit]

cpu mode

if you're going to clock down to system bus level, go all in

just use your cpu for genning at that point

a gpu only benefits speed when using the dedicated vram that is hardwired to the gpu

[eternal fog]

how much VRAM do you have @delicate grotto

[delicate grotto]

6gb, its not a vram problem its a ram problem

seperately i can do almost 2048 using the sdxl

[eternal fog]

What are your lanch parameters

[delicate grotto]

none

xformers, normal vram

i think ill make a 2 stage for the pictures, save them as latent, and later on load them as latent into the refiner

[inner ruin]

what time is SD XL being released on Tuesday?

[delicate grotto]

usually they broadcast the same day so ~between 18 and 20 europe time?

[sharp robin]

#✨｜sdxl message

[sharp robin]

idk if it didnt link right @delicate grotto

[normal cliff]

howdy audi

[eternal fog]

Yeah I think its this as well

[delicate grotto]

must be it, ill play with it tomorrow

for now without the refiner it is fine

thank you all for the tip

[paper phoenix]

ive noticed this myself. is it trying to figure out depth by itself as part of composition?

[eternal fog]

I keep getting this annoying blur over the images when I use "Movie Still"

Like the whole thing looks a bit smudged

[uneven dove]

me, staring at arron's blurry images

[civic sigil]

Maybe cause all the motion blur from slow shutter speed in film cameras

[eternal fog]

Yeah it might be, or film screengrabs

Trying to work out a better style word

Cinematic Camera Shot seems to work a little better

That's without refiner though

And it looks a bit plastic like

[uneven dove]

movie still doesn't seem to do that here

[trim orbit]

lol jlaw super saiyan's hair is even limpy

[eternal fog]

lmao

I had 1 where it was big

[uneven dove]

cigarette smoking intensifies

[eternal fog]

If you go too far with super sayian weighting it just turns into a dude

[fresh path]

bro got bluetooth

[eternal fog]

argh lol, it's still doing the blur stuff

[uneven dove]

i think it's just your config or something

[eternal fog]

I don't see how, these are just a single shot through the base

It's got to be the prompting somehow

ok a little better

[delicate grotto]

overwatch icon

[eternal fog]

A snake with a gun

nice

[delicate grotto]

i wanted it to climb the rifle, but it insists on holding it hehehe

i like this model

[fresh path]

video game lighting

[eternal fog]

Yeah used RDR2 as a style for that and the wild west

[lusty raptor]

the bot produced a NaN

[west breach]

Yeah

[winter raptor]

Thanks! So far I’m loving your workflow, it’s given great results. Are you planning on updating it as time goes on with new features?

[winter raptor]

And thank you for making it publicly available!

[eternal fog]

lmao

[autumn forum]

im surprised at how well it does porche

[uneven dove]

it's overbaked actually

it's hard to get it to do anything special with a porsche

[eternal fog]

close but her leg has fallen off

Is it just me or does that Porsche look a bit wide lmao

[upbeat summit]

well maybe it thinks it's shoulders 😉

[upbeat summit]

I remember trying to do hybrids with dall-e. it never worked

[eternal fog]

Told it to use scales instead

[eternal fog]

Seems pretty easy with this

[uneven dove]

Will Ferrell XL

[upbeat summit]

nice hat!

[eternal fog]

lmao

Has technology gone too far

[uneven dove]

18th

[eternal fog]

What have I found lmao

[west breach]

Thanks! 🙂 I am constantly fiddling with workflows, particularly a 2x workflow. Pretty much all the images I post will have the workflow in them. Feel free to use them, although I do use some other's custom nodes too. Getting a good image at 2x is tricky. Any ideas or questions, I'm happy to hear them. I mainly created the custom nodes for things I thought would be useful and that I am also capable of coding myself 😄

[lusty raptor]

i just tried yours a bit ago mikey, that's a nice setup. I did hit some sort of fail state with it somehow. The first gen worked, but everything else after that gave me noise. Could be a comfy hiccup, I only solved it by restarting.

[lament rune]

Is your workflow posted on Civitai or anywhere @west breach ? The custom node looks cool

[west breach]

what I've found so far with 2x is more noise added to the upscaled image is good for images with lots of details and small faces, really bad for close up portrait photos. So really have to reduce the amount of noise for close ups or you will end up with extra noses, eyes etc

[inner ruin]

LoRA?

[eternal fog]

no

[west breach]

there's an example workflow on my github for the prompt with style node https://github.com/bash-j/mikey_nodes

[lament rune]

ah nice! thank you

[west breach]

The number of steps I have in that workflow is 92% for the base and then 8% for refiner, if you change it slightly you can get noise

[lusty raptor]

i think my only critique is there's no option to do separate G/L prompting, but we'll see what happens once 1.0 launches. It's still wild territory, but i've found separate prompting to better for certain things, especially anime

[lusty raptor]

that's the weird thing, i changed nothing

[boreal bough]

dragonborn!

[west breach]

yeah, the way that node works is it puts your prompt + the styles prompt from the selection into the G and then just the styles prompt into the L and the refiner

[lusty raptor]

but as far as easy and clean goes, it's great

[west breach]

you can add your own styles to the styles.json. you don't need to reload comfy after you save the file, it should update as soon as you click the selection widget

one thing I want to add is the ability to use wildcards

[upbeat summit]

thanks for all your work in the community! your nodes are super helpful and the latest workflow of yours - that's some clean layout - great work!

I was like - where is the rest? nope, it's just super minimal and compact 🙂

[west breach]

one of the benefits of being lazy I suppose 😄

[upbeat summit]

do you have a twitter or something so I can give you a shout out next time I post some images?

[west breach]

since you are all being so lovely, I'll put together a 2x workflow too

[west breach]

no sorry

[winter raptor]

That 2x would be clutch. I have been trying to figure out how to implement quality upscaling into the workflow (any workflow tbh lol), with no success

[upbeat summit]

ok so I just link to your github if that is alright

[west breach]

it's the hardest thing, I think you need different levels of steps and noise depending on the original image

[winter raptor]

Yeah it’s super tough. The way I was doing it in 1.5 was using ultimate sdupscale with controlnet tiling, but haven’t been able to make that work yet with SDXL in comfyui

[west breach]

would love to be able to lock workflows, so often stuff goes weird and it's because I accidently clicked on a widget and changed the cfg to 20 or something 😄

[upbeat summit]

right click -> pin?

[somber hill]

my newest SDXL tutorial with comfyui : https://youtu.be/FnMHbhvWUhE

hopefully lora training coming for SDXL

tomorrow

[west breach]

how did i not know this?!?!?

[urban fjord]

You can already use Kohya-ss to train LoRA

[upbeat summit]

doesn't work with groups yet

[winter raptor]

Wait huh? Does this lock the fields as well as the location?

[upbeat summit]

yeah the node is locked in position

[upbeat summit]

no the widgets are still free to be changed

[winter raptor]

The more you know

[trim orbit]

anyone ever notice that his eyes look like a butt?

[winter raptor]

Pepe’s?

[trim orbit]

yeah pepe's eyes

[upbeat summit]

the manual is not complete, some stuff didn't work for me but this ComfyUI shortcut list helps if you haven't found them all yet: https://blenderneko.github.io/ComfyUI-docs/Interface/Shortcuts/?h=shortcu

[sour obsidian]

How dare they not have left alt+left click haha

[upbeat summit]

ctrl+drag also missing

[west breach]

ctrl + shift + v is a thing of beauty

[upbeat summit]

yes 😄 I learned this from the sheet

just chain as many samplers as you want and don't have to reconnect everything

[west breach]

2x workflow on my github (also in this image)

[winter raptor]

@west breach so to be clear, that bottom text field is a negative prompt? I'm loving the simplicity of the nodes btw

[west breach]

yep, it should put positive and negative as the default if you add the node

the styles have a bunch of negatives already, which is why i left it blank

[amber fulcrum]

Sad i am getting some tensor errors from this workflow ^^

[winter raptor]

Makes sense. Finally, where did you find realesrgan_x2, can't find it anywhere lol

[west breach]

do you have the manager extension?

[winter raptor]

yeah

[west breach]

It's in the install models section

[winter raptor]

oh sick

[upbeat summit]

woha that resolution

[winter raptor]

decided to just use the 4x ultrasharp, no fancy tricks lol

[eternal fog]

Yeha Ultrasharp works really well for that sort of stuff

[west breach]

the wildcard support for my prompt with styles needs this widget, but I have no idea how to add it

[static prawn]

feel like 1024x1024 is not a good res, getting way better results on 1280x1280

[eternal fog]

Just have a look at the code for that node and copy it

[upbeat summit]

great colors and lighting

[west breach]

can't see it in the inputs

[sage basin]

It shows up automatically when you have an input named "seed"

I'm not sure if there are any other ways to get it

[upbeat summit]

it is only mentioned in the app.js and widget.js

[static prawn]

does refiner also destroy the result for u sometimes?

[west breach]

I called my input wildcard_seed, I'll change it and see if it works 🤞

[static prawn]

Without refiner

with refiner

[upbeat summit]

so there the refiner does work 🙂

[static prawn]

look at the eyes

completely destroyed after refiner 😄

[eternal fog]

The eyes look fine in both to me to be honest

[upbeat summit]

yeah.. well kinda. they look proportional better but they are now flat - no life

[static prawn]

look at the eyecolor

one grey/green eye and one brown eye after refiner

before 2 brown eyes

feel like the refiner is a hit and miss 😄

[upbeat summit]

the refiner is super finicky

[west breach]

thanks, that worked 🙂 not really a 'seed' as such, more of an offset but at least it's working 🙂

[winter raptor]

The refiner doesn't like eyes, I have noticed that

[winter raptor]

This is really good

[winter raptor]

Damn, I didn't know bulbasaur was chill like that

[eternal fog]

It struggles with him

I like the vapourwave style

[static prawn]

is there a database for different styles?

[eternal fog]

Not that I know of, I'm just typing shit in

[static prawn]

allright haha 😄

hopefully i get a better gpu soon

[eternal fog]

Warning, Mermaid sometimes produces nsfw results lmao

[quasi remnant]

i meaaaan; that should be a given

most paintings/depictions of mermaids are bare chested

even in family friendly restaurants if there's mermaid pics they're bare chested lol

[eternal fog]

That's one scary looking monster

[static prawn]

still extremly bad face dunno xD

i never get any consistency

its really hard 😄

[eternal fog]

I love mixing nonsense stuff together lmao

Same prompt that made that but add Sonic into it

[autumn forum]

thats actually not bad lol

[eternal fog]

lol

[urban fjord]

Nothing is wrong here.

[eternal fog]

lmao

[static prawn]

are u guys scared of tomatoes?

[eternal fog]

No?

[static prawn]

im asking honestly

[eternal fog]

Are you?

[static prawn]

normally not, but i changed my mind

[urban fjord]

Maybe...

[autumn forum]

i hate tomatos but im not scared of em lmao

[static prawn]

i dunno which i would get a better face , empress in this pic

[eternal fog]

This is your fault void

[static prawn]

xD

any negatives u use, or quality tags?

[eternal fog]

for the moment Ive been using

melted, compressed, artifacts, earrings, deformed, amputee

earrings because they always look shit

compressed and artifacts seems to help a tiny bit, but not much on quality

[urban fjord]

Strawberries are scarier.

[static prawn]

dunno i just dont like the human on the tomato, but i guess thats a general model thing

[static prawn]

would it would look more 3d

love*

or like an actual character, not so flat

but im really limited with generations anyway 😄

[static prawn]

its probably bec. i miss inpainting, i mean faces on this distance looked way more worse on sd 1.5

so it seems to be fine

[eternal fog]

You can still in paint, it's just more of a hassle to do in ComfyUI I think

[eternal fog]

no

lmao

[static prawn]

oh ok, i didnt know that 😄

[urban fjord]

SDXL is great.

[eternal fog]

it is

[urban fjord]

I just hope the 1.0 has some real improvements from 0.9, has anyone seen improvements in the bot?

[raw prairie]

A bit

Yes

[static prawn]

wtf

[raw prairie]

Tomatosaur

[raw prairie]

Interesting

[urban fjord]

Tomatosaur

[raw prairie]

Lmao

[static prawn]

but faces kinda look generic 😄 when u just type in empress

[trim orbit]

g'over hurrrr

[eternal fog]

yup

[urban fjord]

Yeah SDXL makes a lot of the same face.

[static prawn]

OK its definetely going to far xD

[azure oxide]

i think the same is true of the original 1.5 model too, do the usual tricks work? [person1:person2:.6], alternating person names, using the name of 1 person or person of a country, etc?

[steady grove]

mildly nsfw but mostly just cute. using roop with sdxl from faces made with sd1.5. wish i could get gfpgan wired up. thought i'd share cause i liked this result. clothes that don't look super skin tight

[civic sigil]

Her feet look weird but very nice overall

That shadow is kind of suspect too now that im looking at it

[raw prairie]

The car looks ok

The car looks good tbh

[steady grove]

i agree i didn't even prompt for a car either. sdxl's extra details are so fire

[raw prairie]

Yeah

[steady grove]

feet are probably weird cause i prompted for sneaker high heels

[raw prairie]

And the hand looks like she had an accident with a meat grinder

When is 1.0 releasing btw?

[eternal fog]

18th is the date they gave

[autumn forum]

the anticipation is killin me

[raw prairie]

Cant wait

I remember when 1.5 came out

1.5 is a legend now

[urban fjord]

Apparently, there will be a new refiner. I hope to see a lot or improvements there.

[autumn forum]

a new one?? hmm

[raw prairie]

The majority of the models on civitai are based on 1.5

[raw prairie]

What's a refiner?

[static prawn]

i really like doing horror stuff with the model