✨｜sdxl

[steady grove]

without --medvram it just piles EVERYTHING into vram all at once, so it's a mess

[native knot]

That time at that step is normal.

[solar bone]

ah okay

[solar bone]

interesting

and what is a sampler exactly

ive read stuff on it but theyre all so confusing

[steady grove]

well, there's steps right? each step is trying to remove some of the noise. the sampler determines when and where

i think

i'm confused to bruh

[solar bone]

ah i see

[native knot]

sampler and schedule is essentially the formula that reads through the data to infer the image from the latent space based on the prompt

[steady grove]

i liked computerphile's video but i'm still confused

https://www.youtube.com/watch?v=1CIpzeNxIhU

[molten gull]

https://www.youtube.com/watch?v=T1oWgIICSkI

is it possible to use the quality-of-life suit with a local LLM instead of openai?

[steady grove]

i think i need to train it more. or with different captioning. adjust the dataset a bit. making styles is tricky but it works if i wrangle with it

[steady grove]

kinda getting the hang of prompting this beast. not bad not bad

[steady grove]

noticing a LOT of teal so gona do another round of training with that setting color augmentation... i think it's supposed to shift the colors of the image randomly. could be neat?

[dim steeple]

Hey guys I have a question about img2img

Does a 1.0 denoising value mean it fully ignores the img color?

[steady grove]

sort of, but it still does affect it. since it's adding noise to those initial colors to denoise it, the initial color affects that 100% noise level,still

it'll be different than if you use txt2img

[bright valley]

For anyone doing civitai's 404 contest: https://civitai.com/models/204953/the-404ra-add-on-for-harrlogos

[steady grove]

speaking of harrlogos, combined with the madball lora i'm experimenting with

[half ivy]

hello team, i am here
any idea why im getting a black image with lcm?

[steady grove]

maybe a precision problem. full vs half should work

full is slower though

[half ivy]

yea just found a thread on that online, will try removing it from command prompts, thanks bruh

[south horizon]

Nice. Make a game. 😄

[dim steeple]

Is it possible to do text2img controlnet but have it only render part of the image? Does that make sense?

[half ivy]

MAYBE IMGTOIMG INPAINT?

[jolly brook]

I would very much appreciate an @ when you're done training 👀

[half ivy]

yo bro's,
LCM
The config attributes {'requires_safety_checker': True} were passed to LatentConsistencyModelPipeline, but are not expected and will be ignored. Please verify your model_index.json configuration file.
ANy1 dealt with this before

[dim steeple]

yeah except I am using text2img so that it it doesn't pick up on any of the color data, just controlnet

[half ivy]

IMG2IMG
CN: 1
Denoise: 100%
MAybe? This tends to change the colors for me

Or txt2img but only have the CN for the part you want

so do it in pieces basically

🤔

[dim steeple]

I did that where I masked controlnet for part I want only, but it still creates the image in the rest of the canvas in text2img

[dim steeple]

even at denoise 100% it still seems to pickup on color

If I do it in smaller peices I fear I will lose consistency

[bright valley]

Wow this is awesome!!

If you want to try another one that works really well with it @steady grove , try this one https://civitai.com/models/201098?modelVersionId=226328

[steady grove]

nice. its really cool how that lora can blend with concepts so nice!

[bright valley]

Facts! People have figured out so many awesome combos too. That's how I was able to make the 404ra too it as well

Made this one with the Ghost style lora

[steady grove]

i entered some entries into a buncha categories then realized i was making logos for real games and movies cause i got carried away

gonna go for 404

[zinc cargo]

smart move 🙂

@bright valley it works awsome with my Infernoframes 🙂
well, at leats most of the times

[bright valley]

Looks very cool man! It's not a perfect science, so you may have to either try a few seeds or parameter tweaks.
But the saving grace is that when it hits, it's worth it (imo )

[zinc cargo]

it's such a great lora. i actualy adding these to all my loras

not always hitting the right spelling, but it's still cool 🙂

[bright valley]

hey thanks a lot, I really appreciate that! Can't wait to see more of what you can do using it with your models!

[zinc cargo]

why didnt you allow using it in civit create?

[bright valley]

@smoky patrol ^^^

[bright valley]

I'd really like to, but unfortunately CivitAI requires that you grant commercial rights to have your model in the on-site generator

[zinc cargo]

commercial rights to civit or to the user?

[bright valley]

They are looking into changing that if possible, so if they do, it will be in there

To the users

[zinc cargo]

is it rights on the lora or on the images?

[bright valley]

In this siuation, the images

[zinc cargo]

you wouldnt want that?

[bright valley]

I currently do not grant commercial rights for it, no. The model is trained on all my own artwork and graphic design work, which is what I get commissioned for most often.

[zinc cargo]

cool, i totaly respect your choice. but i'm intersted to learn more about this stuff

[potent mica]

is anyone else getting issues with the stability API for v1 xl? My results are all pixelated all of a sudden from last week?

[dim steeple]

Hello, is there a good way I can do inpainting on a large image using SDXL?

AND using controlnet

[carmine gale]

What is the closest thing to regional composer or latentcoupling/composable Lora from auto1111 in terms of comfyui workflow nodes?

[shy kelp]

yes, reported by many users, there is a discussion on this topic this over at #1042896447311454361 but no word from stability.ai yet.

[smoky patrol]

Hmm?

[pure crystal]

I've got into a workflow lately, I start with a well formed specific prompt I paste into gpt4, then I say try again, ... with specific details and changes and have it print the prompt back out. then running with these and ending up with a notebook about the idea

[half cedar]

Awesome. I remember the toy like this in the 80s

[deep locust]

Yo! You guys think it will be possible to network render at some point. Like, combining two machines to achieve batch rendering lets say in comfyUi for exemple?

[carmine gale]

hey guys, does anybody know how to break a segs mask of two persons into individual masks, left and right, in order to prompt each person separately and assign them each their own lora? Here's my mask. Person left, prompt1 + lora1 & prompt2 + lora2

[indigo carbon]

just did some comparing between ComfyUI with AIT and A1111 with TRT.
batch 4 with ComfyUI+AIT is 2.31it/s for me

batch 4 with A1111+TRT is 2.07it/s

[steady grove]

@jolly brook @half cedar thanks! i htink i'm going to publish v1 tonite. it's a little rough but its fun to play with.

i was trying to train with color augmentation, but it requires cache latents to be off, which more than doubles my training time. might try it instead on v2

going to finish refining this one off then adjust the set. maybe use the model to create more variety in the image set

[weak wagon]

Idk where to ask this so... how do i out this to work? Ehich folder should i introduce this to?

[peak dove]

I am using Harrologos2_LoRA and a Papercut SDXL_LoRA model from HF - it is crisp and clear and almsot everytime spells it right!

[shy kelp]

Chester Bennington?

[dapper dragon]

Yes

Posted it on Tensor.Art

so you can play with him for free 🙂

[bright valley]

Sorry there was a bunch of spam posts in the channel, was just trying to alert somebody who could remove them. Looks like that already happened before you saw it though. 😅

[bright valley]

That's awesome dude!

[smoky patrol]

Oh, it's all good! No need to worry---you can always ping me!

[bright valley]

You're a legend!

[smoky patrol]

Oh, ty!

[final root]

reproduire un robot identique qui est sur l'image et ajouter une pillule sous son bras. Il faut que ce robot avec la pillule soit au centre l'image. L'idée est d'en faire un logo pour promouvoir un site sur l'intelligence artificielle et facile à comprendre

/prompt

[nocturne dove]

I like the Frieza a lot there 🙂

[bright valley]

Tbh I think that was mixed with Cooler or something, so I just now combined with frieza

[sharp robin]

what is dropping tomorrow?

[wise smelt]

请设计一款类似的，但更年轻化的女士拖鞋

[steady grove]

testing is going splendidly

[crisp owl]

Must weigh a ton, completely broke that last guys thumb!

[steady grove]

these are the heavy boies

get the waifu test over with

[steady grove]

its up https://civitai.com/models/207437

@jolly brook heres your @

[fleet harness]

Used to collect those back in the day. Wish I still had them. Nice work.

[steady grove]

thanks! i had a couple but never collected many. its a cool project. i learned a bunch

[hoary saddle]

looks awesome, thanks for the creation!

any idea why i keep getting the pointy head thing?

[hoary saddle]

wow, you outdid yourself on the quality

[steady grove]

Thanks! The base model has some understanding of mad balls already . Many of my training images were 480x480 pixels scaled up haha. Not a lot of quality scans made it . I think my model evokes that existing knowledge.

Got a couple ideas to improve on that

[steady grove]

Some bucketing issues over fit a few things I think

Cyan particularly haha. Going to have to do some color balancing on the set and maybe captioning adjustments

[supple wind]

what is the best model rn?

[native knot]

Art is subjective. There is no "best model".

[peak dove]

...although Linda Evangelista comes close!!! 😉

Cannot get my Version of OenAI to marry up with my ChatGPT link to ComfyUI ... so learning Python, GIT, SVN, Bash, Ruby, Cobol etc (well,certainly Python 3.9)

[pale canyon]

Hey 👋
Does the SDXL bot in the discord use all the default parameters for the API? 20 steps, and 7.5 guidance scale?

[rustic garnet]

The sdxl base model is really good. Custom models are better in specific styles (e.g. photorealism) but in general you can use the sdxl base model

I still like to use Dreamshaper XL a lot, because it gives me often slightly better results than base (i.e., less blurry backgrounds), but in contrast to most other custom models it is not overtrained and stays close to base

[cyan crown]

SDXL Base has everything inside but it needs better prompting.. Combine with xl_morte_artist lora and it's incredible

[crisp owl]

Anyone have a workflow they commonly use for regional prompting?

Trying to prompt for a norse giantess standing with a shorter god of the seashore, but it keeps just making the male the dominant tall figure.

[steady grove]

The base is actually not as good as 0.9 beta release was. Oddly. They went a neat way with it and iit's commendable. The base, while great, is better if refined slightly at all. If any little bit of training is applied to it, suddenly the quality leaps ahead. They did something during the finalisation of the model in order to make it easier to train on for the community.

Lora's work towards that refinement too. Even using the sample noise offset lora it was released with, lend towards a higher quality generation.

[cyan crown]

Well QRCode XL works, but in some way gives fewer good results than 1..5 version

[jolly brook]

http://batbot.ai/generation/1176610449630961704

I am loving it 🫶 thank you

[steady grove]

What a meatball!!

Lol madball I meant

Meatball works too

[old cradle]

Is that the more artful Lora or something else? Can you link please?

[molten gull]

and one that i really like:

"the green universe cube" 🙂

[upbeat summit]

very nice images, @molten gull . great mood and atmosphere. nice work

[molten gull]

thanks 🙂

[steady grove]

i wasn't sure if the photos of hands squeezing madballs would work in the training set. this is a pretty great result though. there was a "sick series" where if you squeezed them they'd pop like the second real image does

i think if i caption it better, "slime balloon" or something

"tumor balloon" lol

[worthy orbit]

indeed

[sinful falcon]

The rose is such a nice touch

[crisp owl]

heh, altman returning to openai

[lilac wren]

SDXL ❤️

[rustic garnet]

the more_artful lora is almost always making things worse for me. It can help sometimes (used with very low weight), but it usually adds a lot of "trashy details" to the image,

I know, some people want to keep their prompts secret, but I miss the days when everyone used ComfyUi and you could get the prompts from the images. I would love to know some of the styles of the great images posted here

[worthy orbit]

yes, I stop using it as it added bad detail and fought the style. I think after some updates the models can do better than that extensions. It helped with the inicial models, but no more.

[high skiff]

A 2.0 release of my SDXL workflow may be on the way. With native 1536x generations in 9 seconds on a 3090, or native 2048x in 18 seconds. Some sneak peeks at what to expect in quality:

Tiger is native 1536x, and the man is native 2048x

it uses a tonnn of node trickery to get this fast and clean, so some issues are expected

small sneak peek at what the 2048x one looks like node count wise

Also demo'd in this is a sneak peek at a LoRA I am working on for realism in collaboration with a company

trying out a 3,072x workflow

at 3072x, VAE becomes the problem

ok yeah, 3072x does not look worth while pursuing

another one of my 2048x gens, took 19 seconds

its not perfect but its damn good for that speed haha

[high skiff]

at 2048x, you can do batch size 4 on a 3090 and still have plenty of VRAM left over. Full text encoder to 4x images VAE decoded is 63 seconds, or a whopping 15.75 seconds an image

getting 4 results like this, all 4 2048x res

[heady vale]

all except the tiger seem to have unintended noise on the right and bottom of the images. what is that?

[high skiff]

no idea, its solid image diffusion, no tiling

though my new changes I have made since are even faster and higher quality

16 seconds

[heady vale]

small noise on the right of that one too, although less

[high skiff]

I am not sure what you mean

this?

if its that, then thats likely just an artifact from my LoRA

ahhhh, I see it now, yeah

its an artifact from the LoRA

[heady vale]

basically the top half right side

[peak dove]

Werf!!!

Did anybody get Omar92's ChatGPT node QualityOfLifeSuite to work at all? It's on Civitai ...

[high skiff]

holy shit this workflow with my LoRA's is insane

no negative prompt

17 second image gens

[high skiff]

some more gens from a further refined workflow

Please feel free to zoom in and pixel peep

some crops

no negative prompts, no inpainting

The company sponsoring me is interested in potentially using this workflow with some in house mods for one of their projects, so I am not able to share it as of now

[upbeat summit]

looking good and good luck with the project

[high skiff]

Thank you!

Its a company open to sharing advancements with the community, so if all goes well, expect some form released for the public

I have some hopes for changes to get these 2048x image gens below 10 seconds on a 3090. I know, very lofty goal, but I think I can do it

also, this is not only a workflow demo, but a sneak peek at the realism LoRA I am making

[crisp owl]

Not amazing, just liked it as I sat at my pc. From a batch I ran overnight

[hoary saddle]

an old Michelle Pfiefer (she's always been my #1 in "the" list)

[crisp owl]

hehe, loki disguised as an old lady, deceiving

[tawny leaf]

If i use a SD1.5 checkpoint as base, can I refine it using SDXL refiner, or do both checkpoints need to be XL? Basically, my fav checkpoints are still not XL, so is there any way to use the 2 together
(I use a1111)?

[hoary saddle]

in comfy yeah, dunno about a1111

[lusty wolf]

You can do Image to image in SD or in ComfyUI you have one workflow that does it all in one go.

[crisp owl]

Not sure if a1111 would be able to switch vae's like that. Been a while since using a1111 now

[tawny leaf]

Is comfy better?

[lusty wolf]

My answer yes... dpends on the person.

[tawny leaf]

I may give it a try. A1111 lets you select a refiner and at what step its introduced, but what I am unsure of is if models of XL and SD1.56 are even compatible.

[crisp owl]

directly, latent to latent, no

[pallid path]

any alternative to gligen?

there's no model for sdxl

[meager canopy]

SDXL can be better at sticking to the prompt and making higher res images, so why not use that as a base and refine with SD1.5?

[fading nexus]

prompt: highland cows frolicking under a cloudy sky deep in the Scottish highlands

Promt

[cyan crown]

How many steps? (thanks for your work)

[cedar tusk]

Anybody have any guidance on merging SDXL models in Comfy? I'm trying to do an add difference (inpaint - sdxl-base) + (another model), but I'm getting an error

Looks like the error happens when it tries to load the sdxl inpainting model

Nevermind, needed to use unet loader apparently

[rustic garnet]

they don't have to be compatible. you can use the refiner simply as an img2img pass. The only reason why refiner and base are sometimes coupled is to make it a bit faster

[crisp owl]

Did it gen with those faint eyes you can see through the eye holes? Or did you inpaint/photoshop those in?

[cyan crown]

gen without inpaint or photoshop

[crisp owl]

Nice, really like how faint the eyes are

[cedar tusk]

Are you using the refiner?

[cyan crown]

yes

[cedar tusk]

Base SDXL or a third party model?

[cyan crown]

base model with different styles

[cedar tusk]

Wow, looks great!

[cyan crown]

I use xl_more_art to have more stylish images, but also without it's great

I found that using artist or photographer names works better than normal prompts

i..e. this is base sdxl withou art lora and it's still great

this is fun

[cedar tusk]

@cyan crown So I had stopped using the refiner some time ago and want to get back into using it. What's your steps looking like for base and refiner?

[hoary saddle]

setup Comfy, start with the default workflow, break it, fix it, memorize it, clear, built it from start, learn how the nodes connect, what goes to what, slowly add in a lora, an upscaler, then move into loading other peoples workflows

it's intimating at first if you haven't worked with node based software like blender and others, but once you get it, you likely won't go back to A1111, just my 2 cents...and the API is great (not restful like A1111 but very powerful in scripting)

[crisp owl]

I avoided Comfy, then SDXL was about to be officially released soon and I wanted to try it. Haven't gone back to A1111 since

[zinc cargo]

is there a way in comfy to pause the workflow?
or bypass the queue from a different tab?

[crisp owl]

Not sure if this is what you're getting at, but there's this button to push a new generation to the front of your queue

[nimble heart]

SDXL Latent HSV controls because fuck your understanding of color theory.

it just works™️

"but how does it work?"
How do YOU work?
fucking magic.

apparently the CIE guys were 47 years ahead of their time

hypothetically if i mapped out the min and max ranges across the 4 channels it could be used to directly convert to and from sRGB with hopefully not too much chunkyness.

[pure crystal]

Having too much fun with Suno

[lusty wolf]

Glitter bot?

[terse pewter]

hi, i didnt know sdxl can create video image generations too

[halcyon tusk]

here on the bots 1 to 10, you can prompt and then press "tab" and select format video

[shy kelp]

Hi! Question for expert SDXL ComfyUI users -- If you use the SDXL ultimate workflow: https://github.com/BeTheRobotSD/sdxl-comfyui-ultimate-workflow/blob/main/SDXL-ULTIMATE-WORKFLOW-v3_2.json

Have you gotten AfterDetailer to work after txt2img if everything else is disabled? No upscaling, only AD?

I found this issue but the instructions there aren't helpful for me: https://github.com/BeTheRobotSD/sdxl-comfyui-ultimate-workflow/issues/4

[native knot]

Just for giggles:

Steps: 66
CFG: 6.66
Prompt: 666
Model: unsafexl_v31
LoRAs: add-detail-xl: 6.66, wowifierXL: .666, xl_more_art-full_v1: 6.66```

Same as above, but steps: 666:

[native knot]

Mama mia!

[lusty wolf]

Meow...

[tawdry crow]

Sing me a good sea shanty.

[lilac wren]

Guys, look a this :
https://civitai.com/models/176555/harrlogos-xl-finally-custom-text-generation-in-sd?modelVersionId=214296

[native knot]

Yeah...just scroll back far enough or do a search and you'll see that it's well known here.

[rustic relic]

Anyone here with a rtx 4070 and can share Performance?

[glad grove]

https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html here just type 4070 in the search bar

[soft bone]

dreambooth training is amazing. sadly the eyes STILL give me problems. annoying af

[upper totem]

Very interesting ! I'd love to know more and see if you can achieve this state or realism in less "shutterstockish" subjets.

[high skiff]

It can for sure, though the demos I am doing are along that line.

More to come for certain, new tests being run. New tools being made. Hope to bring something truly remarkable out soon

[upper totem]

It's amazing, I'm truly fascinating. It more or less confirms that Comfy UI is more capable than A1111. If one day you sell teaching, I'm more that interested (working in the high end imaging field and theses competences are now highly sought after).

For the moment I'm at the struggling stage to make this look realistic 😄

[mellow tendon]

Buxom

[icy brook]

New LoRA: Aether Aqua!

Makes stuff into water.

https://civitai.com/models/210754/aether-aqua-lora-for-sdxl

[uncut steeple]

Dont try to ping everyone

[icy brook]

that was a mistake

which I instantly removed

[uncut steeple]

We dont let it through anyways

[uncut steeple]

Have you tried feeding those into the video model?

[icy brook]

Nope. Just Runway so far.

[uncut steeple]

Can I try?

[icy brook]

Ofc you can. The LoRA is free 🙂

[uncut steeple]

Meant with your images, I am very lazy

[icy brook]

Go ahead. Share everywhere 🙂

It’s cool if you refer to the LoRA. That’s all.

[stone fossil]

❤️

[zinc cargo]

this is exactly what i was asking. never notied that.

[peak dove]

Comfy Manager - said all was up to date - but I ran Git Pull on 20 different nodes, and 14 updated!!!

[zinc cargo]

does everything work now?

[clever verge]

Dreambooth as in the method with reg images? Or something else?

I think the main problem training at this point is that we need at least the dataset the model was trained on plus the improvements we like to make and the hardware. This without making the model worse at other things as we do now.

Sure we can improve certain parts now but it will most often be at the expense of others.

This is why using reg images works as a mitigation for retaining learned data. I see it as non optimal biased way to handle it.

[crisp owl]

Yeah manager seems to miss a lot of pulls somehow, so I usually go through and manually do them

[icy brook]

How did it go? 🙂

[uncut steeple]

Not so well sadly, doesn’t seem like it understands they are made of water

[icy brook]

Ok 🙂

[livid axle]

┬─┬ノ( º _ ºノ)

[soft bone]

yes the dreambooth tab of kohya.

[soft bone]

this is what makes DB a million times better than Lora

[soft bone]

this is why I use ground truth images and not generated reg

a Lora could never make these

without a lot of help

[clever verge]

You could train a LoRA with dreambooth method so I guess you are referring to both dreambooth and finetune of the full model weights.

[soft bone]

yeah, all kind of a spectrum tbh. ground truth regularization is normally for real fine-tuning (everydream2) so using it with traditional DB settings is already a mixture. but the advantage in quality is definitely from training the full weights

[clever verge]

I just wish we could get the dataset that the model was trained on, it would be neat.

[pine seal]

I am fresh starter with sdxl. what settings or extension should I install to improve performance?

[pine seal]

seems like my speed is about half then others.

[heady vale]

Looking online, I think you have a Ryzen 3500x? If so your cpu is too slow for the gpu most likely

[pine seal]

aren't mostly GPU handle render part?

[heady vale]

it does but also relies on the cpu to feed it the data

[pine seal]

I have 3950x

[heady vale]

its very single thread dependent

[rotund surge]

That cpu is a beast

[heady vale]

its a beast at multithread

[rotund surge]

How much it/s do you have on sdxl? @heady vale

[pine seal]

it take only 30% CPU usage while render it.

[rotund surge]

Have you tried with other model? @pine seal

[pine seal]

even slower with other model

[rotund surge]

oof

[heady vale]

I havent tried that benchmark on sdxl

[pine seal]

now i am getting 1.3 ~2.5 which from 4.0~4.5

well, result is beautiful. however,

[heady vale]

have you tried using comfyui?

[pine seal]

i don't like comfyui style...

[rotund surge]

I think comfyui is faster in sdxl and slower in 1.5

[pine seal]

which is relate with I don't like 3d modelings.

[rotund surge]

Or at least is slower in my PC (i only use 1.5)

[rotund surge]

x2 U.U

[heady vale]

you could try comfy to see if the slow gen is only related to A1111

[pine seal]

I have vladiffusion. would it make different?

[heady vale]

vlad is slightly different than A1111. comfy is different again

[pine seal]

and yet, this is freash installed version of auto1111. haven't optimize yet. and I want to know how optimize the a1111

yea, almost same speed with a1111 and vlad. I am going to try comfy if I can do that.

[pine seal]

even more slow and beside it/s only 20 steps take 78 seconds to render....

[rotund surge]

What driver version are you using? @pine seal

[pine seal]

latest version?

[rotund surge]

This is old, but still, I would try using an old driver

[heady vale]

Maybe ask in #🤝｜tech-support to see if they can assist. If all 3 are slow its something else

generally comfy is faster than the others

[rotund surge]

I have this version, when I installed the 32 one it took like a minute to a single 512x512 image

[pine seal]

ok

[heady vale]

out of curiosity I ran the same settings as I have the 3080 10gb. I am on 546 driver

[hoary saddle]

Anyone know what this style would be called and if there is an additional Lora that would help create this type of image?

Tried vector and a few line art variations.

[ivory blaze]

how is comfyui any relation to 3d modeling? cos it is node based?

[pine seal]

kinda yea.

[ivory blaze]

there's a million different node based things, just cos a few 3d modeling tools use it, doesn't really

it is not at all like the 3d modeling node based stuff anyhow

it is much faster, and really isn't hard to get used to.

[pine seal]

ok, let saying I just don't like node system lol.

[ivory blaze]

so you prefer modular python scripts?

[pine seal]

yes.

[ivory blaze]

that you have to hopefully get into the right directory...

i assume you just don't like the actual process of having to put it all together to make it work, lol

which i get.

[pine seal]

I prefer vhdl ove circuit drawing lol

[ivory blaze]

I mean vhdl does its job, but you still gotta put the circuit together once you have the schems

if you actually building it

so in the end...

[pine seal]

like this one?

[ivory blaze]

plus, just drag someone's json noodles into comfy and be done with it, hide all the nodes under a plate, and just have your inputs, and outputs visible

i'd use a more readable font, but usre

sure*

[pine seal]

yea, for this little lines be ok, but if it goes more and more... my eye will blind on it.

[ivory blaze]

that's not so bad, looks like one of my older comfy workflows.

i just keep adding crap... :\ and it keeps getting bigger. I should probably try just starting all over with it .

[pine seal]

this.... kinda ok. I can do.

but not this...

[ivory blaze]

that's a mess though,

but i gave up when I grabbed a group and it noodled everything all over,

[west breach]

https://civitai.com/models/141592

[dim steeple]

Does using high res fix actually render the image at that higher resolution? Or does it just upscale it?

[west breach]

It upscales it, adds some noise, and then denoises the image using the model to get the final image

[dim steeple]

Okay thanks. Can you recommend a good one to use and the settings?

[native knot]

This image comes with a bit of dialog...read it first, then check out the image:

||"They knew I was dead...
They knew that the Empire had won...
They knew that all hope was lost.

They were wrong."||

[west breach]

I've been using this upscale model, along with denoise value of 0.2 to 0.4 depending on the image. Portrait photos you have to go lower denoise to stop weird things happening. https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/001_classicalSR_DIV2K_s48w8_SwinIR-M_x4.pth

[still veldt]

Aye whats a good generation time for 25 step sdxl ?

[west breach]

5 - 60 seconds, for best GPU to entry level GPU, latest generation

[still veldt]

ok, I'm getting 25s for 25 steps so 1:1 not bad with a 3070 ti

Wish I had more vram for video tho :^(

[crisp owl]

Anyone using an integer eval (python eval) or something that's not from "Efficiency Nodes"?
Seems he's not going to maintain his pack anymore, and it's stopped working

[crisp owl]

oh nvm, just looked in manager and saw this note.

[zinc cargo]

hoous heating is gpu based

[native knot]

Been trying to play with the new Kohya Deep Shrink node. What I haven't yet wrapped my head around is the block_number field. I read the blog entry, but it doesn't explain that. I don't know where else to get more information on it, either.

The results at some points can be pretty amazing. I've had some gens that were kinda out of this world...like a single shot generation using the node that got me to a really detailed 1272x2048 image. But then the same settings with another and it ends up looking like a Picasso. Worst part is that I've had a lot of images come out nearly flawless except for like one eye is a bit too high up and everything else is spectacular, which is pretty peculiar.

[dapper dragon]

Amazing stuff @stone fossil

[stone fossil]

Glad you enjoy broski. 🙂 ❤️

[stone fossil]

Pure txt2img no edits, Paradox 2.0 soon. 🙂

[vale eagle]

https://huggingface.co/kandinsky-community/kandinsky-3

Even bigger than SDXL parameters

[stray warren]

"In comparison to its predecessors, Kandinsky 3.0 incorporates more data and specifically related to Russian culture, which allows to generate pictures related to Russin culture. " 🤔

[eternal fog]

It's funded by Sberbank, the largest, majority state owned, financial institution of Russia, so it's not surprising.

[still veldt]

Anyone know good sdxl model for porn?

[stray warren]

I wouldn't trust their stuff on my pc

[eternal fog]

Same

[still veldt]

so cyute

[shy kelp]

is the bot different from sdxl 1.0

[old shard]

Whats the best checkpoint for realistic doggos?

Xl of course

[lime estuary]

/deram

[pale canyon]

How do you use SDXL styles?

[still veldt]

with your huge brain

you just put in an sdxl checkpoint instead of an sd checkpoint

and have enough vram

[crisp owl]

https://github.com/twri/sdxl_prompt_styler

[low meadow]

What do you guys use exactly?

[crisp owl]

Huh?

[low meadow]

i cant get anything accurate

[crisp owl]

First make certain you're using right image size, 1024 x 1024 variants

These are what it was trained on

[dim steeple]

Can someone help me undestand why my image always looks so messed up when I use img2img SD Upscale?

Is it not compatiable with XL?

[dim steeple]

Which Cany model do I download for SDXL on the list? https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0/tree/main

[peak dove]

how to install huggingface-hub>=0.19.4

[peak dove]

Just started d/loading ... it is a refreshingly diferent look, and has many aspects of Russian Culture available (as stated)

[molten gull]

some pictures from tonight 🙂

yours is nice @stone fossil 👍

[stone fossil]

Urs also love the cat. 🙂

[molten gull]

thanks 🙂

i got some pretty scary ones, too 🙂

this one here is kinda funny 🙂

and one last in portrait mode 🙂

or maybe one more 🙂

[zinc cargo]

we have to thank them for puttin the effort...

[stone fossil]

Bro, no free VEGA?

[still veldt]

Anyone know how to get rid of these artifacts, I started using zendaya lora and it's causing facepaint / glasses to appear constantly

sometimes the facepaint just looks like more glitches, I've tried raising and lowering lora strength but that doesn't seem to change anything

[still veldt]

looks awesome brooo

[vale eagle]

My fine tune, OpenXL, updated to version 1.1.
Mainly improved hands and merged with anime model.
Around 70% good hands, exactly 5 fingers, without post process in 100 generation of waving hands.
Appreciated for image and comments.
https://civitai.com/models/160738

[carmine gale]

Best skin model for extremely high enlargements and creating high frequency skin pore textures and irregularities through the latent space. Testing any models that come well recommended

[still veldt]

Do I need more than 8gb to train an sdxl lora?

[glad grove]

12gb just to be safe but u could probably do it with 8gb but with 768x768 imgs

[still veldt]

hmmm, yeah well I can wait a couple days I guess then. Wanted to start training, but I got a 4060 ti arriving on tuesday

so I wait 🙂

[slender coral]

Do you guys have any suggestions on tiling a upscale in comfy?

I'm close but there's tiling issues

[glad grove]

probably because u are using euler a

[still veldt]

do I load multiple loras like this, or is there a node to do multiple together?

[crisp owl]

Yeah you can just string em together like that

[still veldt]

ok cool! thanks

[bright valley]

There is also nodes that can load multiple at once, if that's what you prefer, such as Efficiency Nodes LoRA Stack

[still veldt]

ty

[crisp owl]

efficiency nodes pack (luciano) has stopped updating his repository, so use that pack as your own risk.

[slim wren]

play around with seam_fix options and bluring the edges

[stone fossil]

xD

[steady grove]

Call him guano man cause he be drippin?

[slender coral]

What would you reccomend? I can't seem to get a good result.

[slim wren]

don't have much more idea than what I've told you already. I don't play much with tile textures nor comfyUI. Try asking in #📝｜prompting-help

[ivory blaze]

pfffft random seeds are overrated.. .

[upbeat summit]

nice - undo/redo in ComfyUI! https://github.com/comfyanonymous/ComfyUI/commit/a6676384423b39dee66bc3833ef3415cd089602a

[vale eagle]

WOW

[night bramble]

muppet brandishing a knife

[bright valley]

awesome work

[uncut steeple]

Couldnt make it work the other way around yet unfortunately

[uncut steeple]

There we go, kind of

[peak dove]

ComfyUI ChatGPT Simple_O - where are those ChatGPT Models stored at all?

[vale eagle]

my fine tune, openxl updated to version 1.2. Mainly adjusted realistic shadow and improved eyes generation.

[zinc cargo]

hey guys who's got a good quick lcm xy workflow?

[high skiff]

Wanted to come in and show off what my latest realism LoRA mixed with my 2048x ultra fast workflow can do. More improvements to come.

All images shown here are rendered in just 22 seconds on a 3090, with single sentance positives, and no negatives at all

still working on some hand improvements, but thats not my biggest focus at the moment. getting ready to try a distructive negative LoRA merge to push it even further

so much more to come from here

[lusty wolf]

Still a secret workflow?

[high skiff]

Yes. The workflow is likely not to be released, as its for the company sponsoring me. I am trying to prioritize making some money for once, as all of this AI stuff I have done has made me almost nothing, and I am severley hurting for cash 😅

Although, there is a high chance of a watered down version to be released at some point. We will have to see how that all goes down

[lusty wolf]

Yes I hear you, I wish I had a company like that... I also have a pretty good workflow going....

[high skiff]

Of course ideally I would continue doing these types of improvements to my own tools and release them free to the public, however at some point I do have to prioritize being able to make a living, and I have basically no money to my name at the moment lol

[lusty wolf]

I hope you make lots of money... I'm sure it is a song. 😂

[high skiff]

The new workflow I have is very purpose-made for a specific task as well, something that the company directly wants, so it wouldn't even necessarily be the most ideal version I could output for consumers

[high skiff]

It also utilizes some custom nodes, of which I am not necessarily allowed to give out in general.

With all that said, consumers can still expect a realism LoRA from me at some point in the future. Of course it is not going to be as good as what I'm offering to said company, but based off of my current tests of where I am right now, compared to all other realism LoRA's I have been able to scrape off of civit, it is already significantly ahead of the competition I have seen.

Perhaps I will even release the version that I am working on right now when I produce higher quality versions for said company

All of the competition I have seen requires several prompt triggers, and obtuse negatives in order to get any form of decent result out of them, whereas my training philosophy has specifically been trying to avoid any and all trigger words, and completely removing the need for negatives at all

As of right now, the realism LoRA I have (which is a combination of 14 different ones I have made over time), is better than any of the major realism models for SDXL by a considerable amount from my testing

in terms of ease of use, lighting, texturing, and various other things (been working for MONTHs on it lmao)

[glad fulcrum]

where can I download the LCM scheduler for auto1111?

nevermind, it seems that updating animatedDiff gives me the scheduler

[supple knot]

Is SDXL Turbo version 512 or 1024 outputs?

[high skiff]

512

[supple knot]

it must be 512 .that why my outputs blury

[high skiff]

@wicked frigate Any info on how to use SDXL Turbo without CFG as suggested in the paper?

[supple knot]

that a confusing naming convention

[high skiff]

perhaps a new comfy node?

[high skiff]

Its SDXL, but smaller and turbo speed lol

oh, there is an SDXL turbo channel

[wicked frigate]

i believe if you just set cfg=1 in comfy it will automatically disable processing of cfg (since mathematically that disables it anyway)

[high skiff]

that is very much not working. I will try the image that is provided

cause any CFG seems to make it get huge and blocky

damn, I alrwady have like 300 SDXL turbo gens... lmfao

[supple knot]

yeah new channel ,I was told its 512 and was confused because its called XL, when it should be call SD 1.5 turbo

[broken crater]

It is the same architecture as XL, so it is XL

[supple knot]

Ok

[eternal fog]

Kohya's Deep Shrink seems to sort of work with SDXL Turbo

10 images in 10 seconds

[hoary saddle]

forget, do we need them both?

[eternal fog]

no

I'm using the FP16 one because it's smaller

[hoary saddle]

haha, was gonna do the same, thought i'd ask first

[eternal fog]

IPAdapter sort of works with it as well

[wicked frigate]

only need the fp16

[eternal fog]

Ok yeah so Kohya Deep Shrink works well if you want stuff that isn't just 512x512

You can play with the settings to get it "close" to what it was trying as well

[cyan crown]

but do sdxl turbo need refiner?

well it's 512x512 so it's a no for me

[eternal fog]

Use Kohya Deep Shrink

Still messing with settings, but these are all SDXL Turbo

[hoary saddle]

i'm only interested in Turbo for 1 thing...discord MSPaint Bot LOL @noble shoal

[eternal fog]

This works really well without even using the Turbo Scheduler

[carmine gale]

sorry, guys, do you know how to get the Add GetNode and SetNode in the context menu?

[carmine gale]

looks like this https://github.com/diffus3/ComfyUI-extensions/blob/main/setget/setget.js but not working

[eternal fog]

You can use the ModelMergeSimple node with Turbo and other models to pull to quick render time across

Regular model 4 steps

Mixed with Turbo, 20% Turbo, 80% Model

then 40-60

[heady vale]

This is LCM lora at 5 steps. Of course these are cherry picked because LCM creates a lot of incoherent images. Some are OK

[eternal fog]

LCM Lora

Mixing with Turbo

both took the same amount of time

Changed the mix values a bit and made it less blurry

[sharp vale]

this actually looks pretty good

[cyan crown]

working on a new Lora. I think it will be ready tomorrow

[crisp owl]

"beasts in bars"

[cyan crown]

it's a gothic lora not focused on monsters

These results are with 50% of training.

[crisp owl]

very nice dreamy look to them

[cyan crown]

thanks

[crisp owl]

I've just been playing around today stuff like this.

[candid jasper]

that's crazy

[still veldt]

man turbo is insanity

zoom

0.2 second generations on a 3070 ti

Anyone figure out how to upscale it without deepyfrying it?

[eternal fog]

If you are using Comfyui try using Kohya Deep Shrink. It lets you make larger native images

[still veldt]

this ?

also do loras work with this? trying to load one and I don't think its working

[native knot]

No. It's in the built-in nodes. Look under "_for_testing".

[eternal fog]

This

[still veldt]

ty

also wdym look under _for_testing ?

[eternal fog]

If you right click to add a node, it's under that category

[still veldt]

there's categories??

sorry I started learning like 3 days ago

[eternal fog]

Yeah if you right click in space you can go to add node and then there's a big list

[cedar tusk]

I noticed that in Comfy you can now switch between 1.5 and SDXL without changing the conditioning nodes. How does it handle SDXL parameters like POS_G vs POS_L and the height/width & target height/target width? Are those no longer necessary? Is there a benefit to including them still?

[crisp owl]

All these ninja speed things lately don't do anything for me as they all reduce realism quality too much. Great for artistic or anime stuff, or just testing, but I can't use em

[still veldt]

ninja speed?

[upbeat summit]

some TurboVision XL images. @delicate kelp already fine-tuned the SDXL Turbo model with amazing results. Already does higher resolutions, coherence and fidelity.

1344x768, 5 steps, cfg 2.5, dpmpp_sde_gpu karras (not using the ComfyUI Turbo scheduler). takes ~8 secs on a 4090 per image

[still veldt]

now that's popping

[upbeat summit]

trying some latent manipulation next - sharpening, contrast, cfg rescaling

[still veldt]

can you run this in comfy?

[cedar tusk]

@still veldt Yes,

See here for Comfy workflow

[still veldt]

sorry yeah got it, I just had some wrong stuff plugged in

its pretty good, but can't get my lora to work with it let met try some different ones

[vale eagle]

I guess it just merges the model with turbo with limited release hours

[crisp owl]

I'll have to try the finetunes, didn't realize there were people doing stuff off the base of it, thanks for the insight 😊

[upbeat summit]

I mean SDXL Turbo was just released like only a couple of hours ago 😄

[crisp owl]

Socalguitarist on it like hotcakes

[upbeat summit]

yeah we did some experiments earlier

[shy kelp]

i already have SD 1.4 installed locally, how do i add SDXL to that

[upbeat summit]

and it's already much improved because of the fine-tune

[shy kelp]

ive been trying to look this up myself but im pretty confused

[crisp owl]

What UI are you using?

[shy kelp]

uhhh webui?

[crisp owl]

Automatic1111?

[shy kelp]

i think so?

how do i tell?

[crisp owl]

Screenshot

Been a while since I've been in it lol

[crisp owl]

Just a min, heading down to pc soon

ok, yeah that's A1111, so you just need to download an SDXL model (from CivitAI or wherever you prefer) and then just select that checkpoint.
You can use the refiner option as well if you want, or just do 100% from the main checkpoint.

[upbeat summit]

Using PixelWaveTurbo by @west breach. Also a SDXL Turbo model.
1344x768, cfg 1.8, steps 6, dpmpp_sde_gpu - takes about ~7-8s for one image. less in a batch

[crisp owl]

@smoky patrol

[bleak relic]

Hey guys

[crisp owl]

hi, get a life. blocked

[still veldt]

That's dope, make the bricks yellow and ya got wizard of oz

[bleak relic]

I can try it

[still veldt]

niceee

[bleak relic]

This looks, well.... interesting 😅

[still veldt]

looks epic mayne nice

also kinda like cheese

[bleak relic]

True haha

But i would recommend to not eat it 😁

[broken crater]

Do you want to share your workflow on these? For some reason with this model I am getting really bad images even if i set the same sampler and scheduler. Edit: All good, I was using dpmm_2m_sde_gpu sampler, that gave bad output

[peak dove]

I am impressed by the speed of SDXL Turbo - and slightly disappointed - as I often enjoy waiting until the end to see the finished product!! 😄

But that's just me!!! LOL

[west breach]

just increase the batch size to make it go slower 😄

[stone fossil]

https://media.discordapp.net/attachments/1097191636448055370/1179157218285863132/image.png

I'm looking for a flow that does this, anyone?

[vale eagle]

why do you mix base 1.0 with turbo?

[native knot]

It looks like a subtract, not an add.

[vale eagle]

oh, didn't aware that is subtract

[stone fossil]

If anyone could share would be very much appreciated.

[eternal fog]

What's the purpose of subtracting base SDXL from turbo first before adding juggernaut?

[vale eagle]

To minimize the sdxl base weight affect in the merge\

[delicate grotto]

did anyone with 6gb vram manage to run sdxl-turbo yet?

[west breach]

works on my laptop with a 6gb 3060

[eternal fog]

I've been using the modelmergesimple node with 80% of the weight on the model I want and 20% on turbo and that works well.

[stone fossil]

Anyone can share a workflow for it?

[vale eagle]

I tried above subtract method. It seems better than directly merge with turbo

[eternal fog]

I'll have to have a look tonight

[stone fossil]

Can you share your flow?

[vale eagle]

directly merge required more steps to get a decent result, like 10

[eternal fog]

I was getting good results with 6

[vale eagle]

merge with subtract keeps 4 steps decent

[eternal fog]

Much lower and it would get blurry

[vale eagle]

yes, I could but it is just an implementation of your graph

[stone fossil]

Would be very cool.

If you like drop it here or dm it.

[rocky cloud]

hi i would like to generate a modern bedroom without touching the view on the window. what is wrong with my workflow? can anyone teach me please

[topaz sleet]

I want to combine SDXL inpainting with a model controlnet but I get this error:Given groups=1, weight of size [320, 4, 3, 3], expected input[2, 9, 128, 128] to have 4 channels, but got 9 channels instead
here are the models I use:

CONTROL_MODEL = "diffusers/controlnet-depth-sdxl-1.0"
VAE_MODEL = "madebyollin/sdxl-vae-fp16-fix"
MODEL_NAME = "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"

is this combination possible?

[bleak relic]

My feelings in a nutshell

[delicate grotto]

great! thank you!

[peak dove]

Turbo SDXL takes about 90 minutes to d/load!!!

AUTOMATIC1111 SDXL via Dynavision plus Refiner and ADetailer.

[tender island]

anyone have any tips on why my LoRA isn't showing up in SD webui? I have the correct base model selected to apply the LoRA to but nothing 🤨
(lora trained on sdxl using dreambooth)

[cyan crown]

well dumb question.....lora is in the right folder?

[tender island]

Yup, same folder as the other Lora I downloaded from civitai

And my Lora is valid, I've tested it in the colab I used to train it

[tender island]

(the other one DOES show)

[cyan crown]

maybe a problem of output name?

not filename

just output name embedded in the lora

if you have 2 loras with same outputname

it doesn't work

[tender island]

Oh I see, I'll check that when I get back

Cheers!

Hopefully works

[cyan crown]

I had similar problem and was that

[floral island]

1 second render -> join the LCM TURBO merge hypergang!

[vale eagle]

does it really improved by merging LCM + TURBO?

[floral island]

that's my model, merged with lcm + turbo, 6 step render

it's fckn amazing

[vale eagle]

why would need to merge LCM if Turbo is enough to improve speed?

[floral island]

increased detail/contrast at the same render speed (this was just testing the possiblities)

[vale eagle]

I think the left one is better

[floral island]

well, because the right one didn't have the right setup yet

that has a bit too stronk lcm lora weight

but the only two which i didn't change other settings

[vale eagle]

what I get from lcm is improved speed but downgrade of quality. It is why I really considerate of merge it with another tool which also improve speed and downgrade of quality.

[floral island]

one with, one without lora

[vale eagle]

I still prefer the left one

[floral island]

uhoh

you selected the lora

[vale eagle]

Interesting

[floral island]

lets do it again

the lora doesn't lower quality. it improves quality, because you can get MORE resolved in the same amount of steps

meaning more detail, less errors

not to mention, increased contrast

however, you do need set it up well, and i'm not decided yet if base model has influence too, if the lora works well or not

[vale eagle]

hmm, other than speed. The lora learned some extra thing

[floral island]

it's basically weights, that the lcm sampler can use

still, i'll take any "free" upgrade i can get

[vale eagle]

Thanks for explains.

[floral island]

that's why non-lcm sampler cannot use strong lcm lora weights, but the lcm sampler can -> it uses the weights presented by the lora a lot better

[vale eagle]

I personally would not merge the lora for the quality improvement. But it actually improved.

[floral island]

nah, i'm in the process of merging 1.3 of my model

i won't merge in the lcm lora either -> if people want to, they can load the lora just fine

but that way, it can still be used without lora, for traditional sampler

once lcm lora is loaded, a lot of normal sampler just won't work properly anymore

[cyan crown]

So lora finished

[crisp owl]

private use or posting?

[floral island]

not for me, "it just works" for whatever it is i'm doing

[cyan crown]

will post on Civitai after dinner

[floral island]

just don't merge into models

oh, youMADE a lora

[cyan crown]

eheh yes

[floral island]

sorry, interpreted the remark wrong

[peak dove]

SDTurboScheduler missing in my ComfyUI SD Turbo?!?!?!

[eternal fog]

If you don't have it you've not updated

You don't need it anyway, it works with the regular KSampler

I'm right on the edge of getting nice outputs at 3 steps

[floral island]

6 for mymodel, but i'm doing stuff it doesn't like that much (anime)

[vale eagle]

great job

[cyan crown]

https://civitai.com/models/216323?modelVersionId=243787

Have fun

[hoary saddle]

i have a square image as an image2image input, is there a node or a set of nodes i could use that would add some space on both sides of the 1:1 image to create a more landscape image?

is that what this is for?

i just wierd stuff like this

[fierce hollow]

that's because it's just filling the padded area with black color

you should use noise or gradient or something, no idea if there are nodes for that