✨｜sdxl

[boreal bough]

depends. style loras remain mostly unhurt, as long as the negative prompt doesn't cause it to go to extreme lengths (though you can just remove the negative input in refiner if this is happening)
existing concept lora remain either completely unhurt, or mostly unhurt - assuming a denoise value 15~25%
some specific concepts are hammered into the refiner (face/hands/body proportions/anatomy) - so those get obliverated in even just a few refiner steps - for those situations base only is your only option

^ also this is only true for default sdxl usage. if you're using sytans workflow, then you're out of luck - since it adds extra noise, which kills all lora options other than style, or prettier concepts

[sharp robin]

this is pros and cons of refiner, wished we didnt go this route even tho sometimes refiner ads that pizzas.

[quasi remnant]

let's goooooo. no need for standalone vae anymore xD

[sharp robin]

[proper goblet]

Where can I get sdxl_base_pruned_no-ema and sdxl_refiner_pruned_no-ema?

[ember cargo]

this one has the same workflow with ultimate upscaler (Iceland lora is mine I trained it today with my own photos) @sharp robin

[soft zealot]

Hmmm I have some questions to come back to there when I get off the sofa

[wicked frigate]

it's SDXL 1.0 but without the older vae that some people prefer baked in

grab the new copy of SDXL^

[quasi remnant]

i didn't even need to zoom in most of the time. i noticed it on nearly every pic, was very visible around edges

[sharp robin]

thank you much appreciated, but in the end it appears that if we wan lora basically we forget ab refiner, THIS picture is stunning btw, thank you for shairing it, so pretty

[ember cargo]

I know, i used the same WF with a person lora, and just a few refiner steps was *almost * fine, but like made the person many years older lol and thanks!!

[spark bear]

so its 1.0 with .9 vae and 1.0 licence for both?

[dark jacinth]

I have a Winston too he is smaller and more white hair

[sharp robin]

REFINER does leave a ton to be desired, trying out person lora and its like tossing a grenade into itjust base works well surprisingly in fact

[ember cargo]

I suppose we need some tuts on how to make a lora for the refiner

[sharp robin]

turning into such a mess, and barrier to entry for older cards that much higher

[leaden marlin]

What is a Lora?

[spring fulcrum]

What is condensation

[dense chasm]

two prompts combination is interesting like # prompt will be passed to OAI CLIP-ViT/L-14
prompt = "award winning photograph of elephant"

prompt_2 will be passed to OpenCLIP-ViT/bigG-14

prompt_2 = "award winning photograph of octopus"

[dense chasm]

Our experiments were conducted on a single 40GB A100 GPU. https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sdxl.md

[quasi remnant]

wait did the upload get messed up? how can they have the same sha256 and i still get chromatic aberration T_T

[wicked frigate]

huh?

[dense chasm]

Lora asks for high VRAM and GPU

[wicked frigate]

@hard fractal that's the wrong file lol

[quasi remnant]

and i still get bad vae on it ;_;

[swift acorn]

Hi. I have few newbie questions.

• I need some custom VAE to SDXL? Link please, if yes.
• Why SDXL is so popular? Why this is not like every other model? Some tl;dr?
• Can I render this well at GF 4080? I need to know smth imporant about this model?

[midnight shuttle]

Confirmed files are identical. Either a mistake or a placebo A/B test.

[midnight shuttle]

Why SDXL is so popular? Why this is not like every other model? Some tl;dr? Better training set and larger native image size.

[quasi remnant]

i'm sure it was an accident, i don't even need to zoom in to see something went wrong with my gens xD i didn't even think to check the hash or anything til i saw it mess up T_T

[indigo carbon]

\

[boreal bough]

@wicked frigate
could you maybe ask about this internally?
Cause currently this is the only newbie friendly option I feel confident posting on reddit/civitai :/
like training settings will only get you so far without a significant amount of knowledge if you go base only + lora.

#✨｜sdxl message

[wicked frigate]

simple mistake yeah, please hold, getting fixed atm

[wicked frigate]

you can merge loras into the base model and the train new loras atop that

[white iris]

Does anyone know what the style setting in dreamstudio does? For e.g. when selecting realism. Is it a positive/negative prompt? And did they ever show what the prompt/negative prompt is when that style is selected?

[wicked frigate]

it's pinned in this channel

[swift acorn]

Can SDXL create NSFW content? Didnt test it before.

[wicked frigate]

^

[eternal fog]

Are you guys able to tell us why the new vae is the way that it is?

Just curious of the change

[quasi remnant]

i think devs aren't allowed to discuss why; unless that's been lifted.

[spark bear]

im sure there will be a way to train the refiner reliably at some point, might even be beneficial, you can train the refiner with a cherry picked subset of the dataset from the base lora with only the highest quality/best textures

[white iris]

Love this. thanks!

[eternal fog]

It's a bit odd, because it just looks like they've done something wrong with the new one. But if they aren't allowed to talk about it for some reason I'll leave it.

[quasi remnant]

yup; latest I know is they are not allowed to discuss it. don't think that policy has changed.

[sly jay]

why is the refiner doing this? entirely changing face and imo making the image look much worse

[quasi remnant]

the only worse part i see is the vae over the face, making eyes look weird

[eternal fog]

Either too many steps, or you are doing a straight img2img and you aren't feeding the noise from a partially finished base generated image

[quasi remnant]

i think the face structurally looks better in the 2nd one at least

[eternal fog]

It's got that over smooth look

The refiner does that if you just img2img with it

[earnest herald]

Maybe it thinks thats a kimono and is inferring that she should be more asian

[quasi remnant]

oh

they're using automatic1111

i recognize the png name format

[hard fractal]

oh oops

telling Robin to reupload now

[azure oxide]

no way LMAO

[sly jay]

how do I do that in a1111? i know how to in comfy but am setting up a1111 on a more powerful rig

[quasi remnant]

i think automatic1111 doesn't really have a way to use refiner yet @sly jay

comfy is the best UI for sdxl imo

[eternal fog]

You can't do it in A1111 at the moment. It's incapable of using the refiner in the correct way.

[violet fossil]

anyone have a good comfy ui sdxl setup i can use to study. I'm new to using it.

[quasi remnant]

mine is gonna change once the sdxl 1.0 0.9 vae weight thing comes out but otherwise it currently looks like this with a vae loader:

[eternal fog]

So many steps

[quasi remnant]

i want fully denoised images

sytan's workflow doesn't give enough detail for meeee

[wicked frigate]

ps this a good demo of occam's razor. It could be a secret conspiracy testing on unsuspecting users... or somebody just clicked the wrong file lol.

we're all human yall

[sly jay]

I know but i hate using comfy lol :/

[quasi remnant]

and there's been like charts & stuff showing around 50-60 steps of 2m gives best quality from what i've seen

[azure oxide]

(insert conspiracy theory of secretive, evil greedy corporation here grumblegrumble)

[wicked frigate]

StableSwarmUI can use refiner!

https://github.com/Stability-AI/StableSwarmUI

[wicked frigate]

you might love swarm owo

[eternal fog]

I think it's decent at 20. But I've chosen 40 for now for a balance between the quality I want and speed.

[sly jay]

is it similar to comfybox?

[boreal bough]

rip those few users I told to get the new model :/

[eternal fog]

Just another 6gb to download. Let's see if they even notice.

[spark bear]

at least we wont have 100 people a day asking why their gens look weird and how to load .9 vae

[boreal bough]

prob not. but poor interlaced mouths of whatever they generate

[quasi remnant]

i... noticed in the very first pic i generated and again i don't even need to zoom in xD

[wicked frigate]

kinda similar ye. Not the same

[quasi remnant]

if there's lots of whites or like edges on the picture

you notice something "off" immediately

[sly jay]

does it have image to image, impainting. Those are the two I miss most from a1111

[quasi remnant]

cuz the whites turn rainbow

and the edges turn really jagged

[eternal fog]

There's plenty of people that somehow don't see it

[wicked frigate]

.> <.<
image2image, yes, via a placeholder Init Image field.
inpainting, not yet :(

[shy kelp]

Hi. Mind if I ask what's a more powerful rig? What's the specs?

[wicked frigate]

that's definitely something i want it to have, but it's gonna be a big project to do that right

[azure oxide]

if this is in reference to just the UI, its essentially the same thing for end users. you can ignore the comfy nodes and use the new UI

[orchid oar]

Shocked Monster as it sees New York City

[wicked frigate]

i'm not of the mind to settle for the bare minimum inpainting interface auto has

[sly jay]

3090ti, its not mine lol

[shy kelp]

he doesn't look shocked. looks more upset. 😄

[azure oxide]

stablestudio has a beautiful inpainting gui

[orchid oar]

true, but that was the prompt i gave sdxl

[upbeat tundra]

Is there a comfy UI workflow that has Loras and upscale? The ones on civitai either don't have lora or are broken

[orchid oar]

This one looks more shocked

[wicked frigate]

yeah dude one of these days ima convince @knotty lotus to cross-integrate studio and swarm

[shy kelp]

slightly more shocked yes

[boreal bough]

I think kaj's ui already has inpainting working, right?

[orchid oar]

There are not many loras till now for SDXL, thats why i think

[boreal bough]

or is that still in preview mode

[quasi remnant]

i'd love some kind of marquee selection, because circular brushes don't always get the exact area I want; as a random suggestion

[sly jay]

ahhhh so many ui's

[eternal fog]

Setting up a LoRA is easy. It's just the LoRA loader node. And you feed the model and clip through it.

[boreal bough]

It's gonna be a while until it gets simple.
If you wanna learn it properly, start with comfyui, since all the cool ui's use that in the background - so any knowledge you gain here won't lose its value in the future

[upbeat tundra]

What's offset lora and why are people uploading UIs with that instead of simply load lora?

[shy kelp]

Man....I worked my butt off to get a used 3090, built a 5800x3d and 64GB ddr4 system, a few TB of storage, just so I'd be ready for sdxl 1.0. And as it turns out the reality is that I need you guys to help train it. It's a good model but it needs training. It will take a few months or sooner depending on the ambitions of certain nerds 😄

Geeks not nerds....to be more accurate.

[quasi remnant]

also does stableswarmUI have support for the refiner in the generate tab or do you have to use the comfyUI workflow editor? i'm looking at the example screenshot and i don't really see a way to use refiner? ._. https://raw.githubusercontent.com/Stability-AI/StableSwarmUI/master/.github/images/stableswarmui.jpg

[eternal fog]

The offset LoRA was provided by stability, it would be selected in a load LoRA node.

[boreal bough]

watcha making?
anything 5k or under should be done in 2 days or less

[finite socket]

I am newbie here 🙂 Is there any good tutorial for beginner?

[shy kelp]

Nothing really important. I wanted to dabble in robot soldier design and femtosecond handheld laser systems. You know....to take over the world with. Was hoping SD could help with that. 😄

[sly jay]

i've used comfy I just don't like the workflow

[boreal bough]

https://github.com/Stability-AI/StableSwarmUI/releases/tag/0.5-Alpha
get this, let it install everything, then use the default settings for now - will get you through the first few days of usage, by then you'll know yourself where to continue

[swift acorn]

Can SDXL creater NSFW content? Or its censored model?

[finite socket]

Than you so much!

thank

[sly jay]

sdxl is kinda hard rn because unless you have a really beefy PC you have to mess with comfyui which can be difficult

[shy kelp]

From what I can see.....yes. But it needs training.

It's like a gifted child that doesn't know anything yet and needs to go to school.

[azure oxide]

#✨｜sdxl message based on this i think we have to use the custom comfy workflow for now

[shy kelp]

But it's gifted

[boreal bough]

oh right! we did the prompt search for the cool mech soldier 😄

[shy kelp]

like that Chris Evans movie with the little genius girl

[quasi remnant]

even if you have a beefy PC comfyui is better (than a1111, stableswarm uses comfy and therefore is equal i believe?) i can do batches of 3 images on my 3090 on comfyui but only 1-2 on a1111.

[boreal bough]

yeah, definitely doable, without any big issues

[shy kelp]

Oh you remember that. 😄

[white iris]

The reason why i asked, is cause i noticed theres a difference between the output from dreamstudio, and the one from clipdrop. I used the prompts from the pinned .txt file but it still doesnt come out as good for some reason. Not sure what the reason is.

[boreal bough]

you just need to get the captioning right. but training will be the easy part - I can give you the settings for whatever dataset you have, once the captions are done

[sly jay]

comfy is faster, a1111 is much easier

[shy kelp]

Yes.....SDXL 1.0 is the little genius girl from that Chris Evans movie. Gifted, bored from all the dumb people around her. But still doesn't know about life yet. She needs training.

[boreal bough]

since you have a 3090, your settings will be identical to mine

[knotty lotus]

slightly different wrappers (same core though) & different hardware

[quasi remnant]

faster and more efficient.

efficiency is huuuuge. a1111 takes like 4-5x longer

or it might just be 2-3x longer but still xD

[boreal bough]

ah perfect! you're online c:

since multiple users keep asking for inpainting,
is your UI ready for that? Can I refer them to it?

[quasi remnant]

For basic inference, anything involving txt2img and upscaling, comfyUI is not difficult, you can use a community workflow.

[wicked frigate]

it's below init image, so scroll down (or git pull if you haven't in a while)

[civic sigil]

Would you be willing to share your settings with me too?

[quasi remnant]

A1111 is more convenient for img2img/inpainting

[boreal bough]

sure!

[still dove]

can you share that workflow we discussed?

[civic sigil]

Just got my 3090 build 🙂

[visual glade]

inpainting is horrible on a1111

[sly jay]

but no impainting or img to img no?

[quasi remnant]

i totes just covered that xD

[wicked frigate]

that's outdated as of earlier today :D

[knotty lotus]

StableStudio? it's not fully integrated with img2img or masking just yet. very busy with a ton of projects interally currently, will probably have time to spike that out this weekend.

[boreal bough]

just depends on your dataset/caption style/if its a concept or style

[azure oxide]

niceeee

[quasi remnant]

well, it's good enuff. i like the interface with the painting & what not instead of uploading masks, though I wish there were different selection methods.

[shy kelp]

What!!!! pray tell, is the advantage of workflows/nodes in the creative process here? Why should I spend any energy on comfy?

[boreal bough]

alright. thanks for the update ❤️

[bold osprey]

img2img is pretty intuitive but I didn't seem to get to an inpainting workflow

[quasi remnant]

aight, i'll keep that in mind if I try it out sometime.

[sly jay]

where? am i being dumb brained lol

[quasi remnant]

here @sly jay

lol

[civic sigil]

Yeah I usually only do full finetunes for styles, loras for subjects

[bold osprey]

you can chain multiple checkpoints and set up a feedback loop if you want to

[sly jay]

damn I am too dyslexic for this shit lmao

[bold osprey]

loads of stuff surpassing a1111 limitations

[white iris]

What do you mean by wrappers?

[quasi remnant]

btw comfy DOES have img2img, but for inpainting you need to use masks

[shy kelp]

It negates the need to merge custom checkpoints?

[quasi remnant]

Regular img2img works fine, but you need to put pics in a specific folder i think

[spark bear]

still waiting for everydream to come out with a 1.0 trainer..but there's no way we'll be able to do full finetune with te on a 24 gig card, so will have to see how unet only looks.

[quasi remnant]

So it just comes down to convenience for a1111 for that specific part

[shy kelp]

Interesting!!!

[boreal bough]

styles are currently by far the easiest to train in sdxl. even if you get most of the captions wrong it still works. loras now work amazingly for it.
concepts are fine as well - as long as they dont touch the face

[sly jay]

so is a1111 the only impainting option? I mean hypothetically I could gen in comfty img to img in a1111

[visual glade]

you can do live checkpoint merges: https://comfyanonymous.github.io/ComfyUI_examples/model_merging/

[bold osprey]

well it's more like img2img but you can also define areas that your prompts affect and have much more control over it

[hasty smelt]

Hi everyone, I'm facing an issue. The bot achieves a much superior result, very close to what I imagine. I'm sure it can still be improved, as I've done before. However, when using SD with the deliberate_v2 and lora:Drawing models, the outcome is far from what I expected using the same prompt. I would like to obtain a model or lora capable of achieving a similar result to SDXL, which is excellent with vintage illustrations.

[runic lark]

How many model.safetensors are you supposed to download? 4th time in 2 days about 20 gb now each time

[quasi remnant]

i'd still prefer using comfy for img2img. but inpainting is a whole other thing

[visual glade]

you can do inpainting in comfyui by right clicking the image in the load image node and "open in maskeditor"

[orchid oar]

You need a higher resolution i think

[shy kelp]

Live on the spot merges.....okay okay you got my attention. lol 😄

[quasi remnant]

oh, i didn't know this

[azure oxide]

if you want to do this then i suggest you use invokeai instead of a1111, invokeai's inpainting UI is by far superior over a1's

[orchid oar]

What model are u using?

[boreal bough]

bot is running sdxl 1.0 - which is why the results are so different

[orchid oar]

The Bots use sdxl 1

[sly jay]

okay! will try!

[bold osprey]

damn I wasn't aware

[ember cargo]

the image contains it, just drop/load it into comfy

[bold osprey]

feels way better than a1111 because zooming in doesn't reset your ui

seems more responsive too

[tender timber]

Thank you! My 3060 12 GB model does 1024x1024 32 steps at about 1.5it/se, so they seem similar. Even though the it/s is always 1.47-1.48, the runtime ranges from 27.45 to 59.65 seconds. I'll check out the 6900 XT line, or the new ones, I'd like to take advantage of larger VRAM.

[leaden marlin]

Is it possible to go from sketch to ai image in comfyui

[sly jay]

anyone have any good starter workflows for sdxl in comfyui? specfically for using refiner and inpainting

[visual glade]

and you can right click copy (clipscape) and paste (clipspace) images from the save image node to the load image node

[shy kelp]

le sigh I was hoping that learning comfy wasn't going to be necessary. But perhaps I need to schedule in some learning time.

[bold osprey]

this helped me get into it https://www.youtube.com/watch?v=OdMtJMzjNLg

[boreal bough]

comfyui is absolutely worth it to learn - as you can do some pretty creative things
will take a while to get used to though, as it doesn't provide shortcuts, and expects you to do things the right way - and also gives you settings that you shouldn't touch or combine

[azure oxide]

im sure theres popular shared workflows around already for the basic functions we're used to, you might not need to learn anything and just download those lol

[quasi remnant]

so what I'm hearing is that comfyui just totally eclipses automatic1111 in nearly every single way xD (except maybe for things like adetailer), still lots I gotta learn

[hasty smelt]

Does anyone know of a model or Lora capable of performing well with vintage illustrations like SDXL does?

[wicked frigate]

i didn't know that was a thing, awesome!

[orchid oar]

Why not use SDXL?

[eternal fog]

There are custom nodes for adetailer, although they are a bit complicated to setup.

[boreal bough]

comfyui can do everything XD just not very fast if you have a blank slate

[quasi remnant]

setting up my first workflow that i'm still using & refining today didn't take that long tbh.

[hard fractal]

Today I learned.

[sly jay]

the more I learn about comfy ui, the less I wanna use a1111, even with the gap in ease of use

[quasi remnant]

yeeee

i seriously had no idea inpainting was possible

[shy kelp]

FINE!!! But I'm not bothering with invoke though! lol

[hard fractal]

ditto 😐

[hasty smelt]

What is SDXL, and can I already download the model?

[orchid oar]

Yes

Its Stable Diffsuion XL

Its a newer Model

[bold osprey]

so this fills the area with the average color. do you need separate mask nodes so you could use a1111 inpainting options like latent noise or using the inpainted area as a reference?

[sly jay]

sdxl is stable diffusion but bigggg(and better)

[quasi remnant]

wait a little bit i think. they're uploading the 1.0 weights with a better vae

[wicked frigate]

you want comfyui with more ease of use? https://github.com/Stability-AI/StableSwarmUI

[boreal bough]

^ also read this!
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

[shy kelp]

Dude walks in on the launch party of the space shuttle, it can still be seen in the air in the distance flying to mars, but he asks what's a space shuttle. 😄

[bold osprey]

didn't convince me either from the looks of it

[quasi remnant]

i think the vae weights are coming soon™️ thouuuugh xD

[sly jay]

does this has impainting? If not I am just gonna bite the bullet and learn comfy

[wicked frigate]

it has comfy inside it, so technically yes

[visual glade]

for true inpainting use the VAEEncodeForInpaint like on: https://comfyanonymous.github.io/ComfyUI_examples/inpaint/
for img2img with the mask use latent->inpaint->Set Latent Noise Mask instead

[quasi remnant]

it has comfy built into it

[bold osprey]

I am way too nerdy to pick the easier option

[still dove]

it did not work

[sly jay]

if it has comfy built in why not use comfy?

[wicked frigate]

why not use swarm* you mean

[boreal bough]

It still needs a new user guide XD
I hope your new prompt book comes with an additional swarmui book - which will be userfriendly enough so that even a first timer can gen a pretty raccoon

[hasty smelt]

I'm sorry, I thought the beta version was available only for selected individuals. Thank you very much for everyone's help; I'm already updating my files

[quasi remnant]

oh how far we've come... invokeAI used to be the one that was the best/everyone used, then automatic1111's webui took over, and now comfy/stableswarm xD

Invoke AI used to be Lincoln Stein's version IIRC?

was like the best cmd line version of SD

[bold osprey]

was there ever a colab version for invoke?

[quasi remnant]

idk

[ember cargo]

it does, i do it myself. Be sure to open the original image

[bold osprey]

looked like a1111 was the most prominent one for colab

[sly jay]

also for swarm do I need a regular comfy install. I already have one if that is helpful

[quasi remnant]

swarm installs it automatically i think

[azure oxide]

a1 was always good enough that i never swapped off to try invokeai but i still always daydreamed about invoke's inpainting ui whenever i tried doing so on a1 😭

[shy kelp]

We need more YouTube teachers that can explain this stuff in a user friendly way.

[still dove]

how can I find the original image? 🙂

[bold osprey]

especially the comfy community needs to be more vocal than it is

[wicked frigate]

you can use your existing one, or install one automatically, your choice

[visual glade]

if you want to use SDXL + refiner on free colab, the comfyui colab notebook works: https://github.com/comfyanonymous/ComfyUI#colab-notebook

[sly jay]

cool beans

[vale eagle]

What captioning tool is recommended for lora training? Blip?

[boreal bough]

ufff. on one hand I want this. on the other hand the spam & false info outperforms the quality
even if interest goes up, it just causes the false info to increase exponentially as well

[bold osprey]

I can gladly run it locally but having a colab is nice too!

[shy kelp]

Here's your assignment: Build a giant....and I mean a GIANT space station above earth. Something that is designed to service starships with crew and tenants in the thousands. Go!

[azure oxide]

idk vlad fanbois were so vocal to the point that they were increasingly toxic towards anyone not using it lmao, i dont want anyone to become like that

[eternal fog]

For best results, manually.

You don't need many images to make a decent LoRA now.

[wicked frigate]

sebastian kamph said he might make a video for swarmui (you should go tell him you want it to make sure he does :3)

[orchid oar]

Will do

[hasty smelt]

Guys, I'm downloading the XL model. I wanted to know about the available styles in the bot. Are they already available for download?

[bold osprey]

lmao I meant in the sense of sharing and explaining

[shy kelp]

I can't disagree. Most of them are like 20 year olds trying to create traffic with very little value.

[bold osprey]

https://drive.google.com/file/d/1IZq_0CGTbfxlAdIMsjz3VwRzHjl3JL-n/view
maybe you are looking for something like this

[quasi remnant]

#✨｜sdxl message

[vale eagle]

hmm, using common sentence to descript the image?

[quasi remnant]

these are the styles used by the bot

[shy kelp]

Maybe. 😄

[eternal fog]

Describe them the same way you'd prompt for that image.

[quasi remnant]

it's just additional prompt stuff

[vale eagle]

Ok, got it.

thanks

[crude sparrow]

anybody have other fintuned xl models i can use?

[sly jay]

oh yeah do styles work for comfy/swarm?

[bold osprey]

well damn

[quasi remnant]

you can save workflows with styles

[wicked frigate]

yep! They're called presets because you can configure more than just prompts with em if you want

[shy kelp]

"any other"? is there already a custom model using sdxl 1.0 as a base?

[wicked frigate]

if you have a .csv you can just import it directly in the Presets tab of the ui

[crude sparrow]

dreamshaper xl (in alpha)

[quasi remnant]

and yea; like, you don't wanna use same cfg/sampler/scheduler for an anime image as you would a cinematic photographic image for instance

[spark sinew]

thanks a lot for help and sharing,

[wicked frigate]

if you like anime there's https://huggingface.co/hakurei/waifu-diffusion-xl

[quasi remnant]

so you wanna save more than just the prompt, so saving the workflow/presets makes much more sense

[sly jay]

neeto

[hasty smelt]

thanks buddy

[shy kelp]

I'll look for it on civitai.

[sly jay]

where can I find some starter presets?

[bold osprey]

here

[sly jay]

also thanks for all the help everyone here has been super helpful!

[quasi remnant]

gonna be a bit before the 1.0 finetune is out. the 0.9 is good at pushing towards anime style but I think you can get better anime-style images in SDXL with proper prompting for now.

[orchid oar]

if you want anime fast positive:" anime artwork {prompt} . anime style, key visual, vibrant, studio anime, highly detailed"negative:"photo, deformed, black and white, realism, disfigured, low contrast"

[hasty smelt]

Guys, thanks for all the support. I'm downloading the XL now.

[sly jay]

good luck!

[quasi remnant]

i actually don't really like the default anime prompt from the styles list very much; usually ends up with messed up faces/some other features being a bit messed up

[tight crown]

Is this recording going to be uploaded? 🙂

[wicked frigate]

#✨｜sdxl message

[quasi remnant]

meanwhile you can get something like this with a super simple prompt xD

[orchid oar]

Yea the faces are not quite as good. For me its enough with the refiner

[boreal bough]

damn those look legit XD

[orchid oar]

Niceee

[shy kelp]

I'm going to immerse myself in comfy tutorials over the next few days.

[sly jay]

oh I already have those on my a1111 styles! Is there anything I need to do specfically to get that working with swarm?

[quasi remnant]

that anime style won't get u to this level

(last one is 0.9)

[orchid oar]

Is this waifu diffusion xl?

[quasi remnant]

No

it's sdxl 1.0

except last one which is 0.9

[boreal bough]

just use a preset someone made, and go from there. for any questions feel free to ask here ❤️

[quasi remnant]

pure 1.0/0.9

[wicked frigate]

just grab the styles.csv file from auto and yeet it into the Presets -> Import

shove the file in and click import

[boreal bough]

@shy kelp here's a good comfyui setup to get you started,
made by the lovely @quasi remnant ❤️

[orchid oar]

The workflow looks interesting. Im still new to Comfy so i just made myself a simple one with examples. Good job👍

[quasi remnant]

the workflow won't need the vae loader soon™️

[orchid oar]

Thats quite hard (for me)

[quasi remnant]

just waiting on this to become official xD

[boreal bough]

XD

[brazen patrol]

is there a recommended range for how many steps one should use the refiner relative to the base? i.e. 20% more steps with refiner or other?

[boreal bough]

#fake-news

[orchid oar]

(not the best)

[bold osprey]

I am using 32+8

[frozen sundial]

Why is my automatic1111 running super slow with SDXL? >90s/it
I was running SDXL just fine yesterday with normal speeds (<1s/it)

[brazen patrol]

for example, out of curiosity I tried 40+10 with 2m, then tried same seed with 40+30 and second was much worse, so wondered

[boreal bough]

same steps as base, but denoise value between 0.2~0.5
smaller for less changes, bigger for more detail, but at the cost of some other details

there are other ways to use it, but this is a good starting value for base+refiner

[orchid oar]

Can you share your prompt for these?

[quasi remnant]

it should be attached to every single one

just open the pic in a new tab and drag it to comfyui

[orchid oar]

I tried, the kimono girl worked but the figures not. Wait a sec im gonna try something

[quasi remnant]

o

the figures

is one i'm doing rn

just a simple prompt

OHJ

because

yea i know why that doesn't work

sorry xD

[boreal bough]

fyi, you can use https://discord.com/channels/1002292111942635562/1100167411493257218 as a source of inspiration for prompts

[orchid oar]

Thx <3

[quasi remnant]

i was testing comfy's new launch argument... lol

the disable metadata xD

[orchid oar]

Ahhhh

Thats why

[quasi remnant]

i was seeing if it saves any time on gens or anything

turns out i might save 0.1 second an image.

...not worth it tho just for a timesave lol

[quasi remnant]

also i was seeing how much less bloated it'd make the file size

but i think it's just like .1 mb

or not even that

a lot less

xD

[orchid oar]

Thats interesting research

yea its only text so imagine

150 Steps was too much

[quasi remnant]

o yea this isn't like a normal model

you

uhhh

[vocal stream]

a classic in 1.5 when doing higher height images, I guess training in multiple aspect ratios hasnt helped that much

[quasi remnant]

the refiner makes it so you don't ever wanna do too many steps i believe

you wanna pass it over with some noise left over, and have refiner finish last 20%

i found 48->60 to be pretty decent overall

[orchid oar]

Im normal doing 25 and i like a picture i set it high, sometimes to max like in this picture

[quasi remnant]

now the nendos have their metadata xD

[orchid oar]

Thanksss

[quasi remnant]

i c ur using a bad vae in that image

[orchid oar]

It works now

[boreal bough]

I wonder if there's any speedup from using the highest end ssd for model loading XD

getting my 12gb dataspeed ssd tomorrow 😄

[orchid oar]

I thought i was using the offical one

[quasi remnant]

are you connecting the vae node from your refiner/base model?

[orchid oar]

Indeed

[shy kelp]

Helluva good start! The two on the left is what I was after especially the one in the lower left.

[quasi remnant]

yes, that's the issue

[quasi remnant]

set up a load vae node, and use this https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors

for any vae decode

[boreal bough]

props to good note usage

[orchid oar]

i copied the workflow lol

[quasi remnant]

ye, it's the official workflow

[orchid oar]

I just added somethings

[quasi remnant]

official workflow doesn't have a load vae node

[orchid oar]

Thx will do that

[orchid oar]

Glad you like it

[quasi remnant]

put that in models/vae btw

just a note since you said new to comfy i think

xD

[zealous viper]

Didn’t they update the models on hugging back to the .9? Just get the retrofitted model

[quasi remnant]

Not yet.

[brazen patrol]

I tried and for me at least this was way worse than a small refinement. Here are three images. first is 40 base + 10 steps refinement. Second is 30 steps and third is 40 + 40

[quasi remnant]

When they do it won't be necessary

[shy kelp]

Now was that done on comfy?

[upbeat summit]

very cool!

[orchid oar]

Thanks, im new to Comfy but i know sd. Still thanks👍

[orchid oar]

Yes

[shy kelp]

Okay, gonna invest the time and energy into learning it. 😄

[quasi remnant]

They are doing this but it's not ready yet; they accidentally uploaded the same weights (as in no changed vae) at first.

[orchid oar]

Best luck👍

[quasi remnant]

np; soon this won't be necessary, they're releasing the weights with 0.9 vae baked in soon.

but for nowwww

we need a load vae node xD

[zealous viper]

Ah thought I saw it posted earlier it was up, MB

[orchid oar]

I already had that vae installed lol, just didnt use it

[vast narwhal]

Thanks. Those are made 20 days ago from 0.9 tho

[visual glade]

one trick for doing anime with SDXL is using something like WD1.5 as a second pass model

[quasi remnant]

lol

[quasi remnant]

Oh?

[brazen patrol]

so this is 40+10 refinement, then 40+30 refiner and last is 40+40

[orchid oar]

Didnt know that, thanks👍

[quasi remnant]

How many steps do you give it?

[last crane]

hello there, I just stared using stable diffusion and I want to make some awesome art and my quesiton is how can I make something like this image:

[quasi remnant]

Do you still give the second pass model just like 20% of the steps or do you give it more since it's a completely diff model architecture?

[zealous viper]

Taking that to mean load a 1.5 model instead a refiner? Same steps?

[quasi remnant]

i am very interested in this xD

[visual glade]

in that workflow I latent upscale then do 8 steps with cfg 12 dpmpp_sde

on WD1.5 illusion (which is a SD2.1 768-v model)

[quasi remnant]

Oh is the workflow on that?

[visual glade]

yes

[quasi remnant]

is this part necessary? pokerface.png xD

[visual glade]

not really

[quasi remnant]

ooo ic u give it about 35% of the steps

so for me... hmm that'd be...

[orchid oar]

Now using the vae 👍

[sly jay]

ngl still unsure what the point of stable swarm is. If I make workflows in comfy and use comfy, why would I bother with adding another layer on top of that? May as well just learn comfy fully

[quasi remnant]

oh and then you also use a different prompt for 2nd part (danbooru tagging instead of nat lang)

[visual glade]

yeah completely different model so that's why

[quasi remnant]

yup

Oh is this more of an img2img?

[visual glade]

yeah kind of

[quasi remnant]

ahhh, no wonder it works so well

i'll try this out with the anime models i like to use

[clever verge]

If you do a batch_size larger than 1, is it possible to generate one of the images again? Seems like all images have the same seed.

[visual glade]

so that way you can get the way better overall composition from SDXL with the style and details of your favorite anime checkpoint

[quasi remnant]

o lol you named your vae nearly the same thing i did

[quasi remnant]

infact i probly could've just... not changed that at all

xD

mine wasj ust capitalized i think

[orchid oar]

Your Workflow is amazing

(The Model i used was just bad)

[visual glade]

you can lower the cfg I use on the second pass a bit

[high skiff]

@visual glade Do you happen to have a way to implement DDIM inverse into comfy UI? I have seen some papers on how it works, and I think it could be extremely beneficial to my upscale workflow

[visual glade]

the unsampler? https://github.com/BlenderNeko/ComfyUI_Noise#unsampler

[orchid oar]

Thx, will do that

[high skiff]

I am not too sure, but I will have a look. Thank you!

[fleet harness]

Thank you to all who share workflows! Now I have Mega Persian Cat X lol

more or less lol

[spring fulcrum]

What should the cfg be for the upscaler? can that be lowered as well or should that maybe even be higher than say 7.5?

[quasi remnant]

Nice!

[orchid oar]

Cant wait for Loras so i can try nendos of saber or something

Thanks for sharing these pics

[strange thorn]

Does anyone have a ComfyUI workflow for SDXL 1.0 image2image + refiner?

[quasi remnant]

first quick attempt, looks reaallly good.

[wicked frigate]

#🐝｜swarm-ui owo new channel get!

[high skiff]

its about time lol

[quasi remnant]

ty for sharing comfy <3

i see lots of potential with this since sdxl can make more varied scenes

[fleet harness]

again, thank you people for sharing. Now My Beamer is cruising Tokyo in style

[orchid oar]

SDXL sadly doesnt really know the mercedes c123 coupe :(

[fleet harness]

still works xD

[hearty ginkgo]

It also dosnt know car carriers

Hardest thing ever to reproduce

[orchid oar]

Yea sadly :(

[mossy lodge]

How do I select DPM++ 2M Karras in comfyUI?

[static prawn]

i dont know why my sdxl images always look do pixelated its weird, even more than 512x512 in 1.5

[uncut fiber]

@static prawn probably you have not sdxl vae in proper folder?

[static prawn]

i have, its super weird

[sly jay]

depends on your work flow but should be the sampler and the scheduler

[uncut fiber]

can you post some image?

[static prawn]

i cropped this

shows it really well

looks weird pixelated more disorted than in 1.5

sure 1024*1024 gets pixelated in crop

[uncut fiber]

are you using hires?

[static prawn]

but its way more messed up than on 1.5

nono i dont use

but it looks messed up in general

[uncut fiber]

yes

[sly jay]

SDXL is really good at drawing pretty things but seems kinda meh at doing what I want

[mossy lodge]

I put this node for samplers, sorry I'm not using stable diffusion right now but in the sampler_name there is not DPM++ 2M Karras and other similar samplers.

[autumn forum]

Omg this is great, I miss my e30’s 😦

[zealous viper]

Did you switch to the .9 vae? Instead of the baked one?

[fleet harness]

[autumn forum]

Is that a model/lora or sdxl output?

[boreal bough]

you sure the denoise is set correctly there? O_O cause those look like denoise set to 1

[ionic dragon]

Can we now inpaint in SDXL?

[static prawn]

using the new 1.0 vae but i tried both

[fleet harness]

SDXL 1.0 with 0.9 VAE only

[brazen patrol]

These are the settings I have:

[autumn forum]

Wow it knows e30’s so well already that’s insane

[sharp robin]

i love how that prompt can generate something completely different with just a few variables, nonetheless thank you for ideas.

[brazen patrol]

lol. I have had fun testing all settings using the exacts same seed, to be sure that was not a factor, and yes, a lot of variety

[zealous viper]

Use the 1.0 model but make sure you manual set the vae to .9 like mastadon said

[quasi remnant]

another example with comfy's anime workflow

[quasi remnant]

daaaang it's suuuch a difference

switched into using kl-f8 vae loader last step tho

(yea there's the random hand but ... lol)

[brazen patrol]

Same prompt, same seed

[static prawn]

as i said i tried 0.9 vae and 1.0 vae , its always kinda pixelated

[paper phoenix]

whats the deal with this

[static prawn]

have no idea what im doing wrong

[boreal bough]

ah you're using sytans workflow.
yeah that reintroduces noise. good for some styles, less good for others.
either max 60 steps, with the second one ending at 60
that or changing samplers to one that's more friendly towards artwork, like dpm2m

[quasi remnant]

Wrong one was uploaded.

They are fixing it

[paper phoenix]

aaah, cheers, tysm

[quasi remnant]

The base model has right upload right now if you wanna use that

[paper phoenix]

the right vae baked on the base model, yeah?

[vernal cloak]

The same for me too btw

[quasi remnant]

yes

[zealous viper]

Using a1111 or comfy

[brazen patrol]

I can disable the noise. Do you have a different recommended workflow?

[vernal cloak]

I've also seen multiple images uploaded here with the same pixilation issue.

[vernal cloak]

ComfyUi m8

[rustic garnet]

I mean, just use the standard workflow

[visual glade]

do you get pixelation with this workflow? https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[rustic garnet]

sytan got great results with his workflow, but it is totally unclear if it improves different style of images

[vernal cloak]

Errr I will try this out. thx btw!

[brazen patrol]

my only purpose was to enable the Refiner

[rustic garnet]

in particular this "linguistic" promt is highly experimental

just do either img2img with refiner or stop the txt2img early and continue with refiner

[static prawn]

i have absolutely no idea what im doing wrong bec. of thr pixelation

[brazen patrol]

THat sounds like a massive hassle

[static prawn]

im using auto1111 but also used comfy with

same results

always a pixelation

[rustic garnet]

its how the refiner is supposed to be use

[paper phoenix]

link to a pixeled image?

[vernal cloak]

Are we seeing the same thing? Like Edges of objects are not being anti-aliased ?

[boreal bough]

the official workflow. umm ask around, dont have a link on hand

[brazen patrol]

Except his workflow does the second automatically. That is its purpose

[rustic garnet]

no it doesnt

you have to build the workflow once and then you can use it as often as you want withouzt hassle - same like sytans workflow

[brazen patrol]

The first is base, the second is refiner

[zealous viper]

@vernal cloak @static prawn you guys are using comfy's base workflow then adding the vaeloader node to the end right?

[rustic garnet]

the latter works as following:

1.) you use the KSampler (advanced). You set the number of steps to K and the "end_at_step" to K-(Kstrength). You set "return with leftover noise" to enabled
2.) You add a second KSampler (advanced). You set "start at step" to K-(Kstrength) and "end_at_step" to "K". You set "return with leftover noise" to "disabled" and connect this second KSampler with the refiner model

the other variant is:
1.) The first KSampler works normally as without refining
2.) The second KSamper is same as above, but with "add noise" is enabled

[vernal cloak]

yeah. I've also experemented with image to image and also used the refiner, still getting blocky edges on objects. It's very minimal in some cases, but on some results it's rather noticeable.

[rustic garnet]

both methods work but might look slightly different

[brazen patrol]

which is what this does, no?

[rustic garnet]

This is the variant: "1.) you use the KSampler (advanced). You set the number of steps to K and the "end_at_step" to K-(Kstrength). You set "return with leftover noise" to enabled
2.) You add a second KSampler (advanced). You set "start at step" to K-(Kstrength) and "end_at_step" to "K". You set "return with leftover noise" to "disabled" and connect this second KSampler with the refiner model"

my point is rather that the "linguistic prompt" is very strange and I would be careful with that. For some images it looks better for other images it looks worse

[brazen patrol]

I am failing to see how your description does not match the image

[zealous viper]

dm me the workflow with a saved prompt and seed your using ill look at it and see if im seeing anything different then with mine.

[rustic garnet]

I'm sometimes a bit worried that so many people treat linguistic prompts as canonical, although they are totally different from how SDXL is supposed to be used

[vernal cloak]

ok m8. one moment.

[thick mist]

which one is the good one?

[zealous viper]

the newer one

[boreal bough]

the refiner, in theory, is an expert at generating the last 20% of the required steps. Hence, this hack option makes use of that, to significant success. It has some (few) downsides, but for any workflow where you don't experience those downsides, its the superior method to go

[zealous viper]

While im thinking of work flows, for Inpainting atm is it just better to run a 1.5 inpainting model

[rustic garnet]

erm, its exactly what I described lol

[hard oar]

Is there a recording of the SDXL announcement from the other day? I wasn't able to listen to much of it, but it seemed like there was sharing of of some details that would be useful to know. I checked Stability AI's YouTube channel, but didn't see it there.

[rustic garnet]

really? I heard that SDXL is really good in inpainting, but haven't tried yet

[brazen patrol]

ok, I was scratching my head thinking: what am I missing? lol

[static prawn]

idk i mainly use auto, but also used different workflows with comfy

[zealous viper]

i tried it a couple of times and it was kinda jank

[zealous viper]

i stopped using a1111 when it took 5mins to run a single image with 8gb vram and it took 30 seconds in comfy

[rustic garnet]

For refiner you can use both, img2img workflow or the early stopping workflow. But note that img2img takes a bit longer, but has the big advantage that you get both images (unrefined and refined) which is very useful when you sometimes experience that the refiner makes things worse

[zealous viper]

havn't messed with vae for a1111 with xl so i dont have an exact workflow

[boreal bough]

its because its using the advanced sampler, which hides the denoise value, and shows you the raw steps.
Essentially 1 - ending divided by starting step = denoise %

[static prawn]

found a fix for that, since i use -medvram i went from 20 minutes to 2 minutes so im fine with auto 1111 now

[brazen patrol]

Look, I'm willing to change this if there is a benefit. However I don't really plan to tweak the denoise from image to image.

[boreal bough]

the thing is you're using a highly optimized workflow - which brings out peak performance for photographs, and some artstyles
but as a downside, it ruins some other artstyles unless you change it to a less optimized workflow

if you just want the best 'average' workflow, then use the official one provided by comfyui. it will give you the best possible results if you wish to make no changes to the ui itself

[zealous viper]

have you updated your model to the one posted 45 mins ago on the hugging face page, listed with the .9 vae. if not do so without trying to change vae run it with the correct one baked in and see if that problem goes away

[tender timber]

@visual glade Thank you so much for placing notes in your updated workflows (https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), I know they'll be really helpful as I go through and learn this system!

[boreal bough]

this is the official workflow btw
https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[static prawn]

always having the problem, even with 0.9 but i can try again

[boreal bough]

is it the new new 0.9 now? xD

cause they accidentally uploaded the same 1.0 vae again

[zealous viper]

reupload was an hour ago assuming that its correct atm

[shy kelp]

any a1111 experts? Cause comfy seems to have taken over. On the latest 1.5.1 a1111 build I don't see the drop downs for the VAE or clip skip. Did I do something wrong?

[vocal stream]

look in settings, you can enable the dropdown from there

[brazen patrol]

ok, thx

[boreal bough]

those need to be added via quick settings, if memory serves me right

[elfin flare]

Question.

Should I download the normal 1.0 base

or the 1.0 base 0.9 vae safetensors model?

[boreal bough]

the new one with 0.9vae

[elfin flare]

Thank you.

Do I need the refiner downloaded at all?

I have no clue yet how to set up SDXL on Vladmantic's.

[thorny frost]

can somebody do a furry octopus?

i cant

[molten gull]

i try furry octopus

gimme a moment

[elfin flare]

Wait a second.

Where even does the refiner model go in automatic vladmantic's WebUI?

[inner ruin]

what's a good scheduler for SDXL 1.0? I prefer fast over super good!

[thorny frost]

that is the most furry i can get. Some tiny white hairs

[molten gull]

1 second 🙂

[elfin flare]

Wait, do I even need to download sdxl_vae if it's incorporated in the main model?

[molten gull]

funny 🙂 aint so easy 🙂

[elfin flare]

Or is it incorportated.

I'm confused.

[shy kelp]

Hi guys, does anyone uses a gtx 4090 card. I want to know the it s of it

[thorny frost]

lol, it isnt. it is a very weird thing the model cant even imagine

[molten gull]

i m trying 🙂 and yes, got a 4090 here @shy kelp

[shy kelp]

Is it takes like 5-6 sec for you?

[wary bay]

people are suggesting you use the 0.9 model's VAE with the 1.0 models

[shy kelp]

To generate a image

[wary bay]

the 1.0 vae is packaged into the 1.0 model so to use the 0.9 vae with it you'd need to separately download it

[elfin flare]

https://huggingface.co/stabilityai/sdxl-vae/tree/main This?

[elfin flare]

What's the link to the 0.9 vae?

[wary bay]

I don't have the link lol someone here probably does

[molten gull]

in the standard 512x512 test i get about 40it/s

[shy kelp]

with xl with 1024x?

[molten gull]

nah 🙂 that wouldnt be a 512x512 test, i guess

i dont even know the standard test for 1024

[shy kelp]

Oh can you test a 1024 image with XL 😭

I try to find best card

It takes with 4060 17 seconds

[soft bone]

the new "sdxl-base-1.0_0.9vae" or whatever is the base model with the correct VAE baked in. If you get it, you do not need to also get sdxl_vae.

[shy kelp]

with a100 it takes 3 seconds

[elfin flare]

Thank you for telling me.

[wary bay]

they added a version with 0.9 vae baked in? Last I saw they were saying it was all subjective 😋

[molten gull]

at 1024x1024 i get 3.3it/s

[shy kelp]

Thats nice

[thorny frost]

Nice!

[shy kelp]

Around 7-8 seconds is it

I get 1.6 with 4060

[sand pier]

same with 4090, 3.3

[soft bone]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

added 1hr ago

[molten gull]

7-8 sec aint a value that makes sense, since it depends on the number of steps you go

[sand pier]

it takes a bit longer if you use a refiner as well though

[shy kelp]

4090 is good

It is close to a100 I guess

[molten gull]

a100 got a lot more vram, so thats a big plus there

[fair tartan]

are you renting or buying cause the a100 is like 5x the price

[sand pier]

i'm curious how the 4060 ti 8 vs 16 gb benchmarks

[shy kelp]

A100 is like twice faster than 4090

[fair tartan]

doesn't make sense to look for few second savings

[shy kelp]

Price performance seems 4090 wins

[visual glade]

A100 should get about 8.5it/s on SDXL base at 1024x1024

[molten gull]

if you do stuff with LLMs you may run into problems that require >24gb vram

then an a100 makes sense

[visual glade]

and that's with just basic pytorch

[wary bay]

Nice, does this change it back to the 0.9 license?

[molten gull]

i got indeed trouble with a furry octopus 🙂

[thick mist]

why are my images looking like this?

[sand pier]

using an upscaler?

[noble shoal]

Jesus Christ. Well its a furry octo-thing

[sand pier]

seen results like that with it

[molten gull]

can you post your comfyUI setting @thick mist ?

[thick mist]

i dont as far as i know, they get generated likle this

[thick mist]

i use automatic1111 1.5.1

[violet merlin]

what a monstrosity... now if you'll excuse me i'll be in my bunk

[soft bone]

nope

[molten gull]

can you post your a1111 settings ?

[thick mist]

How can i export them?

[sand pier]

ctrl + s

downloads json

[molten gull]

screenshot would be enough:)

[thick mist]

this?

[noble shoal]

Might as well post that in #1019361238234443776 😄

[void skiff]

What they say "SDXL 1.0 can make amazing good images", what I get when trying.

[sand pier]

anything that involves a beer is inheritly good

[soft bone]

pretty amazing imo. artists used to get payed big bucks for work like that

[zealous viper]

took me a second to make sure seeds were locked down but just so everyone is on the same page. new baked in .9 vae appears to be working without the artifacts noticed with the 1.0. Know this sounds obvious but,

this is the 1.0 model with a crop showing the artifacts (the slight green and pink line under pure black parts)

same shoulder with the 1.0 with .9 manual loaded

same part again with the .9 baked vae

[sand pier]

most important thing to do with new model

[lilac wren]

Guys, how do you activate in the settings to have this at the top of the page?

i don't remember

[zealous viper]

just select it from the menu and it loads isn't it ...

[soft bone]

isnt it called quick settings

[zealous viper]

oh nvm misread ... yeah something like that

[lilac wren]

ty bro

[vast narwhal]

Settings, user interface, quicksettings

[lilac wren]

Yeah, thank you, its for a friend

[wraith apex]

what promt. looks very nice style and its dreamshaper or base xl?

[hearty ginkgo]

what are you guys using for the cfg for comfy ui in sdxl 1

[autumn forum]

6 Base 6 refiner works well for me

[wary bay]

has anyone done any comparisons between outputs on 0.9 vs 1.0, same seed and settings?

[soft bone]

anyone using x/y grids in comfy? the only way I've found to do it is crazy complicated

[zealous viper]

Man I don't think I can thank stability enough for these new models, It's so much easier to prompt now

[soft bone]

I like 4. But I've heard 2/13 is useful. 8/8 works good for noobs

[sand pier]

12 gb

6 gb base, 6 gb refiner

[wraith apex]

whats the prompt?

[azure oxide]

grab the bottom one

[zealous viper]

it's not getting the hatchet right yet

i lost what i had were it was giving me a melee arm, properly

[wary bay]

1.0_0.9

[soft bone]

1.0_0.9 is new

[sand pier]

if model updates yes

[soft bone]

do you mean your auto1111 webui?

[fair tartan]

Is there a commanline flag I should be adding when running comfy w/ XL on linux? I'm getting 18+GB ram usage compared to 6gb ram usage when I run the same workflow on windows. The VRAM usage is about the same just the system memory differences.

[sand pier]

yup

[vast narwhal]

Base XL and revanimated (1.5). It's a workflow comfy himself posted few hours ago

[sand pier]

and ouch

[soft bone]

rip i just got 600mbps

[wraith apex]

oh mix of those and can mix 1.5 wow

[wraith apex]

cool

as im trying base sdxl and seems underwhelming

[vast narwhal]

The workflow is on the image

[azure oxide]

ya comfyui is cool in that you can use more than 1 model in your workflow

[wary bay]

Can I host my comfyui instance on a gradio link like with auto

[wraith apex]

sadly im running on runDiffusion, because i have AMD GPU and that does not even run on automatic1111, as says not enough vram with 8gb 😄
and confyUI no support on windows for AMD as far as i understood

im thinking to buy 3060 12gb for it 🙂 but testing in RunDiffusion now to see if i like sdxl

[sand pier]

how is the 12 gb cards working with the sdxl model?

[wraith apex]

nvidia vs amd 😄 1.5 i runned fine, sdxl does not want to run, no matter i add lowvram, medvram, set 1024 or 512 😦

[drifting sigil]

Alright, no one asked but I made a clean comfyUI workflow inspired by others that may or may not be good (didn't test it that much)... If anyone is interested

[sand pier]

other then no refiner, it looks nice

[drifting sigil]

There is though

[sand pier]

oh i see it now

you just take both to same

[tender timber]

Sure, please share. I like clean ones over noodles of lines. I tired to load the PNG file from here into Comfy but it did not work

[drifting sigil]

yeah its really compacted

[sand pier]

getting used to comfyui 😄

[drifting sigil]

it's okay, i'm pretty new to it too haha

[slender timber]

so clean

[raw prairie]

Looks cool

[sand pier]

and pelase share 🙂

dump that stuff to json

[tender timber]

Ditto (new to Comfy, not SD) , don't understand 95% of it but slogging through

[drifting sigil]

Oh yeah yeah I don't know if the image works so here it is:

[raw prairie]

Thanks

[drifting sigil]

The parameters may not be the best, I didn't really know what was good

[sand pier]

thanks

is there any way to disable nodes in comfyui?

[slender timber]

Ctrl + M

"Mutes" the node

[sand pier]

thanks 🙂

like stuff like upscalers i usually use for testing

[pure hazel]

Thought they fixed the 1.0 vae?

[lilac wren]

I guess ControlNet is not functional with SDXL yet?

[wary bay]

I haven't heard anything about 1.0 vae being fixed but I was told they added a version of 1.0 with the 0.9 vae now

[hot rover]

I really hope some workflows get shared for segment anything masks and face restoration, those elude me right now.

[craggy valve]

is there anything like controlNet that takes in vectors instead of images?

[fleet harness]

https://tenor.com/view/lmao-deep-fried-mad-raging-gif-14229167

[craggy valve]

like instead of controls needing to be 512x512x3 images, could they be just 128 dim vectors etc

[rose vigil]

What does the experimental Lora do?

[vocal stream]

that's how GANs work, no idea what you are trying to do and actually asking about though

[shell oasis]

wheres my 5 steps is plenty gang at

[craggy valve]

for example you could make a network that encodes character designs as a feature vector

and some sort of control network that takes in this vector and diffuses images of the character

so even if you had a single image as example you could recreate that character as if you had a lora

[slender timber]

they are right in front of you, you might just need to squint

[shell oasis]

lmao

[vocal stream]

I dont know what makes you think that would be better for handling single images

[wary bay]

I think that's the offset lora for adding contrast back in?

[rose vigil]

Thanks!

[slender timber]

helps w/ text too according to the stage yesterday

[wary bay]

👀

good to know

[fleet harness]

When SDXL walks in

[wicked frigate]

#🐝｜swarm-ui SwarmUI can grid your comfy stuff

[quasi remnant]

spent a while modifying the workflow comfy posted earlier for my own purposes but it's sooooooooooooooo good for anime.

[slender raft]

How is swarm to get running? Is the Linux flow any good?

[zealous horizon]

highres workflow with 3 LoRAs on BASE
BASE -> REFINER -> BASE (variable upscale) -> 4x upscaler

(this is just for testing purpose, to include multiple LoRAs and perform upscaling while keeping the subject recognizable (even improve it)

[still prairie]

I know dreambooth cant use sdxl today but can you create loras in automatic1111?

[molten gull]

use https://github.com/derrian-distro/LoRA_Easy_Training_Scripts for loras @still prairie

[spice patio]

Anyone had luck generating 16:9 aspect ratio images with SDXL 1.0? It generates weird + broken images on wide size.

[raw prairie]

Make sure it is exactly 1 mega pixel

[spice patio]

Noted. Let me focus on that part. Thanks 😄

[arctic bloom]

what resolution?

[raw prairie]

Try 1366×768

[sweet fog]

I used 1333 x 749.

[spice patio]

I tried 1920 x 1080

[spice patio]

Perfect

[still prairie]

Isnt that link Kohya, not automatic1111

[sweet fog]

I used excel to solve the equation of with two values in a 16/9 ratio multiplying to 1 million, and 1333 x 749 was the solution

[static prawn]

anyone having that pixelation issue too?

or know how to fix it?

[zealous viper]

did you try with the new model?

[arctic bloom]

hmm, what kind of pixelation issue, like jpeg artifacts?

[sweet fog]

I actually used the 1333x749 on the 0.9 version of SDXL

[static prawn]

especially on the mouth it does some rainbow like artifacts

i dont even had that on 512x512 with sd 1.5

[alpine pine]

Im getting this issues aswell

It may be the VAE?

[zealous viper]

bruh we've said get the new model off the hugging face site that was uploaded like 2 hours ago

[static prawn]

i used 0.9 vae and 1.0 vae

[spice patio]

Perfect. Trying that as well

[static prawn]

will try it, sorry i didnt get all the messages bec. i was away for some hours

[arctic bloom]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/tree/main

[alpine pine]

ah!

[static prawn]

but i had the problem with 0.9 aswell

[zealous viper]

all good, this one is the one you want

[sweet fog]

they updated the 1.0 model? I guess I'll try downloading that

[zealous viper]

if this fixes it there was an issue with the vae not loading even when specified in your UI

[static prawn]

will try it directly and report

[zealous viper]

but the artifacts your talking about have been identified as the vae

[static prawn]

mh i also had this problem with sdxl 0.9

[violet merlin]

When training LoRAs, do you expect the loss value to go down steadily? In my current training, it seems to be hovering around 0.08 without dropping much over time. Is that a sign that the LoRA is doomed or not necessarily?

[static prawn]

do i just go for none vae then?

[azure oxide]

training is destructive

[fresh path]

am I stupid bc wheres the download button

[static prawn]

bec. its baked in

[pure hazel]

Hmm I'm getting images fried to a crisp using SDE sampler...

Used to work great in 2.1

[slender timber]

i gave up with sde

[zealous horizon]

yeah, I get those too.

[arctic bloom]

press on file name and

[zealous viper]

the model in the image i posted you

[fresh path]

I want to download all the files

[static prawn]

huh?

but thats the full model

[zealous viper]

yeah

[static prawn]

so i choose vae -> none

[zealous viper]

it has .9 baked in

[arctic bloom]

git clone https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/

[zealous viper]

vae the one baked into the model

[azure oxide]

click on this down arrow to download it

[static prawn]

yes i know

thats why i choose none 😄 okok

[zealous viper]

oh yeah

slightly distracted mis read

[static prawn]

nono dw 😄 , im sometimes overwhelmed by all the messages in general

[arctic bloom]

1600:904

[static prawn]

im happy i found a fix for the 20 minutes images in auto 1111, now its fast

i have nearly the same speed as in comfy now

[sweet fog]

There are t wo files, use the one that states it was updated 2 hours ago, not the 3 days ago version

[candid spindle]

How do I make this human hand plant look less photoshopped? Any good workflows for something like this?

If you try to grab the flower, it grabs you

[fresh path]

thanks

[static prawn]

i hope those comfy features especially with the multiple prompts will be added to automatic 1111 too in the future

and a better use of the refiner

[zealous viper]

@static prawn what was the fix, prob not gonna use auto again because I got comfy set up in a nice setting and I'm needing to learn to get comfortable with node based programing. But so i can pass info along

[static prawn]

i just used -medvram

[zealous viper]

k

[static prawn]

i never got any memory erros

but it got super slow without -medvram

went from 20 minutes to 2-3 minutes now, which is normal for 1 image for me

[static prawn]

i wish i could use it too but the unstructured way its displayed is just kinda cancer in my eyes , it disturbs me as hell

[zealous viper]

yea, i wanted to switch into using invoke because i liked the integrated canvas but I couldn't find a way to run the normal launch options if they even exist anymore

[static prawn]

and i dont get how i can add all those custom nodes and stuff

[zealous viper]

just move em around to look pretty. and you just drop the custom node git code into the folder ... pretty easy

[candid spindle]

clone the custom node repo into the "custom_nodes" folder

[static prawn]

but my biggest problem is i dont understand what all this extra stuff is doing and i dont find a detailed or simplified explanation for it

[static prawn]

yeah it starts with i dont know what a scheduler is

for example

in auto 1111 u just have samplers, which are together with the schedulers

and i feel it kinda weird that there are 100s of workflows and everyone brings a different result, so u never know if u got the "best" one

[autumn forum]

Maybe read a bit on this? https://stable-diffusion-art.com/samplers/

[static prawn]

read a bit of that, but i still dont get what the benefit is combining euler with karras now

[boreal bough]

because its all new territory. everyone in the community is still experimenting, and there is no definite guide - since a lot of settings change in combination with other settings.
but the best 'explanation + good simple working worflow' will be the one from official comfyui, which has as much documentation as possible, without going overkill on details
https://comfyanonymous.github.io/ComfyUI_examples/sdxl/
(just save the photos and drag them into comfyui)

[hoary saddle]

Best way to prompt for a word or letters?

[static prawn]

is that basic comfy workflow using the refiner aswell?

[boreal bough]

add to prompt: with the word "sdxl"
you can make variations of that. but the word 'text' is off limits, as its more of a garbage collection prompt

[zealous viper]

yeah true, I'm wanting to do some custom batching with things like prompt weights vs cfg scale so im wanting to be able to build the whole thing manually

[boreal bough]

yep - with explanations for most things

[static prawn]

https://github.com/SeargeDP/SeargeSDXL i just tried this one but it was "a lot" for me

[static prawn]

ok might try this one again

[zealous viper]

yeah comfys base one explains alot in its notes

[boreal bough]

xD yeah thats using all the advanced settings + custom nodes

bit overkill for learning the ui

[static prawn]

just wanna use a highresfix and refiner

thats probably the most important for me atm

[fast vector]

So far, I am sticking with example one. By the way has anyone tried adetailer

?

[static prawn]

and those split prompts i wanna learn about, seems to be pretty important

i see workflows and there are like 4 /6 prompt fields

[azure oxide]

honestly idk if hires fix is necessary with sdxl given it being trained on 1024x1024 images

[sweet fog]

I've been using comfyui and I really like all the flexibility with how you can connect the nodes. I put in the 4x enlarger and then passed that through a sharpener at 0.5

[zealous viper]

there's prob an easy way to patch in highres into comfys, running refiner on comfy out the gate in a single press got me on the bandwagon so I didnt have to learn how to double run in A1111

[static prawn]

i think the refiner is a hit and miss sometimes

sometimes it makes stuff even worse for me, sometimes it adds a lot of detail

but im probably doing sth wrong

[boreal bough]

first try it with the default comfy workflow. searge was doing a lot of experimental things that worked better in 0.9 (unless he updated it, and I missed that) yep its updated

[static prawn]

ok allright

but i think in general any sdxl images look kinda blurry, sometimes great but compared to 1.5 there are more blurry

im always missing some detail

[boreal bough]

that's sampler + prompt fault

[azure oxide]

put blurry, depth of field, and bokeh in your negative prompt to counteract that

[boreal bough]

all ancestral samplers cause the blur to go brrrrrrrrr
also many prompts have a lot of bokeh baked in

[static prawn]

someone posted an awesome picture here with a workflow, its really great, but if u look in the details it looks kinda blurry imo , especially for a 2048x2048

[cold mica]

Our new weekend challenge is live if you'd like to test your sdxl skills! #1087493421209485393

[fast vector]

what sampler do you usually use?

[static prawn]

always use dpm2 m karras

[eternal crystal]

Im trying to use ComfyUI locally, but keep running out of memory, yet i've heard reports of people using the same workflow with 8GB vRAM

[fast vector]

same

[boreal bough]

you can fix it via negatives, and by not having positive prompts that increase bokeh/blur (like "cinematic" or "4k")

[cinder aspen]

Is SwarmUI the most simplistic way possible to use SDXL?

[hoary saddle]

Quotes around the word needed?

[indigo carbon]

is this "backrooms" worthy?

[boreal bough]

that error shouldn't show up, even if you did run out of memory. are you on an old nvidia driver?

[static prawn]

This is not my picture! that one was posted here i think a few hours ago . when u look closer its imo kinda blurry

in general it looks imo absolutely awesome

[eternal crystal]

might be, ill update and report back

[static prawn]

but when u crop it , i miss a lot of details

[boreal bough]

depends. in some instances it works better, in some it doesn't. I usually have a quick run with and without, since it depends on the word

[static prawn]

oh i thought 4k / 8k makes stuff more crisp

[boreal bough]

that was good bias in the 1.5 models. in sdxl it's the opposite.
basically any prompt that would result in images with extreme blur on google, will now also cause that blur in sdxl

[wicked frigate]

install it, follow install instructions and make sure you check the download sdxl button
run it
select the model in the models list
type a prompt in the prompt box
click generate

I thought you asked how not is it oops

is it: yes probably lol

check the pins in #🐝｜swarm-ui

[indigo carbon]

is this like the backrooms?

[static prawn]

mh ok. i dunno i feel the overall images on sdxl always looks more blurry. but its probably because i always used trained models with 1.5

guess i might be to biased because of that.

and the comparison is kinda unfair

[shy kelp]

Hi, is it allowed to train XL

[azure oxide]

i thought all my gens from the first night were bad but after looking thru them, this one aint awful

[azure oxide]

yes infact its encouraged =]

[shy kelp]

Do you know a good tutorial that I can learn how to do it best with large dataset

I am not good in this so need to learn first

[boreal bough]

its really a matter of what prompts you use

[azure oxide]

finetuning's the one thing ive yet to dabble with sorry. #🔧｜finetune could probably help. you could also look up a video on kohya's gui

[solemn scroll]

Am I the only one?

[cinder aspen]

Well that one was helpful too xD thx for both 😄

[boreal bough]

a few examples from my generated images

basically no blur, unless I wanted it, with the burning man

[boreal bough]

large datasets are a lot harder to train. expecially if you're getting started, as each failure on a big dataset results in 20~40 hours of lost time. While on small datasets its around 20 minutes currently

[cunning geyser]

So training loras with large datasets is a pain?

Or models

[boreal bough]

best to practice with small datasets, then progress to bigger and bigger as you get better (and also understand when you even need bigger datasets, as bigger rarely means better in sdxl training)

more angles = good
dataset size =
1~9 <- hard
10~29 <- good enough, if you tag properly
30~60 <- ideal. Will help correct some mistakes you could make while tagging.
60+ <- also good, but training takes longer. good if you want to teach multiple different things at the same time.
1k~10k <- approaching finetune levels of detail with a lora - really hard for beginners, average difficulty if you know your stuff

Tagging:
• always caption the background!
• have a trigger word, shuffle the rest
• dont tag the parts that you want to show up when you use your trigger word
• do tag everything else, especially background

[lost ridge]

I've heard ti training should work the same as before. Haven't seen any attempts at it yet though. Has anyone tried it?

[sweet fog]

I just tried using 1920x1080 with the updated SDXL 1.0 with 0.9 VAE, and it now looks better than when I used the 1333x749 with the SDXL 0.9. I'm using Sytan's workflow with 2x upscale and then sharpening

[slender hinge]

idk if im still doing something wrong, but the quality just does not seem there for me? seems textured / low res but blown up.

[azure oxide]

are you using the new vae that was updated today?

[slender hinge]

i dont think so, where can i find it

[azure oxide]

that could help. the refiner should also help with details if you arent using it

[eternal crystal]

updating the driver seems to have worked, thanks!

[boreal bough]

you're using the new vae.
either get the new upload of base model with 0.9 vae
or download only the 0.9 vae and then put that in.

also blur/pixelated comes from prompts in your case. It looks like you've stacked multiple blur/bokeh inducing words in your positive prompt

[azure oxide]

here

[lost ridge]

Does your prompt have anything that would make the fuzziness on it own? Sometimes it stuff like soft light etc. It does that to photos irl too

[slender hinge]

i literally just copied a prompt i saw which was 'photo of beautiful age 18 girl, pastel hair, freckles sexy, beautiful, close up, young, dslr, 8k, 4k, ultrarealistic, realistic, natural skin, textured skin' no DOF / bokeh inducing words at all

[boreal bough]

👍
if it takes longer than 2 minutes to gen an image, report back. means its working but you still have overflow - in which case it can be fixed to generate images fast. just need to tweak settings

[boreal bough]

yeah - lots of bad prompts going around from 1.5 experiences.
dslr, 8k, 4k, ultrarealistic, realistic <-- all cause extreme blur
also 'realistic' is a word used to describe artwork, not photos - so that will just generally mess up a lot of things, as it will push your generated image into this weird artwork photo mix

'natural skin' I don't know. I'd try it with and without this prompt, see what changes

[eternal crystal]

Took about 5 mins to gen, what kinda settings should i tweak

[slender hinge]

'photo of beautiful age 18 girl, pastel hair, freckles sexy, beautiful, close up, young'

[slender hinge]

yeah idk, i feel there is just some strange texture to all the images I create that makes it worse than older models. Im downloading 1.0 with 0.9 vae now see if it helps

[nimble heart]

why not use the clip retrieval thing to kinda get an idea of what you're actually asking it for? know it's not 1:1 with SD but kinda helps visualize how the model might see your tokens
https://rom1504.github.io/clip-retrieval/

[slender hinge]

minimised prompt down to just 'photo of beautiful girl, pastel hair, freckles'

Still it has this very strange almost moire pattern

[boreal bough]

oohhh that's practical. I always tested everything manually

[static prawn]

how much denoise for the refiner u recommend

and same sampler?

[cunning geyser]

....

I just realized

Sdxl is basically the new best photorealistic model

[zealous viper]

wait for finetunes

[nimble heart]

yea, for example things like "anatomy" give anatomical charts so I'm not sure adding "Bad anatomy" to negatives does what people think it does. Maybe with the rest of the prompt it can figure it out but I wouldn't bet on it

[boreal bough]

first try with the 'tiled vae' node, and have it replace the 'vae decode' node. high chance that that is all you need

[rancid junco]

how do i use a vae with automatic1111? don't see any vae options under sd xl in settings

[nimble heart]

vae is in sd settings

[rancid junco]

yeah but is that gonna.. uh i guess it is

[nimble heart]

sdxl settings are only specifically for xl

vae is universal

you can use any vae for any model, it just won't work that well

[boreal bough]

there's a joke here to be made... about that choice happening for you automatically XD
it should be enableable in the quick settings

[high skiff]

Fuck yeah, just got the reverse transaction confirmation from the seller of my GPU

[nimble heart]

pro tip add the vae changer to your top menu thing like checkpoint already is so you can change it easily

[boreal bough]

congrats! 😄

[boreal bough]

not entirely as bad as on your side, but yeah - something in the prompt is still adding way too much bokeh

[slender hinge]

yeah interesting, you also have a bit of the texture issue, but not as bad. Are we supposed to run the refiner as img2img after using the base model? or are we supposed to be able to use it by itself?

[wary bay]

Is there any way for me to host my local comfy instance to be remotely accessed, like how auto can run a gradio link?

[boreal bough]

🤣 I think its the pastel hair + close up combo

[static prawn]

how u would work with highresfix? Base Model -> Highresfix -> Refiner , or how are the steps?

[slender hinge]

crazy & annoying aha