✨｜sdxl

[stuck bobcat]

one of my favs

[trail bay]

willing to share the prompt? wondering if my setup is wrong because it messes up eyes/hands so much still. but the release said it should do those fine (hands I mean). but still I like it when it can produce these nice pics locally that I'm seeing on this server. but no worries if you don't want to

[kind pewter]

????????

[stuck bobcat]

which prompt?

[deft coral]

how do i get rid of block patining

[trail bay]

are you doing post-processing to get it to look that good?

[high skiff]

just like a normal image

[high skiff]

sorry, i am in a chaotic mood lol

[deft coral]

how can i get rid of this lol

[high skiff]

tree kisser moment

[kind pewter]

now that's better

not that straight

[quasi remnant]

never get rid of that

[high skiff]

why, cause a man is kissing wood?

[quasi remnant]

that is amazing

[high skiff]

O a, sp sprru

WHAT HAPPENED

*I am so sorry

[kind pewter]

Queer/20

[high skiff]

[stuck bobcat]

for the butterfly ones the prompt is ""beautiful glowing realistic blonde haired woman, realistic hazel eyes with an iris in her eye, wearing a bikini, glowing moon behind her, tunnel like background behind her

blonde girl, blue eyes, anime girl with bioluminescent makeup, at a forest with a lake, up close, detailed portrait, up close of upper body, cute anime girl, wearing a Masquerade Mask, wearing a dress, peacock feathers, background forest, chained up, surrounded by glowing rainbow butterflies and flames, up close, detailed portrait; close up of face, beautiful hazel eyes, anime girl wrapped in vines, glowing clothes, dripped out, random colors, detailed portrait, close up, close up of upper body!!! Space background with melting liquid , detailed portrait, close up of face, close up of upper body!!!!, anime girl in black orange green cyber armor, techno girl, detailed portrait , stars, galaxy, and roses, flames, chains crosses, and graveyard tombstones ,anime boy with a rope around her neck, burning fire and chains, detailed portrait, melting water on her clothes, dripping colors, in the galaxy, flowers, detailed portrait, close up of upper body!, growing from the ground, as the brightest flower, detailed portrait, in the galaxy"

[high skiff]

oh my goodness gracious

[fleet fern]

-.-

it is watermelon

[high skiff]

watermelon?

WATERMELON?

[kind pewter]

it's racism!

[high skiff]

Ah, you said it before I did lol

[azure oxide]

watermeLON

[shy kelp]

Inappropriate watermelon

[kind pewter]

Water me ELON

[eternal fog]

Are you saving your images as jpeg or PNG?

[dense chasm]

just generating on the API,1.0 is awesome

[deft coral]

1.0 has so much potential

[high skiff]

remember, Racoon is blacklisted in the SDXL bots because of the racial slur contained in it lol

[high skiff]

gorey I guess

[trail bay]

I got this locally! nice

[high skiff]

wait, let me do it

[deft coral]

@eternal fog png

[dense chasm]

process speed improved as well

[fleet fern]

ah could be. it doesn't always happen

[azure oxide]

i was actually going to generate some cubers to finally see if sdxl could make them

[stuck bobcat]

nice @trail bay !!! with my prompt?

[trail bay]

@stuck bobcat nice prompt

[edgy otter]

You can point at things, sometimes :3

[high skiff]

its likely a gore filter, since its very red and lumpy

[eternal fog]

Hmm not what I thought then if you really aren't getting the weird marks on images.

[kind pewter]

pointing is a big word

[deft coral]

but

[dense chasm]

celebrity is more realistic than 0.9

[autumn forum]

im diggin that offset noise lora they added

[deft coral]

@eternal fog none of my actual gens get watermarks

[fleet fern]

changing the width and height doesn't change the number of credits used anymore?

[deft coral]

im gonna die

[high skiff]

I am quaking

[lean granite]

Just watched a video on comfy and damn it really does seem miles better than 1111! I'm jumping boat right away 😂

[eternal fog]

Every single image I've done with the 1.0 vae gets the weird green fringing

[kind pewter]

let's play a game, GUESS THE PROMPT!

[deft coral]

i haven't had that issue

[stuck bobcat]

a couple more with 1.0

[eternal fog]

It's very subtle. Not at my pc so I can't zoom in enough on your images to see.

[deft coral]

but for real

how can i get rid of that art style

[jade folio]

Anyone know why everything I generate looks like absolute garbage? It's nowhere near what everyone else is getting

[urban fjord]

JSR: Don't prompt for that art style

[honest flint]

A: 20.88 GB, R: 34.98 GB``` dang this thing can take up alot of vram

[kind pewter]

make the prompt more gay

[trail bay]

might be workflow/config or maybe the prompting

[edgy otter]

try "photo of woman, pointing her hand at sky" instead of "photo of woman, pointing at sky"

[lean granite]

Does refiner mean the same as vae or am I trippin?

[trail bay]

you're going to get hands if you do that. and then you are in trouble 😄

[jade folio]

I'm assuming it's workflow related but I saw a video with the default workflow and it worked fine for him but looks trash for me

[edgy otter]

You can use controlnet to fix them tho

[inner maple]

No way!

[trail bay]

yeah I'm looking forward to more models/tooling to help with that because it is still disappointing (that aspect, not overall)

[deft coral]

how can i get rid of thie blocky art style

[deft coral]

i wanna put it in negative

[inner maple]

Do you guys think there will be a way to retrain sd 1.5 loras into sdxl rather than starting from scratch? There are a couple loras I like and I don't know if the creator will retrain.

Some type of conversion process?

[deft coral]

@kind pewter

[west breach]

The vae converts the latent into the final pixel image. It's not hte refiner

[kind pewter]

😳

[flat ibex]

Building a style dataset to dreambooth my own model (well, fingers crossed lolol)
Building out different aesthetics to train into my style training images.
I did find it interesting that the prompt, "artistic" generates mainly portraits of women.
I'd like to think it's a romantic concept rooted in art itself - but then again, how else would SD be able to generate such realistic waifu's? 😆

[clever fable]

yep its definitely biased towards women lol

[outer oxide]

seems unlikely. will definitely be better to re-train.

[urban fjord]

Make sure to run at 1024x1024 resolution.

[inner maple]

That's what I'm thinking......

[fleet fern]

great. and it is blurred through the API too.

[lean granite]

Thanks Mikey!

[jade folio]

oh. That would by why...

[midnight shuttle]

Does anyone know for sure how much VRAM is needed to train SDXL embeddings and hypernetworks (yes some people use those) at 1024x1024? Will 16 GB be enough or does it need 24 GB? Or more?

[jade folio]

Does it have to be 1:1 AR?

[thick goblet]

Can this run on rtx 2060 6gb variant , Intel i5 8400 and 32gigs of ram? ,

[jade folio]

if you mean sdxl yes. I'm running on an i7 8700k and 1080. It takes a hot second to generate but it works

[deft coral]

is there a way to get rid of this art style

[urban fjord]

Just keep your resolution to about 1 megapixels. There is a list of supported ones, but just do normal fractions and you should hopefully be fine.

[thick goblet]

And does the GPU makes noises when. It generates a little or overheating

[lean granite]

@high skiff where are you gonna post your workflow when it's ready? Is it gonna be on here or somewhere else?

[high skiff]

I have a github where I organize everything

It will be posted here, likely on the official Comfy wiki, reddit, and on my github

Along with some documentation I am working on

[hoary saddle]

am i the last person on earth using image2image?

[glad fulcrum]

Is auto1111 working already for 1.0 can it load both models at the same time like comfyUI?

[deft coral]

comics

[shy kelp]

i've update my local branch, i have no refiner selector

[thick goblet]

Post some realism and text pics

[shy kelp]

now im testing the base model

[thick goblet]

I still don't know what refiner model is

[lean granite]

U got a link to either so I can bookmark it for when u do? 💕

[high skiff]

sure thing

[deft coral]

@kind pewter

[kind pewter]

HELL YEAH, now make em kiss

[deft coral]

[high skiff]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[edgy otter]

better than nothing xD

[glad fulcrum]

Refiner model is meant to be used with low denoising on img2img. It gives more details or better end result. But it depends in my experience. Sometimes it makes it worse. Sometimes it is clearly better for realism for example

[lean granite]

Thank ye!

[high skiff]

@eternal fog I can't seem to get any good prompts for skin texture in SDXL 1.0, do you have any recommendations so I can test with my upscale workflow?

So far its doing damn good on things like hair

[deft coral]

@high skiff

[deft coral]

can i get rid if this art

[high skiff]

get rid of what art? lol

[deft coral]

i want photorealistic

instead of this

[high skiff]

upscale workflow is worken for hair

[shy kelp]

i've generate a realistic photo, looks amazing!

[high skiff]

lol

[lean granite]

What's your prompt

[deft coral]

i just got deadpool on it

[kind pewter]

I'M CRYING MY PANTS AAAAAAHAHAHAH

[dense chasm]

do u update your workflow for xl1.0 yet?

[azure oxide]

negative prompt for cartoon or comics mabey

[edgy otter]

ahahahahahaha

[lean granite]

Not yet

He's working on it

[midnight shuttle]

You probably need --no-half-vae. I never needed that until I used SDXL.

[high skiff]

in the works

SDXL 1.0 behaves differently than I assumed

Changing some things, and potentially main merging an upscale workflow as well

[shy kelp]

ok you need to turn of half vae still

[azure oxide]

honestly it looks like SAI is working on it from the looks of it lmaoo

[shy kelp]

yeah, ik

just testing if it's working in the latest branch

[midnight shuttle]

1.5 and 2.1 never needed it for me. But SDXL needs it. I would like to understand why.

[deft coral]

i got no half vae on

[hoary saddle]

and potentially a good i2i workflow 😉 /beg

[shy kelp]

same question

[high skiff]

I have not messed at all with it honestly

I could give it a look sometime

[hard fractal]

Hey everyone...

Thank you.

SDXL would not have been as good as it is without your feedback.

[autumn forum]

your welcome!

[oak arrow]

u can't spell feedback without feed and u can't feed ppl without pizza so where is our pizza

[kind pewter]

Best one so far 🤣

[cobalt kernel]

Did we get the proper 1.0 workflow ?
I assumed that was going to be part of the release ..

[urban fjord]

Yeah it looks good and yeah I can see from finetunes that finetuning is working really well on SDXL

So good job!

[edgy otter]

IT WORKS

[wary wren]

delete that

ugh

[deft coral]

hes holding a cup

whats wrong with that

[oak arrow]

is there any info about training sdxl yet? Do we train at 1024 source images? How much vram? Does it work with automatic1111?

[wary wren]

ugly

[lost glade]

And additional question, sorry if this was asked a lot, but what sampler name and scheduler have you guys been using for SDXL?

I'm using comfyUi

[azure oxide]

all of them

[urban fjord]

You can train on lower res images depending on the LoRA you want to train, if you've got 12 GB VRAM you should be fine, yes it works with automatic.

[lost glade]

Fair, anything work best? 😂

[fair tartan]

7360x3520 output now in 51sec on my 3080Ti

[west breach]

I'd prefer a million dollars

[broken hare]

sde normal and ddim with ddim universal for photography, pretty much anything else for artwork is what i do

[urban fjord]

I'm using DPM++ 2M Karras with whatever scheduler it is using in Auto1111

[soft blaze]

this is using the DreamshaperXL 20seconds using A1111 webui and 1254x836 resolution, no refiner

and DPM++ 2m karras

[midnight shuttle]

I tried Comfy but I miss the "inpaint masked only" from A1111 and I really miss the "send to inpaint" option for iteratively inpainting images. What is a good way for me to do these things in Comfy?

[soft blaze]

GPU 3070 8gb

[wary wren]

looks kinda mid

[lean granite]

I +1 this and am also wondering the same

[urban fjord]

The best if you want to do inpainting is really to use Auto1111

[broken hare]

go back to auto for these things.....it doesnt give you as good of results, but image to image and inpainting is so good for fixing

[midnight shuttle]

OK, I do a lot of inpainting so I will stick to A1111 for now or maybe try SD.Next. Thanks for the answers.

[soft blaze]

image at that distance with 1.5 were really bad , that one is without any retouch

[outer oxide]

I look forward to the time when I don't have to worry so much about sampler selection anymore. The analysis paralysis is real!

but DPMSolver++ seems to still be a good staple. (maybe with Karras at lower step counts.) and there's always the ol' reliable DDIM fallback if things get too weird.

[ionic dragon]

@high skiff any keytake always from your today's testing

[wary wren]

Oh yeh if we're judging based on 1.5 standards its amazing

[lean granite]

So you would generate the picture in comfy and take it into a1111 for inpainting?

[broken hare]

using refiner in image to image is actually working very well on the 1.0 release

[deft coral]

@wary wren

F1 day

[ionic dragon]

@cold mica @smoky patrol or any dev please ban him

[broken hare]

today im actually creating it in auto with base, then running image to image with refiner

[midnight shuttle]

I've tried that and it works for me too. Just inconvenient to have to keep changing models.

Most seem good enough even without refiner.

[broken hare]

yep....it sucks, and i always forget to change it back to base in text to image and it creates nighmare fuel

[high skiff]

A few things

I need to look more into the latent sizing for 1.0, as it seems to do considerably more than previously

I need to work a little bit on my prompting guide, as it seems as though the text encoders behave a little bit differently now

There is potential main branch merge hope for my fractional offset, as well as my upscale workflow for my 1.0 release

CFGs are a lot less finicky with 1.0

Aesthetic score seems to do even more extreme effects with 1.0

1.0 seems to have a little bit more support for resolutions over 1024 compared to 0.9

The 1.0 VAE introduces some weird artifacts, of which I assume are some form of digital watermark

Mixed diffusion AKA my workflow requires additional steps compared to 0.9, but produces better results

Likely some other things I'm not remembering at the moment

[midnight shuttle]

Hoping for SDXL inpainting model soon.

[exotic kestrel]

All images generated by SDXL 1.0 have a red/green defect if you zoom it. Just check your images. Almost all of them has that "mark". It was not present in SDXL 0.9. Has anyone seen it?. Why is there?

[edgy otter]

For me, the best thing about SDXL is the faces. It almost everytime gets them right.

[hoary saddle]

decent i2i from default Comfy workflow with added image input found on reddit lol

[azure oxide]

thats interesting, the report to mods selection isnt working again

[midnight shuttle]

I don't think this is the "watermark" that everyone is so concerned about. I think it is just a bug that they can fix.

The real watermark is invisible.

[high skiff]

I find that goes down very often

[half ivy]

do you guys vae decode with base or ref?

[lost glade]

Alright, thanks for the tips. I'm trying comfy cause auto1111 gives me a "size mismatch for model" error, and I can't fix it for the life of me haha

[azure oxide]

ive noticed it down once before, ya

[cobalt kernel]

I'm pretty sure 0.9 has watermarks as well ..
You just need to look real close ..

[ionic dragon]

So you are using 0.9 vae?

[heady vale]

just add this emote to the message ⚠️ and mods will see it

[azure oxide]

ya true

that should still work

[deft coral]

@high skiff do not use DPM++ 2M SDE

[lost glade]

I tried to go right to 1.0, maybe I need to add the vae manually?

[midnight shuttle]

Why not?

[lost glade]

I'm trying to remember what all these parts do

[urban fjord]

There is no need for one as SDXL can do it well as is.

[jade folio]

What is with the jpeg artifacting effect?

[exotic kestrel]

I have generated a lot of images using 0.9, none came with visible defects like SDXL 1.0... is really notorious.

[midnight shuttle]

Oh that is good. I thought inpainting model meant that some additional context was given to the model so it can make the inpainting fit better.

[deft coral]

i got that with DPM++ 2M SDE

[exotic kestrel]

Exactly what i was refering too... every image in SDXL 1.0 has this weird artifact.

[midnight shuttle]

I think maybe there is something wrong with the VAE. I never needed --no-half-vae until SDXL 1.0 but never tried 0.9.

[deft coral]

This shot is good @west breach

[wary wren]

im a casual can you point it out?

[high skiff]

yes

[lost glade]

How do you use --no-half-vae in comfy?

[kind pewter]

green and purple lines

[exotic kestrel]

This is the artifact i was refering to.

[languid schooner]

Cars and people don't look great when doing native 1920x1080 but man do landscapes look good. Upscaled it in Gigapixel.

[edgy otter]

it was really difficult to generate indian people before, now its not. I used to get the same faces or "HDR" images

[wary wren]

oh i see it here

[honest flint]

someday we will get a model that can do cats

[wary wren]

prob the data set they used

[honest flint]

this just looks like chromatic aberration, might try adding that to your negatives

[jade folio]

Almost got the RB-19

[exotic kestrel]

Just zoom any image generated with SDXL 1.0 seems something wrong with VAE. Did they release a crippled model?

[jade folio]

Except it's driving sideways

[deft coral]

its so good at pics for f1

[autumn forum]

use 0.9 vae then

[kind pewter]

I don't find it anywhere

I use the one you posted here, and I still have em

[exotic kestrel]

Lets do some test... now... where do we find that specific vae...

[jade folio]

you get it to produce a decent car?

[edgy otter]

yea wait

[deft coral]

ya

[hoary saddle]

https://huggingface.co/stabilityai/sdxl-vae/resolve/f85c10757eb0bda0222090261de54c5fe17a7f4f/sdxl_vae.safetensors

[honest flint]

did you. try adding chromatic aberration to your negatives?

[midnight shuttle]

I don't see these artifacts when generating from A1111. But that's with no refiner.

[kind pewter]

it's not related to the prompt or anything

[autumn forum]

you can search discord you know xD #✨｜sdxl message

[honest flint]

did you try it?

im not having this issue fwiw

[kind pewter]

it updated 16 hours ago

it's 1.0

:(

[hoary saddle]

says a month ago

[nocturne dove]

There is one here you can get if you need it https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[midnight shuttle]

If you have the artifacts, what UI are you using and are you using --no-half-vae?

[hoary saddle]

wierd

[autumn forum]

#✨｜sdxl message <<<<<< 0.9 vae

[kind pewter]

checksum equals the one that was updated 16 hours ago...

[deft coral]

@jade folio not the best

[autumn forum]

oh okay

[ionic dragon]

Is it b'cuz it's pruned

[edgy otter]

photo of a car, high resolution, reflections, fast shutter speed, 1/1000 sec shutter

[wary wren]

let us how it goes

[high escarp]

Sampler and steps and cfg

[kind pewter]

how did you get to this branch's page??

[deft coral]

@jade folio

[edgy otter]

Euler, CFG 7, 20 steps (32 steps work better in my exp), I have an AMD 6900XT.

[high escarp]

Thank you

[cobalt kernel]

— Re : cars I get 50s ones a lot ..
Though it has to do with my prompt :

[solar tusk]

I don't think that house is to code

[glad flint]

Is there a place for benchmarks?

[cobalt kernel]

" House of leaves "

[exotic kestrel]

Using ComfyUI not shure but --no-half-vae seems setting of Automatic1111

[shy kelp]

Been messing with SDXL in search of anime style images.
Even without a finetune or hell even the refiner! It's quite good, for a base model.

[midnight shuttle]

Maybe it is caused by ComfyUI. I see no artifacts in A1111.

But I don't use refiner yet.

[kind pewter]

Yup, confirmed, they changed something in the VAE Weights

[high escarp]

It seems the softer group of sampler’s are doing a lot better

[shy kelp]

Which are?

[jade folio]

better than most

[high escarp]

For instance, Euler compared to heun Euler is soft, and heun is sharp

[high escarp]

For instance, Euler compared to heun Euler is soft, and heun is sharp

[solar basalt]

why annouce with broken link ?

[kind pewter]

it was the live chat

?

[shy kelp]

yay batman

[deft coral]

@shy kelp prompts

[shy kelp]

Photo of a batman standing on the roof of a building illuminated by the moonlight

[glad flint]

im getting ~1.3 it/s on comfy UI with a RTX 3060 (12gb vram) and 16GB ram, on 1024x1024 images, euler
My system ram usage keeps peaking to 100% then dropping slightly then peaking 😓 never happened before lol

[cobalt kernel]

— It does reflections well ..
This is from 0.9 model :

[high escarp]

Same on 3090

[shy kelp]

Same on 3080

[lilac raven]

Just got it running on a similar setup, it was crashing because of the memory

[austere bay]

A catgirl with orange hair and stripes wearing a sundress painting at an easel outdoors, warm lighting

[urban fjord]

That looks right.

[glad flint]

euler is definitely working better for my stuff on SDXL

[honest flint]

if you are using windows with recent nvidia card it will page out of vram to ram instead of oom'ing to some extent

[shy kelp]

that's not a photo...

[gloomy totem]

I get some weird memory leak sometimes during rendering. Have to restart the whole application.

[glad flint]

i couldn't reply to anyone because my browser was locked up because RAM 😂

[lilac raven]

I still can't use both models, as there's not enough memory

[shy kelp]

huh
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

im getting this error for the first time

[high escarp]

Also, using Hi-Rez fix latent anti-alias resize to 1.01 at the noising strength 0.7 using Euler seems to produce better results also.

[jade folio]

@high skiff when I change the step value in the int box for your SDXL workflow it distorts the image terribly. It also doesn't appear to affect either pass's step value. Am I doing something wrong?

[midnight shuttle]

I get that a lot when switching models after training or using textual inversion. I know what it means but I have no idea how it ends up in that state. Just restart the UI.

[deft coral]

@jade folio

now the shots are getting there

[jade folio]

This is what Callie Mayer should've looked like

[visual glade]

I can run base + refiner on 16GB ram + 8GB vram on windows but that's without anything else open on the computer

[high skiff]

you need to change the steps on the lower Ksampler as well

That is an issue with the current version, and will be fixed with the new one

Also, make sure the first one ends, and the second starts on the same step

I recommend trying to keep a 2/3 ratio, so 20 steps base, 10 steps refiner if you can

[jade folio]

Seems to be the better the human looks the worse the car looks and visa versa

[peak dove]

Its doing weird double heads!!!

[honest flint]

almost

[visual glade]

I'm eventually going to improve memory more though

[high skiff]

sorry to pester, any progress on the fixed primitives?

[median mulch]

so i dont have to get a 4090? :3

[visual glade]

no because I have had 100x other things to do

[high skiff]

fair enough

[cobalt kernel]

So mister comfy , little confused about " swarm " being written in c# for multi - threading versus " comfy " in python. Though the latter remains the back end - and " swarm " a front end ? I thought you'd want the multi - threading on the back ..

[visual glade]

yeah I'm probably going to add a way to execute multiple workflows at a time eventually

[lilac raven]

Is it possible to unload the first model once the first latent is available to the second model?

[visual glade]

but for now it's like this

[cobalt kernel]

Thanks for your work !

[glad flint]

just closed chrome to free up RAM and realized that's where I was prompting comfyUI from 🤔

[cobalt kernel]

Nice back light ..

[kind pewter]

Left : 1.0 VAE (artifacts, blurry even!) Right : 0.9 (crisp and no artifacts)

0.9 VAE Link : https://huggingface.co/stabilityai/sdxl-vae/resolve/f85c10757eb0bda0222090261de54c5fe17a7f4f/sdxl_vae.safetensors

[high escarp]

snow globe momma bear baby bear
Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 32, Sampler: Euler, CFG scale: 7, Seed: 1476289861, Size: 1024x1024, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Denoising strength: 0.7, Hires sampler: DPM++ 2M Karras, Hires upscale: 1.05, Hires steps: 50, Hires upscaler: Latent (antialiased), Version: 1.5.0

[median mulch]

where do you get the 0.9 vae? cause i think the hugging face one is the 1.0 version

[kind pewter]

I've reposted the link

[jade folio]

whats the best sampler for photorealistic humans?

[tribal moat]

just tried Comfy, dam it's fast. Same image took almost 8 minutes to generate in Auto1111 but comfy did the job in just 30 seconds. Almost 15 times faster! Thanks buddy for recommending comfy.

[shy kelp]

his face is something

[exotic kestrel]

YES!... i can confirm artifacts are gone using VAE from SDXL 0.9!

[median mulch]

so how do we use the vae in comfyui? do i need specific workflow or it can be any workflow?

[kind pewter]

I wonder if devs can explain what has been messed up and why, cuz that looks like a mistake or an incompatibility implementing the VAE

[vocal rapids]

isn't it VAE is embedded in the base and refiner model?

[vital wolf]

my installer is stuck when trying to install swarmui

[kind pewter]

yes, but you can use an external one

[shy kelp]

Batman is watching you

[inner maple]

Does SDXL know who big poppa pump scott steiner is?

[median mulch]

oh i assumed i needed to download base + refiner + vae from huggingface, you telling me i actually dont need the vae?

[wicked frigate]

if it stays stuck, close it and click the desktop icon that should've been generated

[exotic kestrel]

Double click on canvas and write "VAELoader", there is a node (you must have said VAE in models-vae folder in COmfyUI installation).

[wicked frigate]

(or the launch-windows file directly)

[vital wolf]

should i close it first?

[median mulch]

ah ok ty

[vocal rapids]

so you can just add a VAE loader node and connect with the VAE decoder node in case you want to use a special VAE

[wicked frigate]

yes

[honest flint]

i prefer the one on the left myself (1.0)

[vital wolf]

desktop icon closes upon launching

[median mulch]

but just to confirm, we need the vae from hugging face right? it's not "baked in" correct?

[shy kelp]

i hate this

[kind pewter]

yup, you need the VAE from the link I posted, which is a specific commit

[deft coral]

wish i could train 1.0

[jade folio]

anyone figure out which sampler is working best on SDXL?

[median mulch]

cool thx

[kind pewter]

the main branch is 1.0 VAE, it got updated about 16 hours ago

[honest flint]

he doesn't need it, it has a vae baked in

[kind pewter]

if he uses 1.0 model he needs the 0.9 VAE

cuz the 1.0 VAE is messed up

[honest flint]

1.0 model works fine with the 1.0 VAE

[visual glade]

can I see your workflow?

[kind pewter]

no

[vital wolf]

ok, i launched the windows file that helped

[honest flint]

i disagree, I think the 1.0 vae is better then the 0.9 vae

yes

[deft coral]

i need DB to be fixed to i can train 1.0

so

[shy kelp]

oh, i changed it right now when you asked

[kind pewter]

@honest flint then explain this maybe?

[shy kelp]

wait, i'll get it back

[median mulch]

i mean idk... i saw couple people say 0.9 vae is better

[visual glade]

or at least give me the full traceback

[honest flint]

I already commented on that and I said I think 1.0 looks better

[marble dew]

gonna push more modular, change around VAEs, refiners, upscalers, base models and more. This will be really important as the next models of other architectures come through, my ensemble model predictions will become true muahaha (also a reaosn we ❤️ ComfyUI)

[midnight shuttle]

People are seeing different results with the 1.0 VAE. I don't see any artifacts but other people do. It's not a simple answer just to say that 1.0 VAE is broken for everyone.

[kind pewter]

you like blurry and artefacty results? Ok I guess lol

[shy kelp]

this one

[marble dew]

0.9 VAE and 1.0 VAE will be subjective in some ways

[inner maple]

Is torch 2.0 recommended for this bad boy?

[median mulch]

Emad doing god's work

[marble dew]

just as model can produce better images without refiner sometimes

handily we open sourced all

[shy kelp]

can i send it to you in local chat?

[marble dew]

[midnight shuttle]

The artifacts some people were posting from 1.0 VAE were very obvious and not normal variations. But not everyone sees these artifacts.

[shy kelp]

just to not spam here

[boreal bough]

so the license for 0.9 is no longer research only?! 😄

[marble dew]

eventually what will happen is that there will be feedback loops so you'll have a bunch of student teacher and intermediate models with full control and composition

[kind pewter]

k but isn't there a small issue regarding blurr atleast?

[azure oxide]

there are threads popping up about the vae now lol, people are guessing that its due to a watermark

[median mulch]

we just have to wait for SDXL 1.5

[marble dew]

no it is but there is a MIT licensed VAE you can swap out if any issues

[lilac raven]

I'm pretty happy with SDXL and ComfyUI, I got another computer to run SD with them

[marble dew]

this, train MOAR

or more efficiently so much to do

[honest flint]

I think 0.9 is too sharp in that image comparison you gave, 1.0 looks more natural

[marble dew]

we need some super fast models now we have bulked

[exotic kestrel]

As far as i can see vae from SDXL 1.0 produces a weird artifacts on final image, so we are using VAE from SDXL 0.9 wich has not the said issue.

[midnight shuttle]

I don't think it's due to the watermark. All models have watermarks and they are invisible.

[honest flint]

cough larger text encoder cough

[kind pewter]

it is zoomed to show really clearly the problem

[marble dew]

we had it it doesn't do much

RLHF will have a bigger impact

[azure oxide]

yes that was the case before but this seems to be different. idk, its all conjecture tho.

[marble dew]

as will fine tuning

[exotic kestrel]

Problem solved using VAE from SDXL 0.9 .... is not related to watermakrs.

[compact obsidian]

I tried to quanize it down to 2 bits, 6 is about the lowest i could go with good results

[median mulch]

yea i mean people are still experimenting with settings and stuff, but il try some combos :3

[marble dew]

fine tuning super easy the fine tuning API will be out of preview shortly

[spark bear]

I love how emad just pops in with a barrage of posts out of fkin nowhere

[marble dew]

6 is abotu right vs 4 for language models

[heady vale]

can XL be converted to TensorRT format?

[visual glade]

of course it can I did it

[marble dew]

yes most people won't even need refiner tbh

[heady vale]

very nice

[marble dew]

using delayed prompts also can have an interesting outcome

[alpine pine]

What is the benefit of tensor RT?

[marble dew]

its just faster like AI template etc

[heady vale]

heaps faster and more efficient

[cunning geyser]

Is there any anime models rn

[marble dew]

NVIDIA have model

[cunning geyser]

Is the base model even good for training anime models?

[marble dew]

couple of days will be a few

yeah really good

[cunning geyser]

Oh ok

[kind pewter]

you prefer left or right on this one?

[cunning geyser]

Thank god i'll just stock up til then as i get ready asap

Yall think sdxl will replace 1.5?

[dull cairn]

any comments on the "jpeg" artifacts that so many people are experiencing? my SDXL output looks terrible

[median mulch]

@marble dew im curious, how are devs generating images? is it some private/custom UI or is it something that is released out there?

[wicked frigate]

https://huggingface.co/hakurei/waifu-diffusion-xl

[azure oxide]

working with a 3090. would you say that the perfomance +%s good enough to bother with the whole model conversion process?

[magic dawn]

The vae thing is not subjective, the rainbow/lines artifacts are clearly visible

[marble dew]

my tired eyes

[midnight shuttle]

Anyone know the VRAM requirement for training SDXL hypernetworks?

[marble dew]

0.9 vs 1.0 VAE is more noticeable in peoples

[visual glade]

I didn't try it on my 3090

[boreal bough]

will we get a fp32 version of sdxl1.0 base for finetuning with full precision?

[exotic kestrel]

Rigth!... the left one has that weird green/purple artifact.

[visual glade]

it was on an A100

[marble dew]

https://twitter.com/EMostaque/status/1684322470726189056?s=20

Emad (@EMostaque)

Happy to announce the release of #StableSwarmUI open source!

Multi-GPU platform for swarm generation integrating ComfyUI and much more, go use & contribute.

Ideal for #SDXL

https://t.co/LPGSULpT7y

Likes

251

Twitter · Jul 26, 2023, 9:59 PM

look at all the nobs

[midnight shuttle]

Agree. It is not subjective. I clearly see the artifacts. But I also do not get these artifacts when I generate. So it is a more complex problem.

[marble dew]

this was our inspiration

[magic dawn]

@marble dew Will there be any official guides on finetuning SDXL base/refiner?

[marble dew]

I remember when we had a leaky VAE before that was bad

[marble dew]

yeah being worked on its really easy compared ot before

and we have fine tune API in preview

rolling it out rapidly

[kind pewter]

speaking of tired eyes...

[marble dew]

as Joe noted LoRAs where it at

[high escarp]

Out of curiosity, what was the sampler CFG and steps used on the bots during training when we were generating pictures

[median mulch]

cant wait for controlnet and more greatness

[trim orbit]

this guy pieds

[honest flint]

other then the slight chromatic aberration i prefer the one on the left

[deft coral]

i just dud a cute pic

[deft coral]

did

[boreal bough]

7hours in ^^

[vital wolf]

Does it redownload comfyui if I select comfy as backend? @wicked frigate

[deft coral]

@boreal bough what gpu

[plucky prairie]

I want my textual embedding training!

[lost glade]

What have you guys done to get eyes right on SDXL?

[boreal bough]

4090

[cunning geyser]

how big are sdxl loras

[boreal bough]

will take about 10 more hours

[cunning geyser]

@-@

[deft coral]

how so long

[cunning geyser]

Cause theres 1.7gb loras on civitai and its like jesu schirst

[kind pewter]

understandable, tho look at the right eye, which is more on focus, the shiny aspect of it is quite dulled by the blurr from the VAE

[wicked frigate]

in the installer, yes. If you already have a comfy install, select none and add it in the main settings interface after the installer is done

[deft coral]

i can get 100K in like 3 hours

[inner maple]

Does anyone know if multiple loras will perform better with sdxl? In 1.5 if you added too many loras the image would start to get distorted.

[marble dew]

iirc SAMPLER_K_DPM_2_ANCESTRAL 50 steps total, cfg 5-7

[boreal bough]

big dataset go brrrrrrrr

[lavish sinew]

And 1790 epochs

[plucky prairie]

I always got better results from training TIs than LORAs. Probably user error, but still. I love my TIs.

[marble dew]

yeah they better

0.9 loras work with 1.0 better too

which indicates something

[high escarp]

Thank you

[honest flint]

yeh to me that looks more natural, if i was generating an image of a doll I would probably prefer the sharpness of 0.9

[deft coral]

romance pic

[cunning geyser]

Question, how big are loras for sdxl usually

cause the 1.7gb worries me i might not have enough space x-x

[median mulch]

time to get a 8TB SSD :3

[crisp owl]

storage is cheap

[midnight shuttle]

Are Loras similar to hypernetworks in the sense that you can train a lot of similar subjects (example, train many images of cats) and then it uses those cats to give more variety than the cats already built in to the model? That's my current application for hypernets.

[boreal bough]

43mb

[deft coral]

im good for now

[visual glade]

check the size of the example lora: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_offset_example-lora_1.0.safetensors

[lost glade]

I keep getting lazy eyes and crooked teeth, dpm-2-a with karras, and the basic comfy 0.9XL workflow

[cunning geyser]

oh so 50 mb thats fine

[boreal bough]

any lora above 50mb is doing it wrong, because they think they're still working sd1.5

[cunning geyser]

i got a tb but i mean

[honest flint]

sort of, hypernets are like a scaple, lora's are like a cybernetic neural mesh woven into the brain

[cunning geyser]

ima go get a better ssd soon

ngl

yeah ima order it rn

[lost glade]

Just me?

[midnight shuttle]

What is a scaple? As I understand it a hypernet is just another set of weights added on top of the main model. Not so good if you want to train one exact item or character but works very well to train more of a certain object or scene.

[cunning geyser]

Internal or external ssd?

[lavish sinew]

I'm starting training on XLeanBeefPatty to run overnight. We'll see how it goes using the same settings as 0.9

[kind pewter]

NVMe for the win

[pastel spire]

what is the best way to use sdxl with the refiner and lora's, comfyUI is really fiddly and difficult, does a1111 support using both the base and the refiner correctly? I can't get impact to work right with it.

[inner maple]

I'm getting this issue too. 1024x1024

[median mulch]

does civitai have any sdxl stuff yet?

[lavish sinew]

Yes

[west breach]

my loras are coming out at 891mb

[cunning geyser]

Yeah

[brazen patrol]

so what is the recommended proportion of Base and then Refiner? I have it 3/4 Base and 1/4 refiner. Ex: 32 ste Euler with 0-24 Base and 24-32 Refined

[kind pewter]

:o

[marble dew]

do you get it in clipdrop

[midnight shuttle]

I know I won't be able to train with 12 GB. But is 16 enough? Is even 24 enough to train 1024x1024 TI?

[urban fjord]

For some maybe, but the ones Ive made performs worse at 1.0 than 0.9.

[marble dew]

you can look at some of the pantheon here

try matching

[urban fjord]

You can train LoRA with 12 GB VRAM.

[austere bay]

sdxl

[midnight shuttle]

And does that work like a hypernetwork in the sense of adding more of a concept to the existing model rather than training on one character or exact style?

I guess I just need to experiment.

[high skiff]

@upbeat summit Do you mind if I Dm you?

[urban fjord]

Yes with a proper dataset.

[spark bear]

hmm 1320x1472 base gen, still surprisingly coherent, did the sai team ever try to push 1.0 past 1024 base? seems like it could go quite a bit higher

[lost glade]

No, just comfyUI, I havent tried clipdrop

[boreal bough]

16gb vram always works for training.
12gb vram worked on some machines, with 0.9 - in theory should now work on all machines, thanks to the pre-pruned nature of 1.0

[crisp owl]

Using Sytans flow, I notice the "Positive Base/Refiner" and "Negative Base/Refiner" are set at 4096 for width/height, but the default node "CLIPTextEncodeSDXL" is set at 1024.

Trying to make sense of that, but my tired mind can't right now

[visual glade]

loras are a patch on the weights so it should be able to do pretty much anything a trained checkpoint can do

[hard fractal]

I'm not sure i love the idea of not feeding in the subject to CLIP L

[midnight shuttle]

So also the same as anything a hypernetwork can do?

[sour obsidian]

I personally use dpmpp_SDE or dpmpp2m in comfy. The best performance on the bot for both models was 40 steps (32/8) split. 32 base, 8 refiner

[lusty raptor]

same, and same, lol

[azure oxide]

https://civitai.com/models/112902/dreamshaper-xl10-alpha2
looks like the dreamshaper creator took the approach to just ignore the refiner altogether for their sdxl model. i wonder if others are going to also ignore it going forward 😅

[lilac raven]

So close, the perfect glowing cube, but then the hand was SD'ed

[crisp owl]

Care to expound on that?

[honest flint]

the loss functions for training lora's are actual loss functions vs what the one hypernetwork trainer i looked at was using (a1111) which means hypernetworks tend to be a crapshoot as well... success rate on lora's is way higher

[reef sky]

where can we find sytans workflow?

[west breach]

Do we have a SAI lora config file for kohya gui?

[cobalt kernel]

Hey @hard fractal was an offical 1.0 workflow released today as well ? Is sytans the closest that we have in " comfy ui " ..

[crisp owl]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[shy kelp]

Kitty

[hard fractal]

We're still exploring the best pipelines.

[cobalt kernel]

Thanks

[quiet vine]

Uhh... lit? On fire. Lit on fire. Your cat is literally on fire.

[azure oxide]

sytan is our god

[shy kelp]

Where's my bucket of water

[west breach]

here's a super basic workflow

[median mulch]

stay hydrated

[dapper current]

@hard fractal is mine

[lavish sinew]

12gb is enough for training with 1.0, I’m currently getting 2s/it with a 3060 12gb

[hard fractal]

aww

comfy's default workflows are great

[silk geode]

Did any of the SAI team ever show us the examples that precipitated the surprise delay of 1.0?

[autumn forum]

I also have a workflow working with sdxl 1.0 pretty well. feel free to use it

[visual glade]

loras are easier to deal with than hypernetworks because they don't actually modify the model itself just the weights

[cobalt kernel]

Cheers I'll give it a spin ..

[tough vigil]

ComfyUI works for me at least, A1111 not so much

[midnight shuttle]

But they still have the same overall capabilities? If so that sounds like lora is basically just as good. I'm only trying to understand if lora is a substitute for hypernet.

[hard fractal]

A1111 is dead.

Long live comfy.

[urban fjord]

Auto1111 works really well for me so it's subjective.

[tough vigil]

I only use A1111 for Deforum

[boreal bough]

refiner causes a lot of issues for lora training. some of which can be overcome with effort - but some simply can't be solved, due to the nature of what the refiner does.
• for example: if you train a specific head accessory, or eye accessory that didn't exist in sai's dataset, then then the base model will learn it via the lora, but the refiner will not understand it, and will do its best to delete it away.
• if you train a specific face, then the refiner may or may not change it, depending on if the prompt is biased
• certain anatomy information (4 hands from indian gods) will also be deleted by the refiner if possible

@hard fractal is lora support for refiner planned? or is using base only the official sai endorsed method for now?

[lost glade]

Thank you for taking the time to help 👍

[azure oxide]

i wonder if there ever was any updates to the LORA creator's question of "how can i make it better?"
https://old.reddit.com/r/StableDiffusion/comments/1223y27/im_the_creator_of_lora_how_can_i_make_it_better/

[spring fulcrum]

Is anyone using SwarmUI... What is planned for that?

[visual glade]

it should but hypernetworks can be very complex so they could be more powerful, it's something you would have to test

[hard fractal]

The hope is to eventually remove the need for the refiner.

[quiet vine]

I don't think I could tell that from a photograph. Some bits look a little off, but only off in the ways that cats can be off. Now do an opossum xD

[boreal bough]

thanks 👍

[urban fjord]

I might try the new SwarmUI but not until it is more stable in terms of ease of install and got LoRA support

[midnight shuttle]

OK that is also my understanding. Hypernet is the most powerful but also the hardest to work with. Thank you for the info.

[tough vigil]

Agreed, I was really surprised how good it came out.

[cobalt kernel]

Have that on my todo as well ..
What's wrong with " swarm " installation ?

[hard fractal]

Comfy is its backend, so it does have LoRA support

[wicked frigate]

A lot of us internally in stability are making use of it, a bunch in the public are trying it. I've been scrolling between discords looking at people trying it.
You can see some of the planned key features in the readme https://github.com/Stability-AI/StableSwarmUI#status

[cobalt kernel]

I thought hypernetworks ..
Weren't as efffective as using " lora " :

[urban fjord]

Oh, I guess the readme on Github is just outdated.

[dapper current]

It's extremely useful if you run it on a multi-GPU system and run batches or grids - soooo fast

[sour lodge]

I compiled a modified version of pytorch to take advantage of CUDA Unified Memory and it seems to use full GPU speed until the VAE stage, when the VRAM usage spikes up and overflows into my regular RAM (instead of throwing an OOM like with official pytorch)

[lusty raptor]

Can I get a hint on the crop values? Are they useful at all during inference? They certainly do stuff, but I haven't nailed down a positive use for them yet

[sour obsidian]

Honestly they can prob be removed from the ui haha. They should always be 0 during inference but you can have fun toying with them, will just offset things

[spring fulcrum]

I have a single 4090 is it still fast with that setup?

[lusty raptor]

cool thanks, annnnnd how about the other h/w values?

[dapper current]

Yep! Just as fast as good ol' Comfy would be

[midnight shuttle]

The question is "effective at what". I know a Lora is good if you want to train in one particular person or object and re-use that person or object in generations. But what if you simply want to add more examples of cars or trees or waifus or whatever? For that I've always used hypernet. But it seems hypernet training doesn't work with SDXL right now anyway so it doesn't matter.

[vast narwhal]

I don't know why people uses a1111, comfy ui is so easy. ComfyUI:

[spring fulcrum]

awesome thank you

[sour obsidian]

target_width/target_height should be plugged into your gen size, so whatever you plan to output should also be target. Width/Height there I reccomend 4096 in both always, though you can try other square values here (1024, 2048, etc, we found 4096 best generally. yes I realize its a bit odd haha)

[spring fulcrum]

I'll have to brush up on my Italian... That looks like quite a plate of pasta there

[languid schooner]

Have a really shitty Dante holding pizza.

[lusty raptor]

thanks again! I had been leaving them all at 4096 while messing with everything else

[lavish sinew]

ComfyUI makes me feel like a moron, but I have no experience with nodes. Fortunately I’m able to rip off others’ workflows with their pretty pictures.

[wicked frigate]

StableSwarmUI (using comfy as a backend so you can hide the spaghetti from view)

[west breach]

Tip for comfyui. Stop adding more nodes! Just build a different workflow and keep it simple. You can have multiple workflows, it takes 1 second to load them.

[midnight shuttle]

Same for me on SDXL. Generation uses VRAM but VAE overflows and uses system RAM so it's slow.

[digital shoal]

how do you use the refiner in stableswarm? I wasn't able to figure it out

[lusty raptor]

flipside, you can save a lot of time by not switching around a bunch of stuff by hand, or diving in and out of photoshop or whatever.

[honest flint]

thats slick, just needs a preset button to store prompt snippets

[inner maple]

WOW

Finally got it working in automatic

[dense chasm]

anybody tested this one https://github.com/Stability-AI/StableSwarmUI, developed by C#, GPU swarm and multithread performance improved

[inner maple]

This is without the refiner.

[wicked frigate]

not yet a direct native impl but you can set it up in the Workflow Editor tab (or download a workflow that uses it), and then click the button to use that workflow in the main ui and then use as normal

[jade folio]

anyone know how to import the workflow from a civitai picture? on automatic1111 you could use png info to get it from the images but drag and drop doesn't work for same images on comfy

[brazen patrol]

Unless you are using the Refiner too

[boreal bough]

yes. good stuff. would endorse.

[wicked frigate]

the presets feature is really powerful and fully working!

[boreal bough]

made my some cool guy

[west breach]

It's amazing what SDXL can achieve in just 8 steps

[kind pewter]

oh, a chess antheusiast?

[digital shoal]

i tried that earlier and it just gave me a bunch of toggles on the main screen. What workflow are you using?

[honest flint]

where is it? I didn't see it in the screenshot

[wicked frigate]

along the bottom

[sour obsidian]

https://tenor.com/view/mario-spaghetti-italian-gif-14634187

[honest flint]

oh, hrmm does one of those buttons give you a dropdown or something?

[cobalt kernel]

Also for any " stability " devs ..
— Just curious about this warning :
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}

[plucky prairie]

Comfy reminds me of early in my career as an engineer, I would run into guys who thought the harder their code was to read, the better it must be.

[wicked frigate]

there's a bunch of handy lil tabs down there like image history

[honest flint]

oohhh

i get it now, thank you

[brazen patrol]

and chess writer and photographer and.... lol. https://en.chessbase.com/post/study-shows-chess-is-a-powerful-tool-against-dementia-video (my latest article)

[wicked frigate]

and Tools with options like grid gen

[honest flint]

slick

[kind pewter]

what's your lichess handle?

[wicked frigate]

what do you mean, "just gave you a bunch of toggles"?

[west breach]

comfy can also look like this

[civic sigil]

Or this

[digital shoal]

a bunch of toggles that were off by default and wouldn't let me submit prompt, it was really weird and I couldn't figure it out D: probably me just being dumb ill check again on it later

[quiet vine]

Yes that looks more familliar.

[brazen patrol]

I don't really play online much.

[wicked frigate]

the code inside comfyui is actually really really good and clean

[dapper current]

Very. This is coming from two people who coded for Automatic1111 quite a bit.

[wicked frigate]

if it won't let you submit, it will give you an error message in the top-center of your screen

that should explain what you need to fix (and if not, screenshot it and ask)

[digital shoal]

alright thank you!

[lean granite]

Ooc, what kind of result do you get with these many nodes?

[plucky prairie]

It's the UI I have a problem with. There's a lot of zealots who think that if you don't like it, it must be that you aren't smart enough to understand it.

[civic sigil]

More coherent animations

[brazen patrol]

All I can say is that Swarm with Comfy works remarkably well here. It eats up my Ram as needed, but that is fine

[midnight shuttle]

Comfy seems useful and isn't too hard to understand if your workflow is mainly prompt -> image. But if you intend to do prompt -> image -> inpaint -> inpaint -> inpaint -> inpaint then A1111 is still easier.

[digital shoal]

Also, my friend was trying to set up stableswarm on his local machine as well, however it just loops the installer whenever he tries to launch it. any ideas what could be causing that?

[lean granite]

O so it's a workflow for making animations?

[urban fjord]

SwarmUI seems to have solved some of ComfyUI's UI problems but from the images I've seen I still prefer Auto1111.

[civic sigil]

Yeah with a lot of functionality thats convenient so I dont have to manually pre process each shot for SD

Just autuomatically does everything for me

[wicked frigate]

er... would again need to see more exactly what's going on to really say

[honest flint]

I personally still use both atm, automatic when im goofing off, comfy when i need to have a more involved workflow for a series of images

[high skiff]

damn, my upscale workflow is showing so many pomising qualities

[quiet vine]

Stable diffusion still doesn't know what an opossum is. Maybe I need to make an opossum lora.

[honest flint]

though ill give swarm a try, it looks like it might solve the goofing off use case

[high skiff]

but its having some hangups

[glass wolf]

dreamshaper xl 1.0 alpha

[azure oxide]

it fixed it for me but not for another user: i downloaded and installed net7 (coulda sworn it was already installed but i guess not?) and it ended up working

[honest flint]

i can't believe dreamshaper released a fine tune already

[high skiff]

look how good it does with eyes and hair

[digital shoal]

there isn't much to see, he gave me this screenshot. the installer ends and upon double clicking on the icon on the desktop it just runs through the installer again

[azure oxide]

here's the download link

[plucky prairie]

Same. I don't think auto1111 is very good with sdxl right now unfortunately. Hoping that changes.

[high skiff]

this workflow, if I can get it working right, could likely go reliable 2048x image gens

I just need to iron out some issues

[civic sigil]

Definitely looks promising

[high skiff]

it does extremely good for non human face things

[digital shoal]

i can try letting him know to try this, but he literally installed dotnet7 right before setting it up for the first time lol

[high skiff]

like my owl

[plucky prairie]

Nice hooters

[high skiff]

G

GUMMY

CATS!?

[azure oxide]

another user had this suggestion

[agile quarry]

is twopercent the bot maker? any chance this might happen soon or not soon im patient 🙂 dreamstudio credits in a bot that i could take to my own server would be awesome

[high skiff]

lmao, prompted black panther?

I have had that issue

[lean granite]

That's so awesome! I wouldnt go as far as to ask u to share your workflow but if you have some tips on getting started w an animation work flow like necessary nodes it'd be greatly appreciated c;

[quiet vine]

Swole puma.

[fair tartan]

almost

[autumn forum]

so... musculer...

[high skiff]

oh my god, too cute

[high escarp]

sampler comparison Illustration, a sailboat, maneuvering through an archipelago of cloud formations, classic wooden design with black pirate-like sails, dreamlike yet adventurous style trending on ArtStation, an 8K, highly detailed masterpiece inspired by the works of N.C. Wyeth.

Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 2889460738, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Version: 1.5.0

clik will be big

[pastel spire]

so I'm assuming comfy is the only place to be able to have a fully featured UI with refiner and everything working correctly.... but it comfyUI.... so I guess I'll wait...

[azure oxide]

just use sytan's workflow if you're too lazy to make your own like i was, it's as easy as using a1111 if you do so

[civic sigil]

I use custom nodes that crop to masked subjects, letterbox them into 1024x squares, assemble them into grids along with their guidance images for controlnet, processes them with img2img and then pasted them back onto the original render. Can do multiples passes on the same subject too. Then I plug it into a tool similar to ebsynth

[agile quarry]

comfy is kinda fun and you can just use someone elses workflow if you dont enjoy messing with the nodes

[high skiff]

guys, the potential is real

[azure oxide]

yeah but so is the laziness

[spark bear]

love how it interprets "black panther" as either the animal or the character

[cobalt kernel]

I learn a lot more with nodes ..
Than with installing random extensions ..

[high skiff]

SDXL 1.0 upscale workflow may be releasing with my official drop

[peak wraith]

I do like to load other people’s png files too. Lot of them have the workflows

[autumn forum]

good! it should

[spring fulcrum]

That would be amazing.... Those images are looking flawless

[lean granite]

Is the tool similar to ebsynth directly implemented as a node into comfy or is it a separate software?

[high skiff]

I am trying, I just need to figure out some small issues

[azure oxide]

i made an upscale workflow in yours by literally just appending two nodes to the output

[civic sigil]

Sadly it's seperate, that would be convenient lol. That's the only reason I can't call it a one-click solution

[pastel spire]

except I have to figure out how to wire up lora's

[high skiff]

its also very light to run too

[azure oxide]

no its already all wired up oh wait loras

[cobalt kernel]

Yeah nice thanks

[boreal bough]

@wicked frigate is there a way to load loras via your ui?
either by text or model selector directly?
and are there any preset workflows other than basic?

[high skiff]

oh wait, I see unforseen issues

shit

[honest flint]

sdxl refiner - sdxl base + refiner - sdxl base (same prompt and seed)

[high skiff]

ok, i have an idea

[lean granite]

Ahh rip, may I ask for the name of the software? Or simply why u prefer it over ebsynth and then I'll try to search for it

[wicked frigate]

not in the pure basic UI atm, but you can via the Workflow Editor tab

[urban fjord]

Tried gummybear cat with my 0.9 LoRA, not that good, but will retry with 1.0 and with different weighting.

[high escarp]

happy to help making xyz plots is my favorite part

[vale eagle]

loading mutiple loras in comfy seems a bit annoying

[high skiff]

the level of promissing here is huge

[crisp owl]

just string em together

[midnight shuttle]

XYZ plots are a great way to determine if "magic words" like masterpiece and best quality really do anything.

[civic sigil]

Idk the exact fork I used but it's based on https://github.com/OndrejTexler/Few-Shot-Patch-Based-Training

[thick goblet]

where the models at

[compact obsidian]

seems like theres specific words that randomly have a ton of impact, but synonyms hardly do anything

[cursive warren]

Is there somewhere to report a bug in clipdrop?

Negative prompts don't seem to be working there (for SDXL).

[cobalt kernel]

Yeah I saw that " siggraph " talk it was great ..
His real time demo is awesome ..

[compact obsidian]

idk ive also been having issues with negative prompts, and im running it on CoreML

[agile quarry]

https://huggingface.co/stabilityai

[compact obsidian]

maybe my implementation is 🤏 🧠

[pallid sail]

I had some issues with negative prompts a bit earlier today whilst trying to remove depth of field from an image using my Unreal plugin. Didn't seem to like the negative prompts

Used "blurry, depth of field, out of focus" as negatives. Could just be how I've implemented the compel library though

[lean granite]

Cool, I'll try using it later! Any reasons why you prefer it over ebsynth?

[dusk mica]

how do i make a styles csv in a111? it always errors out

[inner maple]

Do we need all kinds of crazy negative prompts with xl 1.0 like we did with some 1.5 models?

[cobalt kernel]

" Interactive style transfer " ,
From realtime live a few years ago :
[ https://www.youtube.com/watch?v=DDlYIzhoXfI&t=610s ]

[boreal bough]

@wicked frigate if one uses a config that doesn't work , such as: sampler + unspported scheduler, then your ui just freezes up after the error, and forever errors out until the webpage is refreshed - despite comfyui in the background still being fine

[compact obsidian]

my negative prompts here are "painting, artwork"

[civic sigil]

From what I understand ebsynth makes you composite the final results with alpha blending, with this software I can take the results and paste it onto my render to be used for another pass in SD without having to do anything like that. I also feel like it gives good results

[high escarp]

CFG 1-10 Illustration, a sailboat, maneuvering through an archipelago of cloud formations, classic wooden design with black pirate-like sails, dreamlike yet adventurous style trending on ArtStation, an 8K, highly detailed masterpiece inspired by the works of N.C. Wyeth.

Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: DPM++ 2S a Karras, CFG scale: 1.0, Seed: 2889460738, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Version: 1.5.0

[languid schooner]

Not a bad effort. Haha

[pallid sail]

What if you add "painting frame"?

[cursive warren]

Can you clarify "didn't seem to like"?

What I'm seeing is simply that negative prompts aren't doing anything at all (eg. A negative prompt of "hat" doesn't suppress hats, "monochromatic" still results in B&W or sepia images, artist names don't affect the style, etc.).

[cobalt kernel]

I usually put " symmetry " in negative prompts too ..
Along with " frame " so on ..

[civic sigil]

Yeah it's funny he could get it to run in real time since it can take an hour for a single shot for me lol

[compact obsidian]

that worked 😂 I suppose it wants very specific instructions

[civic sigil]

But it can be a lot faster if you dont train the network for too long

I just like to give it time for my tests

[cobalt kernel]

Yeah it was pretty impressive ..
I have tried running " fast style transfer " on ipad pro ..
It gets about 3 to 4 FPS with barracuda ML and burst processing ..

— Using " unity " engine ..

[wicked frigate]

ooo wasn't ware of that, noting down for when i'm awake

[civic sigil]

Is that the same technique? I think the actual slow part is the training of the network for each shot, inference is probably possible for realitime with a low enough framerate and resolution

[lean granite]

Fair enough, I'd love to see examples if you got some laying around! Always love to check out animations

[high crescent]

dragged some images into comfy ui to look at workflows and some of you dont seem to have the VAE on btw

[civic sigil]

Waiting on SDXL controlnet lol, I dont have any examples of this specific workflow until then

[hard fractal]

HWAT

[azure oxide]

bokeh is the most important one for negativing out the blur ime

[hard fractal]

wow

[hard fractal]

Drop that workflow, fam

[cobalt kernel]

Nope this was just inference ..
That patch based work is much more advanced ..

[azure oxide]

(workflow not included)

[thick goblet]

safe tensor or ckpt where that at ?

[fair tartan]

IronManCat

[midnight shuttle]

Always use safetensor.

[azure oxide]

we
love
pickles

[thick goblet]

ok do i have to request access from my work email id

[cursive warren]

BTW, I'm really impressed with how much more capable SDXL is. Various prompts that SD v1.x and v2.x couldn't really handle now produce very nice results.

For example, "A cowboy riding a dinosaur by N. C. Wyeth":

[spark bear]

its mostly sytan's workflow, just doing 35 base , 35 to 52 refiner , 11 fcg on refiner and high base resolution

[cobalt kernel]

What happens when you try " cadillacs and dinosaurs " ..

[agile quarry]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/blob/main/sd_xl_refiner_1.0.safetensors

[agile quarry]

.9 you did 1.0 should be public

[spark bear]

1.0 can pull some insane texture detail with just base gens

[outer junco]

This model seems really promising https://civitai.com/models/112902?modelVersionId=126688

[rustic surge]

So do you have to run comfyui to get the best results?

[cobalt kernel]

Uh no its the same model ..
Should be similar results if the code is correct ..

[thick goblet]

is nsfw censored ? in xl 😏

[languid schooner]

Ahem...I present to all of you.....

Duck

[cobalt kernel]

Yes you need to fine tune the knobby bits ..

[outer junco]

wdym?

[outer oxide]

the latest InvokeAI (3.0.1rc2) is working for some of us, but if you're feeling wary the safe thing would be to wait until there's a final release instead of "release candidate" https://github.com/invoke-ai/InvokeAI/releases

[thick goblet]

ok other people will do that i ain't gonna go in that

[outer junco]

oh wait nvm

[agile quarry]

boobs and butts are ok i think but main genitals dont show

[outer junco]

#✨｜sdxl message

[lean granite]

o mb, wasnt aware that sdxl didnt have controlnet yet! Even older ones would be interesting to see tho just to see how few-shot-patch-based-training works

[peak dove]

^..^<

[sonic lantern]

i have a 4090, what's the best way to train or make a lora with SDXL 1.0 right now? (sorry if this is answered somewhere)

[vocal rapids]

finger is still suck , lol

[outer junco]

Has some promising starts to nsfw content

[thick goblet]

but is it based on what they released yesterday

[trim orbit]

meanwhile, deloreans

[outer junco]

Its based on SDXL 1.0 and recommends vae 0.9

[thick goblet]

inpainting outpainting built in ? or seperate models on the way ?

[high escarp]

prompt please

[civic sigil]

Not sure if I have examples that are worth showing, and that part of my workflow I could always switch to something like Ebsynth or similar methods anyway. But there are examples of people using SD with it already on the internet

[cobalt kernel]

I like the tall format too ..
It seems to get more dynamic layouts - hit or miss ..

[outer junco]

looks like on the way currently its only using 10% of their dataset for finetuning, this is the SD 1.5 model the dataset is from https://civitai.com/models/4384/dreamshaper

[austere bay]

The first: A mechanical >^..^< hunting in the dark forest. Adventurous, new adventure, forest, rocks, stream, ripples,angron,angry,phoenix,sad dragon, by Jeremy Mann and Greg Rutkowski
The owl: A mechanical owl hunting in the dark forest. Adventurous, new adventure, forest, rocks, stream, ripples,angron,angry,phoenix,sad dragon, by Jeremy Mann and Greg Rutkowski

[high escarp]

i bow bless the

[pallid sail]

I'm surprised at how well the model is at creating consistent looking characters from specifying how many of that subject occurs in the scene along with img2img hints

[reef sky]

A mechanical UwU noticing ur O.O wuts this

[outer junco]

what ui is that beauty?

[austere bay]

That is not IronManCat, this is xD

[pallid sail]

My Unreal engine SD plugin

[latent zodiac]

SDXL on Mac M1 = snail lol

[civic sigil]

Actually that's IronCatMan

[timid basin]

love it

[cursive warren]

Have Stability mentioned anything about SDXL's training data? Is it still just LAION, or have they expanded it with additional sources?

I'm asking because it seems as though SDXL recognizes more artists than the previous models do.

[fair tartan]

oh shoot

[civic sigil]

The Clip could play a role in that too

[outer oxide]

there was a comment during today's session that it's no longer just LAION, but they did not elaborate on that.

[austere bay]

Monster hunter...

[wary bay]

gn

[cursive warren]

Okay, cool.

[high escarp]

Roll for initiative

[cobalt kernel]

Mystery meats then ?

[autumn forum]

he is. the. goat.

[high escarp]

did a prompt Matrix on The stylization or enhancement words Illustration, a sailboat, maneuvering through an archipelago of cloud formations, classic wooden design with black pirate-like sails, dreamlike yet adventurous
Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2889460738.0, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Version: 1.5.0

[autumn forum]

you need to post all these somewhere. this is great!

[strange mist]

late to the party fellas, do I need to use the offset lora? 😄

[high escarp]

you have my permission

[outer oxide]

heck no this is the fully open source Stability AI latest and greatest we don't do no mystery meats here! /s

just because it has been developed behind closed doors with no published data set does not mean you have to call it a "mystery" 👻

[cobalt kernel]

I think thats just to control contrast ..

[strange mist]

ah, ok. thanks 👍

[urban fjord]

Ironman cat with Lora + DreamshaperXL without refiner.

[cobalt kernel]

Ah right so you're using " unreal " to create the scenes ..
Then feeding that into control net to generate a 2d image then ..

[high skiff]

ok, i don't wanna call it too soon, but this 2048x upscale workflow I am testing for SDXL 1.0 is yielding... incredible results

[cunning geyser]

whattya mean

[high skiff]

like, really damn good (these are crops)

[digital shoal]

man i'd love to try that lol

[high skiff]

I am still working it out, but so far, mannnn

[pallid sail]

These images are with SDXL so I'm not using ControlNet here but I can with earlier SD models - just img2img here

[high skiff]

look at the difference here

[reef sky]

Texture

👀

[high skiff]

and here

[outer oxide]

deloreans, deloreans, deloreans!

[high skiff]

and here

results are looking very promising, just saying

if all goes well, this will be released with the 1.0 drop of my SDXL workflow

[urban fjord]

More IronMeow, not so happy with these but where trying to make it wear a helmet.

[high skiff]

i did a raccoon before and after upscale too

[autumn forum]

thats very good. i dont see any of the artifacting of normal upscalers

[high skiff]

its a very simple process, but I have to tune it jusssstttt right, and also work on a third prompting field (which will be very easy to do)

[boreal bough]

do you get any artifacts on cfg 5?

doing the megagrid right now - and saw that most of my issues came from too high cfg values

[high skiff]

but its giving good result after good result for now

[earnest anvil]

anybody know how upscaling works in A1111?

[pallid sail]

Can't wait till we have SDXL ControlNet. This one is with depth

[cursive warren]

One of the ways SD v2.x was worse than v1.x was interpolating between styles that are too different from each other (eg. adding a B&W artist can result in some really weird color palletes). SDXL not only fixes this regression, it now does a much better job than SD v1.5.

OTOH, the problem with one artist style overwhelming the other seems slightly worse (ie. 'strong' styles are stronger than ever).

[violet merlin]

Has anyone noticed that with SDXL 1.0 LoRAs, very simple prompts yield good results that look a lot like the subject, but as the prompts get more complicated, they drift away from the training images? Is there a param I can tweak to help with this?

[boreal bough]

bad captioning during training causes this

[autumn forum]

im using 6. i just have a really crude way of using an upscale 4x model thats really good then downsizing to 2k. it works well but you zoom in and see the artifacts a bit

[violet merlin]

Ahhh interesting, will work on that next, thanks!

[high skiff]

I have just one more thing that could help with this upscaler, and I might try and source from some friends

[boreal bough]

20/35/50 Steps
cfg 5
cfg 7

[urban fjord]

Sadly a bit cooked.

[autumn forum]

oops guess i couldve zoomed in

[dense chasm]

swarmUI is more customised than bots here

[austere bay]

Is that sdxl? How did you use depth?

[pallid sail]

No no, this image was with SD 1.5 with a depth controlnet. I dump the depth + normals straight out of Unreal

[boreal bough]

iron raccoon baby. the new hit movie!

[boreal bough]

uni_pc sampler giving... interesting results XD

[pallid sail]

Soon as depth and normal control nets are released I'm going to slot them straight in and see what I get

[royal fern]

am I understanding it right that you should use the external vae for better results?

[vale eagle]

I think swarmui need to refine the ui part

[autumn forum]

that image looks like its 0 degrees lmao

[nocturne dove]

I couldn't get good results with uni_pc because it takes so few steps that I couldn't really get the split for base/refiner right.

[high escarp]

The owl: A mechanical owl hunting in the dark forest. Adventurous, new adventure, forest, rocks, stream, ripples,angron,angry,phoenix,sad dragon, by Jeremy Mann and Greg Rutkowski

Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2889460738, Size: 768x1040, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Denoising strength: 0.5, Hires upscale: 1.5, Hires upscaler: 4x_NMKD-Siax_200k, Version: 1.5.0

[sharp robin]

Has anyone compared .9 refiner vs 1.0?

[austere bay]

Ah, ok, I was already looking for the controlnet repo xD

[cursive warren]

Is it just my imagination, or does the order of terms in the prompt seem to matter much less for SDXL than it does for the previous models?

eg. The prompt "By $ARTIST_1 and $ARTIST_2" doesn't now produce results much different from "By $ARTIST_2 and $ARTIST_1".

[vital wolf]

@wicked frigate Unknown parameter type's data type? LIST" getting this error

[vale eagle]

Same

[languid schooner]

Well, I tried. XD

[high skiff]

ok, I need some community help

I need a pixel upscaler that is very sharp and high detail, but does not destroy noise

It needs high detail, and to strengthen noise

[earnest anvil]

I have used both with base 1.0, not a big quality difference but they produce different content results

[high skiff]

current issue I have is the pixel upscaler I use is sharp, but it smooths out some textures

[cursive warren]

"Do, or do not. There is no try."

[high skiff]

if I can get that goal, then I can ship this workflow

[sharp robin]

Is one objectively better?

[meager isle]

Has anyone been able to create image variations with 1.0?

[reef sky]

have you tried 4x ultrasharp?

[fair tartan]

4xUltraSharp and NMKD-Superscale i've had the best results keeping textures

[high skiff]

yeah, unfortunately, it smooths things out too much

its the sharpest tho, and it works good for non fine textures

[wicked frigate]

errr what are you doing to get that

[sour obsidian]

new one performed better over thousands of votes on the bot if that helps haha

[high skiff]

ultrasharp is what I am using now

[high escarp]

4x_NMKD-Siax_200k and LDSR

[high skiff]

do you happen to have a link to either?

[vital wolf]

ok, its the scorer, getting the error ehrn i turn it on

[high skiff]

if I can find the right pixel upscaler for this workflow, I can include the full upscale workflow in my 1.0 release in the next few days

[earnest anvil]

Don't think so. Here's a comparison, 0.9 on left and 1.0 on right

[sharp robin]

Helps tremendously im on potato pc and just confirming the vae issue took me a while. Much appreciated

[high escarp]

https://huggingface.co/gemasai/4x_NMKD-Siax_200k

[autumn forum]

care to try mine?

[high skiff]

thank you!

oh hmm

[autumn forum]

actually that one is very similar to the one i have lmao

[high skiff]

I need a pixel upscaler, the rest of my workflow is working good ATM

[lean granite]

im confused about the refiner, does it come included with the base model? Is it a separate dl? I have the offset lora and the base dl rn and dk where the refiner is

[high escarp]

https://huggingface.co/utnah/LDSR/tree/main

[winter verge]

Monet pastel of a [cute|ethereal|beautiful|elegant] [ghost|rusalka] girl with glowing blue eyes [singing|singing|praying] in a [swamp|lake|puddle] under a silver moon. [Trees|Plants|Swamp growth] surrounds the mire, lending it [an eerie|a dark|an unnerving|a mysterious] presence.
([masterpiece|realistic|ultra realistic|highly detailed|64k|detailed|highres]:2)

Neg: close, closeup bokeh, cartoon, render, octane, (worst quality, low quality, normal quality:1.5), (signature, text:2)

[pallid sail]

I've been checking daily as well XD. Diffusers just added support for a ControlNet pipeline for SDXL but no checkpoints are available yet. I'm going to implement it ASAP so I can just drop the checkpoints in as so as they're releasd

[dusk mica]

what UI do you guys use?

[wicked frigate]

stableswarmui :D

[earnest anvil]

A1111 for me

[high escarp]

automatic 1111

[urban fjord]

A1111

[sharp robin]

Damn. On the left helmet looks more real and face a little better but small details like the flag patch look better on right

[pallid sail]

Unreal Engine

[sharp robin]

Thank you.

[sour obsidian]

🎮 🥅

[high skiff]

this one looks like it might work better-

[earnest anvil]

yeahh it was like that for other stuff too. Difficult to pick a winner

[high skiff]

I need to run more tests, but I might need to credit you as a contributor to this haha

[high escarp]

Wow barely a inconvenience with inpainting

[high escarp]

true

[cursive warren]

Well, sure, but that would still be a reflection of the training data, surely?

[lean granite]

did u delete the msg or someone else xD

[autumn forum]

if that one doesnt work try this one. name nmkd just different by a little lol

[high skiff]

I can try both

[tribal moat]

hey guys, does anyone here uses ComfyUI? if so do you know how to use deepbooru in it?

[reef sky]

Link got disappeared 👀

[sharp robin]

Nothing is ever easy. I appreciate the comparison.

[high skiff]

if I get what I am looking for, then my SDXL upscaler workflow will be out at the same time as my official 1.0 workflow

[cobalt kernel]

No I deleted it - the links are also pinned ..
If you look at the " pin " icon at the top right corner ..

[earnest anvil]

guys, from what I understand, comfy generates much faster than A1111. are there any other benefits?

[lean granite]

easier to iterate with and understand what nodes are changing what in your picture

[high escarp]

Unnecessary confusion

[fair tartan]

ya running all cat prompts

[sharp robin]

Workflows be like need more nodes. And then nodes for the nodes

[earnest anvil]

hahaha

[cobalt kernel]

I feel " comfy " is a bit more solid for me ..
Then its modular and can use " comfy box " " swarm ui " as the front end ..