✨｜sdxl

[orchid ledge]

they should be the same, check hash?

[pearl geode]

holy cow

[shy kelp]

Can someone send me their StableSwarm Comfy work flow for SDXL?

[pearl geode]

@candid spindle the best i saw here so far.. will u share the prompt? 😄

[true wing]

alright thanks. I just wasn't sure why they were posted as separate projects

[candid spindle]

(a (gritty and dark:1.2) photography) of a (lonely astronaut:1.1) inside a (crumbling and eerie:1.3) (abandoned destroyed space station:1.4) with (sparks and wires:0.9) hanging from the ceiling and (a glimpse of Earth:0.8) in the background.

[orchid ledge]

seperate vae is nice

[candid spindle]

I use Bing to create the prompts, I can share the bing prompt

[honest flint]

im happy they decided to release the base and refiner models as seperate safetensors, thankyou

[pearl geode]

outstanding! thanks

[edgy sierra]

how

[tulip kelp]

Bing can help create prompts?

[candid spindle]

Hello Bing. I will give you some words and you will add details to them. The details should match the words and make them more descriptive, they should also always be visually descriptive and make them much longer than the initial words, mix your own ideas in there. You will also mark some parts of the line as important or unimportant. For example, “a (blonde girl in rainy weather:1.3) in a grass landscape” means that the rainy weather is more important than the other details. To mark a part of the line, you will use brackets and a colon like this: (part of the line:importance score). The importance score is a number from 0.5 to 1.5. The higher the number, the more important the part of the line is. Please use 0.5, 0.6, 1.4, and 1.5 sparingly. Use the importance scaling only when it makes sense and when you really want to make something important or unimportant.

Please search for “StableDiffusion prompting guide” and “Midjourney prompting guide” on the web and summarize them for me in less than 3 words each. This will help you know what kind of details I want.

If you understood the instructions, say “Roger” and I will give you some words to add details to.

[supple bronze]

prompt: eco city, apartment with plants, futuristic, headquarters in background settings : sampler - DPM++ 2S a Karras, steps 20, 1024x1024, cfg scale 8

[tropic turret]

Nice

[tender birch]

has any1 figured out how to use any kind of face fix in ComfyUI?

[candid spindle]

have fun with the bing prompt yall lol

[void parrot]

I'm getting grainy images at 1024x1024 in Auto1111. Any thoughts?

[runic lark]

Now do "Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch"

[smoky pewter]

When I generate 512 x 512 images are ok but 1024x1024 are like this, anyone has an idea?

[bold osprey]

on it

[orchid ledge]

https://github.com/melMass/comfy_mtb this has a face enhancer

[echo stratus]

Yes I have, 1.0 is better. Same prompt same seed, 0.9 is the first

[true wing]

I don't mess with VAEs too much, I usually just use whatever comes with the model I'm using or the base stability VAE, what's the benefit to offering the VAE standalone?

[orchid ledge]

people don't need to supply vae's with every checkpoint - but you can continue as you are

[bronze arch]

first one looks nicer

[orchid ledge]

only unet+clip changes with training

[mortal mural]

my pagefile went from 9gb to 30gb is that normal

[echo stratus]

first one doesn't follow the prompt

[true wing]

cool, thanks for the info

[bronze arch]

whats the prompt

[hasty cipher]

dont even know how to use the gosh darn refiner

[thorn lion]

attempting this

[wary bay]

you might be running the Refiner model instead of the base?

[tender birch]

looks good, need to find the install guide first xD

[hasty cipher]

so no one knows how to use the refiner model on top of base? please help

[tulip kelp]

I'm getting around 5it/s, is that roughly right for Nvidia cards?

[thorn lion]

worked

[smoky pewter]

you are correct, base model wont load, and no one provided me information how to fix that, i had RAM not VRAM out of memory message when loading model.

[echo stratus]

"Satan standing in a bloody portal to hell, horror art, ominous atmosphere, inspired by H. R. Giger and H. P. Lovecraft, towering figure, sharp horns, glowing red eyes, detailed demonic features, surrounded by shadows and flames, dark crimson blood dripping from the portal's edge, eerie lighting, gothic, macabre style, intense emotions, unsettling background, desolate landscape, tortured souls, intricate details of suffering and horror lora:offset_0.2:1" yes I used the offset lora on BOTH.

[astral knoll]

@bold osprey @pearl geode

[wary bay]

that would be your issue, the refiner doesn't do the 1024x1024 alone it needs to be fed an input

[green python]

damn new dalle 3 bout to drop!!!!

[pearl geode]

@astral knoll nice!

[warm hazel]

What's the optimal sampler to use for SDXL? I heard DPM++ 2M isn't good.

[smoky pewter]

if I would like to load base model what could I do

[edgy sierra]

what is the prompt?

[cyan crown]

well i rarely use vaes.....

[pearl geode]

this prompt is good. you are a prompt magician.

[wary bay]

if you're on auto1111 unfortunately I can't help, I only use comfyui

[bold osprey]

lovely

[echo stratus]

#✨｜sdxl message

[smoky pewter]

its alright, I do use a1111

[high skiff]

honestly, more legible than the original lol

[green python]

bro uses auto1111💀

[candid spindle]

Oh it's mostly Bing doing the hard work lol

[bold osprey]

lmao you are right

[high skiff]

lmao

[exotic path]

[sudden cliff]

My 2c is that SDXL1.0 is rarely quite a bit better than 0.9, sometimes a bit better, and mostly quite similar. Not the biggest jump but overall I'll probably stick to 1.0 most of the time.

[high skiff]

bloud

[pearl geode]

there is no difference.gif

[bold osprey]

my brain stops midway through

[vale phoenix]

Does SDXL work with controlnet in comfy?

[shy kelp]

So, SD XL 1.0 doesn't have censoring ? :O

[hoary saddle]

hey sorry if i missed any answer over the chat, this workflow is fire! was curious if you could add in a load image node for i2i? i tried a few things and they failed miserably 🙂

[shy kelp]

Yup, for 0.9, it worked

[orchid ledge]

we need new controlnet models

[olive notch]

What VAE are you using?

[exotic path]

Depend things, this is not NSFW but there is no nudity

[proper cape]

Hi everyone, am trying to created a seemless pattern but when I check the result, is not seamless... Any idea how to make SD a perfect tiled texture?

[high skiff]

I will be looking into it. For now I am working on some compatibility with 1.0, as some things seem to have changed

[azure oxide]

are there any optimizations i should be using in comfy with a 3090? --highvram or anything?

[exotic path]

The one included in the model, i didn't download a separate vae

[shy kelp]

Ah ! Is nudity removed ? x)

[hoary saddle]

all good, appreciate ya!

[exotic path]

Yes

[thorn lion]

got your workflow working on StableSwarmUI/ComfyBackEnd. my page file is mad at me LOL i need more RAM for model loading. or I could just close chrome. meh.

[shy kelp]

damned x)

[smoky pewter]

i would like to stay with it, it works fine for my friend, so i would like to use it for myself also but it does not work for now

[hoary saddle]

for what's it worth, the default workflow you released works great on 1.0 🙂

[exotic path]

I use Automatic1111 and it works just fine

[bold osprey]

love the stache

[olive notch]

Oh okay, I asked cause you had a separate VAE node in your workflow

[pale mountain]

well that was a funny test run. back to v1.5 now, lol.

[shy kelp]

anyone actually managed to get it working on their PC yet using the Swarm UI?

[high skiff]

works even better with some changes I am making :p

[shy kelp]

anyone actually managed to get it working on their PC yet using the Swarm UI?

[bold osprey]

I am afraid to do that

[hoary saddle]

tease

[cyan crown]

just fine, but at the moment 1.5 models are still better

[orchid ledge]

if you add 1+2 you have not removed 4... it is not trained on pornography. it has good understanding on what a body looks like but it does not know genital detail if that is important to you

[trim orbit]

disagree

[green python]

it uses to much vram and is slow and difficult to set up, comfyui 20x better

[bold osprey]

where can I find this and the workflow that was mentioned?

[shy kelp]

anyone actually managed to get it working on their PC yet using the Swarm UI?

[pearl geode]

have to test how good it is now with creating places like Berlin Alexanderplatz 🤔

[high skiff]

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

[unreal plover]

New planet discovered

[cyan crown]

not talking about base models, but trained

[lilac wren]

Hey, i tried your workflow but i hav ethis error :

[high skiff]

I will be releasing 1.0 of my workflow hopefully within the next couple days

[shy kelp]

not suited for adult gaming dev, then ^^

[trim orbit]

1.5 has very clear ceilings that sdxl just completely smashes through

[pearl geode]

oh, its your workflow! awesome. but when i go higher in steps, eversything looks like half steps

[orchid ledge]

wait for your community training

[trim orbit]

refined includes doing better prompts. SDXL already is refined if you're a smart prompter instead of cherry picking bad results to confirm your preassumed biases

[shy kelp]

🥂

[high skiff]

its got some additional values you have to change

[thorn lion]

I did

[high skiff]

will be setup on macros for 1.0

[exotic path]

Of course they are, they have been finetuned on specific subjects, and finetuned SDXL models will be far better from SD 1.5 IMO (you can compare SD 1.5 base model and SDXL)

[shy kelp]

Prompt outputs failed validation
CheckpointLoaderSimple:
- Value not in list: ckpt_name: 'SDXL Refiner.safetensors' not in ['OfficialStableDiffusion\sd_xl_base_1.0.safetensors']
CheckpointLoaderSimple:
- Value not in list: ckpt_name: 'SDXL Base.safetensors' not in ['OfficialStableDiffusion\sd_xl_base_1.0.safetensors']

[soft zealot]

WTD?

How come these are so different

The LH image was generated in 0.9 earlier, just dropped it back in comfy and set for V1 so all the settings are the same but totally different

[bronze arch]

does this workflow have refiner aswell

[cyan crown]

Of course

[high skiff]

my SDXL model names are different, just have to load them in as they are named for you

[high skiff]

yes

[simple citrus]

damn generating 10 images takes 1 minute with a 3090 ❤️

[indigo carbon]

started tuning SDXL, opinions?

[unreal plover]

When the Inpainting model will be release ?

[thorn lion]

I may have overreacted and Sytan's solution above is right but if it isn't.....

Close terminal, open terminal in StableSwarmUI folder, git pull, launch-windows.bat, try again

[warm hazel]

Hmm, my tests aren't turning out optimal, but I'm getting something that works at least.

[pearl geode]

is it enough to change in "SECOND PASS LATENT" box the start_at_step:13, end_at_step:20 .... to end_at_step:50 ?

[hard fractal]

already!

[shy kelp]

I'm using ComfyUI with the template it listed for SDXL and noticing a sort of 'square edge noise' or 'fine line jigger' across a variety of prompts. Anybody know what I mean? Tough to describe

[orchid ledge]

anyone like salt

[hard fractal]

July 18

[trim orbit]

from what i understand there will be a lot of same seed comparisons of 0.9 and 1.0 where it seems 0.9 looks better. Part of where they took 1.0 is allowing it to be a lot more refinable for loras but now it sort of doesn't do as good images all the time

[upper sage]

You're asking a Discord server if they like salt?

[tulip kelp]

@high skiff I'm trying to undertand how the workflow repo fits in with StableSwarmUI. Is your work going to be basically moving quicker and being more experimental with SDXL?

[cyan crown]

anyone knows if a1111 uses refined model?

[shy kelp]

Ayo sytan <3

[unreal plover]

that is a salty image

[limpid rose]

Do

A gecko holding a sign with the words "SDXL" on it.
At least that has been working for me.

[shy kelp]

I think you 've done it

[cold cargo]

28?

[upbeat summit]

could be seed dependent, but the faces look good on the right

[candid spindle]

looks kinda like Serana

[buoyant flower]

Tried to train an anime character LoRA, got shit results lmao

[high skiff]

you wanna try to keep a 2/3rd ratio

So base does 2/3 of the steps, and the refiner does the last 1/3rd

So at 50 steps, try 30 steps for base (start at 0, end at 30), and try 20 steps for refiner (start at 30, end at 50)

[soft zealot]

its not so much saying it looks better in one than the other, itsthe huge difference going from what is almost a cinema still to a toay action figure

[median mulch]

does this" https://huggingface.co/stabilityai/sdxl-vae
go into the comfyui models vae folder? and is it used automatically or i have to load it?

[upbeat summit]

I take it with parmesan instead, please

[indigo carbon]

same seeds on both sides, obviously

[hard fractal]

You don't have a new one?

[pearl geode]

that was my thought, some kind of ratio. Thaaaanks!! ❤️

[high skiff]

working on a new one for 1.0, as some things have changed

[thorn lion]

StableSwarmUI is a UI front-end for any(?) Stable diffusion back end. Comfy is just one of those back ends

[high skiff]

finding better results with some small tweaks

[urban fjord]

In case no one answered you, there is no need for an inpainting model.

[orchid ledge]

anyone have a screenshot of the recommended comfyui flow - i just want a quick check if its any different

[trim orbit]

can still tell theyr'e the same seed

[urban fjord]

We really need a FAQ that can answer all the repeated questions.

[high skiff]

I wish haha

[echo stratus]

people will always complain, I like 1.0

[high skiff]

well, we do have the FAQ bot

[shy kelp]

what sampler are you guys using and how many steps in general?

[urban fjord]

I'm using DPM++ 2M Karras with 20 steps because I'm too cheap for any more steps.

[wary bay]

0.9 vs 1.0, same prompt and settings... interesting difference

[soft zealot]

oh I dont disagree, this upscaled version looks better than its non upscaled version IMHO though

[bold osprey]

try 32 once bro and you'll see

[high skiff]

I find that 1.0 seems to have less severe color processing

[upbeat summit]

dpmpp_sde_gpu 25-40 steps (normal or karras)

[tulip kelp]

Ok thanks, and from what I understand, Comfy is a sort of shadergraph, but for block level components for ML operations? text processing, transformers, upscalers ? Is that roughtly right?

[high escarp]

Just got sdxl working on automatic 1111 curious what Samplers steps and CFG people are using

[urban fjord]

32 steps means 12 steps more waiting for each image so no thanks.

[primal nimbus]

how

[zealous horizon]

How do we load LoRA VAE etc in SwarmUI? THere is no instruction on how the folders should be named or where to put them. I tried the structure (names) from auto1111 and the ones from ComfyUI but neither show up anywhere (using Comfu Backend)

[high skiff]

current thoughts, people?

[bold osprey]

32+8 and 4.5 -7.5cfg depending on the prompt

[high escarp]

update to 1.5

[thorn lion]

I've considered this. The repeated questions are all gonna be pretty similar but slightly different so just live scrape all messages, ask LLM if this is a question, and if it's a question, respond by pulling the answer from the FAQ.

[soft zealot]

Its feeling like YMMV when it comes to recycling old prompts.
This is Stable Diffusion , This is the Way

[limpid rose]

50 samples, 7.5 CFG is working fine with me

[trim orbit]

i dragged my sdxl file into the models folder and loaded auto

and bam it worked

[primal nimbus]

ANYONE FINETUNED ALREADY

[halcyon plank]

It just seems like they're not using noise offset anymore

[quasi notch]

having endless fun with this

[placid edge]

minimal prompt and effort attempt lol

[high escarp]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.5.0

[polar epoch]

That one took me 3 weeks to learn, and a few months to nail :P

[primal nimbus]

just the base model is needed right

[high skiff]

the 1.0 base is way better than the 0.9 base, I will say that

[upbeat summit]

also scheduler normal or karras - depending on the look I'm going for

[limpid rose]

yeah its out on the release branch. just pull the latest. i didnt even have to nuke my venv folder.

[quasi notch]

its better at listening to the prompts

[high skiff]

but the difference between 0.9 with refiner and 1.0 with refiner is smaller

[warm hazel]

Any recommendations for the base SDXL settings?

[high escarp]

base in text to imge refiner in imag to image

[polar epoch]

Where can i find configs for comfyUI? No idea how to setup nodes yet, and don't know if SDXL 0.9 will work just as good for 1.0 :P

[limpid rose]

in the meeting today they said you could just drag and drop - no json file needed

[warm hazel]

We still doing CFG 7?

[tulip kelp]

When I'm using StableSwarmUI, the only file in the models folder is sd_xl_base_1.0.safetensors so I guess so?

[urban fjord]

If you're just going to use the baes then Auto1111 is the easiest.

[trim orbit]

yeah you can use the refiner in auto too though

[high skiff]

I found CFG 2.5-9 is typically good to play with

[limpid rose]

it works for me.

[trim orbit]

its a seperate generation is all

[high skiff]

7.5 is my base shipped value for the workflow

[thorn lion]

Shadergraph......a node based system for calculating shaders is my guess....in which case, kinda? Yeah. Good enough to understand to start fiddling and get a better more accurate understanding.

[limpid rose]

img2img i usually do 12.5

[primal nimbus]

how so

[bold osprey]

don't change a running system xD

[trim orbit]

img2img

[shy kelp]

Working steps for windows (and an rtx 4080),

• Have git command line installed
• https://github.com/Stability-AI/StableSwarmUI#installing-on-windows
  -Once installed, make sure refiner and base models are in this folder StableSwarmUI\Models\Stable-Diffusion\OfficialStableDiffusion
  (For me just the refiner was missing https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/blob/main/sd_xl_refiner_1.0.safetensors)
  -Once in the UI, head over to Comfy Workflow editor, clear the workflow and load, enter this url as a filename https://github.com/SytanSD/Sytan-SDXL-ComfyUI/raw/main/Sytan SDXL Workflow v0.5.json
  -Move the workflow till you see two orange boxes below that are named base and refiner, make sure they are selected for the corresponding base and refiner models (refresh if they are not showing and you've moved them into the stablediffusion model folder)

[high escarp]

what sampiler

[placid edge]

where does one find the refiner ? 😮

[echo stratus]

whats a good 4:3 res for sdxl?

[primal nimbus]

but where to load the refiner in

[warm hazel]

Not sure I understand your point?

[high escarp]

https://www.reddit.com/r/StableDiffusion/comments/15afd3y/omg_its_out/

[stray depot]

Search for it

[trim orbit]

copy it into the same model folder, and load it like a model, since it is one

you can generate images with the refiner too but meh

[placid edge]

no i mean within that UI

[urban fjord]

Hmm, what is really the point of the off-set noise LoRA? I did some tests with it and either got weird results or worse results than a similar kind of model I made as a joke for 0.9

[upbeat summit]

cfg 12 but that is a very rare case
miniature sailing ship sailing in a heavy storm inside of a horizontal glass globe inside on a window ledge golden hour, home photography, 50mm, Sony Alpha a7

[bold osprey]

using it with text sometimes

[high escarp]

nice

[jolly creek]

1152x896

[limpid rose]

shoot thats awesome.

[clever verge]

Think the right one is best (most developed), the one in the middle is also good.

[shy kelp]

tell more ? :O

[echo stratus]

ty

[tulip kelp]

Oh yeah sorry I'm neck deep in Unity most of the time so it's the graph editor my brain is filled with. Yeah, so who authored the Comfy graph I'm currently using in StableSwarmUI ? And I guess anyone can write the graphs, hence Sytan's repo ?

[bronze arch]

didnt SAI team mess around with the CFG and statistically determine which is the "best"/favourite CFG to use?

through votes

[high skiff]

being blown up omg

[cerulean valley]

@⛧ Sytan ⛧ hey Buddy thats SDXL 1.0 in ComfyUI?

[upbeat summit]

have fun 🙂
miniature sailing ship sailing in a heavy storm inside of a horizontal glass globe inside on a window ledge golden hour, home photography, 50mm, Sony Alpha a7

[thorn lion]

The only edit I would make is not to say that the models just need to be in the Stable-Diffusion folder and that your OfficialStableDiffusion folder is for personal organization

[candid spindle]

is this nsfw?

[bold osprey]

nah it's for science

[high skiff]

thats my workflow :p

its faster than img2img still :>

[upper sage]

https://tenor.com/view/community-ken-jeong-ben-chang-ill-allow-it-allowed-gif-4468552

[high skiff]

just had to tweak some things

[high skiff]

that is so cool OMG

woooow

[thorn lion]

Comfy authored comfyui. And yeah, anybody can write graphs. For the next month's, that's gonna be currency round these parts

[high skiff]

no idea, I have yet to see any of that info besides some small stuff in DM's

[shy kelp]

i wonder about the original picture teached to the model ...

[pearl geode]

OK weird. i use comfy the whole time, with the setup from @high skiff ... and now i got this error out of nowhere:

Error occurred when executing CLIPTextEncodeSDXL:

Sizes of tensors must match except in dimension 2. Expected size 77 but got size 154 for tensor number 1 in the list.

File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 145, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 75, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 68, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_clip_sdxl.py", line 44, in encode
cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 468, in encode_from_tokens
cond, pooled = self.cond_stage_model.encode_token_weights(tokens)
File "C:\Users\User\Desktop\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\comfy\sdxl_clip.py", line 75, in encode_token_weights
return torch.cat([l_out, g_out], dim=-1), g_pooled
Queue size: 0
Extra options
Running
Pending

[trim orbit]

https://tenor.com/view/rock-dj-robbie-williams-anatomy-piel-gif-16668117

[thorn lion]

And definitely don't apologize. It's only by context switch and multidomain thinking that we will be able to push the bounds of the tech

[high skiff]

[candid drum]

Is there an official work flow for comfyui? Did it change from 0.9 or are those still pretty optimized for 1.0

[high escarp]

If any of the moderators or Devs are on here what or the steps and Samplers used cFG use for the Bots

[shy kelp]

#✨｜sdxl message

[orchid ledge]

if you're getting 'expected size' errors check you're using the right vae for the right pass

[urban fjord]

Weird, the offset noise LoRA seems to ruin any image of Pikachu.

[pearl geode]

can someone try this prompt in comfy ui with Sytans setup?

A (dramatic and action-packed:1.2) scene of (Tracer:1.4), a zeitspringende Abenteurerin, and (Winston:1.4), a genialer Gorilla-Wissenschaftler, fighting against (Reaper:1.4), a maskierter Söldner mit übernatürlichen Fähigkeiten, in the (futuristic and colorful:1.1) city of Busan, painted in a (bold and expressive:1.1) style inspired by (Roy Lichtenstein:1.4)

[dry plinth]

What node do you add in comfyui to randomize the seed? And does the seed of the model and the refiner need to match?

[jolly creek]

Someone one day posted these, might come in handy.

[high skiff]

working on it

[thorn lion]

on it

[wicked frigate]

correct, but: Comfy is the most tightly integrated backend, and direct workflow tools and all are really powerful features from that tight integration

[upbeat summit]

you add a Seed node. you can choose to randomize, increment etc your seed. the base and refiner seed do not have to be the same

[clever verge]

Might have been me. 😄

[orchid ledge]

i'm not seeing the same

[tulip kelp]

Crazy, so your prediction is that folks will be able to get more performance out of SDXL via extensions to the core architecture? Or will it be a little bit like the incremental arrival of sampling methods. Not all better, just different?

[hard fractal]

middle, for me.

[upper lantern]

does anyone know what precisely is supposedly not working in A1111? the contrast lora seems to fail, but the base and refiner both work as expected on 1.5 from what i can see

[wary bay]

0.9 vs 1.0, noticing some pretty clear stylistic differences

[clever verge]

@high skiff Might have same issues as you found no? It's like a rainbow.

[wary bay]

1.0 seems less "animated"

[pearl geode]

Thanks! No idea why it gave me a error. everything else worked.

[void parrot]

Had this too

[placid edge]

where in the A1111 UI are you finding the refiner ?

[candid drum]

Awesome! Thanks! I had been using your one from 2 weeks ago and a couple others and they seem to work okay. Just wasn't sure if the steps or actual work flow was going to be the same or not ♥️

[soft zealot]

someone mentioned my name in that prompt 🙂

[high skiff]

yeah, thats 1.0's VAE, I suggest going back to 0.9's VAE

[slender raft]

Does discord strip metadata from images these days?

[warm hazel]

prompt:geode spaceship, outer space nebula steps:50 width:1024 height:1024 guidance_scale:7.0 sampler:DPM++ 2M Karras seed:2727916641 data_model:SDXL

[echo stratus]

I haven't had any issues with this "rainbow" 🤷

[upper lantern]

you load it as a model and run your txt2img in img2img at .2-.3 denoise

[high skiff]

mine is changing a bit to beter utilize 1.0

[zealous horizon]

where did you get the LoRA and how did you load it? I'm puzzled.

[placid edge]

ah gotcha cheers mate

[high skiff]

its on the 1.0 release page

[urban fjord]

It is on huggingface along with the other models and I loaded it with Auto1111.

[high skiff]

^

in comfy, loaders, load LORA, patch in clip and model, load the LoRA, then patch the model and clip to where they would normally be

[thorn lion]

my prediction is that the dev team at StabilityAI have done a lot of work to get us to 95% and it's time for the community to do the next set of hard work, have thoughts the devs couldn't, try things that haven't been tried, and get us the rest of the way.

at least until we move the goalpost and need to aim for 2048 or 4096 base generations

[orchid ledge]

pich\ku comes out fine for me with the lora

[urban fjord]

Look at the background.

[thorn lion]

so, I was given 1 prompts, 2 text encoders, and no specifications. But, here ya go

[slim plover]

Does a 12 GB VRAM GPU need --medvram for SDXL? thinking about upgrading from a 8GB card

[trail bay]

@visual glade in that example workflow, I see it use CLIP text encode, but under add node, advanced, conditioning, I see a CLIPTextEncodeSDXL and also one for refiner. and I've seen some pple use that in their workflows. what is the difference?

[urban fjord]

But it is kind of pointless to show that the LoRA isn't doing anything weird with Pikachu if you do not show how it performs without the LoRA.

[pearl geode]

thank you!

[upper lantern]

i'm using --medvram on a 16gb A4000 to get live previews lol

[visual glade]

detailed control over the input parameters of the text encoders

[urban fjord]

The prompt is a mess and is too long.

[thorn lion]

not all better. just different. but some will consider subjectively better. and it'll flesh out from an averaged foundation to the variability of human experience, which different little communities gathering where they feel most comfortable

[thorn lion]

I copy pasted and didn't judge

[minor violet]

can i dreambooth SDXL ?

[placid edge]

i am using with half vae and xformers on A1111 with a 3090 and using about 10gb of vram with 1024x1024 80 steps

[clever verge]

Yeah, it's gone with 0.9 VAE.

[past raptor]

@visual glade is there a wiki or something that explains text_g vs text_l?

[hard fractal]

everyone realizing that their 0.9 LoRAs work even better on 1.0:

[ember cargo]

when can we expect some lora training "guides" with recommended settings for sdxl 1.0? please no videos

[high skiff]

yeah, seems to be an issue with 1.0 VAE. Look slike its cooked

[zealous horizon]

sorry, I'm a bit stupid, so many pages. I checked github and HF but couldn't find the LoRA or instructions on using them in SwarmUI
Do you have a direct link for me? ♥️

[high skiff]

0.9 VAE fixes all the issues

[lost blaze]

Getting nice impressionist paintings, running in ~1 minute with A1111 Macbook Pro m2 DPM++ 2M Karras 1000x1000 18 steps CFG 7

[visual glade]

SDXL has two text encoders, those are to give a different prompt to each

[urban fjord]

I don't know if my 0.9 LoRA work better on 1.0, but they still work well.

[unreal plover]

inpainting dont work very well

[high skiff]

for Swarm UI, I have no idea, I am sorry about that

[thorn lion]

you're not stupid. you're learning. don't put yourself down

[high skiff]

^

this

we are all always learning

[tulip kelp]

That's great. Not seen much painterly stuff so far

[zealous horizon]

coll, I just learned that.. +1 on my IQ

[arctic swan]

how to make text on image with model?

[soft zealot]

ok in my own workflow which is based on @high skiff 's

[zealous horizon]

A sign reading "this is written text"

The text in quotes will be written on the sign

[arctic swan]

thanks

[past raptor]

Is there a document or something I can read about the different encoders? Trying to understand haha

[urban fjord]

But I'll test my 1.0 LoRA against no LoRA with both 0.9 and 1.0.

[cerulean valley]

SDXL 1.0 on ComfyUI work like charm

[cyan crown]

hires fix seems not working

[gleaming owl]

Why cant I choose the VAE in auto1111?

ops, didnt mean that to be a reply

[dry plinth]

How do you add a Seed node? I dont see any option for that when I double click. How do I connect it to the KSampler advanced once I do?

[shy kelp]

what kind of steps and cfg are you using in the refiner if youre using it

[soft zealot]

A1111 or COmfyUI?

[placid edge]

on the huggingface list it only shows 1 SDXL vae with no versioning from what i can see

[cyan crown]

A1111

[gleaming owl]

how do I choose the VAE in auto1111? Every time I try to choose it, it reverts to my old VAE

[urban fjord]

How is it not working? Does it crash, produce bad results or what?
It works for me at least thought I didn't test for 1.0, just 0.9

[lost blaze]

No refiner, just base model

[urban fjord]

Remember to save your settings in Auto1111

[pearl geode]

i feel like i understand a bit what supporting terms means,.. but the longer i think about it, the more it doesn't make sense to me 🤔

[soft zealot]

drag back from the input on the ksampler, release and select Primitive

[limpid rose]

I noticed "restore faces" improves the quality of text a bit.

[soft zealot]

is the simple way

[cyan crown]

wich sampling method is the best for sdxl?

[thin nova]

anyone able to get the tauri branch of StableStudio working? I built the latest off of github but comfyui is complaining that StableStudioNode is not found

[primal turret]

Soo how do I install SDXL? Do I just drop the base into the normal models folder like all the others

[soft zealot]

the one that produces the image you're happy with

[trail bay]

if I have 24 GB Vram and say 100 GB system RAM, should I be using any particular flags when running ComfyUI? like the medvram or other stuff like that?

[urban fjord]

Yes, just remember to update Auto1111/Comfy to the newest version.

[pearl geode]

the reflection in the bike.. wow.

[primal turret]

Thankss

[shell oasis]

sheesh my setup is scuffed but wowee

[warm hazel]

Trying an old prompt out:

prompt:anthro gecko in a white tuxedo, vinyl toy figurine steps:50 width:1024 height:1024 guidance_scale:7.0 sampler:DPM++ 2M Karras seed:2401955551 data_model:SDXL

[quasi remnant]

if ur on local i'd recommend the 0.9 vae

[hot dagger]

why tho

[placid edge]

anyone able to see the .9 vae anywhere ? i can only see 1 version of vae on huggingface ?

[hot dagger]

ill give it a shot

[quasi remnant]

you have tons of noise leftover in lines

that's a bug with the 1.0 vae

[hot dagger]

oh

didn't pay attention to it

[thorn lion]

so here's my thought. devs correct me if i'm wrong and i'll edit the message.

what backend are you using? think of swarm like a jacket. it wraps around whatever method you were using before to generate images. so, if you were using A1111, then however you loaded loras there is how you'll load it. if you were using Comfy, then you load it with a node. I am not certain about the A1111 integration with Swarm and I committed to converting to Comfy upon SDXL release anyways so I won't be going back to learn how cause I wanna hazard saying that Comfy backend is OBJECTIVELY better than any of the others we currently have available for local use.

If you're using Comfy, then I'm gonna assume it's a node. somewhere. and anybody that knows Comfy could tell you how to get there. So far, it seems like Swarm is looking at the Comfy workflow and pulling whatever fields you may want to edit into something easier to digest than nodal spaghetti. That's the front page you'd be looking at when trying to generate. I assume.

The great thing about trying so hard and being wrong is that someone will come correct me if I am, and then I won't make the mistake next time.

[quasi remnant]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/tree/main/vae 0.9 vae is here

[echo stratus]

anyway, time to wait for the fine tuned models

[shy kelp]

so has stable released any loras yet?

[hard fractal]

oh nice! you can't tell at all that that's "Edge of Realism" and "RealisticVision" and "Illuminati"

[hot dagger]

theres one that came with the base

[deep dagger]

https://civitai.com/models/116945/doctor-diffusions-spooky-ghst-lora Will post the CosmicDisco one soon too

[shy kelp]

yeah but what the hell is it

[tulip kelp]

Is ComfyUI expressive enough to construct a DreamBooth/LORA training pipeline? or does that require more low level access?

[hot dagger]

no idea lmao

@hard fractal

[urban fjord]

Alright, tested my white background LoRA on 0.9 and 1.0 and the effect is a lot weaker on 1.0. (I was trying to get morpheus in the Matrix but forgot to prompt certain details)

[quasi remnant]

linked it shortly after u asked. the bottom safetensors is the one.

[placid edge]

fantastic ty

[wicked frigate]

yes lora is a node in comfy yes

it's on the TODO to get native clean lora integration

but for now, make a node to do it with comfy

for auto type it in the prompt lol

[bronze arch]

comfyui is nice but its so annoying for experimenting lol, i cant easily just turn off parts of the workflow

[chrome furnace]

Hey peepz, im a bit confused. I expected that the release meant i can just download a sdxl.ckpt file and throw into my SD folder. where do i get sdxl?

[trail bay]

missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'} should I be worried about this?

[wispy idol]

any update if this works on automatic1111 yet?

[shell oasis]

never use .ckpts always safetensor

[bronze arch]

is it possible to make on off buttons in comfyui?

[wispy idol]

im getting runtime errors

[shell oasis]

its on huggingface

[jagged ferry]

Comfyui doesn’t do img2img?

[urban fjord]

Yes SDXL works on Auto1111.

ComfyUI does img2img

[azure oxide]

ur right tho. its in the pins

[tiny walrus]

we released the offset lora that I created with our release today

[thorn lion]

I have heard it does, unless there are specific details to your issue that could help someone help you

[alpine pine]

if you goto the Hugging Face page, it should be there

[jagged ferry]

Really? How do I do that

[chrome furnace]

kk, thx to everyone for fast reply. ill keep digging 🙂

[shy kelp]

its a regular old noise offset?

[urban fjord]

Instead of an empty latent you load an image and use the VAE encoder node. (Or decoder, I always mix them up, you see which one has the right inputs)

[wispy idol]

i keep getting this error "self.weight += updown
RuntimeError: The size of tensor a (2048) must match the size of tensor b (768) at non-singleton dimension 1"

[jagged ferry]

Ohhh okay thanks

[chrome furnace]

lol, dude u got the sam profile pic xD

[wet nacelle]

I'm away from my computer right now but I'm wondering that if you view my issue you might be able to think of another way of fixing the issue.

If you have anything you can think of please let me know and I'll test it later.

https://github.com/Stability-AI/StableSwarmUI/issues/5

[urban fjord]

For Auto1111, disable all the extensions to make sure non of them is messing things up and remember to update Auto1111 by using git pull and make sure you're on the right branch.

[azure oxide]

i fixed the issue by downloading dot net

[alpine pine]

hell yeah!

[azure oxide]

had the same thing happen

[thorn lion]

check my github issue. looks like the same thing at a cursory glance

[azure oxide]

download it from here

[wispy idol]

other then just updating automatic1111 webUI, and dropping in the SDXL 1.0 model + the vae is it right?

[thorn lion]

i first had to uninstall ALL versions of dotnet

[limpid rose]

I love this. I tried doing wimmelbuilders a while ago to mixed results. looks like its going to get better and better!

[wet nacelle]

Can you link the site you got it from? Can you also send me a screenshot of which type?

I did the same as well I thought.

[thorn lion]

all on the readme: installation instructions

[wet nacelle]

Nope

[azure oxide]

linked it in the reply frommcmonkey right above

[potent summit]

How do you invoke a style in ComfyUI? Everything is coming out "painterly" and I want photographic. I tried putting Style: Photographic at the end of my prompt by no luck.

[thorn lion]

?

[thorny frost]

i am getting excellent results with invoke

[eternal fog]

Do we think we are going to get a fixed VAE for 1.0. It's overall better than the 0.9 one, but it puts horrible lines all over images.

[hot dagger]

use this preset

[urban fjord]

If you're not using the fp16 vae you need to turn on the no half vae parameter. You only need to download an extra vae if you're using one of the two 0.9 VAE.

[wet nacelle]

I'm pretty sure I'm the reason that parts there. I might be wrong but that wasn't there when I installed it.

[urban fjord]

And remember to switch resolution to 1024x1024 and remember to make sure you've not set auto1111 to use a 1.5 or 2.x VAE.

[nocturne dove]

Sometimes I add "photo" to the prompt. Often, I add terms like "painting, drawing, cartoon, render" to the negative prompt.

[echo stratus]

I wonder if this will do anything

probably not

[wispy idol]

can i just run it with auto or default vae?

[urban fjord]

If you want more photo stuff then add camera info at the end.

[nocturne dove]

That works well too sometimes.

[urban fjord]

I think auto/default should work, I don't remember the settings.

[slender raft]

Interesting, I've got the refiner taking flowers on a cup and saucer from the base image and replacing them with actual dead bits of plant.

[thorn lion]

which is fair. do check for updated information before asserting a conclusion. that repo has been updated 10 times in the last four hours and I doubt they're done. by double checking first, you will help eliminate misinformation and aid everyone to getting better results faster

[slender timber]

is there a description of what the offset lora does?

[nocturne dove]

The refiner can definitely mess things up from time to time.

[slender timber]

and should we add it to both the refiner and base?

or experiment?

[urban fjord]

I'm pretty sure the LoRa is for the base only.

[urban fjord]

But I won't stop you from trying to apply it to the refiner.

[wispy idol]

im just going to do a full reinstall of automatic1111 and see if i can get it to work

[molten gull]

how do i have to note the use of loras in prompt in comfyUI ? it's not "in the style of lora:myNiceLora"

[hot dagger]

this is why i use SDXL

[clever verge]

Try the sai-photographic. https://github.com/twri/sdxl_prompt_styler/blob/main/sdxl_styles.json

[thorny frost]

it seems sdxl likes old people

[hot dagger]

dont we all?

[potent summit]

Do I drop that in the comfy directory with the rest of the jsons of somewhere else?

[urban fjord]

To use LoRA in comfy you use a load LoRA node and connect it to the model node.

[polar epoch]

Excuse the "whateverthefuckthisis" lol, just benchmarking, but this is a 2048x2048 on 24GB video memory xD

[hot dagger]

click on load button on the right comfy menu

select the downloaded .json

[topaz bloom]

you need hires

[molten gull]

i dont want to just load it, but use the keyword again, too

[urban fjord]

SDXL is meant for 1024x1024 and aspects ratios with similar pixel counts if you create a lot smaller or bigger images you will get issues.
And now I'm tired of repeating the same answers. Please add a FAQ to this channel.

[steady grove]

very cool thanks for linking

[potent summit]

Awesome! Thanks!

[polar epoch]

I do. No idea how to add it as i have no idea how nodes works in comfyUI lol. Or rather, where to add highres fix

[hot dagger]

np

[fleet veldt]

hugging face Datacenter

[sullen adder]

or hates young

[hot dagger]

civitai after new models come out

[thorny frost]

even specifyjng young in the prompt

[sudden cliff]

The refiner loves old people, at least 0.9 refiner did

[vernal cloak]

On comfyUI?

[limpid rose]

My GPU looking at me right now like I kicked a puppy

[hot dagger]

lemon party when

[high skiff]

back

[sudden cliff]

Turn refiner up to 0.9 denoise

[hasty cipher]

omg what

[hot dagger]

lmao,

[unreal plover]

refiner work on automatic1111 ?

[shy kelp]

Does someone know if the Style they use here (anime, fantasy art, ...) are getable ?

[wicked frigate]

BTW if yall go in StableSwarmUI to presets -> import preset
you can just yeet this CSV that Joe posted (of @blissful shoal prompt presets) and it will import them all automatically for you

[hot dagger]

im still not able to select it, idk

[pearl geode]

no one can generate an image with a person pointing at a thing. Proof me wrooooooong.

[edgy wagon]

i also keep getting this error.

[wicked frigate]

lol i didn't even intend to i just coincidentally answered you right after lol

[azure oxide]

@high skiff did you have any findings on this?

[shy kelp]

eh ? where ? xD

[clever verge]

No problem, it's a simple yet effective node. SDXL responds incredibly to stylings that's quite easy to generate. And well, you don't need the node if you don't want to. Just use the templates and paste it.

[wicked frigate]

#✨｜sdxl message

[green python]

i hope they release dalle 3 for free, it's abt to drop!!!!

[high skiff]

I did not look into it, just blindly trusted and ste my end to 1000 now

[slender timber]

looks like its available with the manager tool as well. trying now

[azure oxide]

nice

[molten gull]

nobody knows ? i remember that there was a way to emphasize/strengthen a lora, besides just loading it in

[magic dawn]

Is there a fp32 checkpoint or just fp16?

[placid edge]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/tree/main/vae cannot get that bottom "diffusion_pytorch_model.safetensors
" to save as default vae in A1111 (newest version), keeps defaulting to the 1.0 SDXL vae, as soon as i click apply settings it says 0 settings changed

[quasi notch]

the fact it gets this close to anime by itself gives me hope for good anime finetunes

[sudden cliff]

So anyone know why that strange pixellation / lines is sometimes happening with 1.0 vae?

[shy kelp]

Oh my, perfect timing ! thx a LOT ! :DD

[echo stratus]

my guess? That's the "invisible watermark" that's mentioned on the github

[slender timber]

not a person. false flag

[bronze arch]

WDXL planning to be trained on 6m images multiple epochs, so im also very hopeful

[pearl geode]

is he really pointing at SOMETHING SPECIFIC?? 😄

[pearl geode]

make him point at something specific!! 😄

[quasi remnant]

oh it gets really close to anime (first is 1.0, second is 0.9)

also you're using the wrong vae

[bronze arch]

damn!

[clever verge]

Yea, think it was committed to manager almost the same day I released it.

[bronze arch]

the first one is really good, if i didnt know id assume its from a finetune

[quasi notch]

i am?

[quasi remnant]

ye, the 1.0 vae causes visual artifacting, i see it in the image u posted

[thorn lion]

this didn't work for me.....i need to take a break

[sudden cliff]

that's the first thing I thought of

[quasi remnant]

like all around here it's easy to see

you should use 0.9 vae

[quasi notch]

i dont have a specific VAE loaded, just the model

[unreal plover]

a person with red jacket pointing at a an other person with blue jacket

[quasi remnant]

ye, the built in one is broken

use 0.9 vae if you can

[spark bear]

hmm getting an odd color pattern/lines in most gens with 1.0 even in monochrome images, mostly in the shadows, looks like moire pattern you'd get in photography with lots of fine patterns messing with the image sensor.

[pearl geode]

oh nice. i think that counts!

[quasi notch]

i did notice that, thought it was invisible watermark. Thanks!

[sullen adder]

long finger

[wary wren]

i got it to work and damn. prompt is legit "cat" thats it

like wtf

[eternal fog]

This is annoying because the 0.9 Vae is worse with eye shape by quite a bit

[wary bay]

not sure if related, but I was getting oddly desaturated results in my 1.0 outputs running my 0.9 prompts

[wary bay]

which reminds me of mismatched VAE results...

[unreal plover]

Dont juge him !

[quasi remnant]

i don't see an issue with the eye shape; but yeaaaaah, really hope the 1.0 vae is updated

[bronze arch]

generating a 7680x2160 image, only using 12 gb vram 🤔 do i have some low vram setting on by mistake or something

[unreal plover]

he have a soul

[wary bay]

however I noticed my base results for 1.0 are less saturated than my 0.9 results so I don't think it's the vae causing my issue

[eternal fog]

Left is 0.9 vae right is 1.0, look around the iris of the right eye.

[violet shale]

so

[karmic dew]

Will ClipDrop support the Watercolor style?

[fleet veldt]

yeah i've got some VAE nonsense doing this rn..

[violet shale]

since my model folder is 112gb

[wary wren]

dont use a vae

[violet shale]

i have to ask

is it worth the storage?

[wary bay]

this is the issue I was having @fleet veldt

I'm not using auto either, comfyui

[wary wren]

yes

[fleet veldt]

i went and set VAE = NONE.. same sorta deal.. hrmm

[wary wren]

download treesize and see if you have random files taking up space

[indigo carbon]

I think I'll stay with my tuned 0.9 for now, it makes slightly better results than base 1.0, but I'm sure if the same process will be executed on 1.0 it will be way better.

[fleet veldt]

wiztree is much faster imho

[wary bay]

what size are those outputs

[wary wren]

Yeh i gave up trying to use automatic 1111 im sure theirs a bug. currently using atroomAi

[fleet veldt]

1024x1024

[fleet veldt]

i'll check it out. ty

[pearl geode]

also gave up on A1111. not working. just error codes.

[wary wren]

it looks so ugly tho

what is all this

[wary bay]

Yeah I won't touch auto until I see some clear indication that it's user friendly again lol

[topaz bloom]

have you updated it do 1.5?

[fleet veldt]

yes.. i'm doing file server stuff.. so... helps find out why someone has uploaded 30gb from their iphone and most are cruise or cats..

[pearl geode]

@topaz bloom i did git pull ...

[thorny frost]

A person pointing at the moon 🤣

[wary wren]

ah perfect

[pearl geode]

hahaha

[hasty cipher]

i juts asked for a stock market crash

[orchid ledge]

looks like a bear market to me

[pearl geode]

i saw your name and created it 😄

[wary wren]

https://cdn.discordapp.com/emojis/1024543017949986856.webp?size=48&name=PepeRealizes&quality=lossless

[topaz bloom]

is it 1.5?

[runic nest]

Recording of todays event?

[wary wren]

depression

[hasty cipher]

it's joever (for low vram lads)

[pearl geode]

uh oh:

[wary wren]

honestly looks cool. its a little wonky but that could be seen as an art style

[trim orbit]

a windirtree chart for ants

[shy kelp]

that remind me, is animation totally free on local setup ?

[trim orbit]

the squares represent files and their relative sizes

wiztree is what i use too actually. its so fast

[fleet veldt]

yeah.. i'll bbiab to work on this.. my automattic on this lab pc is hosed.. i'll pickup a grocery order and try on the office PC who's got a cleaner automatic1111 install and a 3090.. see if i can make it work there.. ALSO.. my computer crashed MID download of this thing and i resumed it when it rebooted.. could i actually just have corrupt model files????? hrm.. here's another "CAT"

[thorn lion]

i like wiztree for that task, if you want an alternative

[polar epoch]

Huh, why is upscaler's k sampler greyed when i try to add prompts to it?

[shy kelp]

I KNEW IT WAS REAL !!

[thorn lion]

oh okay cool someone beat me to it

[sullen adder]

512 wont work, try 1024

[upper lantern]

don't gen on xl at 512^2, try 1024^2

[eternal fog]

You are trying to directly add a string to the sampler, it needs to go through a CLIP node

[unreal plover]

@shy kelp Unclassified files just dropped

[candid spindle]

scary Kerr black holes

[thorn lion]

you could check the hash to confirm

[fleet veldt]

hearoh right right.. yeah. i did a ui restart. here's 1024x2

[fleet veldt]

alright. i'll check that

[pearl geode]

@thorn lion

[clever verge]

You have to pass the prompt to clip text encode

[thorn lion]

idk, my fine fellow. i'm just letting you know how to know for yourself

[pearl geode]

@thorn lion thanks!

[polar epoch]

Thanks :)

[wary wren]

you should up the steps to a minimum 25 i recommend 35

[polar epoch]

Testing upscale to 2048 now

[quasi remnant]

nice steins;gate reference

[polar epoch]

Well.. i might be able to shoot it a bit higher than 2048

[quasi remnant]

but aaaaaaa my eyes, i can tell at a glance when someone is using the 1.0 vae lol

i wonder what happened and why it broke

[clever verge]

4096x4096 should be no problem with that much vram.

[karmic vault]

needs more steps, low steps seem to cause that

[polar epoch]

by the way, does confy UI have a tool to load images to get generation settings from, similar to auto's? As i wanna download a few of you guys's images and make a personal library of prompts to puzzle together to get a realistic image :P

[slender timber]

where can you load the 0.9 vae?

[quasi remnant]

drop the image directly into comfyui if it was made with comfyui

[wary bay]

I can't get the 0.9 vae thing to work at all lol..

[quasi remnant]

for instance you can take this image and drop it into comfy

and you get my workflow

[clever verge]

Yeah, just drop the image on the workspace

It's explained here: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[high skiff]

oh wow haha

[wispy idol]

having some issues with automatic1111 still - full fresh install, only model in it is SDXL with the --no-half-vae agrument, but it refuses to generate, running a 4090

[polar epoch]

Neat! And i think my upscaler needs tuning lol

[zealous horizon]

which app are you using to generate?

[high skiff]

seems intentional haha

[slender timber]

comfy, I meant more of "where can I download it"

[wary bay]

@quasi remnant are you using two checkpoint nodes like this, and just taking the VAE from the 0.9?

[bronze arch]

took 855 seconds for this

[polar epoch]

I have no idea how this shit works lol. Normally i just choose highres fix, and adjust denoise in automatic's

[spring fulcrum]

@visual glade @high skiff Do either of you by any chance have a workflow that does batch upscale... like say put a directory path and then it passes it to the upscaler and then you select an output path and then that folder is where your batch upscales go?

[zealous horizon]

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

[high skiff]

it will work like that at some point

[quasi remnant]

yep

i put a separate vae loader next to my vae encoder just for that

[topaz bloom]

prompt pls?

[wary bay]

do I need to separately download the VAE for that

[quasi remnant]

Do not use this one

[slender timber]

oh, I assumed that was the 1.0 vae

[clever verge]

Not sure if this is the most efficent upscale but works for me.

[high skiff]

I do not at the moment

[quasi remnant]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/tree/main/vae Use the one from this page

The one he linked is 1.0

[wary bay]

using the checkpoint node does not work so whatever you're doing is the better way lol

I'll try that

[high skiff]

yeah, I reccomend using 0.9's VAe, it works a lot better than 1.0's

[grizzled viper]

For A1111 there are scripts for batch txt2img and batch img2img… the latter is a bit harder to find…

[zealous horizon]

my bad 😉

[sudden cliff]

that's what I'm doing

[chrome furnace]

SD only?

[high skiff]

these look insane

absolutely love

❤️

[polar epoch]

4k atm, yeah, lets crank it to 12k at this rate

[orchid ledge]

especially the space lazer

[unreal plover]

@high skiff The magic of this model is you don't need huge prompt, a simple one without any negative prompt can generate this !

[high skiff]

yeah, thats whatmy guide suggests as well

[fiery plover]

ignited the atmosphere

[unreal plover]

a meteorite hits the earth, view from space, extremely detailed,natural lighting,film grain

[clever verge]

Sometimes even better without negatives.

[polar epoch]

With higher res, is it just denoise i have to drop to get the same image, but way higher res? Or CFG?

[fervent orchid]

Is there any recording of the SDXL super stage?

[ocean phoenix]

11 seconds per iteration on an RTX 2070. Color me not impressed.

[wary wren]

what does the xl stands fro in sdxl?

[polar epoch]

Xtra Large pretty much

[wary wren]

no shot

[polar epoch]

¯_(ツ)_/¯

[quasi remnant]

it's getting uploaded somewhere

[autumn forum]

its resolution is 1024x1024 thats what it was trained on so its pretty large

[fervent orchid]

thanks

[trim orbit]

🎻

[sharp patrol]

In ComfyUI, How do I run a node from middle? I have generated an image and now I want to upscale the same using a different upscaler. any help?

[polar epoch]

What's super stage?

[slender coral]

Could I get the promt? Love the env.

[fiery plover]

I am dumb, haven't used huggingface in a minute. Which file am I downloading here?

model.bin VS model.fp16.bin VS model.fp16.safetensors VS model.safetensors

[eternal fog]

Ice cream van anyone?

[trim orbit]

the event where they all talked

[quasi remnant]

#✨｜sdxl message

[quasi remnant]

the bottom one or more specifically https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/resolve/main/vae/diffusion_pytorch_model.safetensors (this is a direct dl link)

[polar epoch]

Ah right, a recording of the event.

[eternal fog]

It can sort of do car liveries now

[leaden fable]

Hello, to make sdxl work on comfy, do i need sd_xl_base_1.0.safetensors or sd_xl_refiner_1.0.safetensors ? and about the vae ?

[quasi remnant]

both + 0.9's vae

this is 0.9's vae https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/resolve/main/vae/diffusion_pytorch_model.safetensors

[urban fjord]

The vae is built into the 1.0 checkpoint so you do not need to download it.

[quasi remnant]

1.0 vae is broken

[leaden fable]

thanks. all in the same folder ?

[hard fractal]

vae is baked into 1.0

and refiner

[lean garden]

sdxl 1.0 deforum

[quasi remnant]

1.0's vae leaves so much leftover like, line-noise style stuff though. ._. 0.9's vae works fine

like in every generation 1.0's vae feels broken

[eternal fog]

Joe do you know about the strange artifact issues with the 1.0 Vae, as compared to the 0.9 one.

[polar epoch]

Color the gpu older gen. But hey, if comfyUI supports tensorRT, convert the models to that, then you can use the tensor cores to speed up generation

[clever verge]

the 0.9 VAE is also here: https://huggingface.co/stabilityai/sdxl-vae/tree/f85c10757eb0bda0222090261de54c5fe17a7f4f

[quasi remnant]

vae is usually in a separate folder

[leaden fable]

so… just sd_xl_refiner_1.0.safetensors ?

[trim orbit]

i saw 2 examples of that and quick conclusions. i'm not sure if that is supported

[vital wolf]

Can you generate nsfw from the released models?

[finite moth]

Mind blown

[trim orbit]

i've not caught any lines on my generations

[sharp patrol]

When I click QUEUE PROMPT, it generates the art from the start. How do I run it post the refiner process?

[hard fractal]

that's the 1.0 vae, yeah.

[vague monolith]

Is there a way to use consistent styling in SDXL (like you normally would with lora)?

[eternal fog]

Zoom in on areas like eyes and compare them with a 0.9 gen

[quasi remnant]

b-but... 1.0 vae... broken... ;_;

[clever verge]

Yes

[trim orbit]

i can see the issue in your examples easily and dont need to compare

it would stand out if i saw it on my own

[hard fractal]

I think it's beautiful, and trying its best ❤️

[vague monolith]

please tell me how 🙏 I guess it wouldn't be jsut prompt magic

[hard fractal]

It's trying to fill the shoes of its dad, but that's hard!

[quasi remnant]

is it intentional to be that way then? for me 0.9's vae completely gets rid of all of the line noise that is caused by 1.0's vae. tl;dr, 1.0 base+refiner mixed with 0.9 vae produces best results is what i'm seeing people say.

[clever verge]

#✨｜sdxl message

[trim orbit]

i'm using a 4080, ada architecture, if that lends towards another diagnostic to do with precision errors

[sullen linden]

So I keep getting this 'failed to create model quickly' issue with sdxl where it's like failing to make a yaml

[south vector]

Guys does sdxl 1.0 is live on stable discord

[sharp patrol]

I want to upscale this same image using a different upscaler now

[clever verge]

Check that out, only some prompting.

[pallid path]

good stuff, maybe the refiner is enhancing the face a little too much but it still looks good

[smoky bough]

https://tenor.com/view/reference-captain-america-avengers-understood-gif-7991943

[trim orbit]

you can look through all the generations here and not see the artifacts that you two are claiming is for certainly due to the 1.0 vae

[lean knot]

I'll have to check the 0.9 vae, because there are definitely some artifacts with my gens

[hard fractal]

You can change where the switch happens. 32 out of 40 was by far the highest voted.

[trim orbit]

most people are not loading in the 0.9 vae

[wary bay]

I hate trying to figure out file stuff because I must've downloaded the wrong vae from your link @quasi remnant

[vague monolith]

I wonder if the style would be consistent in all if say generated 100 images, or, again, it's porbailistic 🤔

[clever verge]

You mean another step? or change the model?

[wary bay]

its not even recognizing the safetensors

[quasi remnant]

i see them, but i have a trained eye for noise-related stuff. in every single 1.0 gen using the 1.0 vae i've seen it

[quasi remnant]

i heard it doesn't work in a1111

[wary bay]

using comfy 🥲

[trim orbit]

artifacts will show up, but are you seeing what they're talking about specifically. thin pixel wide lines like interlacing patterns?

[wary bay]

just dont know what im doing lol

[quasi remnant]

in comfy it should work fineee.... https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/resolve/main/vae/diffusion_pytorch_model.safetensors is the direct DL link for the right one though

you put it in vae folder

models\vae

[trim orbit]

i have eyes too. i lived during the crt era

[wary bay]

I diiiiiid

reee

[lean garden]

Yes it does, and with deforum

[wanton lark]

Trying to generate unusual fruits - sdxl 1.0 is awesome

[lyric grove]

I see it too, though we can wait and use 0.9 vae until there's a fix

[quasi remnant]

the 0.9 vae i mean, i saw someone couldn't get it recognized in a1111

[lean garden]

Oh i see sorry 🙂

[wary bay]

0.9 base vae with the refiner model?

[quasi remnant]

my vae folder just looks like this

[quasi remnant]

No do not use 0.9 checkpoint

use the 0.9 vae, standalone

in the vae folder.

With 1.0 base and 1.0 refiner

[wary bay]

right but did you mean to link the 0.9 base vae instead of 0.9 refiner

[trim orbit]

si that's them. i bet there's a different better fix

[candid spindle]

Ok this one is kinda dark, I reread the rules to be sure lol

[wary bay]

I'm just making sure I do this right lol

[quasi remnant]

Yes. this is the vae that works afaik

[wary bay]

gotcha... so none of the refiner VAEs are worth using lol

[quasi remnant]

i think it was linked by outerside

[trim orbit]

--no-half-vae

[delicate grotto]

when you make an image with 1024, the latent image is divided by 8

so 1024x1024 = 128x128(latent) --> sampler(base) -->img2img with refiner as the checkpoint(model)

[trim orbit]

i use that on my auto and don't see none of those

[clever verge]

Try and see!

[quasi remnant]

idk about if you should use base or refiner, i do know that this one works thooo

[polar epoch]

Ok, confirmed. 4k is the highest i can go

[lost blaze]

Getting better results than Midjourney on photorealism. (left SDXL 1.0 base only 1312 x 1312, 18 samples DPM++ 2M Karras, 18 samples, CFG 6, right Midjourney v5.2 upscaled)

[polar epoch]

Or no, it chuggs on ram, lower than 4k

[trim orbit]

oh nice 2mp renders!

[vague monolith]

thx for the link

[edgy otter]

SDXL is better with faces, I rarely have to use "restore face" now

[quasi remnant]

It's the same vae, i compared hashes

[indigo carbon]

that's bs, SDXL is way better than MJ

[sharp patrol]

Yup, I want to change the model. I have loaded ESRGAN_4x but want to upscale this same image. How do I run the Queue from this point only. I don't want to generate from the base.

[quasi remnant]

so dw about which one u use

[lost blaze]

I mean better results with SDXL than Midjourney

[sudden cliff]

"This guy gave me this really tiny cone. But he says he'll give me what I deserve. Says I just need to come back tonight when he closes up."

[trim orbit]

(getting better than midjourney)

[quasi remnant]

u can use the one from 0.9 base or 0.9 refine

[polar epoch]

Ok, too low denoise, i made high quality squares lol

[delicate grotto]

that's subjective
mj has a certain trick, i think that they use more than 1 model ot generate images

btw, why are all of the SDXL have the square artifacts?

[pallid path]

I love these

[indigo carbon]

it already is.. if it's not for you then you didn't set it up correctly

[pallid path]

look at the anatomy too though

[torpid spindle]

Is an Nvidia RTX 3070 with 8GB of VRAM enough for SDXL, im trying on ComfyUI but it takes forever to generate images

[delicate grotto]

you can see the square artifacts here too

is this normal?

[trim orbit]

you're really coming in so sideways with a huge misunderstanding here

[clever verge]

Then you have to go tiled scaling, but I've got mixed results with that. I've been happy with 4k * 4k.

[indigo carbon]

good luck doing text with it

[pallid path]

the refiner is trying to give it an eye

I'll try the BASE only

its winking lmao

[median prism]

I'm getting this when trying to run SDXL on Auto1111
RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same

[delicate grotto]

i think that mj is just a 1.4 or 1.5 checkpoint merge, alongside some more checkpoints that specialize in certain areas

[polar epoch]

Oh no, i'm experimenting with denoise and CFG to find the right balance for high res stuff

[median prism]

I'm not quite sure what to do

[indigo carbon]

do you really think MJ gets to this level of quality?

[pallid path]

yeah?

[cunning geyser]

Has anyone tried anime with sdxl

[quasi remnant]

aaaaaaaaaa. the 1.0 lines xD

[torpid spindle]

Is an Nvidia RTX 3070 with 8GB of VRAM enough for SDXL, im trying on ComfyUI but it takes forever to generate images

[delicate grotto]

look at the artifact that 1.0 got, we didn't have that in 0.9
what is with the lines?

[slender timber]

I mean, it can

[quasi remnant]

mj def beats those lines.

[high skiff]

can be fixed with 0.9 VAE

[quasi remnant]

yea, i know.

i see it IMMEDIATELY when people post with 1.0 vae tho

[polar epoch]

Also love that comfyUI doesn't regenerate all from scratch per gen, it just redoes the node that has changed

[quasi remnant]

i'm using 0.9 vae myself

[high skiff]

not sure what they did to the 1.0 VAE to make it do that

[fiery plover]

does that mean 40 steps and refiner kicks in at 32?

[high skiff]

same haha

[indigo carbon]

are you seeing something I don't?

lines?

[quasi remnant]

can you not see it? you're using the 1.0 vae lol

[pearl geode]

anyone has a working img2img flow for comfy?

[quasi remnant]

the 1.0 vae is brokennnn

use 0.9 vae

[delicate grotto]

just a sec, ill show you
its weird but there are lines for a reason

[sour obsidian]

looks correct to me, just make sure total on both is 40

[high skiff]

these crusty artifacts

[quasi remnant]

i literally posted a screenshot of it

yes

that is from 1.0 vae

hurts my eyesss

lol

[delicate grotto]

lots of these artifacts

[quasi remnant]

meanwhile 1.0 model with 0.9 vae? none of that

you won't find any of those artifacts here

[indigo carbon]

oh damn, sorry, I'm blind, didn't sleep in a few days

[median prism]

I don't know how to use SDXL with Auto1111

[polar epoch]

Does it matter what sampler i use for upscaler/refiner? Or does that depend on what result i want?

[clever verge]

Can't believe VAE 1.0 got through 😄

[delicate grotto]

its not something that exists on every pic though

[quasi remnant]

...it does.

[pallid path]

perfectly safe

[quasi remnant]

i can see it on EVERY 1.0 vae pic

i've been calling it out when i look at a pic closely

it's LESS noticeable in some pics

but it's there

[trim orbit]

i hate seeing this attitude. especially when it really seems like this line diagnosis is catching a brigade event suddenly

[lean garden]

Is it not the sampler?

[high skiff]

I brought the issue up to 2 dukes a while ago, and it went away in the bots, not sure why its back

[indigo carbon]

how do I downgrade to the last VAE?? I can't unsee it now

[delicate grotto]

here for example i don't see it

[high skiff]

but the solution is just use 0.9 VAE

simple enough

[quasi remnant]

No, i tried all sorts of diff workflows

and diff samplers/etc

[quasi remnant]

i gotchu fam, https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/blob/main/vae/diffusion_pytorch_model.safetensors

[high skiff]

its the VAE, I have done direct comparisons

[delicate grotto]

mmm not sure in comfy, but you can load the vae, 2 sec ill link you to one

[quasi remnant]

my workflow has an example of how to do that. copy this pic to see it

[indigo carbon]

pytorch model.safetensors isn't working as VAE

[quasi remnant]

Wrong vae.

[quasi remnant]

it is though...

it's the exact one i'm using

[eternal fog]

vroom vroom

[pearl geode]

can i use XL BASE for upscaling???

[high skiff]

1.0 VAE vs 0.9 VAE

[indigo carbon]

really? hold up, I'm checking

[delicate grotto]

thanks for the mention, mmm i should find 0.9

[untold crescent]

this probably is 0.9 vae? https://huggingface.co/stabilityai/sdxl-vae/resolve/e6e6fee4ec9c36d1f8e534a0874f7a773fb44bf2/sdxl_vae.safetensors

it's from 1 mo ago

[quasi remnant]

i keep linking it lol

[candid spindle]

astronaut is about to experience that Kerr black hole

[candid spindle]

0.9 is the right one?

[wary wren]

was that the prompt?

[rich coral]

I tried Ultimate SD Upscaler and it did not work well.

[candid spindle]

(a terrified expression:1.2) on (the face:1.1) of (a lone astronaut:1.3) in (a white spacesuit:0.9) as he (floats:0.8) helplessly in (the dark vacuum:0.7) of space, (his tether:0.6) severed and (his radio:0.5) silent, while (a massive black hole:1.4) looms behind him, (distorting:1.1) the (stars:0.9) and (light:1) around it with its (immense gravity:1.2), and (stretching:1.2) his (body:1.1) into (a thin strand:1.3) of matter as he approaches the (event horizon:1.4).

[high skiff]

1.0 vs 0.9

[delicate grotto]

yeah, i wasn't here for the past few hours too much

[quasi remnant]

this is the 0.9 vae https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/blob/main/vae/diffusion_pytorch_model.safetensors

[high skiff]

something with the 1.0 VAE adds some weird artifacts

[pallid path]

also does the VAE improve on anything (if you specifically make Comfyui use the file)

[high skiff]

yeah

[native pewter]

and whitout vae?

[untold crescent]

"Access to model stabilityai/stable-diffusion-xl-base-0.9 is restricted and you are not in the authorized list. "

[quasi remnant]

apply for access

[wary wren]

are you sure its the same region? it feels like youre showing more of the right image

[quasi remnant]

it's automatic

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

[untold crescent]

it's for research

and we now have official

[quasi remnant]

apply for it, you'll be auto approved.

[high skiff]

1.0 vs 0.9 VAE

[thin nova]

just use this one https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors

[slender timber]

its vaery easy to apply

[untold crescent]

yeah but..

[pallid path]

mfw SDXL 1.0 VAE

[high skiff]

I am just crudely cropping

[untold crescent]

this is official

https://huggingface.co/stabilityai/sdxl-vae

and it has older VAE

[quasi remnant]

and that's 1.0

the wrong one.

[untold crescent]

why not use it

no

the one I linked was one month old

[pallid path]

damn this artifacting lmfao

[quasi remnant]

those are all the 1.0 vae

[spark bear]

it looks like moire, some kind of pattern/noise mismatch

[quasi remnant]

the one you just linked is 1.0 lol

[pallid path]

looks like some chroma subsampling madness

[quasi remnant]

this is 1.0 vae

[quasi remnant]

do NOT use that one

[thin nova]

use the fp16 fix version because 0.9 vae will take more vram

[untold crescent]

this one

[trim orbit]

i'm using 1.0 vae and can't get these artifaccts if i try

[delicate grotto]

you can load a checkpoint and just link the vae from it

[high skiff]

here, whole face 1.0 VAE vs Whole face 0.9 VAE

[pallid path]

good for you

[thorn lion]

you're not dumb. you're learning.

[untold crescent]

they had 1.0 VAE up month ago?

[delicate grotto]

shouldn't be a problem with decoding using 0.9 after 1.0 latent

[indigo carbon]

A1111

[trim orbit]

its not about me. its about diagnosing the problem instead of leaning into the first conjecture

[quasi remnant]

u still shouldn't be afraid about applying for research access, they set it to auto-accept anyone who fills out the form now.

[shy kelp]

my comfy workflow. might put on civit later.

[quasi remnant]

there's no problem with using the 0.9 one from the 0.9 page

[pallid path]

the only problem is that technically the 0.9 is for research only, therefore its non-commercial 🤓

[high skiff]

[delicate grotto]

ah, right you use a1111

mimizukari — Today
this is the 0.9 vae https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/blob/main/vae/diffusion_pytorch_model.safetensors

[pallid path]

nah idk if it applies to the VAE xdxdxd

hopefully not

[trim orbit]

i would put a big fat IANAL disclaimer on this brigade of recommending 0.9

[untold crescent]

can someone with 0.9 access just show the SHA256 on huggingface?

is it different from:

[quasi remnant]

1598f3d24932bcfe6634e8b618ea1e30ab1d57f5aad13a6d2de446d2199f2341

Yes

it is

[high skiff]

more 0.9 vs 1.0 VAE examples

[untold crescent]

ok thanks

[quasi remnant]

i posted it earlier too comparing base/refiner vae

being same vae

[clever verge]

Let's see what the fix is, there will be one eventually. Either we doing something wrong or there's a fixed VAE.

[wary wren]

this?

[pallid path]

how do I know the seed of an image in ComfyUI

[high skiff]

base and refiner VAE are the same, but only on their version

[delicate grotto]

how do you change a vae inside a model?

[quasi remnant]

syt ur using one outerside linked too right?

[delicate grotto]

would be easier like that to just make a 1.0 with 0.9 vae

[pallid path]

you don't, you just download a VAE and make it load that

[wary wren]

you gotta take it out for dinner first

[pallid path]

like this

[high skiff]

what? I have it from when I got the research weights

[delicate grotto]

goddaamit ahahahahaha

[pallid path]

I make it use the 0.9 fp16 VAE

[deep dagger]

Here is the CosmicDisoc one too! https://civitai.com/models/116970/doctor-diffusions-cosmicdisco-lora

[quasi remnant]

o, still same one then i think(?)

[upbeat summit]

awesome 😄 thx!

[pallid path]

also does someone know this

[quasi remnant]

was talking about this

these r the ones i'm using and it's fine

[rich coral]

sdxl is very nice but I sure miss control net

[high skiff]

oh, you can just find it online, they officially released it

I can send it here

[untold crescent]

huh

[high skiff]

just a sec

[quasi remnant]

i mean

[high skiff]

SDXL 0.9 VAE, there you go

[quasi remnant]

the one i'm using works fine, it's the 0.9 vae lol

[untold crescent]

even this has the same has on that one I linked https://civitai.com/models/112390/sdxlvae

[hasty cipher]

holyshi i legit cannot tell it's ai