🐝｜swarm-ui

[dim sonnet]

go to server -> server config -> and set LogLevel to Debug, probably the comfy raw debug output will explain what actually happened

[hearty notch]

Hi @dim sonnet been quite a while 😄

[dim sonnet]

did you install comfy manually or something? That looks like it was installed wrong and has an outdated package

[dim sonnet]

hi

[dim sonnet]

ImportError: tokenizers>=0.14,<0.19 is required for a normal functioning of this module, but found tokenizers==0.13.3.
this package, tokenizers, is outdated

[hearty notch]

after trying a custom nod, I cleared the folder and removed it but all went down

i did install a fresh copy of ComfyUI

and now I am stuck in this erro I shared

shall i clear the full folder and try again?

[dim sonnet]

https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_or_cpu_nightly_pytorch.7z <-- should install by downloading the nightly prepacked

[hearty notch]

okie let me try that

i did notice that SSUI is not very friendly with custom nods, is it just me?

getting this now with the nightly pack

[dim sonnet]

custom nodes should work fine - ones that have custom JS sometimes have a specific bug that's easily fixed but isn't uncommon

h... huh

is it... missing a file?

or does that source from an nvidia driver issue? idk i've never seen these error messages before

[hearty notch]

hold on, I will delete the backend folder and start from scratch one more time,

[dim sonnet]

@cunning holly do you have any idea what those errors ^would be about?

OSError: [WinError 127] The specified procedure could not be found. Error loading "D:\StableSwarmUI\dlbackend\comfy\python_embeded\Lib\site-packages\torch\lib\nvfuser_codegen.dll" or one of its dependencies.

[cunning holly]

might be a driver issue or an issue with old windows

[hearty notch]

seems to be fine now,

but i have to delete the comfy folder to get it working agin

looks like the workflow that i loaded with missing nodes was causing that

is there anyway to flush or reset the wf manually?

[hearty notch]

i do not think so as it's back online after the steps i did

[cunning holly]

my script that makes these packages runs them once before publishing them so they should not have any errors on launch

[hearty notch]

can you elaborate more?

[cunning holly]

https://github.com/comfyanonymous/ComfyUI/blob/master/.github/workflows/windows_release_nightly_pytorch.yml#L61

[hearty notch]

Ummm, this is quite odd, as this is not the first time I am facing this issue, but usually when I delete the custom nod I can load the ui, clear it, and things will be back to normal, looks like that workflow has some nasty error in it

[nimble belfry]

@dim sonnet I'm trying to make my model folder on a network drive. Now model loading takes forever. I stream 4k video from that NAS so I don't think it should have an issue with speed.

I'm trying to come up with a solution for the 2 other instanceses of Swarm. What is best solution for them to all share the same folder?

[dim sonnet]

4k video stream is 30-50 Mbps, maybe higher if you got high quality local movies. Loading an SDXL model is basically downloading the entire movie file all at once.
A standard home network is 1 Gbps internally, aka ~120 megabytes per second. SDXL is 6.5 GiB - 6500/120 = 54 seconds to load a model at gigabit speed.

now if you have super high speed tengig LAN, you should be good. (10x faster == ~5 seconds to load the model)

but that's pretty rare

if you have tengig, or loading times are over a full minute, there might be a secondary issue. But otherwise... yeah network speeds be a lot slower than device-internal speeds lol

[hearty notch]

@dim sonnet when I trying to add an image preview to the lora file

[dim sonnet]

oop, that's not meant to be required, one sec

fixed

[hearty notch]

is this normal?

[barren moat]

Ok thanks I will check it when I get back to my pc

Was the remove bg something I installed with that custom node?

This is what it said in the console after I di the debug, is there a log file somewhere?

[hearty notch]

@dim sonnet these warning are getting more and more, and custom nods are not loading anymore

[dim sonnet]

yeah i needa fix that. You can safely ignore it for now.

[dim sonnet]

no that's a new swarm feature

[dim sonnet]

get rid of \ComfyUI-Stable-Video-Diffusion\ custom node, that's redundant, SVD support is built in now

[dim sonnet]

that is the log there

22:33:23.509 [Debug] ComfyUI stderr: OSError: Unable to load weights from pytorch checkpoint file for 'C:\Users\jason/.cache\huggingface\hub\models--CIDAS--clipseg-rd64-refined\snapshots\583b388deb98a04feb3e1f816dcdb8f3062ee205\pytorch_model.bin' at 'C:\Users\jason/.cache\huggingface\hub\models--CIDAS--clipseg-rd64-refined\snapshots\583b388deb98a04feb3e1f816dcdb8f3062ee205\pytorch_model.bin'. If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. 22:33:23.509 [Debug] ComfyUI stderr:
it looks like a corrupt file possibly? Maybe it got interrupted while doing setup?

pop open that .cache folder and delete it and restart and try again, might work after a reset

[hearty notch]

can you share a wf for SVD in comfy?

[nimble belfry]

It's pinned in the SVD channel

#▶｜stable-video-diffusion message

[dim sonnet]

@barren moat @hearty notch the opencv/rembg errors should be fixed in new update

[barren moat]

Awesome thanks! I’ll give it a try now

[barren moat]

I'm still getting there error's. Where is that .cache folder located?

There's a new error at the top line of the console error: The source specified has already been added to the list of available package sources. Provide a unique source

[dim sonnet]

Oh. Uh. The prior method of autoinstalling prevents the new method from working >.> oops

uh

[dim sonnet]

ignore that, that's normal, the autolaunch script does some redundant stuff to be safe

[barren moat]

ok

[dim sonnet]

go find: swarm\dlbackend\comfy\python_embeded\Lib\site-packages and delete both folders named rembg

and that will trigger a reinstall

[barren moat]

ok

@dim sonnet Good news and bad news 😅 the rembg is working perfectly but the segment deal is still not working

[dim sonnet]

why not working?

turn on debug

[barren moat]

ok one sec

@dim sonnet

[dim sonnet]

OSError: Unable to load weights from pytorch checkpoint file for 'C:\Users\jason/.cache\huggingface\hub\models--CIDAS--clipseg-rd64-refined\snapshots\583b388deb98a04feb3e1f816dcdb8f3062ee205\pytorch_model.bin' at 'C:\Users\jason/.cache\huggingface\hub\models--CIDAS--clipseg-rd64-refined\snapshots\583b388deb98a04feb3e1f816dcdb8f3062ee205\pytorch_model.bin'. If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True.
yeah that's still the cache folder thingy

[hearty notch]

it's working for me,

however, I did notice when I try installing some custom nods i am getting this error dialog box asking for -xformers

however I think its a comfui thing @cunning holly maybe intrested in that

[barren moat]

Should I delete the .cache folder?

[hearty notch]

but in full honesty I think the segment option is not working very great as it will wash away the style form the generated image

[dim sonnet]

yes

[dim sonnet]

huh?

[hearty notch]

lert me show you

[barren moat]

@dim sonnet https://tenor.com/view/gil-cat-dancey-gif-18141936135897027445

[hearty notch]

here you can see the change

[barren moat]

It's WORKING!!! Thank you!

[hearty notch]

i used as a part of the prompt segment:face in this gen

[dim sonnet]

oh yeah if you have a highly stylized image you'll have to be careful to make any use of segment inpaint

more useful for photorealistic

[hearty notch]

i did notice that the hard way, as it will ignore all styles, and go for realistic style

hope to see it soon as more style friendly

that will be amazing

[dim sonnet]

that'd be a matter of prompting/model choice

it's just automatic segmented inpainting

[barren moat]

@dim sonnet is there open pose is swarm? For example I get one good image and want to regenerate that same image in a few different poses. Is that possible?

For instance, I need the same character in three different poses to make my sample image of the football player, if that makes sense?

[dim sonnet]

ControlNet in general: yes

a pose editor: not yet

can use any 3d pose editor and screenshot and put that as the input image for the controlnet

and use an openpose preprocessor

[barren moat]

Ok thanks

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7668336

SVD in swarm :D

[barren moat]

Which model is it?

[dim sonnet]

top one

or the XT https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/blob/main/svd_xt.safetensors

[barren moat]

What’s the difference between the regular and the XT?

[dim sonnet]

"regular" is a 14-frame model, "XT" is a 25-frame model

it also was trained longer and is a bit smarter as a result

[hearty notch]

And what’s the difference between the normal model and the image decoder? I can’t find proper info to understand each model

[hearty notch]

aaaaaaaaaaaaaaaaaaaaand anew from of addiction is here

[nimble belfry]

same.

[gray geyser]

ok so i have installed this. how do i switch to video?

oh i see the video switch 😂

does it save gif automatically?

[nimble belfry]

You either make an image or load one in init Inage and toggle video on choose what video model you want

[nimble belfry]

Right now no. I manually throw it in photopea.

[gray geyser]

d'oh. not familiar with photopea but ill look into that next i guess

[nimble belfry]

It will have more options later when the dev has time to add more.

Photopea it just a free web based Photoshop clone

[gray geyser]

so i copy a video model from my comfy and drop it in models > stable-diffusion > officialstablediffusion right

[nimble belfry]

Yes. If you want to make a new image make sure to add any other models to the folder to make images from.

Or just use init Image and load in an existing image. For it to make the video from.

[gray geyser]

just not my day for AI. tells me i dont have a model selected even though i do

[nimble belfry]

You have to select an image model as well

So the tab at the bottom

Just pick SDXL or whatever you want.

[gray geyser]

oh ok. which do you suggest

[nimble belfry]

I like more photo realistic models.

[gray geyser]

me too. can you suggest one

[nimble belfry]

I like Colossus project XL

[gray geyser]

thanks ill look into that

[nimble belfry]

What GPU you using?

[gray geyser]

3080

[nimble belfry]

Not bad. I want to upgrade but I'm broke

[gray geyser]

i got lucky, i have a friend who bought a 4090 that he does not need because he never uses it he just wants top everything you know

gave me the 3080 even. did buy us pizza though haha

i see 50 series coming soon, i told him about it thinking maybe he'll want that too and not need the 4090 😂

[nimble belfry]

Holy crap yeah that's luck. I got a 2080ti that way but I had to pay $300 still a good deal. He was upgrading to 4090.

[nimble belfry]

@gray geyser you get it all running yet?

[gray geyser]

I got it set on comfy now with Dragon's help so I switched back 😂

[nimble belfry]

Swarm uses comfy you should do those exact steps but using the swarm comfy Install. That way you also have the UI. Personally I like to have both.

[gray geyser]

yes i agree

right now Im goign to start some 16.9 images in dalle because so many i have are square

[nimble belfry]

Yeah you can also make them in Swarm just turn off video toggle until you have a good image. DALL-E is better at text and having multiple main focuses in the image without having to create a whole workflow.

[dim sonnet]

ignore the image decoder thing, irrelevant

[dim sonnet]

If you install this comfy extension https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite swarm will add a video format parameter automatically to let you pick

without that extension it can only save as animated webp

[gray geyser]

you da man alex thanks

[nimble belfry]

🤯

[nimble belfry]

It seems like if i choose any format other than webp it does not make the video. No errors or anything just does not generate the video. This is a windows install, installed manually first then used comfy manager to reinstall after it was not working.

[dim sonnet]

er... huh?

should definitely be some amount of message if it didn't generate anything

might need to set LogLevel to Debug

[nimble belfry]

22:01:30.253 [Info] Generated an image in 0.00 (prep) and 13.25 (gen) seconds
22:01:33.648 [Debug] ComfyUI on port 7821 stdout: got prompt
22:01:33.649 [Debug] ComfyUI on port 7821 stderr: ERROR:root:Failed to validate prompt for output 108:
22:01:33.651 [Debug] ComfyUI on port 7821 stderr: ERROR:root:* (prompt):
22:01:33.651 [Debug] ComfyUI on port 7821 stderr: ERROR:root:  - Required input is missing: crf
22:01:33.652 [Debug] ComfyUI on port 7821 stderr: ERROR:root:* VHS_VideoCombine 108:
22:01:33.653 [Debug] ComfyUI on port 7821 stderr: ERROR:root:  - Required input is missing: crf
22:01:33.653 [Debug] ComfyUI on port 7821 stdout: Requested to load SDXL
22:01:33.654 [Debug] ComfyUI on port 7821 stderr: ERROR:root:Output will be ignored
22:01:33.655 [Debug] ComfyUI on port 7821 stdout: Loading 1 new model
22:01:36.420 [Debug] ComfyUI on port 7821 stderr:
22:01:36.808 [Debug] ComfyUI on port 7821 stderr:   0%|          | 0/20 [00:00<?, ?it/s]
22:01:37.197 [Debug] ComfyUI on port 7821 stderr:   5%|\u258c         | 1/20 [00:00<00:07,  2.58it/s]
22:01:37.592 [Debug] ComfyUI on port 7821 stderr:  10%|\u2588         | 2/20 [00:00<00:06,  2.57it/s]
22:01:37.983 [Debug] ComfyUI on port 7821 stderr:  15%|\u2588\u258c        | 3/20 [00:01<00:06,  2.55it/s]
22:01:38.378 [Debug] ComfyUI on port 7821 stderr:  20%|\u2588\u2588        | 4/20 [00:01<00:06,  2.56it/s]
22:01:38.770 [Debug] ComfyUI on port 7821 stderr:  25%|\u2588\u2588\u258c       | 5/20 [00:01<00:05,  2.55it/s]
22:01:39.162 [Debug] ComfyUI on port 7821 stderr:  30%|\u2588\u2588\u2588       | 6/20 [00:02<00:05,  2.55it/s]
22:01:39.558 [Debug] ComfyUI on port 7821 stderr:  35%|\u2588\u2588\u2588\u258c      | 7/20 [00:02<00:05,  2.55it/s]
22:01:39.951 [Debug] ComfyUI on port 7821 stderr:  40%|\u2588\u2588\u2588\u2588      | 8/20 [00:03<00:04,  2.54it/s]
22:01:40.343 [Debug] ComfyUI on port 7821 stderr:  45%|\u2588\u2588\u2588\u2588\u258c     | 9/20 [00:03<00:04,  2.54it/s]
22:01:40.737 [Debug] ComfyUI on port 7821 stderr:  50%|\u2588\u2588\u2588\u2588\u2588     | 10/20 [00:03<00:03,  2.55it/s]
22:01:41.128 [Debug] ComfyUI on port 7821 stderr:  55%|\u2588\u2588\u2588\u2588\u2588\u258c    | 11/20 [00:04<00:03,  2.54it/s]
22:01:41.522 [Debug] ComfyUI on port 7821 stderr:  60%|\u2588\u2588\u2588\u2588\u2588\u2588    | 12/20 [00:04<00:03,  2.55it/s]
22:01:41.916 [Debug] ComfyUI on port 7821 stderr:  65%|\u2588\u2588\u2588\u2588\u2588\u2588\u258c   | 13/20 [00:05<00:02,  2.54it/s]
22:01:42.309 [Debug] ComfyUI on port 7821 stderr:  70%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588   | 14/20 [00:05<00:02,  2.54it/s]
22:01:42.703 [Debug] ComfyUI on port 7821 stderr:  75%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u258c  | 15/20 [00:05<00:01,  2.54it/s]
22:01:43.098 [Debug] ComfyUI on port 7821 stderr:  80%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588  | 16/20 [00:06<00:01,  2.54it/s]
22:01:43.494 [Debug] ComfyUI on port 7821 stderr:  85%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u258c | 17/20 [00:06<00:01,  2.54it/s]
22:01:43.890 [Debug] ComfyUI on port 7821 stderr:  90%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588 | 18/20 [00:07<00:00,  2.53it/s]
22:01:44.284 [Debug] ComfyUI on port 7821 stderr:  95%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u258c| 19/20 [00:07<00:00,  2.53it/s]
22:01:44.286 [Debug] ComfyUI on port 7821 stderr: 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 20/20 [00:07<00:00,  2.53it/s]
22:01:44.287 [Debug] ComfyUI on port 7821 stderr: 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 20/20 [00:07<00:00,  2.54it/s]
22:01:45.692 [Debug] ComfyUI on port 7821 stdout: Prompt executed in 12.04 seconds
22:01:45.992 [Info] Generated an image in 0.00 (prep) and 17.05 (gen) seconds```

[dim sonnet]

22:01:33.649 [Debug] ComfyUI on port 7821 stderr: ERROR:root:Failed to validate prompt for output 108: 22:01:33.651 [Debug] ComfyUI on port 7821 stderr: ERROR:root:* (prompt): 22:01:33.651 [Debug] ComfyUI on port 7821 stderr: ERROR:root: - Required input is missing: crf 22:01:33.652 [Debug] ComfyUI on port 7821 stderr: ERROR:root:* VHS_VideoCombine 108: 22:01:33.653 [Debug] ComfyUI on port 7821 stderr: ERROR:root: - Required input is missing: crf
what

required input crf? wat

oh my god they added new params and the params are incoherent letters very cool okay

[nimble belfry]

those errors dont show with webp

[dim sonnet]

needa just delete that extension and make a swarm internal one

yeah webp uses internal code

[nimble belfry]

giving you more work is my specialty.

[dim sonnet]

it's all self-contained in swarm now, VHS is no longer needed

[upper cave]

Hey mate, loving to updates, one question that you may be able to help me with, trying to tile my upscales. Thoughts on how to make them better or if I'm using the encode correctly here? it's 90% there. You can fainlty tell there's a end to the tile in this tiled 2x2

[potent panther]

So just installed SwarmUI and skipped installing comfyui and added my already installed comfy as a backend

But getting

00:57:13.767 [Error] [BackendHandler] Backend request #1 failed: System.InvalidOperationException: No backends match the settings of the request given!
00:58:19.607 [Warning] [BackendHandler] No backends match the request! Cannot generate anything.

I'm pretty sure I did something wrong but don't know what

[balmy orchid]

that "crf" is a simple quality level setting associated with encoding h264's in ffmpeg, I think you can just hardcode it at 20 and call it a day.

[dim sonnet]

strongly recommend using Comfy Self-Starting, not comfy api by url

saves a lot of trouble

[potent panther]

but why is it an option if it doesn't work?

[dim sonnet]

API-by-URL does work, but requires a lot more knowing-what-you're-doing

self-start manages everything for you

[potent panther]

huh

[potent panther]

Honestly this is pretty nice to use

All you need is to revamp the ui so it's actually readable and doesn't need csgo flicks

[dim sonnet]

Huh? Can you explain more what you mean?

[potent panther]

Most buttons are tiny

I don’t really have any problems with the style
But the model panel has everything shown as big
While the settings panel on the side has tiny toggle buttons there’s also empty space

[potent panther]

See here Gradio has a dropdown object which let you close or open wherever you click on it, it doesn't care about the text.
SwamUI has dropdowns too but you need to click on the text to do the same thing, it's tiny and is less pleasant to use due to this.

[dim sonnet]

oh, easy fix - pushed a commit that makes the clickable area full width as you expect

[potent panther]

nice

oh

[dim sonnet]

wh

[potent panther]

toggles aren't invisible?

[dim sonnet]

oh, fek, that's why i get for testing with a workflow loaded

[potent panther]

lol

[dim sonnet]

okay now it's fixed more betterly

[potent panther]

the moment I did git reset

That's nice qol

[potent panther]

Found a bug

Embeddings are saved as comfyui embeddings in metadata

so <embed:aidxlv05_neg.safetensors>
become embedding:aidxlv05_neg.safetensors

so reuse parameters technically give the wrong embedding path

[dim sonnet]

ooo

negative ones specifically

[dim sonnet]

fixed

already had handling for that for (positive) prompt, just had to duplicate to negative too

[potent panther]

thanks

Yeah yeah scam

That's not how you get a job

Because you spammed it in every channel duh

Still that's not how things work

@uneven moth @thorn harness

[nimble belfry]

@dim sonnet I am not able to edit the metadata on SVD models in the UI. Is this intentional?

[dim sonnet]

that looks corrupted / perm blocked? Check that you have permission to the folder it's in

if that's not the problem, shut down swarm, open folder, delete model-metadata, restart

[nimble belfry]

yep, deleting that fixed it. Thanks!

[balmy orchid]

There are some settings (the ComfyUI settings under the gear icon) that show up when I run ComfyUI directly, from http://127.0.0.1:7821 for instance, that do not show up in the same installation of ComfyUI when I access it from the StableSwarmUI. It's hard to tell, but I noticed Efficiency Nodes, ComfyUI-Custom by Pythongossss, and failfast-comfyui-extensions are missing. Interestingly, these packages all alter ComfyUI's Interface in some manner.

[potent panther]

Reuse parameters doesn't work on resolution if aspect ratio isn't custom

[rapid garnet]

where are my customnodes on the comfyui backend

am i doing something wrong here

[dim sonnet]

there are sometimes bugs in these extensions that make it fail in swarm but in direct usage don't cause problems - pythongossss comfy-custom-scripts I sent a PR to fix that bug a few months ago: https://github.com/pythongosssss/ComfyUI-Custom-Scripts/pull/39 are you maybe just outdated on those extensions?

[dim sonnet]

that looks right

albeit the extension names look weird. Folders like PoseNode sound off

if you downloaded individual node files (.py) you can't put them in subfolders

subfolders should only be for git clones of extension repos

[rapid garnet]

its how they came, all of these were 1 big package

[dim sonnet]

what

OH

i see

yeah no you did it wrong

you downloaded a repo directly into your custom_nodes

you should git clone it, or save it in a subfolder by the repo's name

[rapid garnet]

so like this then

im not sure anymore cuz i checked every tab and dont see them in there

[dim sonnet]

yes, though you accidentally also moved the example for which wasn't part of the repo

[dim sonnet]

make sure to restart swarm

the comfy backend needs a full restart, and you need to refresh the page, for new nodes to show up

[clear herald]

per @potent panther request to post this here

[potent panther]

cake you forgot to mention your issue

[clear herald]

oh
on mobile the interface is a bit crammed
then the tabs (i.e the negative one) cannot be closed unless switch to desktop mode on phone

[nimble belfry]

Image editing tools
Wildcards
Localization / automatic translation
Mobile browser formatting
Self-contained internal mini-backend for simple usages
Simple box grids for batches (eg 2x2 squares of images when you generate a batch of 4)
full detail "Current Model" display in UI, separate from the model selector (probably as a tab within the batch sidebar?)``` It looks like its on the TODO list already.

[potent panther]

btw mcmonkey

[dim sonnet]

oop i forgot about that sorry

[dim sonnet]

at a quick check, it worked for me?

i generated a 16:9 image, then set aspect ratio to 1:1 and clicked reuse parameters and it applied 16:9

oooh, but if you have it on custom, then change it off custom, it goes wonky, i see

fixed

[potent panther]

Yep

[potent panther]

Am I missing something?

[steady pewter]

I wanted to just give a heads up that 2.1-turbo in stableswarm isn’t functioning.

works fine in comfy with same settings

I also have no issues with sdxl-turbo in either, so seems like it’s specific to sd 2.1-turbo in stableswarm

(silicon mac btw but not sure if OS related)

[potent panther]

Did you update your backend comfy?

[dim sonnet]

uhh does it fix itself if you refresh?

[potent panther]

No

It's always like this

I have 3 VAE files in safetensors format

[dim sonnet]

oh, wait

you're editing the User settings, right?

not the Server Settings for Default Users?

[potent panther]

I'm just here

[dim sonnet]

that's the Server Settings

user settings are under User tab

[potent panther]

but why doesn't it let me put a VAE?

are the server settings not allowed to check for VAEs?

[barren moat]

@dim sonnet I'm doing a super fresh install on a new formated Windows 11 PC. When I run the .bat file I'm getting this error. Do I need to install GIT first, I thought SWARM would install everything needed, or no?

Never mind I manually installed it and it's running the install now

Hmmm I think I'm missing something else...

[dim sonnet]

the install.bat should install everything on its own if you followed the readme

which i don't think you did cause those errors can't happen if you did

https://github.com/Stability-AI/StableSwarmUI#installing-on-windows

on brand new fresh win 11 you might need to restart PC

[barren moat]

Let me go check the readme. I don't recall doing the manual installs before

[dim sonnet]

win11 installs some core files after it boots slowly over time and doesn't activate em til restart

you usually don't notice but eg winget which swarm uses has a slight delay

[barren moat]

I rememeber swarm installing everything needed.

ok

Actually I did follow that readme and download the .bat from there. It's probably needs the rebootty.

Hmmm still getting this error. There's some windows updates that need to be installed so I'll do them, reboot and try again. Sorry it's me that seems to stumble on weird issues that are most likly user error.

[barren moat]

@dim sonnet I never had it put an icon on my desktop before either. I get that same error when I run the launch-windows.bat file

Does it need to be on the C drive??

[dim sonnet]

yeah, still, start over - delete the folder, grab the installer file

it's the install-windows.bat that installs the basic requirements and that's what failed for you

so it's missing the basic reqs rn

[barren moat]

ok I'll delete it and redownload the .bat file

after I run that install-windows.bat it creates the swarm folder and and this is the contents.

so now I run the launch-windows.bat correct?

[dim sonnet]

it, uh, should've ran on its own, but - yes

[barren moat]

I've never had it run on it's own . Maybe it's doing somwthing and I'm not waiting long enough??

I'm not having any luck. I'll see if I can find a work around

[dim sonnet]

can manually install dotnet 7 sdk https://dotnet.microsoft.com/en-us/download/dotnet/7.0

[barren moat]

Ok thanks I’ll do that

Is python installed automatically or do I need to install that?

[dim sonnet]

yes that's automatic

[potent panther]

Aspect ratios are kinda useless since they lock you at a precise resolution

And none are good SDXL resolutions

[dim sonnet]

The resolution is based on the model, if you're using SDXL that's 1MiP and the resolutions it uses are very precisely intended to be exactly the correct resolutions SDXL was intended to use

21:9 is the limit of stretch because that's where the model was shown to break down and produce stretching artifactions in testing

1:2 will work but is pushing it

768 * 1536 is also a bit off from 1024*1024 pixel count, 704 * 1536 would be closer, or 768 * 1408

[potent panther]

But why not have a slider to still change the resolution but be locked to a certain aspect ratio?

[dim sonnet]

mostly because it's extremely rare for a model to support that, so i hadn't bothered?

SDv1 models historically sometimes allow you to do a wonky stretch between 512...768, SDXL i've seen one or two to let you get weird with it but most don't

you technically can go out of range on standard SDXL models it just produces stretching artifacts quickly

and the goal of the quick selector for aspect is to be easy very good preset res options

and Custom for anything else

[hearty notch]

just use custom!!!!

[potent panther]

Yeah I do

[gray geyser]

every time i try to generate i have this error

[hearty notch]

your setup is not correct, you need to have a backend to run the UI

[gray geyser]

Friday I make a new install, and try it, Saturday nothing work I only shut down my computer

[hearty notch]

When you go the comfy editor tab, do you see anything?

[nimble belfry]

Did the backend load properly? Do you have any errors?

[gray geyser]

i dont have any error

[nimble belfry]

In the server tab where it shows the backends. Does it say they are all loaded? Do you have it trying to do some custom settings where it's trying to use a different backend?

[gray geyser]

im new with comfy and i dont try any custom thin

[nimble belfry]

click on your server tab does it say your backend is running?

[nimble belfry]

what have you already tried? Have you tried the basics? restart and reset workflow to default?

[gray geyser]

yes i try, but sometime i can make only 1 pic or sometimes the error cam again

this mornin a made a new install and now it work but i only have dsxl in model

[dim sonnet]

when you get stuff like this and aren't sure why, go to Server -> Logs -> and set view to Debug, and there's a much more detailed view there that probably has the explanation

[dim sonnet]

if you have other models not showing up you might just need to redo server config -> model paths

[gray geyser]

i change it and it work 1 hour and the backend error again

[dim sonnet]

well, look in the debug logs, find out why

[hearty notch]

can someone share an exmaple of wildcards that can be used with SSUI?

[dim sonnet]

again,

go to Server -> Logs -> and set view to Debug, and there's a much more detailed view there that probably has the explanation

[dim sonnet]

an example? just go to Wildcards tab -> click Create New Wildcard, and type some stuff in the box

if you want some for actual usage, there's a lot of collections out there, including for auto webui's older wildcards impl, eg https://github.com/adieyal/sd-dynamic-prompts/tree/main/collections is a big collection dump

[potent panther]

Does the image history have a way to sort by date instead of alphabetical?

[dim sonnet]

... name your files with date in it? :D

yeah i know it needs sort options lol

[dim sonnet]

New: LoRA extractor utility https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7772593

[dim sonnet]

(full disclosure this exists pretty much because in the middle of the night I went "I want a turbo lora, why no turbo lora? wait i can just make turbo lora. Wait does comfy have a node about this? no node?? screw it i make node. and UI. lora extraction for everyone weee")

(my turbo lora results in outputs that are very slightly faded but otherwise works fine)

[potent panther]

[dim sonnet]

more new! Tab complete in prompts! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7775593

[nimble belfry]

[barren moat]

@hearty notch Check that out! Pretty cool!!
#🐝｜swarm-ui message

[nimble belfry]

How do i get a list of LoRAs using the API?

Is LoRA a valid subtype of ListModels? Is that how its meant to be done?

[potent panther]

Currently running lora extract but there doesn't seem to be log output during extraction

[potent panther]

It finished and there was no log output

[dim sonnet]

yes

[dim sonnet]

there should be under Debug or under the Comfy backend view

[potent panther]

shouldn't it be under info like images?

[dim sonnet]

could add an info output ye

i wasn't too focused on clear terminal output as i managed to get clear frontend UI output

[potent panther]

Since I'm loading from HDD it takes ages to show the loading bar

I guess I could keep the debug log level

[dim sonnet]

added, 1 when it's starting and 1 when complete

[potent panther]

thanks

since I use custom resolutions I didn't realise but
aspect ratio and custom conflict with how they show values

this is portrait

this is landscape

this is portrait

[dim sonnet]

haha, yeah. Yeah they do.

The aspect ratio selector is target standard ratios (eg 16:9) but then applied with rounding to the nearest x64

when you're using custom it just displays the actual mathematical result of whatever values you used... and the x64 rounding makes it technically different from the ideal standard

it is technically wrong, but it is intentionally left at being slightly wrong

(you're a lot more likely to know offhand what 16:9 is than 7:4 - 16:9 is a super common standard res used for monitors and wotnot)

[dim sonnet]

i added welcome messages to the main page https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7791189

which i don't really need to mention here cause, yknow, you'll immediately see it when you update and open swarm but lol

the best welcome message, announcing itself <3

[potent panther]

Noice

[fickle jacinth]

How do I download controlnet? Is there a way I give it a url and it can download it?

[dim sonnet]

atm, download via your browser and save the file in (Your swarm folder)/Models/Controlnet

in the future i would like to have clean happy automatic downloaders , but don't have that yet

[fickle jacinth]

I understand, it's still in it's infant stages. Having something like automatic would be so nice

[potent panther]

I think it would be a good idea to move these buttons in the top bar

The gallery cog is tiny and quick tools overlap with the top bar

[potent panther]

Also I'm confused as to how to close the edit images panel

[mellow pendant]

This looks pretty good, can this be installed over an existing ComfyUI without breeaking much

[dim sonnet]

can't yet other than refreshing

image editor is still experimental new thing, not quite ready for primetime

[dim sonnet]

yes you can use an existing comfy install fine, just select "None" when it asks what backend in the installer, then on the main interface go to Server -> Backends -> add a comfy self-start backend and give it the filepath to the comfy folder, and go to Server Config and set the ModelPath to your comfy model folder

[mellow pendant]

Nice, thanks, will do.

[worldly oxide]

Can this be used as a front end for my local installed Automatic 1111, if so, what would be the steps for making this work?

[dim sonnet]

technically yes but i would not recommend using auto as the backend

it's... kinda just a worse experience with no real upside over comfy

the one thing auto had going for it in general was having a more put together frontend, but naturally that doesn't count in its favor anymore when you're replacing the frontend anyway

you can install swarm with a comfy backend and using the same models dir / output folder / whatever else you want to share though, that works perfectly fine

[worldly oxide]

Thanks for the quick and clear response, I'll stick to the Swarm and ComfyUI workflow

[dim sonnet]

don't ask why it took 2 months but anyway

it's a user setting:

if you do non-square numbers it'll just

have a hole at the bottom

[nimble belfry]

hell yeah now I can literally cut out a bunch of code

I'm at SeaWorld today but maybe I can play with it tonight.

[last tinsel]

I'm really enjoying Swarm! Just wondering if there's a way to put the Negative text box under the main Prompt box? Thank you.

[nimble belfry]

any chance of having a grid on preview images? also, is there something that distinguishes that the image is a grid? Like in the gen_progress or something like that? I dont see where it shows that in dev tools on chrome.

[dim sonnet]

is there something that distinguishes that the image is a grid?
(A) it's the final output guaranteed rn (B) it has a negative batch index

any chance of having a grid on preview images?
uhhh maybe? Maybe.

I'm a bit cautious about that just cause processing a large image repeatedly and rapidly might be iffy so ima have to test if that's stable or if it'd make a problem

[dim sonnet]

can I ask why you want that?
For context the reason it's separated off to the side is that, especially in modern XL models, negative prompts are often not too helpful - and even back in SDv1 days it was common to use a single static preset negative repeatedly, rather than modifying it often.
I wouldn't expect to ever be the case that you'd be modifying it nearly as frequently as the main prompt

[dim sonnet]

cc @glass nova the docs for multiple backends are much more fleshed out now: https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md#step-by-step-setup-guide-for-multiple-machines

[glass nova]

Awesome! Will take a look. Thanks so much.

[last tinsel]

More a selfish wanting of all the prompt boxes in the same area as I've been used to with WebUI, but I think you are right and I will keep using Swarm more with its intended layout! I'm so impressed with the gen speed and the ability to use Comfy without feeling too overwhelmed. Thank you for your explanation and response!

[nimble belfry]

@last tinsel I feel the same way. I wanted to start learning Comfy and this made it super easy.

[potent panther]

It’s honestly very useful to remove stuff you don’t want

But it could be made easier if you moved the neg box to the bottom of the left panel

It would be next to the prompt box

[lime elk]

any Idea why not all GPUs on backend are generating?
Looks like only 3 out of 6 gpus are getting the jobs

[lime elk]

Ok, I will answer myself 😉
Somehow, models were not autoloaded upon backends starting on that machine
After forcing 1 img to generate on each gpu, it works ok

time to configure another one 🙂

[fickle jacinth]

Can swamui run sdxl turbo in realtime? I can't seem to find anything for it

[lime elk]

Any idea how much ram should I provide for 1 instance of backend [1 gpu]?

[dim sonnet]

realtime previewing? Yes

[dim sonnet]

GPUs only get allocated a model if the prior GPUs in the list can't keep up with the queue - so if you want to force more allocation just run a big queue, -or- click the Load Now button to forceload on all:

[dim sonnet]

the comfy python will spike up above 8 GiB of system RAM, particularly while loading models, but generally idles more like 2 GiB

if you're planning out a large multiGPU build, 64 GiB RAM is relatively affordable and should more than suffice

if you're doing some form of VM software locking of RAM limits, uh -- hard-allocating per-instance is a bad idea due to how variable usage is

[lime elk]

Just stolen 32g from another machine for testing, Im running 6x3060ti

[potent panther]

You’re trying to make your own image gen service?

[lime elk]

nah, trying to have some fun with "leftovers" 😛

Looks promising 🙂

[potent panther]

Bitcoin mining

Or work gpus being sold for a low price due to an upgrade?

[lime elk]

yep, ex mining rigs

with LCM lora it's ridiculously fast

[nimble belfry]

I was looking into building a GPU rack using mining parts. I was worried about the speeds from how the cards are plugged into the risers. Do you have any issues with that and are you using anything special?

[lime elk]

To early for decent conclusions. It's normal rig with ga-h110-d3a mobo, 6x3060ti on risers, 32gb ram, and some crappy cpu - need to swap that. Comparing to mining the cards are chill 😉

I'm not that deep into technicals od SD, but as the model is in vram I think pcie speed on risers is not a concern

I will do some testing soon, but looks like single 3060ti on rig is around half of 3090 speed in my desktop, so I think it's correct.

[dim sonnet]

as long as the full model fits on the GPU VRAM, pcie speed shouldn't make much different for AI usage

if it has to unload/reload model it might hit a bit

3060 (non-Ti) is ideal here btw, more VRAM than the other 30xx cards

[lime elk]

not counting 3090 ;)?

[dim sonnet]

well yeah lol

[nimble belfry]

Yeah I think he meant price performance ratio.

[unreal geyser]

Hello, how to insert an image into a picture so that it looks realistic (in color, light and perspective). instead of this machine you need to insert cybertruck

[nimble belfry]

What do mean by "insert an image"? Do you mean training a model or LoRA? Or do you mean just doing initimage?

If you want that exact image just with a cybertruck you can do a few things but getting exact pixel perfect colors is a bit harder to do.

[tidal crypt]

Is there anyone to know about stablecode?

[dim sonnet]

not just that - it's actively better flatout than a 3060 Ti, 3070, 3080

which all have less VRAM

the cheaper card is a better card

(for AI)

[dim sonnet]

#1098025024541167646 would be the channel for LLMs

[dim sonnet]

image editor in swarm is still a bit of a WIP, but you can drag the image into the UI, click "Edit Image", then on the layer list on the right click the "+Mask", and use the brush to fill in black anywhere you want unchanged (eg the background) and white anywhere you want changed (ie the car), then give a prompt for what you want filled and hit generate

you might also want in some cases to use controlnet to keep the result similar to the start

[nimble belfry]

why the extra step to paint it black? Why not just assume anything not white should remain?

[potent panther]

I think a commit on comfy or swarm messed up the loading of samplers

[dim sonnet]

The mask defaults to white atm (ie paint everything)

ideally there should probably be a button to quick flip the base on that

[dim sonnet]

eh?

at a glance, latest swarm + comfy loads the list fine and generates with a custom sampler fine

[potent panther]

lemme test with latest commit on everything

yeah everything seems fine idk what happened with the list last time

[hearty notch]

Hi @dim sonnet A friend of mine was asking if SSUI will work on m3 iMac

[potent panther]

As long as comfy supports M3 it should work

M1 and M2 are currently supported at least

[potent panther]

The new dropdown is very nice but it doesn't spread far enough vertically in my opinion

[dim sonnet]

tru - they autoexpand now

[potent panther]

nice

[raw laurel]

Great work! I really enjoy Swarm so far and prefer it over A11111/SDNext. My use case is personal with sharing access to close friends and family so I'm really looking forward to the separate user features. I want the system to be very user friendly for those who are not technical persons, but allowing for thorough setup and tweak for me and some friends.

I wish there would be possibility to make a front page for dummies in the future - containing only positive prompt, preset selector and collapsible extra options that are selectable by admin (I would probably add collapsed negative prompt, aspect ratio and not much else). And possibility to view that user's generated images easily.

Looking forward where this project evolves into!

But not all has been smooth sailing. With the default workflow sdxl with refiner I enabled image tiling option once and even after unticking the box it doesn't stop tiling. Anything I can provide to help fixing this?

[dim sonnet]

which tiling option specifically?

[raw laurel]

Seamless Tileable 😄

[dim sonnet]

oh yep, true, upstream change seems to have broken the reset on that -- fixed, update and it should probably unstuck itself

[potent panther]

@uneven moth You’re the only mod online

It’s a scam website

Domain is even .ru

[potent panther]

@thorn harness ?

[thorn harness]

Thanks

[rigid bear]

Tried install StableSwarmUI got the following error

[raw laurel]

Is there any way to monitor terminal output from self-starting comfyui? I'm getting random drops from WebAPI as websocket dropping connection. Judging by the memory usage comfyui doesn't crash totally but is there any way to investigate?

It was during normal sampling steps halfway to the process - also not the first image of the batch

[potent panther]

The error is literally telling you what’s wrong

You should try reinstalling

[raw laurel]

That's not helpful at all. I don't know how windows or ssui installs comfyui as didn't use it myself. He could try installing comfyui manually and then link it as backend

[potent panther]

I did give a solution to try

[raw laurel]

Silly me - the logs page has a small dropdown to view the log. Veeery nice! Just little difficult to spot at the first time. Maybe add some way to direct there from the backends page?

[nimble belfry]

Monkey told someone else in the past that he does not recommend that. He said it was better to let it install the normal way.

[dim sonnet]

oh, oops, my bad, mistake in the recent change to a different comfy backend installer -- it's fixed now, delete your swarm folder and try again and should work fine

[dim sonnet]

also visible if you select "Debug" there (shows all backend outputs in one view) but that's a good idea - added two lil buttons:

(needed a per-backend Restart button too)

[open badger]

i am currently using both comfui local install and stableswarm. I want to use stableswarm exclusivly. how do i set up all my comffy model folders. on swarm. do i move them to swarm model folders or to comfy model folders. I know i need to move the the custom nodes to the dlbackend, but what about other folders and files and models like clip, ipadapter, onnx, roop, etc?

[dim sonnet]

open Swarm -> Server -> Server Configuration, there's a section labeled "Paths" - wherever you put your models, just make sure ModelRoot points to that folder. You can use the old folder and just point swarm there, or you can copy the models into the Swarm default Models folder, up to you.

Personally I keep all my models on a second drive (a very big secondary drive to fit em all lol)

note you may need to make sure the subfolder names match between the folder and the Paths settings as well. Any folder type not listed in settings is one you should leave named exactly how comfy had it named

[raw laurel]

I'm getting funny readouts on generation times. Last row in reality generated in about same time (148 sec with manual timing). Not sure how I could be helpful to solve this.

[dim sonnet]

I'm aware of this and have plans to fix it -- it's actually accurate timings it's just doubled atm

[raw laurel]

Last night I queued about 120 generations before going to bed. I see hilariously long prep times there, what is included in the prep time? Generation time is also elevated

[dim sonnet]

and by "have plans" I mean probably today lol i just got a list to go through

Prep time is how long it was sitting in queue for, so that's including other things queued up before it, time to load the model, etc.

Gen time is the actual time, it just doubles up when you queue several

[raw laurel]

Maybe prep could be subtituted as queue time then? In the top photo it was running only one generate forevers, wasn't anything else in the queue. At least on puprose

[raw laurel]

Oh nevermind, maybe there was something else in the queue as I forgot batch count on before pressing generate forever. Is there any way to view and manage the queue?

[dim sonnet]

atm 'view the queue' is just:

manage is, uh, you can clear your current session's queue, or all session queues:

(other session would be eg open another tab or whatever)

[raw laurel]

Yeah - I'm aware of those, thanks. For admin would be great to see an actual list of what is in the queue and what parameters

[dim sonnet]

a more advanced queue manager might be useful though, esp. for slower gens like the two minute ones you're doing there

(you can view more detailed info about what's running in the server Logs btw, just can't do much management about it)

[raw laurel]

That would be super helpful. I'm running on an older gpu and the queue is a black hole. Especially last night when I threw in 30 batch size generations by 4 times and wasn't sure what's going on afterwards. Apparently the requests were interlaced somehow but not sure on by what logic. I know there's the setting for amount of preloading items into the queue of a backend.

But the batches weren't generated in sequence but rather funnily in seemingy random order

[dim sonnet]

order is only loosely preserved. Especially eg if you have several gens with different models, it will try to prioritize not changing the model for as long as possible (since model loading on most machines is significantly slower than generating)

[raw laurel]

While I'm awake I'm mostly using the generate forever button - very nice to get generations all the time and do adjustments on the go. I'm just little stumped on how closely it is tied to the browser session - I guess the loop is maintained in the browser? Maybe document this behavior somewhere to eliminate guesswork. That logging question was actually tied to this as I saw websocket error in the log assuming it was swarm - comfyui but now I'm guessing it's swarm - browser instead. Would it be possible to make the log entry more verbose to indicate user what connection actually gets disconnected?

[dim sonnet]

yes Generate Forever is part of the browser session

(if it kept going on the server out of control in the background that'd make a mess)

yeah that error is browser disconnected from the server, i can clean that up

• has dedicated custom error message now

[raw laurel]

Wonderful! At first I thought this error was thrown within the backend linking and was trying to investigate at the wrong place. Probably my laptop just dozed off long enough to drop the connection to server. If possible maybe also implement browser notification that the generate forever was interrupted? Keep up the awesome work!

[dim sonnet]

if the browser loses connection (and is aware of it) it should already show an error

that type of websocket fastcrash indicates the browser was effectively dead at the time it happened, so not much can be done on browser side for the browser itself failing

[raw laurel]

Okay. I'm not too familiar with using webockets in browser. I'll test different scenarios of simple internet disconnects and also sleep modes with the laptops. I'm expecting it to be the latter to cause these interrupts

[dim sonnet]

yeah laptop sleeps do bad things to programs that are running when you do that.

I've seen things break a lot worse - often laptops will disable their GPU when going to sleep, which is difficult to recover from without just restarting things

[raw laurel]

Wow now happened something really strange! I started that generate forever with setting 30 images ran through. Then it just looped copying the last image with 0,1 sec generation time. I did change the images to 1 during generate forever.

Edit: couldn't reproduce...

[dim sonnet]

If you Generate Forever and the parameters are static and your queue isn't filled up with other things, it'll hit cached output and automatically avoid regenerating the exact same thing it generated last

the caching is pretty cool. You can also eg change only the parameters of the Refiner stage, and it will cache and reuse the first-stage output and only rerun the refiner

[raw laurel]

Aha - so somehow I had locked the seed by accident? Possibly misclick of reuse parameters or something and it went into looping and saving the cached image all over in very rapid succession

[potent panther]

That's unlucky

[raw laurel]

Is there a way to set default parameters? I see under quick tools how to reset defaults but how to set them?

[nimble belfry]

Name it default. Let me find the exact thing I read 1 sec.

[nimble belfry]

#🐝｜swarm-ui message

There it is.

[dim sonnet]

that info btw is in one of the welcome messages when you first open the page

[nimble belfry]

How dare you expect me to read something directly in front of my face. I'm going to dig through weeks of discord messages Instead.

[raw laurel]

Once again I'm asking questions. Is there any way to offload models from memory after being idle for a certain time? While I'm on a fixed electricity cost I'd like to offload models to reduce idle wattage on GPU (50W to 9W) and free system memory just in case. I'm open to any hacks or cron scripts

[dim sonnet]

Good question. Uh. @cunning holly is there an API route to unload from VRAM on command? Would be useful for both the above request and things like sharing text-gen & comfy image gen on one machine

[cunning holly]

I can add an api for that but pytorch is still going to take ~500MB vram so it might just be best to kill it

[dim sonnet]

API would be handy for just dump to .cpu() so it can load back in at a moment's notice (vs a process restart needing 30 seconds for the process and another 30 seconds to load the model)

[open badger]

keep getting this after a few gens

[open badger]

does the google colab need to be updated for Microsoft.DotNet.SDK.8 ? git is already the newest version (1:2.34.1-1ubuntu1.10).
dotnet-sdk-7.0 is already the newest version (7.0.114-0ubuntu1~22.04.1).
0 upgraded, 0 newly installed, 0 to remove and 27 not upgraded.
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
E: Unable to locate package dotnet-sdk-8.0
E: Couldn't find any package by glob 'dotnet-sdk-8.0'

[dim sonnet]

the text cut off to the right is key info here, but it kinda looks like either Colab or the CloudFlare forwarding are breaking and dropping your network connection?

[dim sonnet]

net8 isn't in use yet, it's just a warning that swarm will be updated to net8 soon. I'll update the colab file whenever swarm is updated to net8 (hopefully colab actually, uh, adds net8 to their packages by then?)

[open badger]

Thanks so much. All of your hard work is appreciated.

[potent panther]

Ubuntu LTS doesn’t have dotnet 8 so it won’t be anytime soon

https://pkgs.org/download/dotnet-sdk-8.0

Only Mantic Minotaur (23.10) has Dotnet8
Jammy Jellyfish (22.04) the current Ubuntu LTS will probably receive it in half a year lol

[dim sonnet]

argh

probably need https://learn.microsoft.com/en-us/dotnet/core/install/linux-ubuntu#register-the-microsoft-package-repository to get it sooner

yeah

https://learn.microsoft.com/en-us/dotnet/core/install/linux-scripted-manual#scripted-install owo? non-admin install?

[open badger]

i am using the google colab in windows 11 and I just want the colab to keep working. All of that is over my tiny windows brain. we just like to point and click. lol

[dim sonnet]

pushed an update, there's now an easier Linux installer using the dotnet non-admin install script, which will autoinclude both dotnet 7 & 8. Also swapped the colab notebook to use it so that'll stay good

[open badger]

the new colab worked perfectly................./content/drive/MyDrive/StableSwarmUI
From https://github.com/Stability-AI/StableSwarmUI
a5f5849..b3f5d6f master -> origin/master
Already up to date.
=========
Welcome to .NET 8.0!

SDK Version: 8.0.100

Telemetry

The .NET tools collect usage data in order to help us improve your experience. It is collected by Microsoft and shared with the community. You can opt-out of telemetry by setting the DOTNET_CLI_TELEMETRY_OPTOUT environment variable to '1' or 'true' using your favorite shell.

Read more about .NET CLI Tools telemetry: https://aka.ms/dotnet-cli-telemetry

---

Installed an ASP.NET Core HTTPS development certificate.
To trust the certificate, view the instructions: https://aka.ms/dotnet-https-linux

---

Write your first app: https://aka.ms/dotnet-hello-world
Find out what's new: https://aka.ms/dotnet-whats-new
Explore documentation: https://aka.ms/dotnet-docs
Report issues and find source on GitHub: https://github.com/dotnet/core
Use 'dotnet --help' to see available commands or visit: https://aka.ms/dotnet-cli

MSBuild version 17.8.3+195e7f5a3 for .NET
== Determining projects to restore...
Restored /content/drive/MyDrive/StableSwarmUI/src/StableSwarmUI.csproj (in 4.46 sec).
StableSwarmUI -> /content/drive/MyDrive/StableSwarmUI/src/bin/live_release/StableSwarmUI.dll

Build succeeded.
0 Warning(s)
0 Error(s)

[potent panther]

mcmonkey would it be possible to have an option to save images before refiner postapply?

also merry christmas

[nimble belfry]

@dim sonnet Using initimage I am not able to set batchsize to more than 1. Is this intentional or a bug? 16:49:55.848 [Error] Internal error processing T2I request: System.IO.InvalidDataException: ComfyUI errored: { "error": { "type": "prompt_outputs_failed_validation", "message": "Prompt outputs failed validation", "details": "", "extra_info": {} }, "node_errors": { "100": { "errors": [ { "type": "exception_during_inner_validation", "message": "Exception when validating inner node", "details": "unhashable type: 'list'", "extra_info": { "input_name": "latent_image", "input_config": [ "LATENT" ], "exception_message": "unhashable type: 'list'", "exception_type": "TypeError", "traceback": [ " File \"C:\\Users\\kaleb\\Desktop\\projects\\StableSwarmUI\\dlbackend\\comfy\\ComfyUI\\execution.py\", line 454, in validate_inputs\n r = validate_inputs(prompt, o_id, validated)\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n", " File \"C:\\Users\\kaleb\\Desktop\\projects\\StableSwarmUI\\dlbackend\\comfy\\ComfyUI\\execution.py\", line 435, in validate_inputs\n o_class_type = prompt[o_id]['class_type']\n ~~~~~~^^^^^^\n" ], "linked_node": [ "100", 0 ] } } ], "dependent_outputs": [ "9" ], "class_type": "RepeatLatentBatch" } } } at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUIAPIAbstractBackend.AwaitJobLive(String workflow, String batchId, Action`1 takeOutput, CancellationToken interrupt) in C:\Users\kaleb\Desktop\projects\StableSwarmUI\src\BuiltinExtensions\ComfyUIBackend\ComfyUIAPIAbstractBackend.cs:line 348 at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUIAPIAbstractBackend.GenerateLive(T2IParamInput user_input, String batchId, Action`1 takeOutput) in C:\Users\kaleb\Desktop\projects\StableSwarmUI\src\BuiltinExtensions\ComfyUIBackend\ComfyUIAPIAbstractBackend.cs:line 611 at StableSwarmUI.Text2Image.T2IEngine.CreateImageTask(T2IParamInput user_input, String batchId, GenClaim claim, Action`1 output, Action`1 setError, Boolean isWS, Single backendTimeoutMin, Action`2 saveImages, Boolean canCallTools) in C:\Users\kaleb\Desktop\projects\StableSwarmUI\src\Text2Image\T2IEngine.cs:line 287

[spiral plinth]

I cannot run the swarm ui with comfy ui backend, I run comfy ui with cpu and don't know if it is the reason that my computer cannot run swarm ui. Can anyone help

[nimble belfry]

you are trying to add a separately installed comfy backend? or the one that auto installs with Swarm?

[spiral plinth]

the one that auto installs

[spiral plinth]

will it work if I add a separate backend?

[dim sonnet]

I mean, technically, you can just reuse seed and turn off refiner's checkbox and you'll get that (thanks to comfy internal caching it won't even take any time to generate it'll just give you the image instantly from cache). But still I added Refiner Save Before Refine anyway, requires Display Advanced be checked to see it

[nimble belfry]

monkey said in the past that is not recommended. So best to use the one that auto installs. Do you have any errors or more info on what exactly is happening?

[spiral plinth]

it just tell me the I don't have gpu that is require

[dim sonnet]

Fixed. Was untested as BatchSize is heavily discouraged vs using the Images param. Has minimal benefit and creates a lot of secondary issues.

[dim sonnet]

Go to Server -> Backends -> on the comfy backend that's there, click the lil ✏️ edit button, and in the ExtraArgs field type --cpu and save

[nimble belfry]

batchsize generates my 4 images at the same time so it looks cooler in my img grid

[dim sonnet]

lol

[spiral plinth]

ComfyUI-0 on port 7821 error: RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx

[dim sonnet]

#🐝｜swarm-ui message

[potent panther]

Thanks

[dim sonnet]

I saw https://github.com/comfyanonymous/ComfyUI/commit/a252963f956a7d76344e3f0ce24b1047480a25af was committed the day after I asked about that, and I did some testing with --disable-smart-memory to have it autounload:
Nothing running: 0.8 GiB VRAM usage
Comfy running: 1 GiB
Comfy generated once then finished: 1.2 GiB

Based on that, I think there's theoretically ~0.2GiB of VRAM usage that doesn't get unloaded normally but could be? idk if it's cached values or what

But it's low enough either way that in my case I'm able to load a 13B LLM and Comfy for image gen at the same time on a 12GiB card (Barely, lol)
So, disable-smart-memory seems to be sufficient.

[cunning holly]

I'm going to add a route to unload models manually, --disable-smart-memory can make things a lot slower depending on the workflow and system

[potent panther]

@dim sonnet What is the right syntax for LoRA weights in the grid generator?

I can't find it in the grid docs

[dim sonnet]

if loras is mylora, mylora2 then lora weights should be 1, 1 ... you can use || to make the grid do grid things and not get confused by the ,

If you want to muck with loras in grids and that format is inconvenient you can also just do <lora:mylora:1> in the prompt or with prompt replace

[potent panther]

Wait so if weight is 1, 1 and there's 2 LoRAs how do I also have the weight to 0, 0 for the next image in the grid

Do the commas matter?

It's painful how tiny the values are in the final grid

[dim sonnet]

if you want to set weight to 0? that's, uh, weird to do, but you would do 1, 1 || 0, 0

if you're just trying to turn the lora on/off i'd recommend the prompt replace method

[potent panther]

oh ok

is it possible to make the text bigger?

[potent panther]

Oh by the way I don't know if it's intentional but I tried SKIP: at one point and it crashed

19:00:04.229 [Error] Internal error processing T2I request: System.FormatException: The input string 'Dpo-SDXL.safetensors:0' was not in a correct format.
   at System.Number.ThrowOverflowOrFormatException(ParsingStatus status, ReadOnlySpan`1 value, TypeCode type)
   at System.Single.Parse(String s)
   at StableSwarmUI.Builtin_ComfyUIBackend.WorkflowGenerator.<>c.<.cctor>b__9_2(WorkflowGenerator g) in /StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs:line 116
   at StableSwarmUI.Builtin_ComfyUIBackend.WorkflowGenerator.CreateStandardModelLoader(T2IModel model, String type, String id) in /StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs:line 1087
   at StableSwarmUI.Builtin_ComfyUIBackend.WorkflowGenerator.<>c.<.cctor>b__9_0(WorkflowGenerator g) in /StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs:line 86
   at StableSwarmUI.Builtin_ComfyUIBackend.WorkflowGenerator.Generate() in /StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs:line 1039
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUIAPIAbstractBackend.CreateWorkflow(T2IParamInput user_input, Func`2 initImageFixer, String ModelFolderFormat, HashSet`1 features) in /StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/ComfyUIAPIAbstractBackend.cs:line 549
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUIAPIAbstractBackend.GenerateLive(T2IParamInput user_input, String batchId, Action`1 takeOutput)
   at StableSwarmUI.Text2Image.T2IEngine.CreateImageTask(T2IParamInput user_input, String batchId, GenClaim claim, Action`1 output, Action`1 setError, Boolean isWS, Single backendTimeoutMin, Action`2 saveImages, Boolean canCallTools) in /StableSwarmUI/src/Text2Image/T2IEngine.cs:line 307
19:00:04.229 [Error] Grid generator hit error: Something went wrong while generating images.
19:00:05.229 [Error] GridGen stopped while running: {
  "error": "Something went wrong while generating images."
}

[dim sonnet]

eh, sure - updated to dynamically scale up a bit if it has room

[dim sonnet]

this error message doesn't look at all like it came from a SKIP:

that looks like you typed Dpo-SDXL.safetensors:0 into the loras input

or rather: into the lora weights input*

[potent panther]

Wait did I

Oh oops copied the wrong error

[potent panther]

Anyway can you try something like this because it crashed

[dim sonnet]

skipping in Grid Image has no point btw

that's an option for when you're doing Web Page stuff mainly

but i fixed the error in generating the Grid Image with skips

[nimble belfry]

How do presets work with the API? Do they just set the perams listed in the preset?

Let me rephrase, how would I get the preset so I could send those perams?

Im trying to have an LLM choose the best settings for whatever the user is asking for by setting a preset.

[dim sonnet]

easiest/laziest way is just shove it into the prompt, <preset:whatever>.
You can alternately add "Grid Gen Presets": "whatever, whateverelse" to the json request
or you can call GetMyUserData and read presets and apply things at your own discretion

[nimble belfry]

oooo nice and easy, I like it. Thanks. I think im going to make a new discord logo contest for my server. I need to make sure it switches to the correct lora for txt.

[open badger]

I modified the colab to use ngrok. So much better.

[dim sonnet]

yee i wanted to use ngrok cause cloudflare quality seemed shifty, but ngrok requires an account :(

[crimson pecan]

Have any of you been able to get FizzNodes to work with your ComfyUI through StableSwarm?

[brave urchin]

Is there a fix for the compatability issue for the canvas_tab nodes? https://github.com/Lerc/canvas_tab
Seems stableswarm wont let me click on the edit, or queue on change buttons. Here's a screenshot for stableswarm, and then comfyui standalone.

[brave urchin]

Also, I'm wondering why I receive two different images when generating the exact same thing, using the exact same settings, to a T, in stableswarm vs comfyui. From my understanding, it should generate exactly the same image, if everything is exactly the same. Is there a setting behind the frontend that I'm not seeing somehow? Also, the variation seed is off

[dim sonnet]

use this button:

to get the exact workflow which should indeed give exact identical results

you'll notice:

SwarmKSampler (which has added options like sigma, var seed, etc)

and the CLIPTextEncodeSDXL with specific values

if you match settings exactly between swarm generate tab and comfy tab, it will indeed produce identical results, but swarm does a lot behind the scenes to perfect the generation as best it's able

eg how resolution conditioning is 1.5x bigger on positive prompt and 0.8x smaller on the negative prompt

(this biases XL models to produce higher quality images)

[dim sonnet]

sent em a PR to fix it https://github.com/Lerc/canvas_tab/pull/16

[brave urchin]

🫡 Thank you!

[brave urchin]

Thank you for this as well! Very much appreciated!

[brave urchin]

Ahhh, okay, I see, so if I run into that error again on another addon node, it would be worthwhile to search through their main.js, add in that relative pathing. That issue makes a lot more sense to me now, thank you!

[dim sonnet]

Yep! If you fix yourself, please also send a PR so it's fixed for everyone else too

[brave urchin]

🫡 Sounds good!

[brave urchin]

I've found a few other nodes that I'm used to working with, that also don't work for stable swarm. Unfortunately these seem to have a different issue, something with where they get their image preview from I think? I tried adding relative pathing to their "/view?filename=$" string, but that didn't end up working. Likely, I just don't have enough experience with JS to really understand what's going on lmfao.

https://github.com/pythongosssss/ComfyUI-Custom-Scripts/tree/main/web/js
The two in here are "imageFeed.js" and "showImageOnMenu.js" Their intended effect is to show previews of the generated images within the UI, however with stable swarm they come back with a broken image icon.

https://github.com/chrisgoringe/cg-image-picker
I'm not too sure what the issue with this one is to be honest. The intended effect is to pause the workflow until I select an image to pass through, then continue on. However, when I use it with stableswarm, what happens is the UI locks up and glitches, and then the whole server needs to be restarted in order to clear out the queue. I'm also unable to choose an image that will continue on to the next step in the workflow

[dim sonnet]

https://github.com/pythongosssss/ComfyUI-Custom-Scripts/pull/39 <-- that first repo was already fixed, might've regressed on a few specific nodes with unusual requirements I suppose. Weird that it hasn't been updated in months but i can literally just go DM pythongosssss and ask lol

[brave urchin]

Could it by chance have anything to do with me having a custom backend directory?

[dim sonnet]

I would expect that a path of "./view? ie literally put a . in front should be the correct way to relativize that path

the other one is probably the same, it has /view? paths that probably also just need a . in front

[brave urchin]

Yeah, that worked perfectly. I previously had put the dot in front of the ' without noticing

[dim sonnet]

yay

(don't forget to PR the fix!)

[brave urchin]

🫡 💪🏼

[nimble belfry]

Whenever you go back to working on the inpainting I would like to make some feature requests. First please let me close it out without refreshing. 🤣 can you make it auto add the mask when you open it? I cannot tell you how many times I started to generate white blobs because I would forget to click that. Lastly, I think I mentioned it before, is there an actual reason it's not automatically set to everything you don't paint it kept?

[dim sonnet]

First please let me close it out without refreshing no, suffer

can you make it auto add the mask when you open it? that would make sense as a default ye

is there an actual reason it's not automatically set to everything you don't paint it kept? i do not understand this sentence

[nimble belfry]

I found where I asked it before, I did not remember you had answered. Its best you did not understand what I said. That way you were not annoyed you had to answer twice.
Just reference when I asked in a more coherent way. #🐝｜swarm-ui message

[brave urchin]

Well, after trying to fix the autocompletion in https://github.com/pythongosssss/ComfyUI-Custom-Scripts for comfyui, I've noticed that significantly more stuff doesn't work lmfao. Here's a comparison (the one with all the errors is in stable swarm, the one with only two errors is in comfy standalone)

Looks like I found something that'll keep me occupied for awhile lmfao. Man I wish I didn't hate coding lmao. At least I wont be bored for a couple nights Just gonna run down the list and start doing PRs after I fix em

[brave urchin]

Well, after fixing everything else, I unfortunately came up empty with these two. Just can't figure out how to make em compatable. The wd14tagger node and its model downloads do work like they're supposed to, so I don't know why it's throwing errors, even after adding the correct pathing.
At a complete loss with the autocompleter though. Also, I noticed he's using " where everyone else is using ' so I don't know what's up with that either. Tried replacing those as well, but to no avail. One of these two also come up with the warning that "clipboard-write" is an unsupported feature.

I don't enjoy JS enough to delve deeper into these lol, so hopefully one of yall who do work with that language regularly could figure something up if you're feeling frisky.
https://github.com/pythongosssss/ComfyUI-Custom-Scripts/blob/main/web/js/autocompleter.js
https://github.com/pythongosssss/ComfyUI-WD14-Tagger/blob/main/js/wd14tagger.js

Here's the ones that I fixed though, so if anyone else uses these and are impatient like me, they're can reflect the changes into their local files before the pull requests are accepted lmao:
https://github.com/sipherxyz/comfyui-art-venture/pull/7
https://github.com/rgthree/rgthree-comfy/pull/97
https://github.com/chrisgoringe/cg-custom-core/pull/2
https://github.com/Feidorian/feidorian-ComfyNodes/pull/2
https://github.com/pythongosssss/ComfyUI-Custom-Scripts/pull/149

Also, just looked through all of those again and feidorian accepted the pull, then reverted it, so it's worth noting something. What that something is, I couldn't tell ya lmfao. I just know everything works on my end

[dapper garden]

Hey everyone, just a starting question before provisioning my GPU Server. Would It be better to go with 4 16 GB Cards or one 48 GB Card? Iirk Stableswarm only uses multi gpu Setups for batch processing, right?

[dim sonnet]

Iirk Stableswarm only uses multi gpu Setups for batch processing, right? correct

Would It be better to go with 4 16 GB Cards or one 48 GB Card? depends on the specifics of what cards. The most valuable things is being current gen nvidia (ie 30xx or 40xx series, aka a100/h100 if using datacenter equipment)

[dapper garden]

I can choose between 4xA4000 16GB, 1x 3090s 1xA6000 48GB Variant. I want to look into training models. this is my first time really diving deep into SD. Everything will be powered by 32GB DDR5 and 8 Core Epyc Server CPUs

I guess the 48 Gig Card would be my best bet

[dim sonnet]

nvidia's naming is annoying but i think those are both RTX 30xx series

[dapper garden]

1 mom

[dim sonnet]

being same gen means probably if you're going to be generating more than 1 image at a time the 4x a4000 will be be better

but the a6000 will be better for single image at a time, and also useful if you want to do training with it

[dapper garden]

https://technical.city/en/video/GeForce-RTX-3090-vs-Quadro-RTX-A6000

Both are ampere cards

I would want to look into Batch Generation of 20 Images, 50 steps.... etc. and slowly move on to training my own models. I can´t really see anything wrong with the A6000 then or is there something I missed?

[dim sonnet]

if you're wanting to do your own training, definitely the a6000 is the better option

[hearty notch]

@dim sonnet Hi and happy new year, quick one when I try to create a new preset and I click on the button for that form the IU nothing is happening

actaully never mind, just reloaded the UI now it's workingh

[dim sonnet]

Thanks to comfy adding the requested API route to unload, there's now a "ClearVRAMAfterMinutes" server setting

system RAM is available too but atm it doesn't quite work right, @cunning holly says PyTorch is holding a lot of memory in a cache that he can't clear

(also the Server Info tab has buttons for it, and there's an API route FreeBackendMemory)

[nimble belfry]

if you clear vram does that mean all models will unload? so it will take longer when the next request is made, but in the meantime it would free up vram?

[cunning holly]

it unloads models from vram to regular ram

on most desktop machines it shouldn't make the next request take much longer

maybe 1 second and that depends if you have loras and other stuff like that

[nimble belfry]

oh ok i understand now. Maybe i can finally run an LLM at the same time now?

[dim sonnet]

yes

hit the API route FreeBackendMemory to unload the image model, give it a second, then load the text model

note that text models are way less friendly to quick unload+reload, i looked into implementing and it's basically a no-go cause the underlying source code is hacked up mess

my bot already has most of this done https://github.com/mcmonkey4eva/SimpleDiscordAIBot except it uses a less efficient way to unload comfy's memory

[open badger]

getting a lot of this using the colab since the update: 15:29:23.850 [Warning] [BackendHandler] No backends are available! Cannot generate anything.
15:29:23.850 [Error] [BackendHandler] Backend request #10 failed: System.InvalidOperationException: No backends available!

[dim sonnet]

Might just need a restart or reinstall?

[open badger]

I went for the reinstall because it happens during generations, not during startup. It finds the backend just fine. It seems to time out. Hopefully the clean install works.

[potent panther]

would it be possible to have the other buttons, have the same rounded corners as the did you know?

it's just so much better than this

[last tinsel]

How do I close the edit image frame? It stays open even when I choose reset page layout. Thanks! (Firefox Win11)

[potent panther]

you reload the page

it's not finished yet

[potent panther]

I thought of this like tabs for settings and each tab can have separate settings and gallery
But the top bar and bottom library thing stay static
I forgot to add a cross to close tabs but I think you get the point

It's just an idea don't take it seriously

[last tinsel]

Thanks! The tabs sound cool. Looking forward to the updates!

[potent panther]

It’s just a random idea

[dim sonnet]

huh? They all have rounded corners

the generate button is a tri-button (3 buttons in one spot)

[dim sonnet]

oo

[potent panther]

I mean less rounded like the did you know buttons

[dim sonnet]

oo

less rounded okay i see

[potent panther]

Yeah also make the dropdown bigger since you can generate by hitting enter

Clicking on a big rectangular button is easier than a tiny round button

[dim sonnet]

you can right click on Generate to get the same dropdown btw

i added the arrow button to make it more obvious

[potent panther]

Oh I thought you separated the button in two

[sacred mason]

does stable swarm UI support multiple visitors to the same stable swarm UI URL (non collaborative / not "multiplayer" but ordinary multi-user web application style), multiple comfyui backends, and then it correctly distributes workflow requests to backends?

does its Comfy Workflow Editor show the comfyui node interface? isn't it going to be from one frontend?

[dim sonnet]

Yep! That's supported out-of-the-box in general, automatically splits jobs across backends and tries to balance them (ie not block out one user cause another is using up the server, optimizes to avoid wasting time reloading models too often, etc). atm by default will share one "user account" (ie all image history / etc goes to the same folder), but support for multiple users is built in just not easily accessible through normal UI yet.

[dim sonnet]

sorta yes, but, there's a button about that:

by default if you have more backends than concurrent users, it will be "Use One" ie each user sent to a different one, but jobs can end up overlapping

"Reserve Exclusive" is like "Use One" but it forbids other users from taking it or queuing jobs onto it

"Use All" will automatically send jobs to the first free backend (this can cause time waste loading models atm as it doesn't prevent that effectively enough yet)

("Use All" also allows you to have one comfy graph hit several backends by color-coding the output nodes)

[sacred mason]

give me a sense of how to configure stable swarm ui to discover pre-existing comfyui backends. I am a sophisticated user. i'm not asking how to deploy it, just where i can discover the config

[dim sonnet]

Server -> Backends

if you want literal config files Data/Backends.fds

(config file can't edit anything the UI can't)

[sacred mason]

does it expect a list of hosts (dns or ip addresses); a dns+srv; or something that gives it a list of IPs?

"Use All" will automatically send jobs to the first free backend (this can cause time waste loading models atm as it doesn't prevent that effectively enough yet)
this is fine, i think it will all be the same

i am going to deploy this all on kubernetes

[dim sonnet]

you can list backends and each backend can have an IP

[sacred mason]

it is a pretty small lift to periodically update the backends configuration to match what is in a comfyui deployment

with the raw IPs (aka a headless service with a list of IPs)

[dim sonnet]

If you're automating deployment you would want to have the remote server be Swarm instances, not comfy, and you can use Swarm's API to automatically add/remove backends when they change via simple JSON POST

(to learn how to format the calls, just open UI + open browser console, add a backend via the UI, and look at the network traffic - the UI uses its own API to make it super simple)

relevant sample code of using swarm API just in general if it helps https://github.com/mcmonkey4eva/SimpleDiscordAIBot/blob/master/Program.cs#L74-L158

[sacred mason]

you would want to have the remote server be Swarm instances, not comfy, and you can use Swarm's API to automatically add/remove backends when they change via simple JSON POST
i guess because i am actually interested in running specific comfyui workflows, do i still want swarm instances?

what is a swarm instance exactly?

if you're familiar with kubernetes, i was planning to do two deployments - one for N stableswarmuis, and another for M comfyuis, and configure stableswarmui to use the M backends.

presumably they are aware of each other's jobs. it doesn't matter though, i would be happy with N=1, as long as multiple users can visit

1 stableswarmui capable of serving however many users, with a bunch of comfyui backends

i am already experienced in running comfyui in kubernetes, but it's more of a question of whether i need to make a bunch of comfyuis each individually visitable, or use stableswarmui as a web GUI instead

[dim sonnet]

a running instance of StableSwarmUI. When doing remote network connections (ie not all on one machine), it's best to connect swarm to swarm's on that machine, and then that swarm instance is itself running a local comfy backend. This ensures swarm properly automatically manages things like model paths and etc., as well as making swarm's nodeset available (swarm adds a bunch of custom nodes to make life easier for general usage)

[sacred mason]

okay, so i can do

frontend swarm deployment
 -> backend swarm deployment
    pod:
     swarmui container configured to use comfyui container
     comfyui container

[dim sonnet]

rather, swarmui container

you don't need to worry about/manage the comfy bit, just consider it an inclusion within the swarm instance

[sacred mason]

i maintain a comfyui fork that fixes all sorts of bugs, and is way easier to install and package (https://github.com/hiddenswitch/ComfyUI), i have custom nodes and models that i want to load

that i don't wanna do from the UI. i kind of need it all to just be there and ready to go without any installation

custom nodes that i will not install via some jank community method

but with a package

[dim sonnet]

0.o a fork

that's, really weird. Can still use that tho

as long as it's connected to swarm via Comfy Self Starting it gets all of swarm's management stuff

[sacred mason]

comfy self starting is, swarm starts it or I start it?

[dim sonnet]

swarm starts it

[sacred mason]

how do frontend swarms get informed about backend swarms? same thing, a bunch of ip addresses?

and do the frontends transfer the models to the backends, or is it expected the backends have all the models

[dim sonnet]

can either prebake a list of backends into the main front one, one per backend, and just enable Allow Idle if the IP is static but won't always be on.
-or- as said earlier you can hit API calls to add em dynamically

[dim sonnet]

it's expected backends will have models available

i plan in the future to add automated model transfer but it will probably not be a great option (sending models over networking tends to be slow even in ideal networks)

[sacred mason]

no it's good there's no model transfer

the dumber the better

[sacred mason]

how hard is it to get the same benefits with "remote" comfyui backends as "Swarm starts it" backends

[dim sonnet]

not terribly hard but you gain nothing by doing it that way, you only add complications for yourself

[sacred mason]

did you author comfyui nodes that correspond to the "easy" ui inputs and outputs?

[dim sonnet]

yes, also add a variety of secondary features

eg recently had a user that was building through vanilla comfy swap to using swarm just for the nodes after discovering that vanilla comfy mishandles seeds in batched generations and swarm doesn't (comfy himself is aware of this and i even made a PR but he chooses to leave the weird behavior)

[sacred mason]

well take a careful look at my fork especially the updated readme here - https://github.com/hiddenswitch/ComfyUI#manual-install-windows-linux-macos - and installing and authoring custom nodes here - https://github.com/hiddenswitch/ComfyUI#authoring-custom-nodes

there's a lot of jank from the community reinventing pip

i'm planning on adding a facade for civitai and huggingface to make the models and datasets compatible with plain old pip. it ought to also support plain old http proxies.

i wish comfyui didn't reinvent job queues

[sacred mason]

i really appreciate your help. i'll try running stable swarm ui in the way you suggest first, and see what happens

ingress for frontend
frontend deployment of swarmui configured to use backend (~1 replica)
 sidecar container which periodically populates the frontend configuration with the
 headless service IPs of the backend deployment
backend deployment of swarmui (many replicas, host anti-affinity)
 comfyui self start

[potent panther]

probably afraid of breaking workflows

[cold flame]

Would it be possible to get an option to specify the path to the venv we want to use for ComfyUI Self-Starting? I use the venv from my A1111 install to run Comfy.

[potent panther]

the stable society spoke

[dim sonnet]

I believe if you just activate the venv before launching swarm (or in swarm's launch script) it will implicitly apply

[cold flame]

What I ended up doing (thanks to Ashtaka's suggestion) is just symlinking the venv so that it's present in the comfy directory too and it works as expected now.

[subtle sand]

how do I do regional prompting in swarmui?

[dim sonnet]

global prompt here <region:x,y,width,height> region prompt here
eg a photo of a catdog <region:0,0,1,0.5> a photo of a cat <region:0,0.5,1,0.5> a photo of a dog
(actual graphical UI for this is planned)

you can also do object instead of region to have it inpaint the object onto the image separately and refine it with regional prompting, or <segment:face> pretty face to do automatic regional inpainting in post (wherein the input is a text matcher for what to paint over)

[subtle sand]

oh thanks. they should add it on https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Basic Usage.md#more-than-text-playing-with-prompts

[sacred mason]

so before i start the stable swarm ui journey, to confirm, if i do

stableswarmui frontends configured
to use a list of stableswarmui backends
 --> backends each using their own instance of comfyui

that i'd observe the following

0. i boot up two frontends, two backends.
1. frontend 1 submits a job A
2. frontend 2 sees a queue of 1 job. frontend 2 submits job B.
3. both frontends should a queue of length 2, job A and B are executing in parallel.
4. both frontends are receiving progress notifications for their
   respective in flight jobs.
4. job A finishes. frontend 1 and frontend 2 both only see job B on the queue.

is this the behavior? this would be fully distributed, multi-consumer multi-producer queue.

[dim sonnet]

if you use multiple frontends they will not be aware of each other

you don't need multiple frontends unless you have a massive expected userbase

swarm is written in efficient multithreaded C# and has been tested with thousands of concurrent generations

even managed to melt a redis-based backend by having swarm submitting jobs to it faster than redis could process, which was pretty fun

[dim sonnet]

(they will work and coexist relatively happily though, just not fully aware of the results of each other's behaviors. Swarm is generally designed for one swarm frontend orchestrating the rest)

[raw laurel]

Why this preset isn't loaded as default? I'm now expanding the ui for a family member and getting it to load every time is much more critical

[dim sonnet]

that looks correct to me - if you're just setting it up, you're probably seeing your recent data overriding default - click Quick Tools -> Reset Params to Default

[raw laurel]

That works, but reloading the page doesn't return all of the defaults. Eg. refiner is left unselected

Is it because the refiner tab (without toggle option) wasn't open in the preset? edit: no it doesn't make difference

[dim sonnet]

ooh, oops, I see why - fixed

[raw laurel]

Thamk 🙂

[dim sonnet]

git pull and reset to default and it should work

[sacred mason]

The front end UI single instance can serve multiple users but show the status of each other’s queues though, correct?

[dim sonnet]

does serve multiple users, does not visually share the queue atm

does internally share the queue

[potent panther]

mcmonkey would it be possible to have the "Upscale 2x" button not care about image amount and batch size?

or have a setting to choose its behavior?

[potent panther]

now that I think about it also having a way to customize the settings of the upscale could be nice

[dim sonnet]

that's a silly button from day 1 of swarm that's just sitting there still despite being very outdated

needs to be either nuked or rebuilt

also! update!

swarm is multilingual

also i published a bulk translation tool for anyone that wants to mass translate stuff the way i did for swarm https://github.com/mcmonkeyprojects/translate-tool

[quick flame]

Hi, does swarmui has batch inpainting?

[dim sonnet]

Yes sortof depending on what you mean.
You might want Tools -> Image Edit Batcher

[quick flame]

I usually work with masks I make in other programs so if it can use the inpaint batch upload option from a1111, that's what I'm talking about.

[primal grotto]

Hello all. Im trying to undestand how I can link my Comfy UI as a a backend. I am able to link it and see my Comfy UI editor in StableSwarm, but it doesnt see my models that I have in my comfy directory. Do i need copy the files to Swarm models folder or can i link it via a path?

[last tinsel]

You can link to different folders in the options

[potent panther]

How should I go about fixing the French translations?

Also I think it would be better to use Emojis instead of the images

🇺🇸🇨🇳🇯🇵🇩🇪🇪🇸🇫🇷🇮🇹🇳🇱🇵🇹🇷🇺🇸🇪🇻🇳

[raw laurel]

Your password should contain numbers, small characters and special letters too

[potent panther]

You didn't install the emojis font did you?

[raw laurel]

Oh, no? What's that?

[potent panther]

I mean these are flags emojis

[raw laurel]

lol I guess that's one good reason not to use emojis!

[potent panther]

here's what they look like

if you don't have google-noto-emoji or something similar that's on you

[raw laurel]

boo 😦 I wish this channel was on IRC

[potent panther]

hell no

Windows comes with emojis preinstalled
MacOS does too
Most Linux distro also do

this should not be an issue except for experienced users that use stuff like Arch

[raw laurel]

I'm on W11 and didn't see those emojis

[potent panther]

🤔

that's odd they should show up just fine

Oh no

Windows 11 doesn't have countries flag

that's every flags in the microsoft fluent emojis

This is the only emoji pack that doesn't have the US flag

[primal grotto]

thank you i fugured it out with this info,

[dim sonnet]

^ this is why I used images. This is also why Discord uses images for emojis normally lol

generally can't trust unicode emojis will actually show up on most systems

[dim sonnet]

make PR and edit the file.

AI is used for the initial generation but won't override edits later

[dim sonnet]

i had to avoid certain pre-emoji-era unicode basic symbols as Ubuntu didn't even contain that full range by default

[potent panther]

That’s odd

[kindred mango]

Hi, are there any plans to integrate kohya's fine tuning extensions? Does anyone know if any community members have made an extension for that?

[dim sonnet]

what now

if you mean kohya's model variants eg LLLite i believe those are natively supported

[kindred mango]

Oh, I mean I want to give the UI images and captions to fine tune my own LoRA/models using kohya scripts

Any recommendations on how to do that on stableswarmUI?

[dim sonnet]

oo

I've been wanting to do a proper training interface but haven't gotten around to it yet

Derrian's UI is pretty nice in the meantime https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

[kindred mango]

Ooh. Thanks!

If you were to integrate a training interface on stableswarmUI, how would you go about it? Add it as an extension? Is there good documentation for writing my own extension?

[dim sonnet]

It would be a BuiltInExtension yeah

https://github.com/Stability-AI/StableSwarmUI/tree/master/src/BuiltinExtensions

Most features are BuiltInExtensions - these work the same as extensions they're just, yknow, built in

so you can use em as a reference to make your own extensions

make the same setup as any of those and store it in a folder named Extensions

proper docs about it not yet written

when it's time to make the trainer I want to integrate it through Comfy - it'd be really cool to have it as a dynamically sent in workflow that you can hack around with if you want to customize the finer details

and also just Comfy Trainer ™️ sounds cool

[kindred mango]

Cool!
Alright, docs for this would be super super helpful for the community to create their own extensions!
I don't know if you saw there's already a comfy UI node for fine-tuning Loras by larryjane491. Given that, would the simplest thing be to have the utility call into this comfy UI node? The comfy UI is quite janky right now, but it's usable. Just needs a better UX

[dim sonnet]

I saw a node somebody made but it was literally a single node that just calls externally to kohya or something

that's not a real "comfy trainer" that's just somebody shoved a node into a place

[kindred mango]

I see. What kind of inputs/outputs do you want in this "comfy trainer" you're envisioning?

[kindred mango]

I might end up building this extension for the swarm UI and it would be helpful to get your perspective 😀

[dim sonnet]

For example there'd be a "DatasetFolderLoader" node, where its inputs are just a file path and whatever
Or you could swap it for a "DatasetZipLoader" that loads from a zip or whatever.

There'd maybe be a "DatasetCaptionLoader" node that you connect it to or a "DatasetCaptionGenerator" that uses clip promptgen or whatever.

There'd be other core inputs like scheduler nodes and etc. (Training schedulers are more complex and replaceable than inference ones so are best in their own nodes)

That'd be connected into the "KSampler" node equivalent, "Trainer" or whatever it's called, which has a bunch of inputs for dataset, sampler, etc. and obviously a model connect in that ideally uses just the default comfy model loader (and thus is also compatible with alternate model loaders, loras, etc).

And then at the end it outputs a standard comfy model, that you can immediately use for inference in the same workflow, or send to a "SaveCheckpoint" to actually save the result. Or, you could feed that model into another trainer with different parameters to get some form of complex multipass training done all in one flow.

There'd probably be a LoraTrainer that looks just like regular trainer but naturally trains loras instead

The Swarm side would work much like the inference workflow generator, just naturally it generates training workflows instead

The most important things on the swarm side imo are:

• Good defaults. If you just open the trainer tab, select "Lora", give it a path to your files, and hit train, it should make something decent without any other setting being edited. This includes for example defaults automatically swapping based on if you're using XL or SDv1 (the same way for example Resolution on the main tab swaps between 512 and 1024 depending on model class)
• Easy setup and intelligent interface behavior. Something that annoys me in kohya for example is the way captions have to be their own separate text file alongside the original file. In the old auto webui dreambooth extension, you just renamed the image files to your caption - it was a lot easier imo. Swarm should let you input your data in whatever format you prefer and just work with it, ideally with autodetection (ie search for .caption or .txt and use those if found, if not use filename). All settings should be simple/automated in this way. And of course be clearly documented on-page the way swarm's generate parameters are with the lil ? infobubbles. The most important parameters displayed by default and advanced/hacky ones under a 'display advanced' checkbox. Make it approachable for noobs while still powerful for pro users.
• Integrated together as one powertool - in Kohyaland, lora training and base model training are fully distinct. In the comfy trainer idea, they share most things but have a split core. In Swarm's friendly user interface, it should literally just be a Select Type and all parameters that can be the same remain the same, and only unique ones move (eg when you select LoRA, a "Rank" parameter appears, that base training doesn't have naturally).
• Also naturally it should have nice convenient things like an interrupt button, previews as it goes, loss graph, etc. I'd love if the interrupt button automatically saved a checkpoint backup when it interrupts and made it available to continue off of that easily - so you can just like pause, change some settings or swap the dataset, and resume, super easily.

[knotty beacon]

Hello, everyone. Can you please tell me if there is anywhere a tutorial or in-depth documentation on how to configure a remote GPU, in my case on RunPod (because I have some credits loaded there), for StableSwarmUI? I read in the official documentation that it is possible but it is not clear to me how. Do I have to start a Runpod Desktop, install StableSwarmUI in it, and then connect what to what? Thanks for the support, this tool is a blast but I'm a newbie! 🙏

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md#step-by-step-setup-guide-for-multiple-machines

[visual magnet]

https://github.com/sm00th/bitlbee-discord

https://cdn.discordapp.com/emojis/1151451446630432828.webp?size=48&name=the_fool~2&quality=lossless

[untold cairn]

How do I close StableSwarmUI image editor?

[potent panther]

You refresh the page

[untold cairn]

How to load image using ComfyUI workflow in Generation tab? Absolute path to image is not working.

[Error] Internal error processing T2I request: Newtonsoft.Json.JsonReaderException: Error reading JObject from JsonReader. Path '', line 0, position 0.
   at Newtonsoft.Json.Linq.JObject.Load(JsonReader reader, JsonLoadSettings settings)
   at Newtonsoft.Json.Linq.JObject.Parse(String json, JsonLoadSettings settings)
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUIAPIAbstractBackend.GenerateLive(T2IParamInput user_input, String batchId, Action`1 takeOutput) in E:\imagen\Stable SwarmUI\src\BuiltinExtensions\ComfyUIBackend\ComfyUIAPIAbstractBackend.cs:line 640
   at StableSwarmUI.Text2Image.T2IEngine.CreateImageTask(T2IParamInput user_input, String batchId, GenClaim claim, Action`1 output, Action`1 setError, Boolean isWS, Single backendTimeoutMin, Action`2 saveImages, Boolean canCallTools) in E:\imagen\Stable SwarmUI\src\Text2Image\T2IEngine.cs:line 235```

[potent panther]

have you tried removing the "

[untold cairn]

Yep. Also tried filename without path and tried disabling "upload" option.

[potent panther]

sounds like a bug

[untold cairn]

For now, using custom ComfyUI workflows in StableSwarmUI looks kinda janky. I experimented for a little, and mostly got various errors.
Sad. I was hoping to use grid generator with custom workflows.

I also hoped using "Save preset" function with Automatic1111 as backend, but failed to connect it.

[potent panther]

janky*

[untold cairn]

thx

[dim sonnet]

er, atm for that you'd have to store the images in the comfyui input folder

bulk alternate images is not a case that's super clean atm. There's the Image Edit Batcher tool that lets you select an arbitrary folder if you just want to do a bulk edit.

[dim sonnet]

It shooould work but ngl auto webui backend is not really worth using unless you're forced into it. The comfy backend is just better on all counts

[dim sonnet]

if you get errors (aside from the wonky input image handling, I know, that's on the todolist) please report the specifics

[untold cairn]

@dim sonnet 👍

[untold cairn]

Most obvious errors were:

• ComfyUI's "Load" button works, but SSUI's "Load workflow" and "Quick Load" don't.
• I press "Use This Workflow In Generate Tab" and got new params in Generate tab. Then I change workflow, press "Remove Workflow From Generate Tab" and "Use This Workflow In Generate Tab" again — but only got params from previous workflow, not new ones.

Also, it's unclear how to use this function. Should I check "[ComfyUI] Custom Workflow" option? Will original SSUI params be ignored?
What "upload" option do in "LoadImage"?

[dim sonnet]

ComfyUI's "Load" button works, but SSUI's "Load workflow" and "Quick Load" don't. huh? If unclear, swarm's workflow saving is different from comfy's workflow saving, that's why the buttons are different. Comfy's save/load is json files, swarm's saves on the server - so to use the swarm quickload you gotta use the swarm save

but only got params from previous workflow, not new ones. that... doesn't make sense 0.o

might check browser console for an error message

Should I check "[ComfyUI] Custom Workflow" option? no, that's for quickusing workflows saved in Swarm, notably for grids of multiple workflows.

[untold cairn]

I am talking about these buttons:

[dim sonnet]

Will original SSUI params be ignored? any params that you don't assign into the workflow will be hidden from view and ignored, yes. You can assign either by just having a recognizable default input (eg checkpointloader will assign to the standard model param) or by adding primitives with title like SwarmUI: Prompt to indicate a param to attach to

[untold cairn]

Looks like it clearly detects *.json workflow files, just choosing them does nothing.

[dim sonnet]

What "upload" option do in "LoadImage"? nothing, that's just wonk from the weird hack comfyui does to have the "upload" button show up in litegraph (it's a fake parameter on the comfy end) that swarm doesn't properly hide atm

[dim sonnet]

oh god if you saved into the folder manually yeah that's gonna break stuff, don't do that

The file format for saved workflows in that folder is not the same as comfy's default save button

[untold cairn]

Got it 😅

[untold cairn]

Can't replicate it the second time 😕

[untold cairn]

Not sure we are talking about the same thing. Here, I turn on custom workflow in Generation tab, but params like "Seed", "Prompt", "Init Image" etc. are still visible, despite not implemented in workflow.

@dim sonnet also, I just want to say how much I appreciate your support. It's a real bless to talk with developer himself!

[dim sonnet]

That falls under You can assign either by just having a recognizable default input (eg checkpointloader will assign to the standard model param)

If you have LoadImage inputs, the first one gets assigned to InitImage (any other loadimage nodes wind up in the general node list). If you have a KSampler or KSamplerAdvanced, that gets assigned to Seed/Steps/CFG Scale/Sampler/Scheduler

[untold cairn]

But what if I got multiple KSamples? Will Seed/Steps/CFG Scale/Sampler/Scheduler still be visible, assigning to the node I add to the workflow first?

[dim sonnet]

yes

[open badger]

When I use Comfy Workflow Editor, the preview images at bottom of screen, never appear, always a broken image placeholder. Is there some setting I am not using correctly. It works if I run Comfy as a standalone.

[dim sonnet]

uhh the what now? Can you screenshot what you're referring to?

[open badger]

It is probably not a Swarm thing, but it has never worked for me when I use Swarm. Works when I use Comfy as a standalone. Was waiting to see if anyone else had this issue.

[dim sonnet]

oh, that's not vanilla comfy, that's an extension - usually that type of bug is a simple path bug that can easily be fixed

[bronze shale]

Hi stableswarmers! I've been enjoying using it a lot, but have some questions about backend issues. I'm looking to setup another instance of stableswarmUI on my server to let some friends use and I want to be able to connect the backend of my current instance to the new one, but I am having issues. I get errors when trying to connect 127.0.01:7820. I am also not able to connect to running comfyUI or A1111 instances. The only way to do it is with self-starting (which can be a pain). I am assuming it's a .net thing and I'm not familar with it. Thoughts?

[open badger]

Yes it's a function of ComfyUI-Manager, thats why I never brought it up here, but it's so annoying. Was hoping someone else had the same issue and knew a fix.

[dim sonnet]

connect the swarm instance itself as the backend

[bronze shale]

this is what I thought and what I tried. The backend starting port is 7820 and the web access is on 7801, but neither work

[brave urchin]

Those extensions (the image feed at the bottom and in the comfy ui panel on the right) are both from a pythongosssss extension (I use these as well). Unfortunately his code is really weird and hard to bugfix for with stable swarm, however it seems as though the pull request I submitted that had fixed the paths for that has somehow been reverted again. Lemme see if I can get it working

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md#step-by-step-setup-guide-for-multiple-machines

[bronze shale]

Oh this is so strange. It seems to be working now, but I had to enter that EXACT string in the other backend config. Just setting 127.0.01:7801 did nothing lol

thanks!

[brave urchin]

Ope, nevermind. The changes are still live in the most recent repo so I don't know why it's doing that again. I've since moved to comfy standalone, but if you mess with lines 1 through 3 & 4 and fix the pathing for the files "web/js/imageFeed.js" and "web/js/showImageOnMenu.js" (https://github.com/pythongosssss/ComfyUI-Custom-Scripts/tree/main/web/js) you may be able to make it work again.
However there's a bunch more stuff from that extension that also doesn't work with stable swarm (text prediction and autofill just to name one), without seemingly refactoring a bunch of stuff (or it could be a simple fix as well. I'm NOT good with js, and his code looks different from everyone elses)

[open badger]

I have not updated my ComfyUI-Custom-Scripts, I will do that and see if that works. Thanks so much for pointing me in the right direction.

[brave urchin]

Ahh, that could be it then. I believe the change went live a couple weeks ago so if you haven't updated since then, it ought to fix it. You're welcome!

[potent quarry]

Where can I find the highres fix for upscaling in StableSwarm?

[open badger]

Thanks again. That was the fix.

[potent quarry]

It looks like it's just the refiner filling the same purpose? I'm not getting particularly sharp lines on illustration-style images.

[dim sonnet]

Yes, Refiner in swarm == Highres Fix in auto webui

(with extra options tacked on naturally)

[brave urchin]

Good deal! I'm glad to hear that it worked out!

[untold cairn]

I can see how this might be confusing. I have a workflow with 13 KSamplers in it, and no idea which one was first. But it is an edge case scenario, I guess.

[potent panther]

@thorn harness @uneven moth

[uneven moth]

Thanks for the ping!

[untold cairn]

@dim sonnet may I leave a tiny-tiny feature request? For Grid Generator to support folder names with substitutions, like grid-[date]-[time]. It's way to easy to accidently override an old grid with a new grid.

[dim sonnet]

Good idea - added

[potent panther]

@thorn harness @uneven moth @leaden oyster

👌

[hearty notch]

Hi @dim sonnet is it possible to add time intervals control to the generate forever? Like having the option to have a break of selected number of seconds between generations, to let the GPU take a breather between generations and avoid overheating

[dim sonnet]

how are you gonna heat your house in the winter that way

[dim sonnet]

added, GenerateForeverDelay in UserSettings

[hearty notch]

Awesome! I will have a look

[hearty notch]

I am using 4090 with what I think a 750w PSU, having the fear to heat it my building instead of my house 😂🤣

[bronze shale]

I feel ya. Dual mi100 cards running close to 90. It's toasty in here

[abstract rapids]

for just face closeups or upper body hi res works perfect, else it magnifies defects in the generation -- full body shot in particular. I have to turn off and resort to comfy workflow for it with adetailer.

[dim sonnet]

use <segment:face> beautiful face in your prompt to do the equivalent of adetailer directly in swarm

can match any part you want ('face' to be identical to adetailer which only matches faces) and then whatever detail prompt you want

[abstract rapids]

Thank you, will do that. Is there a way to delete multiple pictures in the picture grid like pressing cntrl selected files on windows explorer and delete them ?

[dim sonnet]

not currently, other than going to open the actual output folder and doing it there

but that's a valid feature request, will note down

[untold cairn]

To be exact, it's probably not fully equivalent. <segment:face> uses CLIPSeg to detect face, right? While @abstract rapids probably uses UltralyticsDetectorProvider from Impact Pack for ComfyUI.

[dim sonnet]

not exact seed-to-seed identical results that way no but equivalent concept with roughly equivalent results

classic adetailer uses yolov8 for face detection

[potent panther]

on what is clipseg running on? cpu or gpu?

[raw laurel]

@dim sonnet would it be possible to add break into generetate forever loop if the generation took less than 500ms or so? that would eliminate generating dozens of identical images out of cache

and accompanied with a warning that the generate forever is generating from cache (and suggestion to use image count if this is intentional)

((or maybe you have a more sophisticated method of knowing the image output was from cache and no time limit is needed to detect such case))

[dim sonnet]

uhh yes one of those two devices yes :)

probably cpu

[dim sonnet]

ooo i have a thing for that in Generate Previews but not Generate Forever

in previews it only queues a new gen if you change your inputs

gen forever can have the same (and an exception for seed=-1)

wouldn't necessarily account for custom comfy workflows tho

[raw laurel]

yeees thamk you! we had a chat about my experience about this some time ago but as it occured today I had a rethink

[raw laurel]

think that's fine. for me it's so far only when I copy settings from an older image and forget to reset seed to random and click generate forever and come back to too many identical images...

[potent panther]

for some reasons the gpu and resource usage info aren't centered

[dim sonnet]

those cards are a lil shy,,

(fixed)

[dim sonnet]

that's applied now - it'll not make a new gen call if all inputs are unchanged and seed isn't random

[potent panther]

is it normal that the <> options are counted as clip tokens?

[raw laurel]

would it be possible to move the language selection elsewhere? it's not something you change too often (probably once after install) and I'm not too happy at staring at a US flag all day

[raw laurel]

thanks tested, it won't make new gen call but it just gets "stuck", nothing happens, menu suggests "stop generating". if user is not aware why this happens they might be super confused why nothing is happening

[dim sonnet]

yeah sorry it doesn't properly account for them atm. That's gonna be a bit complex to implement properly

[dim sonnet]

yeah yknow i was thinking about that myself. I need to set it up clearly in the installer UI then i can kick it off the main screen

[versed copper]

staring at a german flag as an austrian would be worse, so there's that

[potent panther]

I see

[abstract rapids]

The log window , can we see updates like stability matrix or auto1111. There are no simple views of what's going on..loading models , downloading etc.. all we get are millions of lines in calling port or just one user loaded.

[covert flare]

I see

[native hazel]

/de

/cre

[dim sonnet]

huh?

[potent quarry]

Playing around with a newer SDXL model doing anime/lineart style using StableSwarm instead of A1111 because I did a multi GPU build. I'm not getting the sharp upscaled images that I'm looking for but don't know where to focus my investigation, i.e.

• Is the issue that the original image pre-upscale isn't sharp enough so the "fuzziness" is just being upscaled?
• Is the issue in the refiner/upscaler settings? (I'm using the base for the refiner with a low percentage (.3), upscaling 2x with pixel-bilinear, and have tried with and without hypertiles
• Something else I haven't considered?

[abstract rapids]

sorry for the cryptic message, the options for log window output are many, however in greek. Here is stablity matiix output. tells me what is loading, what failed. simple. If I use comfyui manager and downloading a new model it tells me where it is installing -- etc.

here are swarms-- I don't know what the hell they mean and provide no real info that is usable -- for instance, what model is it loading or crashed on for-- what lora is it searching for it it is not loaded in the model directory. The command window in stablity matrix and auto1111 gives me that so I can correct the problem if there is one. In the gen screen, if a lora is missing -- there is no way to tell if it is as no info is given that I know of, it just keeps on generating images. I copy from civitai, paste -- it generates. However unkown at the time I need xxxxx.pt.

[dim sonnet]

you might try a better upscaler - eg lanczos is a better default one, but you can download trained upscalers https://openmodeldb.info/
or yeah bump the percentage

[scenic jackal]

[dim sonnet]

??

[dim sonnet]

okay slow it down you just listed off 10 separate things there, one at a time

if you need .pt files you are copying things from years ago

if you just copypaste from civit without reading details then yeah you're gonna miss things and there's not often really a way to tell from just copypasted text that there was supposed to be something more

that first screenshot loooks like you have custom nodes that are failing to load

[untold cairn]

Maybe it's model fault? Also, can you share the metadata? It'll help track the problem.

[abstract rapids]

I did not mean to mention separate items each a issue. I simply meant a readable log screen for users.

[mossy sequoia]

how do you install comfyui manager add-on to stableswarm?

[dim sonnet]

same way you do on regular comfy, by default in the Swarm/dlbackend/comfyui folder

[mossy sequoia]

okay, thanks!

[charred geyser]

Why don't you have permission to speak

[potent panther]

#1047610792226340935 this is the StableSwarmUI channel not the discord bot

[charred geyser]

[charred geyser]

Where can I make a picture?

[potent panther]

I don’t know what you mean there’s many ways for it

You should probably in the general channel if you want your question answered

[versed copper]

is there a way to get metadata/images into SwarmUI for models downloaded with Stability Matrix?

short of redownloading everything

[potent panther]

for some reason some of my upscalers aren't detected (yes they are pickletensors because I didn't bother)

I have quite a few more

[dim sonnet]

uhh not sure what format they store it in but probably?

For reference swarm's metadata format is a standard spec that they can and should just use: https://github.com/Stability-AI/ModelSpec

but swarm does also support auto webui format (for compatibility with people who started with auto, naturally)

[versed copper]

this format 😉

SM has a picture and a json with the CP/LoRA name

doesn't help one bit it seems

[dim sonnet]

the picture should just work, that's auto webui format

[versed copper]

nope, no pics

[dim sonnet]

can you screenshot the folder

[versed copper]

the only CPs with pics are the one donloaded by the Swarm installer

[dim sonnet]

iee with model + image visible

so i can see how they're laid out

currently swarm looks for ".jpg", ".png", ".preview.png", ".preview.jpg", ".thumb.jpg", ".thumb.png"

they probably have a like .cm-info.jpg or something

[versed copper]

oh, you probably meant this

jpEg, not jpg

[versed copper]

i put a .jpg copy of the .jpegs in the folder, reloaded Swarm Models tab, no show

[dim sonnet]

oh my god, .preview.jpeg

[dim sonnet]

you'll probably wanna hit this button and then refresh the models list

[versed copper]

have tried that before, will try again!

[dim sonnet]

i added .preview.jpeg to the list, will be parsed once you update

gonna try the jsons too in a sec

[versed copper]

ÝAY

i got previews

yeah, it's the jpeg ending, the model previews now work, the loras don't (wher ei didn't put .jpg copies)

[versed copper]

just run the windows-update.bat?

[dim sonnet]

• added parsing for that json too - just title/description/TrainedWords, and only used on models that don't already have actual metadata available

[dim sonnet]

yep - close swarm, run that, reopen swarm

[versed copper]

lol @ rem incredibly complex script, wow, how do they manage it

[versed copper]

thank you so much - pictures work, lora key words work, descriptions have HTML tags (which was to be expected i guess)

[potent panther]

mcmonkey can you add upscalers here?

[potent panther]

.

[dim sonnet]

atm upscalers isn't even a swarm-managed folder

should probably add that tho

[potent panther]

🤔

[potent quarry]

Sorry for the slow reply, was busy yesterday. Here is the metadata that I used to gen that image.

(Don't mind the model. >.> Aside from its questionable origins, it's kind of wild for generating a whole range of images.)

prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, 1 girl, sfw, long dress, cardigan, holding broom, japanese shrine, negativeprompt: chibi, loli, child, 3d, cgi, render, photo, sketch, realistic, source_pony, score_4, score_5, score_6, bad anatomy, ugly face, images: 2, seed: 1962667544, steps: 28, cfgscale: 6.25, aspectratio: 16:9, width: 1344, height: 768, refinercontrolpercentage: 0.3, refinermethod: PostApply, refinerupscale: 2, refinerupscalemethod: pixel-bilinear, refinerhypertile: 512, model: ponyDiffusionV6XL_v6StartWithThisOne.safetensors, loraweights: 0: .8, , loras: 0: Line Art Style LoRA XL.safetensors, , swarm_version: 0.6.1.0, date: 2024-01-26, generation_time: 2.05 (prep) and 32.89 (gen) seconds,

[potent quarry]

I tried bumping the refiner control percentage from .3 to .6 and it seemed to improve the sharpness of the image but also seemed to generate a lot of... weirdness - a mass of extra fingers, or more extra limbs, or a torso with like five belly buttons in the abdomen.

[dim sonnet]

Might be a case where the quick-n-easy upscaler tools are insufficient, and you'll have to get more involved: generate the base image, then use it as your Init Image, and enable a ControlNet to keep shape, and generate to a larger resolution with that

ps uh

score_9, score_8_up, score_7_up, score_6_up, score_5_up is a really weird thing to put in a prompt?

if you copied that from somewhere else they might be references to TI embeddings that you need to download to replicate it

[potent quarry]

It apparently is a weird quirk of how the model was trained:

But wait, turned out I messed up a bit! What I described above is how PD V5.X used to do things, in V6 I wanted to also be able to say - "hey, give me anything 80% good and up". But score_8 tag would only give us images in range 80* to 90%. Perhaps using both score_8 and score_9 would work but I wanted to verify that, so I changed the labels form simple score_9 to something more verbose like score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up and score_8 toscore_8, score_7_up, score_6_up, score_5_up, score_4_up. In reality I exposed myself to a variation of The Clever Hans effect where the model learned that the whole long string correlates to the "good looking" images, instead of separate parts of it. Unfortunately by the time I realized it, we were way past the mid point of training, so we just rolled with it (I did try to use shorter tags after the discovery but due to the way we train it didn't have as strong of an effect).

So I include the string in the prompt and then use the negative to filter lower scored things back out. It's weird but it works.

[potent quarry]

I'll have to read up on using a ControlNet. That'll be new territory for me.

StableSwarm has been awesome but I've definitely only scratched the surface. I haven't tried ControlNets, ComfyUI, or most anything else hidden under the advanced options.

Thank you so much for making it!

[dim sonnet]

oh, lol, rip, gg. Could've warned them about that - we made the exact same mistake in XL Beta with Aesthetic conditioning

though what's missing from that report above is: sometimes you want "low score" generations. For example with beta's aesthetic cond, you needed to select a low aesthetic value to generate text, since the aes scorer thought text was ugly

[potent quarry]

Do I want ControlNet preprocessing? What do I use for a ControlNet model since the dropdown is presently empty?

I tried looking in the documentation but didn't see anything and the workflow seems different-ish from what I see written about A1111 ControlNets?

[dim sonnet]

You'll probably want this: https://github.com/Fannovel16/comfyui_controlnet_aux#alternative installed into Swarm/dlbackend/comfy, that provides preprocessors - without it only Canny is available by default (I need to set up a convenient installer for that...)

there are good XL controlnet models here: https://huggingface.co/stabilityai/control-lora actual download here https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256 that you can save into swarm/Models/controlnet

(or a larger but messier collection here https://huggingface.co/lllyasviel/sd_control_collection/tree/main)

after you have all that, restart swarm, and then you can just go down here:

when you do that this param will activate

and you can open that to configure more, or just leave it, defaults are fine

and then just have an Init Image uploaded and controlnet activated, and generate and it'll apply