🐝｜swarm-ui

[blissful atlas]

Is stableswarm just a comfyui face lift? if not how does it behave with SDXL models on gpus that don't have more than 4 vram?

[dim sonnet]

Swarm uses comfy on the backend, but does more than just a 'face lift'

performance implications will be the same as comfy

[blissful atlas]

hmmm and how fast is comfy with a sdxl model on a 4 vram card? (rtx 3050)

i currently use webui forge (and i kinda want to switch because it has been ddepricated, and i am thinking of switching back to a1111 or trying out swarm)

[dim sonnet]

forge uses copy/pasted comfy code in its core, so probably same speed as you're used with with forge

[blissful atlas]

a 768x1152 on DPM++ 2M Karras with 30 steps takes around ~1 minute to generate

[blissful atlas]

but doesn't it also use some technique as well?

like

[dim sonnet]

easiest way to know perf for sure is just open up swarm and try it

[nocturne glen]

Can you add support for web image drag & drop for image inputs, such as the ControlNet Image Input? It works for local images, but not web images.

[noble temple]

Would also be cool to have this but for ctrl + v, it works with comfy, just copy the image and ctrl v

[rustic river]

Seems to do the trick. Thanks! 🎉

So, would it be possible for Swarm to not show me a popup for the 300th time, like it knows I interrupted the batch? Or is that a limitation of the underlying tech as well?

[dim sonnet]

Pushed update that hides the error when it was an intentional interrupt

[dim sonnet]

pushed a commit that supports it https://github.com/Stability-AI/StableSwarmUI/commit/cbc1b661e0afdd2009a31732cce3d9cf79466f1f it's a bit cursed as the web is held together by duct tape sometimes, and file dragging handling didn't even get real duct tape it got some knockoff half price duct tape

[dim sonnet]

I'm genuinely surprised to report I have now pushed an update that lets you CTRL+V into the prompt box at least (harder to do for init image cause there's not really a proper way to select that input beforehand)

(surprised because last I looked into it CTRL+V on images seemed to be deeply cursed, but apparently there is in fact a sane way to detect and handle it. Relatively.)

[nocturne glen]

Presets don't have the new samplers like align_your_steps. Not sure if intentional or an oversight.

[rustic river]

Oh, you said Presets.
Well, I got the options there as well.

[dim sonnet]

oo i see what happened

[dim sonnet]

fixed

(what happened was, those schedulers aren't listed in Swarm until a backend has loaded, and the code to update the inputs to recognize the loaded options wasn't updating presets -- so it's not listed if you open the page before backend is loaded, but is there if you open the page after backend has loaded. With the fix applied it's now updates to list it properly same as the regular inputs)

[waxen hinge]

It worked thanks

That's what i was looking for.

[dim sonnet]

now have ^ heavily restricted whitelisted HTML tag support

allows ['a', 'b', 'i', 's', 'span', 'u', 'br', 'br/', 'br /', 'p', 'li', 'ul', 'strong'], which can contain ['href', 'title', 'style'] (and style has a validation/limiter too)

if there's any valid formatting tags that are blocked by it lemme know, can add to whitelist

[versed copper]

brilliant, thank you 🙂

had a quick look, i didn't spot anything except links that doesn't work
but the descriptions got SO much more readable

[versed copper]

<hr /> should be ok

i also found a <br />

[dim sonnet]

links should work? if not gimme specifics

br/hr, will fix

[versed copper]

example <a target="_blank" rel="ugc" href="https://www.deviantart.com/albedobasexl">

[dim sonnet]

ohmygod

target/rel

argh

[versed copper]

argh indeed. afaics, all the links are of that format, and there are quite a few of them

very few [h3] tags also left, but huge font makes no sense in tiny description fields

[dim sonnet]

h3 isn't terrible

it's. huge. but not terrible

... maybe remap it down a step

[versed copper]

<h3 id="heading-34">

not making it easy

<h1 id="heading-2">

[versed copper]

ah, found one more safe one - <em>

<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a6c6269a-2c9e-4c6a-8d5c-
209c1f4f7aa6/width=525/a6c6269a-2c9e-4c6a-8d5c-209c1f4f7aa6.jpeg" />
and those

[dim sonnet]

image has to stay excluded

[versed copper]

yeah, thought so

<em> should be ok though 🙂

[dim sonnet]

for the ids, i can autoremote the id="..." bit and then allow it

and ye em

pushed commit with all the above thus far

[versed copper]

ah --> <ol>

li und ul work

[versed copper]

links work, headlines work 🙂

&

oh, this
link: <a target="_new" rel="ugc" href=...etc>

looks like a typo in html

[dim sonnet]

uhhh that should already work? unless it's in a corrupt text block

_new isn't in spec, but apparently is a thing browsers allow anyway, so i'll add it

[versed copper]

there seems to be a problem with parsing > etc in headlines (very minor thing really)

[dim sonnet]

eh?

gimme specifics

[versed copper]

i had an & in a [h] and now a >

[dim sonnet]

a what

can you just copypaste a full line so i can copy it into my test model's meta and see

[versed copper]

AlbedoBase XL (SFW&NSFW)

[dim sonnet]

it works tho

[versed copper]

not here 🙂
did i miss an update? let me check

[dim sonnet]

pushed current state again just in case

(no relevant change tho)

the only code that escapes & here will only be hit if < is found without a > ie fully invalid HTML so it just goes to the old full escape logic

[versed copper]

ok, i mysteriously still get AlbedoBase XL (SFW&NSFW)

[dim sonnet]

wait is that in not-the-description area?

... no even the Title allows html so that wouldn't be it

is there a stray < lying around?

[dim sonnet]

can you click edit metadata and send me the full description raw text

[versed copper]

<h1 id="heading-123"><strong>AlbedoBase XL </strong>(SFW&amp;NSFW)</h1>

[dim sonnet]

this is cursed, they're using <p></p> for linebreaks

[versed copper]

says so in html for morons 😉

[dim sonnet]

oops, i see why, accidentally extra escape call

fixed now

[versed copper]

<h3 id="heading-206"><strong> <span style="color:rgb(250, 82, 82)">v2.x</span> ---&gt; Main Version | <span style="color:rgb(250, 82, 82)">v3.x</span> ---&gt; Experimental Version</strong></h3>

in this

did you push an update?

[dim sonnet]

yes

[versed copper]

ah, now i got one

AlbedoBase XL (SFW&NSFW)

v2.x ---> Main Version | v3.x ---> Experimental Version

all fixed!

[dim sonnet]

woo

[versed copper]

can i please clone you and put the clone on skyrim modding?

[dim sonnet]

i have looked at skyrim modding before and i promise you that clone would invent the version of teleportation from star trek where it kills you and assembles a replica somewhere else just to end its own suffering

(when LLMs are better and faster I am definitely gonna be involved in building AI-driven RPGs tho)

(or just like, random FPS games but adding cortana-but-she's-real into it for fun)

[versed copper]

not that this isn't already happening in skyrim mods

[dim sonnet]

i've seen the skyrim mods for LLM integration and they're just like... shoddy slow motion ChatGPT conversations with NPCs

[versed copper]

modder friend of mine shoehorned mantella in

[dim sonnet]

i want to walk into whiterun and say "lydia go sell all the shit i made you carry to the merchants" and have her instantly respond "okay asshole at least im done carrying fifty cheese wheels. I want a new thane" and then go over to the merchants area and audibly haggle with the locals for a while

[versed copper]

lol
that's the right spirit

anyway, thank you for going above & beyond
i totally accepted your "it's dangerous" explanation of why no html in cards
but it does look a lot nicer this way 🙂

and so much easier to find trigger phrases for loras, or setting recommendations for models

oh, one more unrelated question: does .8 work as well as 0.8 for lora strength?

[dim sonnet]

pro tip: when you edit metadata of a model, you got these two boxes, which are automatically at the top above description

(and I intend in the future to display those more prominently / maybe even a 1click toggler to automatically add trigger phrases to your prompt for you)

[dim sonnet]

i'm not a fan of writing it that way personally (less clear since the . is tiny) but yes that should work

[versed copper]

with a german keyboard, the . is not on the number block

which makes it all stupid to input

[dim sonnet]

oh, uh -- I never type in the lora weight personally

i just click the weight box and roll my scroll wheel

on Firefox at least that's the most convenient way to edit a number box

not all browsers do that tho

(I need to write some JS to automatically do that, and also do the click-n-drag thingy)

[versed copper]

speaking from the hole i crawled into: oh!

is it possible that mark word, press shift+up to add () is broken? it behaves as if i had caps on, extends selection

[dim sonnet]

it'd be control+up to do that

shift+arrows is move selection

[versed copper]

ok, i need to dig a deeper hole

calendar entry: user complains about windows keyboard working as intended

[dim sonnet]

i did that, now you can mouse wheel scroll number inputs on any browser

[dim sonnet]

++ did the click-n-draggy too

[long swift]

I want to know if this UI is real?

[potent panther]

Yes it is part of StableSwarmUI
Open any image and click the edit image button

[waxen hinge]

Another question, I am accessing swarm comfyUI through <host>/ComfyBackendDirect/ but sometimes i see image mix up from different sessions, accessing the same backend.

3 user's are accessing the same endpoint to run different workflows.

Is there are way to pass session params in this url <host>/ComfyBackendDirect/, so that workflow /output somehow tagged with a session id. So that i don't get the output from other user.

[dim sonnet]

well, A, use the Comfy tab on swarm interface not the raw bypass
B ... use the swarm interface if you want controlled multi-user functionality.
C, some comfy nodes are known to mix things up, eg the PreviewImage node, so don't use those. SwarmSaveImageWS sends an image direct over websocket to guarantee it never gets mixed up. I believe the comfy default SaveImage also avoids mix ups usually unless you have a complex backend setup

[jagged cipher]

Query: I'm trying to make a character in a given pose using open-pose. I made the pose manually, how do I actually import it into controlnet? When I use the OpenPose Preprocessor, it just gives me a black screen.

Or, wait, do I just not run the preprocessor if I already have a controlnet skeleton?

[potent panther]

@dim sonnet what do I do with this?

[dim sonnet]

hahaha I'm not entirely sure why that's even there, but, uh, probably put that back to English? i'm pretty sure that's the example text for negative prompts

might've been included to the translate list accidentally

[potent panther]

okidoki

[dim sonnet]

correct, just set preprocessor to none

[potent panther]

What's OverQueue ?

[dim sonnet]

backend setting - intentionally dispatches more jobs to a backend than it can handle so that it never wastes precious milliseconds with an empty queue waiting on preprocessing/network latency

[potent panther]

I see

[dim sonnet]

let's see if discord lets me yeet a 25MiB gif at it it does hell yeah

i realized i didn't have a convenient way to look at details more closely, so the fullscreen view now lets you scroll to zoom and click+drag to pan

[toxic shale]

@dim sonnet
is there any setting to disable the folder tree from collapsing/resetting on clicking refresh within the gui?

[dim sonnet]

no that's a bug(sorta) i need to fix

(there's no intentional UI-side reason for it to collapse like that, it's just for technical reasons a lot more convenient to collapse it when the source data is refreshed)

[versed copper]

would it be possible to have a filter to show only user-changed settings in the generation left panel?

[dim sonnet]

oo, good idea

[spiral imp]

Is there a way for this to support DPM++ 3M SDE Karras ?

[potent panther]

It should be there afaik

[rustic river]

Shouldn't "model specific enhancements" be included in the metadata of an image?

Another question: I want to compare different models using a grid and a certain prompt, but each model needs some slight variation in the prompt. Is that possible? (Without extraneous prompts in the grid)

Oh wait... Maybe a grid of presets...

[spiral imp]

Restarting server must have updated the list or something D:

[spiral imp]

Is there something like BREAK from A1111 in this?

To reduce bleed between prompt elements (pink hair makes everything pink etc.)

[potent panther]

<break>

[spiral imp]

Nice, there's a bunch of modifiers

[magic ravine]

What are the others ?

[spiral imp]

pressing < in prompt menu shows them

[potent panther]

What's StarNoFolders?

Anyway I'm almost finished

[toxic shale]

@uneven moth needs a moderator

[uneven moth]

Not anymore

Thanks!

[spiral imp]

Does this use xformers by default?

[spiral imp]

Also is it possible to change default 0.1 step for adjust loras?

Don't see that in Parameter Config

[rustic river]

I think it leaves that to the backend, which is ComfyUI by default. I don't think it uses xformers by default. Currently it tells me: 2024-05-05 00:54:12.379 [ComfyUI-0 on port 7821] stderr: Using pytorch cross attention

Someone said:
I've found that Pytorch attention is slightly faster than xformers now anyway. Especially since pytorch 2.3.0

[spiral imp]

I just noticed that I get "WARNING: YOU DID NOT CLONE FROM GIT. THIS WILL BREAK SOME SYSTEMS. PLEASE INSTALL PER THE README." at launch. I used the install.bat from git though.

[dim sonnet]

Karras is the scheduler, auto webui had sampler+scheduler together in one dropdown (they've changed this in a recent update to be separate there too)
so you want:

[dim sonnet]

only if changed from default, defaults aren't listed in metadata usually

[dim sonnet]

yeah presets will work.
My original infinity-grid toolkit let you do YAML files with very complex custom axes for this. That was on auto webui though, which didn't have any convenient preset system. Debatable whether editing a YAML file or presets is the more convenient way to do it, but should both be capable of doing (more or less) the same things.

[dim sonnet]

user setting

(For) Stars, (don't include) Folders

[potent panther]

That should have a clearer name lol

[dim sonnet]

yeah i keep thinking I should add proper names on settings (right now it's literally just rendering the variable name) but haven't bothered thus far (on the assumption that if you're lost, clicking the ? will sufficiently explain)

[potent panther]

That does work

[rustic river]

Woops, it is added, I'm blind.

[dim sonnet]

I pushed a commit that makes the LoRA Weights param under Advanced Model Addons parameter define the min/max/step for that

[potent panther]

no it's just his sampler dropdown didn't refresh after comfy loaded I forgot why that happens

[spiral imp]

It was because ComfyUI hadn't finished loading.

[dim sonnet]

that shows up from launch-windows.bat:

if not exist .git (
    echo "" & echo ""
    echo "WARNING: YOU DID NOT CLONE FROM GIT. THIS WILL BREAK SOME SYSTEMS. PLEASE INSTALL PER THE README."
    echo "" & echo ""

I'm not sure why that'd indicate the .git folder is missing if you used the installer, unless you went and nuked that folder or something

if you downloaded the entire swarm and "ran the install bat", the install bat just refuses and closes itself

you should actually install how the readme says to which is download the single file linked on the readme and then run it

[dim sonnet]

ah yeah.
It's because it reads the list of samplers from comfy, so that if swarm hasn't updated to include new ones, they'll be listed. This naturally can't happen until comfy loads.

I should go update swarm's list to include the new ones

[potent panther]

yeah go ahead :D

[spiral imp]

I moved the install folder. Functionality doesn't seem to be affected though

[dim sonnet]

do you have a .git folder?

if that's missing, it will work for the most part but you won't be able to update

[dim sonnet]

added. This list is really long :(

I wonder if i should like... reorder the common ones to the top, or make the uncommon ones hidden without "Display Advanced", or, something

cause this is a list of silly magic words, where most of the magic words are pointless

ddim/ddpm/heun/heunpp2/dpm_2/dpm_2_ancestral/dpm_fast/lms -> pointless
dpm_adaptive -> interesting concept I guess
_gpu variants -> silly outside of trying to replicate images in specific ways
2s/2m/3m -> all very similar

could easily just be a list of: euler, euler_ancestral, dpmpp_2m, dpmpp_2m_sde, lcm, uni_pc

[rustic river]

Heun pointless?

I think A1111's Restart implementation uses Heun? Also got some nice images with just Heun on certain models.

[dim sonnet]

see and this is why i can't just shove the magic pointless words under an advanced setting for nobody to worry about

[spiral imp]

My preview image sometimes becomes very small D:

top left corner

[dim sonnet]

cause ^ there's at least one person will happily swear by any one random sampler that makes 2 pixels of difference over the next sampler over

[spiral imp]

still works when I click on it

[dim sonnet]

zvxczxcvmzxcvzxcv

what res are you generating at

[spiral imp]

Doesn't matter

sidebar preview works fine

[dim sonnet]

the reason I'm asking there is there's a known bug where under as-yet-unclear-circumstances, you'll get latent-scale images instead of full-scale images

but that doesn't look latent-scale, that looks like an ant

so: is that 8x8 smaller than the image should be, or is the image sane and that's microscopic it

[spiral imp]

and 1024x1024

[dim sonnet]

huh

so latent scale would be 128x128

[spiral imp]

but I've had that happen seemingly randomly with any size images

[dim sonnet]

which... that's way smaller

so... how. What.

do you have any idea when it happens? Like is it when using a certain param, or after clicking things in a certain pattern, or...?

[spiral imp]

for a while I thought that happened when I clicked/switched to image that's still being generated

but restarting tab seemed to fix that

Seems like a browser (Chrome) issue, works normally in Edge

for now

[dim sonnet]

how often does it happen

[spiral imp]

I've only used stableswarm for 2 days

started out randomly, now it's constant

I'll try using Edge, see if it breaks there too

[potent panther]

you could add an option to hide samplers

[dim sonnet]

can you show what parameters you're generating with, and what settings you have on the batch view toggle

the most likely thing to change when you swap browsers is for either an obscure parameter to be forgotten, or the batch view settings to be different

(since those sync in browser local storage)

[spiral imp]

Using same in Edge, still no problems

[dim sonnet]

what about parameters?

these funky things down here

oh also, have you perchance installed TAESD previews to your comfy backend?

[spiral imp]

no TAESD

haven't touched comfy after installing this 😄

[dim sonnet]

jsyk, that aspect there is 9:16, you don't need custom for it

and: does it happen with the initial image, or the refiner step, or both?
does it happen to the first image since you loaded the page, or only after the first, or both, or...?

... also why the refiner on same model with no upscale? 0.o

oh hey hah, i replicated your settings and then refreshed the page and got meself a

latent-scale kitty

not a microscopic one like you were showing, but it's something

okay, update, the reason, at least for the bug i can replicate, is this commit https://github.com/Stability-AI/StableSwarmUI/commit/6c7bbcb5c00a4be02ca1c19269e965b7809e4fdb is misfiring when there's a placeholder image and a sub-1024 width

now the question is... why the hell did i commit that. 8 months ago me, what were you fixing?

pushed a commit to remove it, pending... remembering what its purpose was

so update and you should be good if that was the issue (no idea why yours are microscopic tho)

(also updating will be hard if your .git is missing)

[spiral imp]

I moved the folder back so git works

Tried to get latent-scale girly, seems to be fixed now

[spiral imp]

0.2 refiner does reduce imperfections for portraits

at least in my workflow

jsyk, that aspect there is 9:16, you don't need custom for it
< yeah I was using a SDXL ratio list, it used different ratios for the same resolutions D:

[dim sonnet]

ftr Swarm's aspect ratio list uses the exact aspects&reses that SDXL was trained on

ie most likely to work reliably

[dim sonnet]

ye fair enough

you might want to try <segment:face> some prompt about the face here on the end of your prompt

that will automatically zoom in on the face, and refine just the face, then reapply to the full image

can also swap face to anything else that clipseg can figure out how to match

[dim sonnet]

yay

[dim sonnet]

added this quick tool toggler

which just fills in the filter box with a new <unaltered> value

which automatically hides any parameters that are unaltered (ie either are set to the default, or have a toggler that's unchecked)

[spiral imp]

Yeah this seems p. good

[spiral imp]

Very nice

advanced button still shows up

and would be nice to still have resolution menu available

it's hidden if its set to 1:1 when using the filter

also it should be automatically reapplied when Reusing Parameters

otherwise most menus get hidden

[dim sonnet]

fixed both of those (advanced model addons showing, and reuse param not updating it)

[dim sonnet]

it's, uh, not supposed to be half-showing up like that, idk why it is

nor at all, as that would be counter to its function

oh lol i know why

width/height are hidden if you're not on Custom, but they're basically never set to their "true default", as they're instead set per-model

[potent panther]

btw mcmonkey I'm thinking I'd be nicer to have a </segment> function to make things more clearer especially if you are using multiple

[dim sonnet]

it would be visually clear to the typer ... but on a technical level kinda weird?
Cause like:
wide shot of a lumberjack <segment:face>very strong lumberjack face</segment> waffles
where do the waffles go? the main prompt on the end? A magic additional prompt? A refiner stage? ...??

[potent panther]

main prompt?

but I'd say like it could be useful in making things like this clearer

photo of a cat drinking a beer
<segment:cat head>dog head</segment>
<segment:beer>green sludge</segment>

and adding this logic could make a refiner custom prompt better?

# first prompt 
cat
# segment before refiner
<segment:cat>dog</segment> 
# refiner prompt
<refiner>very fluffy dog</refiner>
# segment after refiner
<segment:dog>cat</segment>

I'm just thinking completely without thinking about how much of a pain it'd be too make this

also just thinking more about it you could add the pill thing to make it even clearer lol

[dim sonnet]

yeah the highlighty thing was a think i've thought about and will probably do

[dim sonnet]

I mean, technically, I could add a </segment> that does nothing and gets stripped from the prompt, purely to satisfy the ocd need to look like clear xml format lol

cause in terms of actual parsing on the technical side it just makes everything much less clear

[versed copper]

brilliant, does what it says on the tin
thanks! 🙂

[potent panther]

teehee

anyway I think an argument for a prompt used by the refiner is needed

[potent panther]

Should be ready for merging https://github.com/Stability-AI/StableSwarmUI/pull/325

[blissful atlas]

sooo i just installed stableswarm (i have no idea how to use it or comfy, i've been using a1111) how do i install those nodes?

[blissful atlas]

I will not...

[rustic river]

@dim sonnet virus, probably

[blissful atlas]

Prob a zip bomb

[rustic river]

Contains a loader.exe or w/e

Screenshot (6 May 2024 14:41:02)

[blissful atlas]

yeah a virus

[dim sonnet]

if you're gonna want to play with a lot of external custom workflows and wotnot, you probably want this https://github.com/ltdrdata/ComfyUI-Manager, it's an addon to comfy that can autodetect+install relevant nodes. Do note this is a third party addon installing even-thirder-party code so it carries a bit of risk

Alternatively you can just google things to find sources, eg "rgthree" is most of those nodes so google comfyui rgthree to find the original repo and install it

Alternately note that a lot of custom workflows are just overcomplicated ways of doing things you can already do without em

if you're new to the space you should probably start with the basics before you dive into complex workflows you found on the internet somewhere

ie use the main swarm Generate tab for things, learn about its features and how to use it effectively, and then when you want to start diving into the comfy noodles, use the Generate tab to set up gens, and on the comfy tab click "import from generate tab" and then you have a spaghetti bowl that contains something familiar you understand, and you can play with it more confidently

[nimble belfry]

If you want to learn how to use comfy I recommend watching Scotts channel. https://youtube.com/@sedetweiler?si=LErIWz1gBPivoLuP

He walks you through how to make the workflow. Just keep in mind if you want to use that workflow in the generate tab you may need to replace a couple nodes with Swarm ones.

You can search this channel for people doing that and read monkey's explanation on how to do that.

[spiral imp]

swarm only uses sdxl clip encode with base sdxl model?

because when I import Generate from custom sdxl checkpoint it uses basic clip encode

left: import using sdxl 1.0, right: custom checkpoint

https://civitai.com/models/143043/starlight-xl-animated the checkpoint

[dim sonnet]

SDXL Base 1.0 with model specific enhancements on uses SDXL CLIP Encode yes

if you're not seeing that, check the model metadata

some people inexplicably have decided to publish SDXL 0.9 models sometimes

(or, well, improperly set their metadata to be 0.9)

if that's the case you can happily just click Edit Metadata and change v0_9 to v1

[spiral imp]

Got it, I see that I have few 0.9 models

like Juggernaut

What's your opinion on using text_g/l for their "intended" purpose (subject/style) or just using same prompt in both? I see that swarm uses same prompt

[dim sonnet]

the "intended purpose" is same prompt to both

the split prompts thing is a clever hack that only kinda works

[spiral imp]

ah

[spiral imp]

Thoughts on UltimateSDUpscale?

[dim sonnet]

iirc that's outdated/legacy stuff

from before the era of having a comfy backend that can automagically tile things for you and all

[spiral imp]

I mean this node https://github.com/ssitu/ComfyUI_UltimateSDUpscale

[bronze shale]

just a heads up even with AMD selected at install time the ComfyUI install done by stableswarm is still installing for nvidia

[dim sonnet]

ye, thread about it here https://github.com/Stability-AI/StableSwarmUI/issues/23

AMD on linux needs somebody who has an AMD card + time + familiarity with linux to get it working properly
AMD on windows is... technically possible but would be extremely painful, as amd drivers on windows are very broken

[dim sonnet]

yes

it's just a refiner-upscale with tiling included, ie redundant/outdated

the tiling might be a tad more complex/capable than what comfy can do natively? not sure how often the different actually matters in practice

[bronze shale]

Time is something I don’t have but I might later. I had it working properly for several months until fedora broke something

[simple swan]

Download this one, plop it into comfyui/custom_nodes, launch comfy, click manager on the right side on the main menu, then click "install missing nodes", then it will fetch what's available.

https://github.com/ltdrdata/ComfyUI-Manager

[digital storm]

Hey! If I'm setting up with a machine that's got an H100, any recs on how to proceed here? (perhaps a missing dep leading to not picking up nvidia gpu?)

[ebon gate]

sponsored machine or self bought?

[digital storm]

bare metal machine (so probably close to the latter?)

[ebon gate]

i mean who bought the H100 - some company sponsored or your own?

[digital storm]

For all intents & purposes, my own. (Not really sure what you mean by sponsored)

[ebon gate]

nice, sponsored means a company purchased a machine

[digital storm]

Oh well then perhaps it's sponsored- it's not in my house lol

[sacred mason]

what is your goal? what are you going to use the image generator for?

[ebon gate]

cool, just wonder since its like very expensive lol

[dim sonnet]

well, first: go with "No" on that question if you have to -- but, probably stop the installer and go back and install nvidia drivers first

it will never mistake for AMD if you have nvidia-smi on your machine, which nvidia drivers always include on both windows and linux for all card classes

[digital storm]

That was it, thanks! Drivers were installed but nvidia-smi reachable

[sacred mason]

an nvidia linux container will have nvidia-smi but no device. not sure if you parse its output.

[dim sonnet]

output is checked to contain any one device (not really validated that it's actually an nvidia device or anything else specific)

for the installer rn there's basically just nvidia or AMD, and only the nvidia installer works properly. I think mac technically works out on its own anyway iirc?

intel gpu gotta do manually, no-gpu ... i mean do the nvidia install and toss in --cpu but ow

[chilly bobcat]

absolutely love the product, any idea when the version where it can be used remotely will be working?

[dim sonnet]

swarm has always been usable remotely

[chilly bobcat]

on setup it says something about being able to be accessed from the internet coming soon, I know it is accessible locally from LAN

[dim sonnet]

the coming soon™️ thing is for shared instances with many user accounts

a personal instance you can run wherever and access wherever, you just have to figure out the network forwarding appropriate to whatever remote server you're using

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Advanced Usage.md#accessing-stableswarmui-from-other-devices

[chilly bobcat]

thank you so much for the response, I will do my own research from here! I have been trying to use moonlight to remote in with zerotier VPN, but it's been flakey

[spiral imp]

Is there a way to see it/s somewhere? Not just generation speed.

[nimble belfry]

Click the logs tab and change it to debug.

[spiral imp]

One model author: "I wager it defaults to 0.9 if metadata is stripped, the others do contain metadata so will set it as 1.0 (if it reads 1.0)"

[dim sonnet]

Correct, SDXL 1.0 requires ModelSpec metadata, otherwise it is not an SDXL 1.0 model

Swarm is configured to autodetect legacy models in a way similar to how auto webui did way back when, so an SDXL structured model without modelspec matches the expected format of SDXL 0.9 models

[lavish glade]

hmm after using swarm for a bit longer. i can say i got a few gui suggestions.

this can either be in the center or left hand side. on a Ultra wide display its kind a pain to go all the way there to get it.

this should have it own line instead of being inline with lora and model info line. especially when i am running 10+ lora's it gets cut off from the display.

like either above or below this.

or right next to the arrow. just push the image down a bit.

also disable page-up/page-down. the ui bugs out when either is accidently pressed.

[dim sonnet]

page-up/down? huh? where/when/what browser? nothing happens when i hit page up/down on main generate tab on firefox or chrome

[lavish glade]

oops just checked, well it did happen a few versions ago lol. using sr ware iron as a dedicated browser for ai web ui's. would have to refresh the page to get it working again when it did happen.

went off a list of things i wrote down as i was using this.

ah found it

its when i when i typing in prompt if i press the page up/down that it break

ignore the actual prompts

[lavish glade]

i would get the big empty gap at the bottom.

srware iron is a chromium based browser iirc

[spiral imp]

can confirm, has happened to me on chrome

[dim sonnet]

oh hey that's weird i can replicate that on chrome yeah. not on firefox tho

... hwhy does that happen that's so weird

[lavish glade]

cuz its chrome

[dim sonnet]

idek wtf chrome was thinking deciding to scroll an unscrollable element if you hit page up inside a different element??? but i blocked the keys + added a force-reset for the scroll on that element in case it somehow happens anyway

[dim sonnet]

see my problem is some people say it should be more prominent, other people say it needs to be removed entirely cause they hate even seeing it :(

left side would make a decent amount of sense

fwiw to open it you can just click any of the tab names (eg click on Models)

closing you need to either click the arrow or drag the bar down tho

[dim sonnet]

heh, never tried running 10+ loras, hadn't even really considered that case

maybe can make it so it automatically expands to multiple lines if there's too much text to fit

[lavish glade]

maybe too much trouble but maybe add a fine tune gui option during install and after install. to customize the gui a tad.

[dim sonnet]

other note as i test with chrome

can somebody explain to me how and why exactly it is that chrome has decided when

• image size = 1024x1024
• max-height = 100% (over 1000px)
• max-width=1000% (over 1000px)
• it has endless free space
  set image size to 480x480

why would you do that

why

firefox just expands the image to 1024x1024 and uses it space

oh omg i know why

[nimble belfry]

Have you looked at it on edge? Windows did an update and made edge my default. I opened swarm and... It was something.. 😂

[lavish glade]

[nimble belfry]

I do have all cookies off on edge not sure if that matters.

[dim sonnet]

seems to be working right to me

[dim sonnet]

swarm very much uses cookies for things, but disabling that should mostly just break all the persistence features

[lavish glade]

i try to stay away from edge as much as possible.

[nimble belfry]

Same. Only when Windows tricks me into it does it ever open.

[nimble belfry]

I will have to replicate and send a screenshot.

[lavish glade]

been 3 months since i reinstalled windows. i even went as far to download chrome off another pc to install it lol.

haven't opened edge since i re-installed windows

[vivid geode]

Does anyone know what is the group input in this node 😦 ?

[dim sonnet]

optional, SwarmInputGroup

for example in

Sampling is the group, sampler/scheduler/tileable are all params

[vivid geode]

Awesome thanks

[vivid geode]

So those group nodes work for organizing and filtering the inputs in the generate tab mostly?

[dim sonnet]

yes

[lavish glade]

Another gui feature that can be kinda useful is a favorite lora collapsible group. Was thinking of a second lora tab next to lora's but a collapsible group in the lora section would look better imo. also a sorting fuction now that i think about wouldn't be bad either like a

favorites
concept
character
style
navigation panel in the image history. just gotta add a "add to" menu in ≡

wait... just noticed i can sort this manually via folder options

in windows

ugh

using stability matrix to keep up to date with lora's and stable swarm updates

[vivid geode]

im getting this error when trying to use a workflow from the comfyui editor tab imported to the generate tab. The node giving me the error is a LoRA loader. Do any of you guys know how to fix it?

[dim sonnet]

so,

• completely clear and remove the workflow from swarm
• close swarm, run the update bat file, relaunch after updating
• import the workflow and try it again

if it's still broken, please post a workflow json so i can test and replicate the bug

[vivid geode]

Thanks alex. I will do so

[vivid geode]

Here's my workflow, i can't make it work in Swarm, don't know what im missing. Thanks in advance

[dim sonnet]

I:

• loaded your workflow
• removed the image load from photoshop node and the thing attached to replace with a regular load image
• clicked "use this workflow in generate tab"
• added images to the two image inputs
• hit generate
  and it worked fine

So... are you sure you're up to date? Swarm v0.6.2.0 (b9b5101c27a7db0b5ca7df67c471cf67de585106)?

I'd be a bit suspicious that the photoshop input node might misbehave but that doesn't appear to be the source of the error

the error looks a lot like one that would have been an issue from a version of swarm that's more than a couple weeks ago

[vivid geode]

So, im trying with a much simpler workflow for debugging's sake. If I follow those steps, the same as you it works okay.
I was able to reproduce the error when I turn this parameter On in the generate tab:

This is the simple workflow I tested with

[vivid geode]

Yeah, I ran the update file as you told me before and the command window said it is up to date

[dim sonnet]

ah, yes, uh, don't enable that lol

[vivid geode]

lmao

[dim sonnet]

That's for if you hook up several workflows to use the same set of parameters, you can eg make grids comparing different workflows

but if parameters aren't identical between workflows weird things happen

[vivid geode]

so i should just load the workflow in the comfy tab, and load it to the generate tab as these steps you did. And avoid that parameter

[dim sonnet]

Yes

You also have the option to enable the experimental new Simple tab, and just use the Save Workflow button and you can select it at any time in the Simple tab, if you prefer that

[vivid geode]

you mean this Enable In Simple Tab toggle?

Or how do I enable the simple tab?

[dim sonnet]

yes, and this server setting:

[vivid geode]

Oh kk

Awesome

I'll take a look at it for now

Thanks Alex

Do you got any idea if I can group my ControlNet nodes inside a SwarmGroup so it looks more tidy u in the generate tab? So far the only thing I've achieved to get inside a SwarmGroup are Input images

[dim sonnet]

if you make SwarmInputs for their values, yes

eg SwarmInputModelName for the load model, or SwarmInputFloat for a strength value

[vivid geode]

Im getting this error. I tried to make a promp a SwarmInput

Like this

[dim sonnet]

ooh, yep confirmed that one, that's a bug -- I pushed a fix, run the updater and it should be good after

[vivid geode]

confirmed, it works now thanks

I will tidy up all my workflow parameters now with SwarmIInputs and if another error like before pops up I will let you know, in case you might want to look it up

[nimble belfry]

Would be great if you shared your Swarm workflows. We could even make a thread or something so everyone can share theirs.

[vivid geode]

Actually I'm just working on a single specific workflow on ComfyUI for academic purpose, but recently i've been trying to implement said workflow on Swarm, so im fairly new to all this SwarmUI thing (hence all the quetions). Though I would be thrilled to take a look at other's workflows and learn a lot

[runic spire]

fresh install of a seperate swarm. I haveswarm running on stability matrix on the same machine and getting dwpose errors and trying to isolate it. Deleted swarm dir, downloaded install.bat, when finished, this is what I get.

[dim sonnet]

looks like corrupt old comfy data stored in browser cache from before the reinstall? probably should fix itself if you hit the param reset

[runic spire]

that fixed it,thank you

[dim sonnet]

swarm 0.6.2 update announcement https://new.reddit.com/r/StableDiffusion/comments/1ct2sj0/stableswarmui_062_beta_released/

[maiden shoal]

On my Linux installation, Swarm appears to install and start a version of stable-diffusion-webui. (at least a process is started that references /opt/stable-diffusion-webui) Is this needed if I only want to use ComfyUI? It appears to be consuming GPU VRAM.
Second question. On the server info tab, one of my GPUs has an asterisk beside it. What does that mean?
Thanks!

[worldly oxide]

Hey, great update do you have any plans to integrate IC-light as part of the interface?

[dim sonnet]

Uh... if you did a standard install? No, not part of that should happen.
If you have an existing old install? You might have done that intentionally, and just need to turn it off.
It's also possible that's just where you put your system's python install

asterisk, idk, I don't think that's something swarm did -- it just reads the GPU name out of your system, and gives Name (ID) where ID is a number

[dim sonnet]

not currently, but
A: seems like an interesting option for the future, probably will eventually add that. Would just need some careful planning around both how to do, and how to best take advantage of automatic light source generation tools
B: you can of course use it in swarm via the comfy workflow tab (and exporting to the other tabs) via https://github.com/kijai/ComfyUI-IC-Light

[maiden shoal]

I don't think Python is installed in the webui folder, but I'll investigate that further and maybe try a clean install. As for the asterisk, it definitely appears to be coming from Swarm. Regardless, huge fan of Swarm!

[dim sonnet]

OH my bad lol sorry, I was thinking about Resource Usage, I forgot the other info part exists -- that * means it's selected as the preferred primary card by default (this mostly is just a hint to the user & selector for the default during install, you can choose what's actually used on the Backends tab) (it's selecting that one because it has the "highest" VRAM) (idk why your 3090 is 24.0 GiB but the 4090s are 23.99 GiB? some nvidia driver shenanigans i guess)

[maiden shoal]

Thank you!

[maiden shoal]

To close the loop on my stable-diffusion-webui question, it turns out an "automatic1111.service" was apparently installed when I was I was trying automatic1111 and something about starting StableSwarm was causing it to also start. It's now been completely purged from my system and I'm all in on StableSwarm. 🙂 Thanks again for the help.

[versed copper]

i have a long name problem in selection popups, using FF

same with the models

affected ones become impossible to click on

[dim sonnet]

uh... using what now? FF implies Firefox but that screenshot is not

which i can tell straightaway from the scrollbar lol

[versed copper]

it is firefox

[dim sonnet]

and also, that appears to be the refiner upscaler list, which, if i give a silly long upscaler model name

i just get a silly long dropdown listing

and i'm testing on firefox here

[dim sonnet]

for comparison btw this is what the firefox scrollbar looks like

firefox notably removed the old style blocky scrollbars quite a few years ago now

[versed copper]

i'm not making this up

[dim sonnet]

I trust you're not making it up lol

the reason i'm pointing all this out, is there's something affecting the way your page is rendering that isn't just vanilla firefox behavior

[versed copper]

some plugin in firefox?

[dim sonnet]

ye

[versed copper]

any suspects?

[dim sonnet]

none of those look relevant

have you potentially manipulated internal firefox settings to forcibly make the scollbars extra wide? eg https://connect.mozilla.org/t5/discussions/scrollbars-disappear-in-firefox-100/td-p/5433

[versed copper]

nope, no custom css or anything like that

same behaviour in 'troubleshoot mode' i.e. all extensions disabled

new profile, same thing

and fwiw, it's the same on two different computers, i just installed swarm on another one (both win10)

[potent panther]

Very odd

I’ll check my Firefox

[strange hull]

How to you use BREAK commands? The <break> feature is not the same.

As the token count should jump when you use it

[potent panther]

mcmonke did you forgor to code that in?

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/discussions/340

i assume that's the same person

[dim sonnet]

is there somehow some obscure thing where windows 10 firefox has the chonky scrollbars but win 11 and linux both have the fancy modern ones

[versed copper]

wouldn't surprise me one bit

[dim sonnet]

i don't even know how to fix that, in css it's set to width: fit-content, which I can't even add +5px to that to compensate or anything (css doesn't let you do a calc on fit-content)

that wide scrollbar is actively cutting the width shorter than what's necessary to fit-content, which, ???

fwiw as a workaround you can type in the search box to reduce the list length, or arrowkeys+enter to manually select in keyboard

[hot bluff]

the new UI theme is alot nicer i think

can you shrink the welcome pane to make the promp section bigger? When i make the promp section bigger the negative box drops out of sight

Also can the aspect rations in the drop down box show the resolutions as well?

[dim sonnet]

wot? the prompt box automatically sizes to fit the text in it

[dim sonnet]

yeah just select "Custom"

[dim sonnet]

(and you can shift+enter to add a newline if you want)

[hot bluff]

Perfect thanks

[versed copper]

yes, i've noticed that if you search for one of the non-clickables, they become read- and clickable
it's no biggie, nice if it could be fixed, but ok if not 😉

[cursive pewter]

does stableswarm support PAG?

[bold belfry]

I have 4 GPU on my PC, How can I use it to render each queue in a single workflow? something like when I press 3 queue it should use gpu0 to render queue 1 gpu1 to queue 2 gpu2 to queue 3.

I already add 3 ComfyUI Self-Starting to the backend but it keep using only the first backend that point to GPU0 "running backend: (0): ComfyUI Self-Starting"

[bold belfry]

Ok it work now, at first I can't see the Multi GPU > use all because my queue panel keep reverse to stacking on the Multi GPU panel every time that I switching between tab, Is there any way to fix that?

[dim sonnet]

not natively but you can use the comfy nodes for it https://github.com/pamparamm/sd-perturbed-attention on the comfy tab

[dim sonnet]

i'm aware of that happening to some users but i have no idea why

(ftr it's a bug in the comfy webcode, not Swarm)

[rustic river]

I have the dropdown box weirdness as well. Always had it, but I expected it was known

I despise these "cool" narrow scrollbars, so I'm actually happy with the chunky one, but it's kind of random when it shows.

[dim sonnet]

are you on Win10? Or like... do you know why you have the cursed scrollbars that are too wide but don't calculate their own size right

[rustic river]

The narrow ones are cursed
But no, I don't know why I have the thick ones, and yes, Win10.
Nothing modified in about:config when searching for scroll
Isn't the problem the <select> box, though?

I mean, doesn't the popup div take it's width from the select box (which isn't wide enough)

Oh wait, all scrollbars in swarm are like that.

My Windows settings are the default (auto hide scrollbars), which works in any other Window. I guess Firefox is just broken.

[dim sonnet]

the reason I say it's cursed is because the CSS section is width: fit-content, but as you're showing in your screenshot, it does not fit the content, ie it's mistakenly sizing for if the scrollbar wasn't there, then adding a scrollbar in the way, such that it cuts into your content width

i can probably do a hacky width extension in JS to workaround it

[rustic river]

Yeah, I get it now, looking at how it looks correct on Linux FF

[dim sonnet]

@rustic river @versed copper try latest update, i added a bit of JS to intentionally overextend the width by 16px

not sure if that's enough or not tbh

nor whether that'll have side effects

[rustic river]

Can't test now sadly, will report back tomorrow 👍

[dim sonnet]

kk

[runic spire]

This is a showstopper for me. If I run comfy stand alone in the dlbackend dir, it runs fine and finished workflows using ipadapter, run comfy in the swarm screen -- this happens. Fresh install of swarm, models in correct dir -- frustrating.

[dim sonnet]

which model, and in which folder?

if you use ipadapter from the swarm interface it will automatically download the correct model to the correct place for you

and if you do that, the same model then works as-intended in the comfy interface flawlessly too guaranteed

(By default that's Swarm/Models/ipadapter/<name>.safetensors where <name> is eg ip-adapter-plus_sdxl_vit-h)

[runic spire]

Are there guides to use ipadapter from the interface? Some of us come from a a1111 auto background with tuts all over the place, plugins with videos how to's. I am just learning comfi-nesse language and mixing a ui as storms and comfi and together to use. I am not using the swarm interface and ipadapter because simply -- i don't know how to use it. I can with use it with comfy cuase I have guides how to do it

[dim sonnet]

drag an image into the prompt box - the ip adapter options will popup on the top of the parameter list to the left

that's why it says to drag an image here

it's funny, auto/comfy have so many tutorials cause you can't use em without it

swarm has the reverse problem: some features are so in-your-face that you don't even process that it's right there

[runic spire]

i remember my first year of coding, my teacher said here is a function and a procedure looking at me with a old dot matrix printed code on paper. He asked me what the function returns and what does that prodcure do. I said teacher, how should I know? you never told the class what a function is or a procedure, what a body of a statment is. Some of us, are not as advanced as say the author of the product, nor even know what revision is. excuse me for my ignorance. Im simply trying to stop beating my head against the wall.

[dim sonnet]

oh yeah, don't worry no blame on you for that. you're not the first to ask about exactly this

eventually either ima convince some youtubers to do more fullerer tutorials about swarm, or ima have to start making em myself

[runic spire]

i respect your work. as a retired coder, i know you wrote a masterpiece. until there are guides on how to use it, it will be beautiful unused software becuase simply nobody knows how to use it. it will be a car undriven becuae nobody knows how to operate a stick and automatic car combined. I come from a assembler background, and the only way i ever learned was from intel chipset manuals. Swarm has sparse documentation on features let alone tuts on how to say combine ipadapter and the svd section to make a vid swarm workflow.

[dim sonnet]

well how to combine features, is, well, even more obvious in the way that if i try to explain it i probably come across condescending

you enable ipadapter... and then you enable video. and then you go

you just do both, the combination is automatic

[runic spire]

explain in your git, more eyes there -- i follow it like a newpaper, daily.

thank you. show that there. it needs it. the more others learn on how to use it, the more exposure it gets.

and thank you for swarm btw.

have a good day. your across the pond, i am states and need sleep.

[dim sonnet]

i'm in california, it's 3 am, lol

[runic spire]

really, san deigo here

3:17 now..

been up trying to get workflow to work

[dim sonnet]

neat. I'm just a lil ways north of you in LA. I should be asleep too, i'm just deeply insane and answering work-related questions at 3 am on a sunday. not sure what it says about you though that you're asking them at the same time :P

also yes i've got some idea of how to improve the docs a bit here

[runic spire]

thank you. it will help a great deal. good night.

[dim sonnet]

Went a little nuts, here's a start at properly documenting advanced features @runic spire https://github.com/Stability-AI/StableSwarmUI/tree/master/docs/Features

[spiral imp]

Advanced usage links to 404

rest is good stuff 👍

[dim sonnet]

oop, good catch, fixed

[rustic river]

Oh noes...

Somehow, this works.

[lavish glade]

why you gotta a call me me out like that

[nimble belfry]

This is what I was talking about how edge looks. Windows 10 and I have cookies disabled by default for Edge because I don't use it.

[dim sonnet]

would you be willing to open in a text editor and edit src\wwwroot\js\genpage\ui_improvements.js
line 45, this.optionArea.style.width = (this.optionArea.offsetWidth + 16) + 'px';
EDIT: Update to latest, then line 19 : this.overExtendBy = 16;
change that 16 to a larger value (eg 32) and see if there's some value at which it reliably works for you?

with any luck it'll just need to be slightly wider

the fact that one menu seems overwide now and another is still short tho implies something's wrong in a more complicated way

i noticed a bug where it was accidentally losing the overextend if you typed in the search box

[dim sonnet]

that looks more like it's just failing to load the page

eg javascript is missing or something

[dim sonnet]

maximized canvas focus makes it look like swarm is a whole different UI lol

[dim sonnet]

been a minute, but: you can sort models by date time

[versed copper]

somehow, this does not work for me 😦

[dim sonnet]

can you try #🐝｜swarm-ui message

omg i just measured your screenshot, scrollbar is 17px ... might be literally 1 short

but try 32 or something

[versed copper]

32 fixes it, popup still looks the same, but no more garbled entries

[rustic river]

this.overExtendBy = 20;```
Seems to be the minimum for me.

[dim sonnet]

okay, i'll push a commit with just slight above that to be safe aaaand shake my fist angrily at presumably this is microsoft's fault maybe or firefox idk, but, angry fist shaking at someone, and then either way consider this bug officially workaround'd

[rustic river]

I'm assuming it's a Windows issue, tbh

[shadow ivy]

nvalid operation: All available backends failed to load the model.

HELPPPPPPPPPP

[potent panther]

That's not gonna be enough to help

Specify your setup and settings

[nimble belfry]

Are your settings pointing to the correct ckpt file? Or did you rename them on each machine or something?

[shadow ivy]

so, i've just downloaded it like normal, and when i select stable diff, it just say this, it won't generate anything, i have all clean settings

[nimble belfry]

what folder did you put it in and is it the same one specified in settings? If you left it default the models folder is in the StableSwarmUI folder. If you have multiple backends that same file needs to be on each machine.

[dim sonnet]

Check Server -> Logs -> possibly Debug

there will probably be more detail about what went wrong there

[shadow ivy]

i've just installed it in download like a monkey bruh

[shadow ivy]

i'll check when i will be on my pc

i will probably just redownload it before bothering you more

[torpid harbor]

Helllo @dim sonnet you wonderful person you! hope all is well!

Quick question: in theory, I should be able to build a comfy workflow and then have it show up in generate and it be like, a little UI for people to interact with it?

I am guessing there is more to it, because I loaded a workflow, told it to show in generate, and it doesnt really show, like all the different inputs the user would need to put in etc. Do I need to build it up like...a widget or something. I have not idea what I am talking about.

TLDR: Are there any tuts on how to take a comfy workflow and make it into a usable UI using this?

also, does this work with animatediff workflows? or is it only good for single image creation

[dim sonnet]

hi amli!

Yes, there's a new feature dedicated to making a nice lil UI designed to be very accessible inexperienced users, the Simple tab:

To use it you:

• go to the Comfy tab, make your workflow
• use SwarmInput nodes to define what inputs you want to make available to the user
• Save the workflow using the swarm save button (top left) and be sure to check "Enable in simple tab"
• go to the simple tab, click your workflow, verify it works and all. The URL in your browser bar is auto-updated to be a direct link into the workflow so you can copy that

[dim sonnet]

in general if you can get animatediff working well in comfy, you can definitely can in swarm too, and it should work fine.
The SwarmSaveAnimationWS node is used to save animated outputs in a way Swarm can reliable recognize. I think VHS will work too but might be weird. I have not actually yet tested video in Simple tab, need to check that

yes apparently video does work on the Simple tab, at least for webp format

[torpid harbor]

OH! Ok the swarm nodes are what I am missing here. So your have to integrate swarm nodes in to the comfy workflow to show swarm what it is supposed to use? Is there any documentation on that?

[dim sonnet]

Yee

• there's an example workflow included you can reference
• for the most part beyond that you can kinda just type in swarminput and look at the options and slap em in

got an integer input like seed, steps, whateveR? use SwarmInputInteger. Got a text input like prompt? use SwarmInputText and set view_type to Prompt

the example workflow is pretty straightforward to see how it works

[torpid harbor]

Thank yoouuuu!!!! Ok, another question. Is there a way to load imageS? Iike a path or batch?

or i dont suppose an "easy" way to convert nodes to swarm nodes. Like, if I am like, "boy boy if only this LoadImagesFromDir could be a be a swarm node and I could use in the UI! Oh wait it can cause I just do this and this and now it is one!"

[versed copper]

on swarm startup, it still takes minutes to read the pic history, when there are enough of them.
and every single folder already had a image_metadata-log file

which i only notice because a restart of swarm loses any settings i had, always (edit: i lied, the prompts sometimes persist)

[dim sonnet]

Just use the "convert widget to input" on any relevant settings

then attach the appropriate Input node to that input

for example a LoadImagesFromDir node presumably has a path input as Text, so you just hook SwarmInputText up to that

[dim sonnet]

Eh?
Settings are stored in config files and persisted
Parameters like prompt are persisted by browser cookies (with a time limit)

if parameters reset often, it's a browser issue. If settings (User Settings or Server Settings) go missing, it's a datadrive/filepermissions type of issue

[dim sonnet]

as a workaround you can just cut these down

if you're using sort by date, the latter should be reduced. If you're using sort by name, the first one

[versed copper]

hmm, shouldn't be, it's all on a drive f

but then again, windows

[torpid harbor]

officially my new favorite person!

[versed copper]

he does that!

[dim sonnet]

Drive F sounds like an external/addon drive? If it's a network drive or Fat32, those do weird things in windows

NTFS is safe, basically anything else is iffy on windows

[versed copper]

no, i just have a ton of M.2s

[dim sonnet]

nice

Also to be clear, can you confirm what exactly is not persisting properly?

[versed copper]

anything on the left panel, prompts mostly stay

[dim sonnet]

okay, so parameters then -- those are what get stored in browser cookies

try closing and reopening a tab of the same page, or closing and reopening your full browser

one of those two actions should yield the same result

(note not the refresh button necessarily, that's different)

[versed copper]

i always close down swarm from the server tab, which closes the swarm firefox tab

then update swarm, restart it, which opens a NEW tab

should parameters persist under those circumstances?

[dim sonnet]

Generally, yes, unless something's going wrong

[dim sonnet]

^ please test these two

knowing which yield the same result or not will narrow down what it could be

prime candidates for reference are (A) just browser settings, (B) a browser bug, or (C) the autolaunched browser window is hardcrashing

[versed copper]

very strange, i just closed the tab, then reopened
it kept my purposely strange refiner percentage (0.21) and the right upscaler
but invents a hypertile setting
steps and cfg are default (were not), resolution is right

[dim sonnet]

... huh

It's possible that it's finding old persisted data that it's thus far missed from bugging out?

so to be safe: Use Quick Tools -> Reset Params To Default, then intentionally make a few param edits, then re-test

(the full param reset button intentionally clears all cookies)

[versed copper]

same with fullo browser restart, except no hypertile this tim
both methods completely lose sampler settings too btw

and i'm running the StabiltyUI package, but (i just tested) this is the same when using standalone swarm

[versed copper]

will do

[dim sonnet]

the StabiltyUI package, the what

[versed copper]

my bad, i meant StabilityMatrix

but the standalone installation on another drive does the same thing

[dim sonnet]

++ I pushed a commit that oughtta load the image history list a decent bit faster in all cases

[versed copper]

fwiw, main reason i use the SM package of swarm is that there is an indicator if and when you pushed something

could the swarmui ui maybe include a check for updates thingie?

[versed copper]

i do have 9794 files there atm, but this was way quicker 🙂
thank you and kudos as always

and if i can be a pest again, i have a question:
you state in the help popup that ClipSkip is never necessary for SDXL
this does not seem to have filtered down to the community, where you see -1 etc on civitai and elsewhere
is there an explanation somewhere?

(i totally take your word for it, i'd just like to know the background if possible)

[dim sonnet]

it has an update checker, but atm only for release builds not dev commits

so when 0.6.3 release is marked, anyone still on 0.6.2 will get a notif on the welcome screen

it could make sense to add a git commit check too

[dim sonnet]

Generally a lot of civitai params like that are from people using auto webui, and in auto webui your clipskip settings get entirely ignored for SDXL models.
In Comfy if you intentionally add a clipskip node, that doesn't get ignored.
In Swarm, same as comfy, if you activate the setting it's not ignored.

The reason why it says that is: in SDv1, it was -1, but NAI's SDv1 implementation used -2, ie for SDv1 models the correct setting changes between specific model selection, so you have to change it on the fly.
On SDv2, there was only ever one value, and on SDXL there's again only one correct setting, and also XL has two textencs that are configured slightly differently, so mucking with that is not a great idea in general.

There has been an argument that sometimes intentionally mucking with it can achieve certain effects, so if you're trying to replicate something where they went out of their way to say "This needs XL on a manual clipskip of -1" and is a swarm or comfy user, then yeah use that -- otherwise probably just ignore it as civit users inputting silly things as usual.

[versed copper]

thanks for answering this
i found it on a model that has blown me away with its quality (crystalClearOneVs1_v10), so those guys must know their stuff, one assumes 😉
and they recommend Clip Skip: 1 (which i suppose means -1)
i tried and did not see a real difference

[dim sonnet]

++ pushed a couple commits to make history load even smoother, both on server and browser sides.

I can't really do much more on the server side anymore as

it loads my entire history folder in less than a tenth of a second for me lol

wait no that's not

it's just stopping a bit unevenly before the max limit of 5000 apparently (EDIT: Fixed)

oh i see why, i have some invalid images somewhere apparently

Listed 10000 images in 0.844 seconds. lol

nvme ssd too stronk

aaah the invalid images rather were video files being mishandled, i see nope. I'm confused where the invalid files come from
OH it's from files that aren't images in the folder okay jfc okay

[dim sonnet]

random other update, Swarm has a full fat Discord of its own now rather than being confined to the one channel at the bottom of the list here (this channel's staying still tho, just adding a dedicated discord so we can have eg commit log channels and all them other fun bits that you can only do if the whole discord is dedicated to one project) https://discord.gg/q2y38cqjNw

[nimble belfry]

Cool. I eventually want to post some of my comfy workflows that are just popular workflows I converted to Swarm.

[nocturne glen]

Damn, the mods are asleep.

https://tenor.com/view/amogus-mods-modsareasleep-discord-meme-gif-23110527

[potent panther]

Crap they actually are

[harsh peak]

This is a virus. Don't download this!

[last tinsel]

My upscalers are no longer showing up at all under Refiner Upscale Method

Started after the latest update

[potent panther]

@thorn harness

Please you’re the only mod online

[last tinsel]

Thank you for the update. Upscalers work now 👍

[potent panther]

Oops @thorn harness this is also a scam

[versed copper]

hum, my upscalers were gone, too, but a SM restart fixed that before an update arrived

[dim sonnet]

yep, comfy updated broke upscalers, i pushed a swarm update today that autoinstalls the new dependency to fix that

also, new Outpainting guide: https://github.com/Stability-AI/StableSwarmUI/discussions/360

[sage grove]

Hi. I'm probably asking in the wrong channel - has anyone installed Stable Swarm on Runpods? I wasted the whole evening and did not succeed. There is very little documentation about setting it up, and using it with multi GPU.

I work on extremely large images, so was really hoping that multigpu rendering could speed up the tedious upscaling / rerendering to fill in the new details process.

[dim sonnet]

singleGPU setup on runpod should be easy, can even just use the Colab Notebook for it

multiGPU setup is gonna require some familiarity with linux command line, but once you get it open on a node with multiple GPUs, as far as swarm goes just go to Swarm -> Backends -> add new backends with incremental GPU_ID values

note that a single very large image can't take advantage of swarm's current multiGPU code, only if you split the parts, those parts can run separately and you can recombine at the end, which currently would be a bit of manual work

[sage grove]

I assure you that it'd be less manual work than I'm currently doing (hence trying to figure this out) --- Currently, I build an image in A1111 at (in this case) 900x1600 - then I upscale it in Topaz to 5400x9600 - then I slice it haphazardly to roughly 1,600x1,600 pieces, and I throw those back into the A1111 and re img2img (usually inpainting) and create more detail. Take the best output, blend that in photoshop, repeat until the whole image is done, Then I kinda repeat the process when that's all done - final resoltion is 21,000 on the Y axis.

Splitting it up into multiple pieces, and throwing them to individual GPUs to work at once would be fantastic.

Though... I spent all night trying to just make it run on runpods, with no luck... ChatGPT was helping me....and then I even had a human help me... Had it working for a few minutes by him adding a Cloudflare tunnel... but that didn't last.

[potent quarry]

Is there a reason both of my GPUs aren't being used now that I'm using ComfyUI in StableSwarm?

[dim sonnet]

if you're on the Comfy Workflow tab, you're only using one comfy instance by default -- you can click the "Multi GPU" menu at the top left to swap to "Use All", see more info in docs @ https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md

[sage grove]

If anyone cares- I took the issue up w/ Runpod support last night, and the guy figured out the issues, and made a template, so it can just be booted..... https://runpod.io/console/deploy?template=u7mlkrmxq3&ref=c6jd6jj0 - Port 2254.

[nimble belfry]

Thank you!

[jagged cipher]

Odd question. Each time I run an image through img2img in order to inpaint something, the stuff that's not been masked also gets redrawn, making it more and more blurry. Is there any way to stop that while still letting the part I'm repainting 'look at' the entire image for context?

Mask Shrink Grow would be my immediate instinct, but that also seems to lock down the 'context' that the inpainter has to work with

Or do I just do Mask Paint Grow with a really large value or something?

[dim sonnet]

hi, sorry, i haven't been home the past few days ---
Is this on SDv1 models? One of the "big improvements" with SDXL is that the VAE became loopable - it could encode+decode repeatedly without altering content. SDv1 had a slow-regression when looped.

[jagged cipher]

It was on an SXDL model, I believe

That being said, I ended up just doing Mask Shrink Grow with a stupidly large value and that's been working sufficiently

[jagged cipher]

(for reference, I was specifically using Animagine XL V3.1)

[dim sonnet]

I pushed a commit that does something close to this minus the redundant logic -- all init image with masks by default now recomposite the changes onto the source image with the mask

(with advanced param Init Image Recomposite Mask to toggle this off if you want)

[jagged cipher]

Thank you!

[dim sonnet]

TensorRT support! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-9641683

[haughty moon]

is there any way to change the 'starred' output folder?

[dim sonnet]

uhh currently no. Could potentially add that - can I ask why you want to change it?

[haughty moon]

no 'good' reason really, i'm just used to having a 'new images' folder and a 'saved' folder in different locations

but it's more of a convenience factor than anything else, i can get used to the current way it functions

in other news though, i think there's a bug with the theme/layout, when the window is wider than the image, it aligns the meta and buttons to the right of the image instead of underneath

if i shrink the width of the window enough it changes the layout

it seems like the second one is the intended layout

[haughty moon]

actually, looking at how the html changes as i change the window size, i guess maybe this is intentional

[dim sonnet]

yes this is intentional - it tries to put the metadata wherever is most optimal for your screen aspect ratio

[haughty moon]

yeah that makes sense, it just looks weird on a larger resolution screen because of the empty space

[dim sonnet]

swarm 0.6.3 https://new.reddit.com/r/StableDiffusion/comments/1d7o9gx/stableswarmui_063_beta_released/

[left lake]

good afternoon, I tried sd swarn yesterday and it is beautiful, user-friendly and simple interface and comfy in backend. If it also has an instantID or an analog, it will be a masterpiece (I haven't looked at the advanced settings yet)
thank you

[dim sonnet]

drag an image into the prompt area to use ReVision, ReferenceOnly, or IPAdapter (incl. IPAdapter Face). InstantID in very particular had a new feature request listed a few days ago https://github.com/Stability-AI/StableSwarmUI/issues/351

[nimble belfry]

How do you add models? I am trying to ssh in but I think maybe there is a config error on the directory location.

[dim sonnet]

New model downloader! In the utilities tab enter a model URL to download the model. If you give a civitai URL it will automatically download the metadata from the civitai page and apply it https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-9717490

[dim sonnet]

... well do i have a convenient answer for you lmao

[fringe dragon]

Hello everyone. I'm hoping someone can answer some questions. I have comfy UI installed. Is stable swarm exactly the same features as comfy but with an interface like automatic 1111 ? Can stable swarm use my comfyui installation for the necessary files it needs ?

[dim sonnet]

yes and yes

in the installer when it asks you what backend to install, select none, and then once the UI is loaded, go to Server -> Backends, and add a new "comfyui self start" backend and give it the path to your existing comfy install /main.py file

also under Server Configuration point it to your existing models folder

[fringe dragon]

nice, thank you very much, will install it now then

[fringe dragon]

hmm i'm running comfyui with a conda env, i'm guessing stable swarm needs to activate the conda env somehow

[dim sonnet]

if you activate the conda env before starting swarm that'll work fine iirc

[fringe dragon]

ok i will try than then

[fringe dragon]

not working for me with my current workflow. I set it up in comfyui to run pix art sigma as base and then sdxl as refiner. I think it needs to be supported for it to work ? Also i can see nowhere to put the path to sigma T5XXL model

seems to be running now after switching all to on on the left side. But i see no progress bar, on the comfyui workflow

it is working fine now, but the thing is. Did i have to put everything on, on the left side for it to work ?

And the prompts needs to be put on the side for it to work. I left them empty and put them in the center as you can see and it did not take the prompt. Is there a way to make the prompts work here

also all these parameters here about the generated image, can we hide them ?

[dim sonnet]

yeah just don't checkmark them if you're not customizing them

[dim sonnet]

When you're doing a fully custom workflow on the Generate tab, you can either:
(A) title a primitive SwarmUI: Prompt to have it use the default prompt param, or
(B) use SwarmInput<...> nodes to explicitly control your inputs

[dim sonnet]

i've never tried running pixart so idk how it works, and it's not supposed in the default/normal way of using swarm, only custom comfy workflows as you're doing

[fringe dragon]

no it works well. the thing is how to use the prompts in the generate tab. I only need to change the title of the prompt node in comfyui for it to work ?

[dim sonnet]

needs to be either a primitive with a title or a SwarmInputText

[fringe dragon]

ok thanks, so the title for the primitive shou be "SwarmUI: Prompt" or install a custom node "SwarmInput<...>"

[dim sonnet]

the title for the primitive shou be "SwarmUI: Prompt" yes
install a custom node "SwarmInput<...>" nothing to install, the swarm nodes are included in swarm

[fringe dragon]

thank you very much

[fringe dragon]

i see the writing for embeddings is different from comfyui : embedding:bad-artist-SDXL, embedding:peopleneg in comfy and in stable swarm is embed:Style/bad-artist-SDXL, embed:Negative/peopleneg. When calling embeddings in the generation tab should i use the new way of calling them ?

[dim sonnet]

whenever using the standard prompt input yes

[fringe dragon]

yeah i only change the title of the primitive as you told me to be able to change the prompts directly from the swarm interface in the generate tab. In that case i should use the new embed:Style/bad-artist-SDXL there right ?

https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fstableswarmui-beta-v0-o7ag9z88njnc1.gif%3Fwidth%3D1920%26auto%3Dwebp%26s%3D71340cd21bebd3523eed10325e052cd81b2a543c

also where are they taking this image editor feature ? I was searching how to inpaint here. Had an inpaiting workflow i wanted to use

[dim sonnet]

drag an image into the center and click "Edit Image". Requires a touch of set up to make the image editor work: ideally use SwarmInputImage twice, once with InitImage and once with MaskImage

full disclosure i've never tried the image editor with custom workflows so idk if anything might go wonky

[fringe dragon]

thank you, it's worth the test. It will be hard i think to configure inpainting with it, but i'll try

[nimble belfry]

How the hell!!! Dang that is very helpful 😂

Also, the person who made that run pod template has just updated it. Seems to work good now.

[dim sonnet]

install instructions here https://github.com/Stability-AI/StableSwarmUI?tab=readme-ov-file#installing-on-windows it is literally download file, run file

once it's running it has an installer that'll ask you basic questions and you can just "Yes" through it all and let it install

and then it'll open a page which will tell you how to select a model and to type a prompt and hit generator

everything else has lil ? buttons you can click on to see what it does

there's also this https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Basic Usage.md if you want a thorough step by step first time guide

the other docs in that folder are available to take you farther with more advanced things when you want to

[verbal junco]

is this automatic1111

chat?

[nimble belfry]

No much better than that. It's a replacement for Auto that uses Comfy as a backend so you can do advanced workflows but still use a nice UI. It also supports multi GPU for batch images.

[verbal junco]

I'm coming from Foocus, and wanted to try auto1111 but does this use node base like comfy?

stablewarm

that is

[fringe dragon]

this can use both, it has a node base back end if you want with comfyui or simpler interface like auto1111

i'm using it with my custom comfyui workflow

[storm mesa]

is this a new tool of stable diffuison ?

[fringe dragon]

yes

[verbal junco]

oh nice

I'll download it

[shadow arrow]

do I just enable do tiling to get tiled upscaling? (Cause SD3 only supports that for now)
so let's say, it upscales 1.5x and it uses the base model

and maybe I increase the control percentage depending on how aggressive the upscaling should be

oh nice this also has seamless tiling built in

whatever's the case, I have stableswarm installed for tomorrow

[shadow arrow]

oops I'm blind

[eternal aurora]

I'm using Pixart-Sigma into SD15 conditioning; then into SDXL, PAG and Face Detailer .. here is a PNG with included w/flow

[gusty solar]

How do you set Clip Skip?

[thick flicker]

I have been wondering the same

[potent panther]

it's the clip stop at layer option

[verbal junco]

is there no inpainting

in this web ui?

on the generate page that is

[fringe dragon]

why are you using sd1.5 and sdxl in the same time, why not only sdxl ?

[eternal aurora]

Errrmmm the PiXart-Sigma I started with outputs via SD15 conditioning - this will not interface with Face Detailoer - so I setup a Base+Refiner (Plus PAG) for input into Face Detailer

... unless you know any different, post me a workflow 😉

[fringe dragon]

i will check your workflow later, i'm at work now. Here is mine to see how i'm using it.

[eternal aurora]

Cannot get AutoCFG to load!!!

[dim sonnet]

use the "Edit Image" button on any image to get a mask interface for inpainting, and the "Init Image" parameter group to configure settings

[eternal aurora]

I am SymLinking my StableSwarm folders to my ComfyUI setup ... wish me luck! 🙂

[fringe dragon]

Should work if you have everything updated, if not simply remove the node and add it again the same way. Works when there are updates that break something

[fringe dragon]

I'm using symlink with my auto1111 in comfyui and have comfyui as backend for stable swarm. Works well. Stable swarm is loading my files from my auto1111 installation with comfyui as backend

[dim sonnet]

you can just set the paths in the Server Configuration tab, don't need symlinks (but symlinks work too if you like em)

[eternal aurora]

Yes, I have had the "blank ComfyUI desktop" from time to time - so a swift uninstall/reinstall does the trick! Symlinks keep all the heavy-lifting afterwards to a minimum!

[eternal aurora]

Symlinks save hours of re-downloading

So now I have A1111, Vladmandic, ComfyUI, ComfyUI_Portable and Swarm all sharing the same folders

I'm finding the text on the Model Descriptions a tad small!

In Swarm ...

Even under BIG CARDS

[dim sonnet]

No I mean click "Server Configuration" tab and you just just set the filepath of your models dir, it works like symlinks but easier to set up

you can at least have an auto model folder, set the path in Swarm and then swarm+swarm's comfy automatically use it

for making vlad work too tho yeah symlink

[dim sonnet]

eh... CTRL+MouseWheelUp? lol

maybe I should add some user settings to tweak that

[eternal aurora]

Can Face Detailer be used on Swarm?

Something from SWARM using a DoRA from ClownShark 😄

[dim sonnet]

i assume by 'face detailer' you mean something equivalent to https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Features/Prompt Syntax.md#automatic-segmentation-and-refining

if you mean specific comfy nodes of some form, then, yeah, you can use any comfy nodes in swarm's comfy tab

[gilded widget]

Is there any way for Swarm to use the models already installed in ComfyUI? If I run it under the "Comfy Workflow" tab, it works, but when I run the same flow under generate tab, I get this...
Edit: Found the server config tab 🙂

[eternal aurora]

It's OK - I didn't realise that you could place your own w/flows into the ComfyUI backend! 🙂

I'm just getting used to "driving ComfyUI blind!!!"

[dim sonnet]

https://new.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/

[slow bobcat]

hey @dim sonnet any chance you can help me pls bud?

im able to load my SD3 model

but if i hit generate i get error

[dim sonnet]

you might just need to restart

[slow bobcat]

i notice also my model selected is the sd3 t5xxlfp8 model and your guid says just to use the sd3_medium.safetensors

[dim sonnet]

if you eg edited the model path in server config that doesn't always update the backend properly atm

[slow bobcat]

yes i redirected my models path to anothe disk

[dim sonnet]

yes you should just use sd3_medium. The other works too, but sd3_medium has less redundant files

[slow bobcat]

ok ill reset the defaul root and model paths and copy the model into that directory to test again

trying now fingers crossed

@dim sonnet same story

it keeps crying that the model isnt selected in backend

[dim sonnet]

uhh that looks like comfy doesn't have the model directory applied

[slow bobcat]

that would make sense for the error

any idea where one can apply it?

[dim sonnet]

it should auto-apply when you restart

can you screenshot your Model Paths settings in server configuration + screenshot the actual folder?

[slow bobcat]

verbose and error filtered

not exactly sure how, but i got it working from the comfy workflow side

[dim sonnet]

... ooh okay there's an error message in there... that's missing details. Argh.

what is that typeerror about

wtf

[haughty moon]

keep getting this error trying to generate using the SD3 model
Invalid operation: ComfyUI execution error: Given groups=1, weight of size [4, 4, 1, 1], expected input[1, 16, 144, 112] to have 4 channels, but got 16 channels instead

[dim sonnet]

someone on reddit had the same error and i didn't know how they did it and they just replied "oh nevermind fixed it" and didn't explain what the problem was ;-;

my best guess is it's something to do with selecting a custom VAE or something

[haughty moon]

good guess, i had a sdxl vae set

now i just need to get the t5xxl model to properly download..

[ornate crater]

im getting an error when trying to load the default fp8 version of the T5 clip model "Invalid operation: ComfyUI execution error: module 'torch' has no attribute 'float8_e4m3fn'"

but the fp16 version works just fine.
Any ideas on getting it to work with the default or is there a way to change the default to the fp16 version without having to import the comfy workflow to the generation tab?

[dim sonnet]

uhhh is this an older install? if so you might need to update torch to make it load the newer features

(or just lazily reinstall to make it easier)

if you want to just do the fp16 version ,download https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/text_encoders/t5xxl_fp16.safetensors and replace the (Swarm)/Models/clip/t5xxxl_enconly.safetensors with it

[ornate crater]

Tried going through and updating comfy’s dependencies but it prob didn’t get the new torch. I’ll just nuke from orbit reinstall then.

[smoky bolt]

I'm trying to run both StableSwarmUI and OpenWebUI through docker. I'm not sure how to get the API endpoint of ComfyUI through the docker container, however. Anybody know, or have an idea where I can start looking to figure this out?

[dim sonnet]

if you have swarm running in docker, just connect to that, comfy is an endpoint within swarm's routing (/ComfyBackendDirect)

[smoky bolt]

Awesome, works like a charm. Just have to fix the default prompt in openwebui for sd3 😄

[rigid maple]

Is there any way to share a link from stableswarm to be able to access it from a different network?

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Advanced Usage.md#accessing-stableswarmui-from-other-devices

[rigid maple]

Legend! Thanks mang

[tulip wasp]

is the tiled refiner supposed to work well with SD3? because I cannot really get any good results from it, the refined image always just looks blurry

the only way how I can get a sharp image from it is if I set the "Refiner Control Percentage" to 1, but that of course breaks the image composition completely

I assume the problem mainly is that it uses a "dumb" upscale between the VAE decode and encode instead of a ESRGAN upscale like what's normal in A1111 for highres fix?

I think I remember that whenever I tried a highres fix in A1111 with a "dumb" upscale, it also resulted in something blurry without way too high denoise values

[dim sonnet]

you can do whatever upscale you want in between. Proper upscale models are definitely better

[tulip wasp]

how?

[dim sonnet]

pick your favorite off this hellscape https://openmodeldb.info/

shove in (Swarm)/Models/upscale_models

[tulip wasp]

I mean, how to select it in swarm?

[dim sonnet]

reload param vals

and select

[tulip wasp]

ah, you mean it just automatically shows up in the list of upscale methods then?

ok yeah, that works, thanks! 😄

I guess it's obvious if you know it...

is there any way to use different seeds for the different tiles?

and the result definitely looks way better now with a proper upscale model 😄

why does the "refiner upscale" slider not go further than 4x, I want more 😂

[toxic shale]

@dim sonnet
halp xD
https://html-preview.github.io/?url=https://github.com/blu3nh/sd3-grids/blob/main/portraitphotoofa40yearoldmanwithredhair/index.html

cant figure out how to successfully publish a megagrid to github

[dim sonnet]

make a repo named yourusername.github.io and put stuff in it, for example this repo https://github.com/mcmonkeyprojects/mcmonkeyprojects.github.io

powers ye olde sample grid https://mcmonkeyprojects.github.io/short_example/index.html

[toxic shale]

works 😄
thanks! ❤️

[rigid maple]

what am i doin wrong?

[lavish glade]

definitely need a lora bar that defualts to a second bar when it runs off a page when running too many lora's

[haughty moon]

after a week or so of using swarm i've found a couple of good (to me) reasons I'd like to be able to set the 'starred' path manually.

First, I didn't like that the 'starred' folder was inside my "new-images" folder, so i added "new-images" to the path/filename for newly generated images and moved the default output directory up one level (eg: D:/AI/Output) but now I have a new-images directory inside the starred directory, which isn't what I wanted.

Second, and more important for me, I have my AI stuff installed on a drive (NAS) and want my saved (starred) images to also live on that drive, but I don't want the newly generated images being saved (and then bulk deleted) from that same drive because of wear and tear and general access speed. I'd prefer to have them on a different (local) drive since 99% of them will just get deleted anyway.

I get why the system works the way it does at the moment, but I like to keep things a lot simpler, I click generate and images go into a folder, I click save (star in swarm) and they go somewhere else. Nice and clean and simple

I could, in theory, set the path to a local space for new generations and avoid the NAS issues, but then I'd also have to then manually move things from the starred folder to the NAS for safe keeping, which to me kind of defeats the purpose of a save/star function.

Beyond that, in my own development experience I've found it a good practice to have a setting for all things of the same type rather than only some, for example if you make a user configurable setting for one path, all paths should be configurable.
I also know, as a developer, that this might be low priority, especially with SD3 being the hot topic right now.

[dim sonnet]

you shouldn't have that * at the front

Either * or 0.0.0.0 individually mean "any connection". Together is just an error that it's kinda funny it doesn't actually break with

[dim sonnet]

now I have a new-images directory inside the starred directory, which isn't what I wanted. haha oops

I have my AI stuff installed on a drive (NAS) and want my saved (starred) images to also live on that drive oh so you don't want a different folder name, you want a whole separate path.
... I feel bad saying "go use a symlink about it" re swarm since usually I try to have easy config options instead of system commands, but, ... a symlink is perfect here: https://www.howtogeek.com/16226/complete-guide-to-symbolic-links-symlinks-on-windows-or-linux/

put your outputs folder on the local drive, and symlink starred to a folder on the nas

(to be able to actually route the starred folder to a whole different place without a symlink, in swarm i'd need to refactor how the starred logic works at all. It just is a subfolder atm it's not its whole own path)

[haughty moon]

yeah, I try to stay away from symlinks, but i'll probably have to in this case. If I have some free time I might just dig into it myself and figure it out, I have a little C# knowledge so I might be able to throw together a PR if I can get it to work

[rigid maple]

alright,so how do i get my link to access from another network?

[dim sonnet]

Depends how you're doing things.
If it's over LAN, your LAN address is automatically displayed on the Server Info tab so just copy that

outside of LAN, you need a forwarder like cloudflared (explained in the docs linked earlier)

[rigid maple]

yeah,how do i setup said forwarder?

like i added the -- line

[dim sonnet]

... yknow i just realized, why is this the one CLI Arg option i never made a setting in the settings UI for

hol up

[rigid maple]

am...am i the first person to try doing this aswell 😭

[dim sonnet]

Usually the people using cloudflared are on colab/runpod/a linux server/whatever

you miiight be the first to do it on a home windows computer lol

[rigid maple]

loool,thats sick

can i get a mention somewhere on the ui for this lol,cool flex imo

[dim sonnet]

hfgddhfg

you're not even in the swarm discord https://discord.gg/q2y38cqjNw

[rigid maple]

i am now,never knew it existed lol

just thought it was a channel

[rigid maple]

pretty please

[dim sonnet]

I pushed a new commit that adds a Server Config option for cloudflared path

just set the path into there and restart swarm, then the public URL will be listed under Server Info

[rigid maple]

awesome,thx man

[dim sonnet]

(note might need to refresh page for it to appear on server info)

[dim sonnet]

you need the full path to the program. You can right click on the cloudflared exe file and click "Copy Link"

er sorry, "Copy as path"

[rigid maple]

ohh,gotcha,thanks

[quasi marten]

Hi there. I have a new vanilla StableSwarm install, but getting an error when trying to load the SD3 models. Other models appear to be working as expected.
[Error] [BackendHandler] backend #0 failed to load model with error: System.AggregateException: One or more errors occurred. (ComfyUI execution error: Given groups=1, weight of size [512, 16, 3, 3], expected input[1, 4, 32, 32] to have 16 channels, but got 4 channels instead)
Has anyone seen this before?

[haughty moon]

I got that when I had a sdxl vae set

turn that off for sd3

[quasi marten]

already off.

[tulip wasp]

it's a bit annoying how at the time 0:00 (now), stableswarm suddenly puts images into a new different folder 😂

[dim sonnet]

User Settings -> OutPath -> Format

can customize it heavily there. maybe there should be an option for 'startdate' ie a date stamp that doesn't update live but only updates when you restart swarm?

[dim sonnet]

uhh that definitely looks like a VAE thing

if it's still happening go to Server -> Logs -> Verbose and post the verbose logs from when it fails to load?

[quasi marten]

logs attached

someone helped me in the stableswarm server discover that the sd3 field is missing here.

this is despite it being a fresh vanilla install. (which I have done twice, just to be sure)

[neat valve]

system:
mac m1 pro
Python 3.10.6
when i install,its error

show error:
22:28:32.318 [Warning] ComfyUI Install (Linux Script) stderr: ERROR: python venv is not installed
22:28:32.319 [Warning] ComfyUI Install (Linux Script) stderr: Please follow the install instructions in the readme!

[dim sonnet]

Okay that means one of two things: either it's an outdated install of swarm - make sure you're installing per the readme instructions https://github.com/Stability-AI/StableSwarmUI?tab=readme-ov-file#installing-on-windows

or it's somehow gotten the architecture for the model wrong - make sure in the model browser that the "Type" entry is correctly SD3

[dim sonnet]

you probably need: brew install virtualenv

i think, not sure

need venv

[thick flicker]

How would I set up Pixart in stable swarm ui?

[dim sonnet]

(above question got discussed in the swarm discord https://discord.gg/q2y38cqjNw )

[slow bobcat]

Hi guys, can someone pls gimmi a tip or trick to get these last 2 nodes to install in ComfyUI
I have spent a full day installing nodes to get this one workflow working but these last 2 nodes are just impossible, i cant crack the install / setup process to get them to work.

i have them placed in the same folder as all the other custom_nodes

[dim sonnet]

check debug logs, they might be failing to boot due to pip requirements

Comfy Manager is helpful for getting things installed easily, might want to use that

[slow bobcat]

as i said they dont have pip requirments, at least there is no requirments.txt, but sure ill try get to the logs and look for clues

*** Appologies, it was another chat where i mentioned the rerquirmets.txt was not present, but i didnt mention it here:

"however, these 2 nodes did not have a requirments.txt installation file"

Comfy Manager is this an extention?

[slow bobcat]

this one?
https://comfy.icu/extension/ltdrdata__ComfyUI-Manager

[dim sonnet]

yeah

[slow bobcat]

ill give it a try now, thanks, and yeah these 2 nodes dont seem to have requirments.txt files so i couldnt find any sign of missing dependencies

😎 COOL

[slow bobcat]

Actually it helped, i managed to get the Upscaler sorted and now there is one left, but unfortunatly Manager cant help this one.

its just a sharpener, if only i could just use another sharpener node and cut this painful one out of the workflow alltogether

anyone know if this is possible / valid?

if i just use a sharpener node that i do have, and wire it up to the same thing the Failed node is wired to, could that be a practical workaround here?

[dim sonnet]

ye

[slow bobcat]

Getting so close now i can almost taste victory
I have one issue that im trying to understand

Do i understand this error correctly:

There is a value set as a string, but when its accessed its actually NULL?

Could this mean the string value is the specified model name, and could the NULL simply be that the model is missing?

[dim sonnet]

yes that error probably just means the model is missing

[slow bobcat]

Thanks Alex

[slow bobcat]

i am trying another workflow today

i am stuck on node setup

for some reason i cannot select the clip models, if i click the load checkpoint model selector, i can select from a few models, but if i click the TripleCLIPLoader model slectors i cannot get an options list

anyone know how to sort that out?

[fringe dragon]

@slow bobcat why are you loadinng a checkpoint that includes clips, than loading the clips again ?

[slow bobcat]

im not sure, i just downloaded the workflow from civitai and trying to get it to work

are u saying i must used SD3 model that does not have clips into checkpoint and then load models that are clips into the other CLIPS loader?

[fringe dragon]

well either use the checkpoint without clips than load them, which would be best, or use the checkpoint you have and don't use the triple clip loader, simply connect clip from the load checkpoint node

[slow bobcat]

yes this makes sense, as i understand my checkpoint includes the 3 components that are split across the CLIPLOADER, so i can bypass the CLIP LOADER, thats what i understood from your message. The other part about simply connecting clip, do you mean i rewire whatever is connected to the triple CLIP Loader and rewire them to be connected to the checkpoint rather?

oh hell yeah thats working

THANKS @fringe dragon

[fringe dragon]

yeah u can remove the triple clip loader now, you don't need it

[honest wraith]

i would use the clips and not the merged checkpoint if your machinery is capable, much better results

[crimson hinge]

@dim sonnet Thanks for answering my question on GitHub. But it does tile on x only or y only too?
does it have batch processing from images?
I use something like this similar to this video to draw my doodles and create images as those as reference
https://youtu.be/Ly6USRwTHe0
Also can you modify file paths for the models loras and etc?

Currently I use A111.
Sorry to ping you but A111 is driving me nuts I can't use SDXL with it or make 1024 images. lost 2 days figuring that out.
Lasty question is can I not use comfy workflow at all but just use word prompts to get similar results?

[dim sonnet]

jsyk swarm dedicated discord @ https://discord.gg/q2y38cqjNw

[dim sonnet]

if you run SD3 in swarm's Generate tab, the first time you run it will autodownload CLIP files for you

[dim sonnet]

But it does tile on x only or y only too? yes
does it have batch processing from images? yes
there's an image editor integrated into swarm if you want to use that
Also can you modify file paths for the models loras and etc? yes
Lasty question is can I not use comfy workflow at all but just use word prompts to get similar results? yes the main generate tab is similar to auto in how you use it

[crimson hinge]

btw kind of offtopic but what courses you took to be able to make something like this

i am studying AI development actually

[mellow pendant]

hey Mr Mcmonkey, thanks for keeping a level head and i wish you success on your endeavors

[autumn forge]

swarm rocks, keep being great

[pine gazelle]

画一个苗绣的图片

[main grotto]

where do i set revision strength. i never set it and yet it's telling me i need to set it to 0 for sd1.x i don't see it as an option anywhere

it's only when i use an image prompt

[dim sonnet]

yes it's only when you use an image prompt, it's a parameter at the top left that pops up when you have the image prompt

[main grotto]

i don't see it
tried searching in the param filter

[dim sonnet]

ohh heck you have legacy promptbox, i haven't tested that in forever

[main grotto]

oh so i should change ui?

[dim sonnet]

... yeah go reset your parameter configuration back to default

the more i look at it the more things i realize you changed that aren't meant to change like that

[main grotto]

how do i batch do it or do i need to reset each grouping on it's own

[dim sonnet]

click the reset button on anything in here that you edited

[main grotto]

i've forgotten which i messed with.... i'll do what i can remember

um... stupid question now. but where do i put in an image prompt now. i don't see it anywhere other than init image

[signal forge]

is there a guide for amd on linux? I confirm my rocm is working and detecting my card, but the launch script still tries to find nvidia

[cold flame]

AFAIK you'll need to install Comfyui manually as a standalone install and create a python venv in the comfyui directory. Make sure all of the comfy requirements are installed to the venv along with the rocm versions of torch, torchvision, and torchaudio.

Once that's set up, using it with Swarm should be as simple as adding a ComfyUI Self-Starting backend in Swarm and in the StartScript field enter the path to .../ComfyUI/main.py

[signal forge]

ah, I never thought to look for the venv, I dont use them since I run docker environments. I'll try deleting that first, and if that fails I'll try your recommendation of doing comfy first. Many thanks

[cold flame]

Coke

idk much about running it inside docker, so you might wait for someone who does if that's how you'd prefer to have it set up