🐝｜swarm-ui

[mystic granite]

I think for just gaming locally

a 4090 and 5950x3D is fine

overkill even

[dim sonnet]

does comfy just run out of the box on a tpu

[mystic granite]

nyo

I had to adjust a bit of it to run on ubuntu

but it was fairly easy

the main issue is that my tpu doesn't utilise ports the same way

as a non server machine would

that's mostly a my server problem

regular pcs shouldn't have this problem

plus im accessing it remotely

from a secure server

in google hq

so naturally they'd be up in arms at me if i ever dared to use open ports

I'd definitely set off some sort of alarm

[potent panther]

Sounds like a hardware issue

32 gb has never left me frozen

Or you just have some kind of ram leak in a software or too many random apps open

[dim sonnet]

too much open is deff a thing

my PC idles at 20 GiB in use

but i have 64 GiB now so i don't mind

... not that i have any clue what's using that, by process list i should only have ~5GiB in use. Probably windows caching going silly

[mystic granite]

@dim sonnet can it segment nonstandard things btw?

like segment:backpack

or something

[dim sonnet]

yep

you can <segment:the dog's face> to get only the dog's face and not the cat's face even (*might not be super reliable depending on how smart clip is about it)

[mystic granite]

ooo

that's real neat!

[cold flame]

Have you considered adding an option to use yolo for face/person bbox detection before ClipSeg so that you could give ClipSeg an already cropped image focused on the desired subject? Similar to how FaceDetailer from the Impact Pack works. Obviously it wouldn't be as flexible as the current method so you'd probably want to have both available, but would probably help with reliability on human faces or full bodies at least. I can send an example workflow with the Impact Pack FaceDetailer setup if you aren't familiar with it and/or I've explained poorly.

[mystic granite]

hi nubbs!

[mystic granite]

if i do segment:hands does it do both hands? or do I have to do each individual hand?

[dim sonnet]

i haven't seen any reliability issues on non-silly inputs so i haven't looked into it

[dim sonnet]

so, gonna respond to this with a general magic trick to use here actually - first, set up your input, in this case just write your prompt

then go to the Comfy tab and click Import from Generate Tab

[mystic granite]

Base XL only?

[dim sonnet]

then scroll in and find the part you care about, in this case the ClipSeg node

then just toss in some previews and run it

[mystic granite]

im working with anime related images

[dim sonnet]

in this case we can see ... lumberjack only had one hand, ow, terrible accident, unfortunate

but also both his hand and his stump were matched

[cold flame]

It might be more reliable on more realistic images, but for anime style at least, <segment:face> fairly often misses at least some of the face for me. Haven't measured exactly, but feels like it fails to capture the full face around 15-20% of the time for me even on images that should be relatively easy, such as a single subject facing the observer on a white background.

[dim sonnet]

this is one of the beautiful things about the swarm<->comfy integration - you can do things in the easy interface but you can also get the detailed graph to muck about the inner workings any time you need to

[dim sonnet]

... i kinda wonder if <segment:anime face> would work better

but also yeah if that's the case then sure we can add a yolo option

didn't they just publish a v9?

[mystic granite]

i think this looks alright

also hi nubbs!

[cold flame]

Oh yeah, I suppose I could take the default node graph and just slap the impact pack FaceDetailer nodes on the end and make the necessary connections to keep it functioning mostly like default Swarm, since I don't usually bother with feeding it a specific prompt anyways.

I'll try <segment:anime face> next time I'm messing around in swarm to see if that's more reliable for me though

[cold flame]

hi blobbs!

YuukaWave

[dim sonnet]

oh for clarity the message you're replying to there was for blob not you

re editing the graph to debug what the segmentation results are

[mystic granite]

does swarm have a tag grabber?

where u chuck images into the box

then it generates tags

[potent panther]

I tried and it's worse

segment:face_yolov8m.pt could be fine as the syntax

[dim sonnet]

not currently no

[mystic granite]

o ok

[mystic granite]

time to add this to the feature list

[dim sonnet]

can't do that right away, so could you please post a feature request @ https://github.com/Stability-AI/StableSwarmUI/issues

[mystic granite]

okey

[mystic granite]

done!

[potent panther]

bruh

[potent panther]

Oh the spam is gone 🤔

[torn grove]

gm

[shadow ivy]

when i'm trying to download it i have this error :

App: C:\Users\user1\Downloads\StableSwarmUI\src\bin\live_release\StableSwarmUI.dll
Framework: 'Microsoft.AspNetCore.App', version '7.0.0' (x64)
.NET location: C:\Program Files\dotnet\

The following frameworks were found:
  8.0.2 at [C:\Program Files\dotnet\shared\Microsoft.AspNetCore.App]

Learn more:
https://aka.ms/dotnet/app-launch-failed

To install missing framework, download:
https://aka.ms/dotnet-core-applaunch?framework=Microsoft.AspNetCore.App&framework_version=7.0.0&arch=x64&rid=win-x64&os=win10```

bruh what have i done :

App: C:\Users\user1\Downloads\StableSwarmUI\src\bin\live_release\StableSwarmUI.dll
Architecture: x64
Framework: 'Microsoft.NETCore.App', version '7.0.16' (x64)
.NET location: C:\Program Files\dotnet\

The following frameworks were found:
  6.0.22 at [C:\Program Files\dotnet\shared\Microsoft.NETCore.App]
  7.0.5 at [C:\Program Files\dotnet\shared\Microsoft.NETCore.App]
  8.0.2 at [C:\Program Files\dotnet\shared\Microsoft.NETCore.App]

Learn more:
https://aka.ms/dotnet/app-launch-failed

To install missing framework, download:
https://aka.ms/dotnet-core-applaunch?framework=Microsoft.NETCore.App&framework_version=7.0.16&arch=x64&rid=win-x64&os=win10```

[potent panther]

You're missing the right dotnet version

You installed 7.0.5 but StableSwarm currently needs 7.0.16

[dim sonnet]

use the installer in the readme https://github.com/Stability-AI/StableSwarmUI?tab=readme-ov-file#installing-on-windows

[shadow ivy]

i've done that already

[dim sonnet]

the install script automatically installs dotnet for you

and it doesn't look like you have any automatically installed .net there, it looks like you tried multiple times to manually install it in different ways

[shadow ivy]

how can i add loras to stableswarm ?

[dim sonnet]

toss the files in StableSwarmUI/Models/Loras and then select them in the "LoRAs" listing at the bottom of the generate tab

[shadow ivy]

but idk where my stableswarm is installed lmao

[shadow ivy]

ty

[dim sonnet]

^ as shown here, C:\Users\user1\Downloads\StableSwarmUI

[shadow ivy]

ty

[dim sonnet]

Feature announcement! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-8736304 Comfy workflow browser!

[dim sonnet]

Also! Swarm is now officially finally in Beta! https://github.com/Stability-AI/StableSwarmUI/releases

[obtuse tinsel]

awesome!

i'm going to have to check out the progress on this soon

[dim sonnet]

https://new.reddit.com/r/StableDiffusion/comments/1bbgsl8/stableswarmui_beta/ announced on reddit

[potent panther]

You might want to ask the owner of this subreddit since it should be yours mcmonke
https://new.reddit.com/r/StableSwarmUI/

[mystic granite]

@dim sonnet can we add DRAGNuwa functionalities?

[fluid ice]

Is there any plan to just distribute the pre-compiled binaries (StableSwarmUI.dll)? Can't run MSBuild because of my security policies and MSBuild is on Microsoft's list of recommended blocked files for such policies 😩

https://learn.microsoft.com/en-us/windows/security/application-security/application-control/windows-defender-application-control/design/applications-that-can-bypass-wdac
"Unless your use scenarios explicitly require them, Microsoft recommends that you block the following applications."

I could build it in a VM but it's kind of a pain

[rapid garnet]

When putting the trigger phrase in a lora/model, is there a way for that trigger phrase to be automatically injected into the prompt instead of me doing the whole "copy paste" routine

like with Pony for example, the trigger phrase is usually score_9, score_8_up, score_7_up, at the beginning, is there anyway if i put that in the "trigger phrase" box for the model, it will automatically be in the prompt at the start?

[mystic granite]

or u could just

unblock it

[fluid ice]

Akshually 🤓👆 you cannot. If you disable smart app control you can't re-enable it without reinstalling windows completely

The idea is it can be assured that no untrusted software has ever run

It's admittedly inconvenient at times but after more than one malware scare I've gone full unhinged on security

[dim sonnet]

ye is listed as planned in the readme

i'm surprised you can't run msbuild but can run python executors

part of the reason it distributes as source code and not compiled binary is to reduce concerns around malware lol

you can trust nothing is shady cause you can see all the source code and know it's exactly what's running

[dim sonnet]

that's a planned feature yes, not right now no - but what you can do as well is create a Preset that has the trigger word and optionally even the lora/model itself set in the preset, and then just select that preset any time you want to use it

[fluid ice]

Yea saw that was planned. I didn't necessarily mean create a whole exe installer, just compile the binaries and have them available for download to copy and paste. You could also sign them that way

[dim sonnet]

that'd be a whole interface on its own to develop - it'd be really cool to have tho ye, could you please post an issue @ https://github.com/Stability-AI/StableSwarmUI/issues to track that feature request

[dim sonnet]

idk if a whole separate subreddit is necessary? if it starts seeing much use i'll talk to the guy that made it

[potent panther]

¯_(ツ)_/¯

[runic spire]

none of my workflows now load with in the workflow browser

gotta use front end for some, others I am dead in my generation tracks using swarm with comfy back end

[potent panther]

maybe a custom node broke something?

[dim sonnet]

what's this bit from

[runic spire]

i found out way,,( i use stablity matrix --- Earlier I installed that olloma package from openarts -- it breaks the swarm and comfy packaging. It would be a cool feature to have, can't use it now.
manager shows up now after i deleted the ollama dir in comfy extend nodes dir, swarm now loads workflows

both working great again.

[potent panther]

you should open an issue there

[runic spire]

i just did

[potent panther]

nice

[spiral arch]

intriguing, thanks for demo

i think what would be nice is something like this combined with comfyui in another tab

[dapper zodiac]

hey all! I've done a couple of stabel swarm installs, all great, except im not sure where to put pre-processors for control net. I only ever seem to have Canny available. I'd tried the windows bat file install method, with both the defaul config and custom, and the exact same issue.

[toxic shale]

I'm angry at civitai, for still not adding the metadata parser for stableswarm
https://feedback.civitai.com/p/support-stableswarmui-image-metadata
/rant over

[potent panther]

if there's no support for their useless hash system they won't care

if the post was brought to their discord mods it might get their attention

[dim sonnet]

huh?

[dim sonnet]

sorry it's slightly nonobvious atm, will be easier/automated in the future -- install this comfy nodepack https://github.com/Fannovel16/comfyui_controlnet_aux to (Swarm)/dlbackend/comfy/ComfyUI/custom_nodes and restart swarm

[dim sonnet]

i'm still upset that they haven't integrated automatic safetensors conversion despite me repeatedly begging them for years across their 3 different feedback platforms https://feedback.civitai.com/p/automatic-conversion-to-pruned-fp16-safetensors

This would limit checkpoint models to 2GiB, as opposed to 4 or 7 GiB. Other model types would also have (relatively) small size. lol the text of that post is so old it predates the existence of SDXL

[toxic shale]

XD was a bit confused till I realized that yeah... this is pre SDXL

[jagged cipher]

Silly quesiton: How do you inpaint?

I generated an image which I ~mostly~ like, and want to have it generate the eyes and such again, but can't figure out how to actually mark the eyes as the specific thing to inpaint

I've tried what was mentioned #🐝｜swarm-ui message but it didn't work - not sure if things have been changed or I'm doing it wrong or what.

[dim sonnet]

that should still be correct

[jagged cipher]

Hm. I think I'm missing something incredibly obvious then.

Ah, poked at it a bit more, I think the masked version isn't getting passed to the init image when I hit Use As Init?

How do you actually 'save' your image edit?

[dim sonnet]

oh yeah Use As Init doesn't work

you just generate with the editor open atm

[jagged cipher]

Huh. When I'm doing that, I just get a black box. Does the 'black' mask layer need to have no opacity, or...?

Yeah, no, I'm clearly doing something wrong lol

Okay, I think I finally figured it out lol, thank you!

I didn't realize that the auto-generated additional layer was an image layer, not a mask layer, so whenever I painted it black, it thought I was using a black box image lol

[ivory pulsar]

StableSwarm is not reading any of the metadata for my models/loras. Im mostly wanting the trigger words of the 100s of loras. I also tried to load them with ComfyUI-Custom-Scripts LoraLoader , with no luck. Hangs up on Loading. New to stableswarm and comfyui. All of these were previously processed with civitai helper in a1111 if that makes any deference.

[nimble belfry]

We talked about that in the past but I don't remember what the end result was. Search in this channel for that conversation maybe there is a hint in that.

[dim sonnet]

Swarm reads metadata in standard ModelSpec format, or in a few common formats

civitai-helper in auto webui probably has its own format

if you send me references for exactly how it formats metadata I can implement it

[ivory pulsar]

Ok still struggling here. Deleted all json files and started over, a few times. latest attempt has been with civitai-browser extention for a1111.

[ivory pulsar]

here are some more examples

Original Civitai-helper just gives a .info
https://github.com/butaixianran/Stable-Diffusion-Webui-Civitai-Helper

Forked Civitai-helper gives json and .info
https://github.com/zixaphir/Stable-Diffusion-Webui-Civitai-Helper

[dim sonnet]

Neat, okay -- I pushed a commit, so update swarm (hit the update-windows.bat) and then relaunch and go to the Utilities tab and hit the reset metadata button

and it should hopefully recognize your civitai model data now

[versed copper]

speaking of which, any chance for the model cards to display SM's info from civitai, which is HTML code, rendered instead of text file style with all tags?

[ivory pulsar]

Fixed! Thank you 🤘

[dim sonnet]

rendering HTML directly would be a security risk - and not a hypothetical one either, one that already happened to auto webui from exactly this source

however it is possible to parse html with a whitelist of allowed tags (ie basic text format) and render only those

[versed copper]

right, can see that's a risk

[ivory pulsar]

How bout some , Sort by : date modified - ascending/descending , in all the model tabs. Really helps find those newly downloaded models in the sea of existing models.

[dim sonnet]

oo, valid point

[potent panther]

for some reasons while segment is running it will automatically switch to previews even though the option is disabled

[dim sonnet]

on a quick test, can't replicate

[potent panther]

🤔

@dim sonnet here's what's happening

it switches to the segment preview every time it finishes a new step

[dim sonnet]

oooo

specifically when save intermediate images is on

[potent panther]

oh ye

I'm refreshing

[dim sonnet]

you have found the cursed result of a messy design choice i made many months ago and i'm trying to figure out how to unbreak it without deleting a large chunk of code and redoing it

atm it doesn't understand that the intermediate preview is a different image

and i need to figure out how to tell it that properly

[potent panther]

ooooh I see

[dim sonnet]

oh wait past me solved this problem before

yep that works perfectly. Great work past me, should've asked me sooner

fix pushed

[potent panther]

nice

that was an easy fix

btw are you gonna rename this at some point?

I think you said you were going to some time ago

[dim sonnet]

already did?

have you not updated in a while

https://github.com/Stability-AI/StableSwarmUI/commit/8ed52d02d222310eb58ed30fde463747953e47a1

[potent panther]

I git pull all the time

hmmm

My swarm install might be broken

restore and reset did nothing

do the config files have something to do with it?

[dim sonnet]

not sure how any configs would be relevant here

only thing even potentially relevant is if you used the parameter customization thingy, but even that I don't think can control groupings

is the git pull working? and pulling latest changes?

[potent panther]

yeah I have everything

[potent panther]

I have used it once but I didn't touch anything other than the swarm internal section

[thin coyote]

Do you guys think it would be a good idea to add a simple text message which will be displayed when the user starts inputting custom folder locations for models so that it will tell him when entering a custom directory whether it exists?

Something of the sort of:
"Directory already exists"
"Directory doesn't exist"

or just add the option to directly select a folder using the file explorer pop up menu

[potent panther]

do web browsers have such an api?

[unreal warren]

Is it possible to mass upscale with the "Upscale 2x" Button/option somewhere?
With that button upscaling is much better than all other i found.
Would be nice when the Tools Section has some Upscale prepared.

[dim sonnet]

this is my pain: Yes but i'm not allowed to use it >:(

https://developer.mozilla.org/en-US/docs/Web/API/Window/showOpenFilePicker

missing from some browsers and requires an HTTPS connection which doesn't apply naturally to localhosty stuff

also "some browsers" is a funny way of saying only desktop chrome has it and nothing else

[dim sonnet]

if you want to do a bulk/automated image manipulation task, use Tools -> Image Edit Batcher

the "upscale 2x" button just sets the res to 2x the input res and feeds the image as an Init Image

[unreal warren]

Thanks for the answer.
I try that like in the pic and set the resoltution to that i want, but he do upscaling ... but not that it should. it generate new ones bigger.
or do i something wrong?

[dim sonnet]

huh?

[unreal warren]

it should use the parameters from the file but the resoltion as maybe textfield (2x)
I use it as in the pic and let the prompt it was make, set resolution to the 2x of the orginal and press batch ... but it dont upscale what there.
Left the input and right the output ... the upscale button make only a upscale and near no changes.

and with the "Upscale 2x" Button it make as many as in Images are ...
I absolut love that software 🥰 so easy to generate awesome photos.

[potent panther]

Their English is very much not there

[thin coyote]

@dim sonnet I have a problem with stable swarm

Specifically with the controlnet preprocessors

[dim sonnet]

...?

[thin coyote]

I use a custom folder on my drive where I store all my controlnet 1.1 models and have given the appropriate folder location

even stable swarm is recognising the models

but it just keeps saying canny or none for preprocessors

no matter what model I choose

everything else is left alone except the modelroot

do I need to manually install all the preprocessors since in automatic1111 they usually get auto installed if you do not have them

[thin coyote]

https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors

here are the models I am using

[dim sonnet]

oh, yes, sorry, making preprocessors easily is on the todolist - for right now you gotta install https://github.com/Fannovel16/comfyui_controlnet_aux into Swarm/dlbackend/comfy/ComfyUI/custom_nodes

[thin coyote]

Would it be an inconvenience to add a small section in the README to mention this method of installing controlnet as a temporary fix?
https://github.com/Stability-AI/StableSwarmUI

[near leaf]

Hello smart people! I'm working with stable diffusion API, and at the moment I'm having an issue with controlnet, maybe someone could help here 🙂
My issue is regarding the openpose unit. When I'm creating my payloads, i'm decoding the input images like so :
`
if N is not None:
totalDict.append({
"control_mode": "Balanced",
"enabled": True,
"guidance_end": 1,
"guidance_start": 0,
"input_image" : encode_file_to_base64(N),
"input_mode": "simple",
"is_ui": True,
"loopback": False,
"low_vram": False,
"model": "control_v11p_sd15_normalbae [316696f1]",
"module": "none",
"weight": NWeight,
})

`

Which works great for all the different units (depth, normal, segmentation and so on)

But for whatever reason, when trying to do so with the openpose module, i'll always have those errors :
`
*** Error running process: C:\ai\stable-diffusion-webui\extensions\sd-webui-controlnet\scripts\controlnet.py
Traceback (most recent call last):
File "C:\ai\stable-diffusion-webui\modules\api\api.py", line 96, in decode_base64_to_image
image = Image.open(BytesIO(base64.b64decode(encoding)))
File "C:\Users\twist\AppData\Local\Programs\Python\Python310\lib\base64.py", line 87, in b64decode
return binascii.a2b_base64(s)
binascii.Error: Incorrect padding

`

Has anyone ran into this kind of issue?
And if this is not the best place to talk about it, apologies, this is the closest i found 😅
But would be happy to go somewhere else where it's fitting more and ask!

[dim sonnet]

why would the Swarm channel be the closest place to ask about ... auto webui's API I think that is?

[potent panther]

I think they have no idea what is StableSwarm and just read swarm and thought api or something

[potent panther]

button

[stoic aspen]

can you use stavle video 3d in swarm?

[runic spire]

The workflow browser losses workflows in the browser. Here I try load a workflow "dreamydiff" and click on the empty space, nothing happens -- quick load works fine. I tried to load two of them to show the bug.

[runic spire]

They were saved without images except one.

[toxic shale]

just to add here. if you do add a parser, copy the regex command from any of the commercial wysiwyg editors, as most parser examples that can be found online are pretty insecure against competent malicious actors that know what they're doing. too many browser specific hacks out there x_x

[dim sonnet]

insofar as it's just an SVD model, yes.

the rest of the 3d stuff you'd need some custom comfy nodes or something for

[dim sonnet]

That's... weird. What makes those unique?

Maybe something with the filenames (eg space prefix/suffix that's confusing filesystem)?

considering they can't even load icons

or were they manually modified files?

or new files or particular old files or...

[dim sonnet]

i am absolutely the hell not going to use regex for something like that, ow

anyone using regex for that is begging for exploits to happen

[toxic shale]

welcome to the world of webdevs. I cried everytime.
I saw "search and replace" in some open source projects x_x

[dim sonnet]

made it not cut off quick so bad

still allowed it to cut off a tiny bit

maybe should make buttons linewrap? idk

have people actually assigned to help with the UI/UX stuff getting oriented soon so ill make it their problem :3

[hoary marlin]

it there any runpod template for stableSwarm

[sacred mason]

what is your goal?

do you want to generate images for all the loras you have?

[potent panther]

lol I got on the channel and discord suggested polls to me

now mcmonkey can make polls

[ivory pulsar]

My goal was to see the metadata (tags, description, ect) for checkpoints/loras in the model cards. My issue was quickly patched and works great.

[ivory pulsar]

I do still have a problem with ComfyUI-Custom-Scripts LoraLoader not reading reading metadata of loras though. It works fine if the loras are in the root lora folder, but if they are in a subfolders it no longer works. Not sure if thats a Stableswarm issue or ComfyUI-Custom-Scripts.

[dim sonnet]

if it's a comfy node misbehaving while in the comfy tab, it's an issue with that custom node

for the duration of time you're in the comfy tab, swarm is 90% hands off

until you interact with the swarm-specific tools at the top

(or use the swarm nodes lol)

[cold flame]

I wonder if that is something similar to the issue I was having with the autocomplete script not working at all in the nodegraph tab of swarm. https://github.com/pythongosssss/ComfyUI-Custom-Scripts/issues/179#issuecomment-1993329105

[noble plinth]

So, any way to launch the swarm with commands like --listen and the like? I got other apps that use SD

[dim sonnet]

you can edit any settings in UI rather than having to hack command line stuff

if you need to make swarm accessible off the same machine just set the host ip to 0.0.0.0. This is selectable as an option to share on lan in the installer screen

(there are also command line args if you need them for a remote server)

[potent panther]

sworm!!!

[dim sonnet]

wait this is very delayed, but -- do you git pull literally? if so you're not updating, the update scripts do more than just a git pull

[potent panther]

oh

teehee

[brave wedge]

i generated a bunch of images in the comfy workflow editor of stableswam ui, but it looks like the "SwarmSaveImageWS" node does nothing to save them?

and it looks like the temp folder got deleted when i launched it?

didnt even move it to trash

[dim sonnet]

Save WS saves to websocket ie it saves over network to swarm. If you ran direct in the comfy tab (not through swarm) then it behaves akin to the PreviewImage node

[torpid harbor]

guuuyyyysss! I was just talking about how this is so desperately needed!

Question: are there instructions for installing it and connecting it to your exisiting Comfy so you dont have to reinstall comfy and everything?

[dim sonnet]

Run the installer as normal and when it asks what backend, select "None", then click into server -> backends and add a "comfy self starting"

and it'll give you a this thingy

fill in "StartScript" with the path to your install, with /main.py on the end, optionally edit other values if you want, none others needed

and hit "Save" and go

OH and Server configuration -> "Paths" header, set the model paths to your existing models folder, and restart swarm to apply that edit

[torpid harbor]

thank yoooouuuuuu

[waxen moat]

yes, feeling the same 🙂

[waxen moat]

it's odd, my outputs aren't making it past the temp directory. they never make it to output

I tried both regular default output directory and overriding it in ExtraArgs, but same result

[brave wedge]

noted 😦

[waxen moat]

re-creating the output nodes seems to have fixed this behavior

[sacred mason]

what are you trying to do?

[floral lodge]

I am getting an error when using from the "Import From Generate" in the Comfy Workflow Editor. If I have an image in the Init Image in the Generate tab it attempts to create a StableSwarm custom node. When running workflow in the Workflow Editor it errors out. I double checked the base64 input is valid. The issue is that the node itself is receiving truncated input. I would like to get to the bottom of this and submit a PR. I thought it best to say hi and make sure such a thing would be welcome.

[dim sonnet]

yes unfortunately the Comfy litegraph itself truncates the value :(

I think the only valid fix from the swarm side would be discard the b64 node and use the loadimage node + autoupload an image

(there's already code to enable this it's just not used for that currently)

[floral lodge]

Interesting. That's a shame, I kinda like the node as I have seen the base64 is a bunch of UIs and it feels right. I'll assume it is pretty involved and I should probably tackle something smaller, but still willing if it is something that should get done.

[runic spire]

In the comfy workflow "Save Workflow - Browse Workflow --MultiGPU - Quickload " section, could we have a Export All feature? If swarm gets corrupted, and or we install on another machine --it would be nice to be able to import all or one workflow instead of image saving and going to duplicate workflows off machine a to b

[dim sonnet]

they're saved as files under Swarm\src\BuiltinExtensions\ComfyUIBackend\CustomWorkflows

[tulip lion]

how would I configure stableswarm-ui to run with a lowvram option?

[dim sonnet]

the comfy backend automatically manages VRAM usage without you having to configure it

[tulip lion]

oh interesting ok

[toxic shale]

how do I load in multiple loras into infinity grid?

from the image, only the first lora list gets loaded in. the second list stays at empty 😦

same is true, if 1 or multiple loras are already selected, and then in the grid I add a list of loras. then whatever was preselected is ignored, and only the lora list is loaded in

I think I see how it was intended now x_x just tried with the || separator... but oh god that would require a lot of manual writing when I need to cross check lora interactions
@dim sonnet could you maybe add this feature, that LoRAs selection can be used multiple times?

[dim sonnet]

uhhh that should probably be added yes but for rn you can do the auto webui style <lora:myloraname:1> <lora:lora2:1> and then Prompt Replace

lil awkward/hacky but does the ting

[toxic shale]

no, thats pretty much perfect ❤️ my solution was to go to A1111 and do it there XD so this works better than that

[marsh glade]

I've got inpainting down. Add a mask, identify areas for new prompt to impact in white, areas that you want to be untouched black. But for the life of me I can't figure out outpainting... how can I extend the image generation beyond the borders of the original image? It won't let me paint beyond the images borders

[nimble belfry]

Someone can correct me if I'm wrong but I don't believe that feature has been implemented yet.

[somber slate]

how to fix this, i install swarm and i replace my comfyui old

[dim sonnet]

Huh? What did you do to the comfy backend?

[somber slate]

i new install swarm

how to install swarm after install comfyui

[nimble belfry]

Best to let Swarm install comfy for you.

[potent panther]

I think you didn't add the right start command

it should be something like \path\to\comfyui\main.py

[somber slate]

thanks sir

[spiral flume]

Does anybody here know, when can the "external access,Public Access " options are live?

These

[nimble belfry]

@dim sonnet I updated and all of my bots broke...What have you changed in the past few weeks that could have caused this? Maybe something with get_session?

3/28/2024 12:40:07 AM [Verbose] Rest: POST channels/1105955639907328101/messages: 479.18 ms
Error in PostJson: An error occurred while sending the request.
Error in GetSession: An error occurred while sending the request.
3/28/2024 12:40:07 AM [Error] App Commands:
Exception: Discord.Interactions.InteractionException: Error occurred executing generate.
 ---> System.Net.Http.HttpRequestException: An error occurred while sending the request.
 ---> System.Net.Http.HttpIOException: The response ended prematurely. (ResponseEnded)
   at System.Net.Http.HttpConnection.SendAsync(HttpRequestMessage request, Boolean async, CancellationToken cancellationToken)
   --- End of inner exception stack trace ---```

[nimble belfry]

wait, it looks like my port changed.... why....

[dim sonnet]

Did it go up by 1? Refer to this server setting:

you'll want to turn that off for instances of swarm used as an API

(and it will just error in your face when you try to launch with an occupied port instead of autocompensating)

[nimble belfry]

Yep that was the cause. I have disabled that now.

I would much rather get the in my face error.

[somber slate]

after install comfy browser, comfy browser just run on port 7822 but in swarn port 8188 cant run. and how to fix this error

[dim sonnet]

update swarm, you're probably seeing a bug i just fixed where one of the ipadapter node packs updated to be incompatible with swarm's ipadapter detection

[verbal robin]

I tried the workflow editor in InvokeAI, in which you can choose what you want to see in the linear UI by right-clicking the fields in indivisual nodes in the workflow; all the other nodes and options do not show up in the UI. Is there any plan to implement the same in the StableSwarm UI?

[dim sonnet]

you can use the old system, which is, make Primitive nodes and title them, and those are your main inputs and the rest aren't

-or- you can use the brand new (literally today, not even complete yet) system, of using SwarmInput nodes to fully customize/control the friendly frontend UI you get

[verbal robin]

How to install SwarmInput nodes in my own comfyUI?

[still crater]

when clicking on edit image in the ui after a generation, how would we go about inpainting there?

[dim sonnet]

as long as you launch Swarm properly and don't go out of your way to enabled advanced alternative options and do a custom external comfy install, any up-to-date Swarm instance has the swarminput nodes

(if you have a heavily customized things and you're not aware of the swarm nodes, you've been doing it wrong and should swap asap)

[dim sonnet]

an issue was posted for that https://github.com/Stability-AI/StableSwarmUI/issues/282 so track that issue, will close that when it's added

[dim sonnet]

image editor is a lil awkward atm --
Click "add mask", paint it all black with the brush tool (ie black for no-change)
click "add mask" again and then paint white where you want to edit
then generate

[still crater]

Okay glad that's the case because I was driving myself crazy trying to figure it out. I

Is there any documentation about these types of things somewhere I'm missing?

[nimble belfry]

not for that I assume because its still a WIP.

Best bet is to search this channel chances are the question has been asked before.

[carmine cloak]

Hi, can i change the "initimagecreativity" Level for the Upscale 2x Button? Cant find it. would be nice if i can change this over the UI

[carmine cloak]

ok if I enable the Init Image Section , the parameter from their is taken. Not really self-explanatory

[silver glen]

In terms of contributing back to this project, are there any guidelines for browser support?

[drifting helm]

i think just "dont break other browsers" or "pls on all browsers"

[silver glen]

Alright, so no old IE support or anything like that? Just the usual "works on all modern browser"?

[drifting helm]

yea

[silver glen]

Cool. Really wanted to toy around with better mobile support if I can actually find the time.

[dim sonnet]

yeee

[dim sonnet]

ooo

i have half a vision for how it'd work on mobile

notably: the lil sidebar thingies, like the parameters sidebar, and the model/history bottom bar, become lil clickables to pop open and push away

ie focused fully on the image + prompt box for main usage

and support for things like long-press the generate button to get the bonus options hidden by right-click normally

[minor jungle]

a tiger in the desert

https://tenor.com/bjX37.gif

[gleaming tulip]

no

[fervent elbow]

idk if this channel is for developer only or not, so i apologize if this is the wrong place to ask

i dont know how to use comfy so just using the autoui for swarm. is there any way to get segment:face to have better detection on anime models? it always gets like the bottom half of the face and not the eyes

[gleaming tulip]

man what's that pfp 😭

[mighty socket]

hey everyone, how do i get more samplers in stable swarm
i dont have dpm++ karras for example

[fervent elbow]

😏

[dim sonnet]

that's done like so:

[gray geyser]

Is there a way to change upscaling parameters? I've noticed that on some images the end result moves away too much from the initial image.

[mighty socket]

found it thanks man, i thought dpmpp_2m karras was one thing. didnt realise karras is the scheduler

[mystic granite]

is it possible to install comfy manager on swarm? @dim sonnet

[dim sonnet]

yes

[gray geyser]

related to that question, is it okay to update comfyui using manager update mechanism?

[nimble belfry]

Comfy updates when you start swarm.

[potent panther]

I git pull manually because I like looking at what the new commits are

[cold dirge]

is anyone around that can tell me where install comfy manger in swarm i dont see a folder for where it goes

🤔

i found it i think

[gray geyser]

Should be similar to E:\AI\StableSwarmUI\dlbackend\comfy\ComfyUI\custom_nodes

[wet locust]

I'm looking to accelerate comfyui workflows by running them on multiple GPUs. Have been hoping to do this via comfyui within Stable Swarm UI, but have been struggling to get it use a second (or more) GPU. Any guidance or suggestions would be very welcome please

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md

[wet locust]

Hi @dim sonnet , thanks very much for the link (was hoping you might answer 😀). I've already read through this documentation.

I'm guessing that the instructions that I want are at the very top or bottom of this page. But I am not quite finding a successful solution. I've added a second backend, but it is always disabled (?). I want to speed up the process of a single image/video generation, not run multiple images.

[dim sonnet]

Unfortunately there's no good way to accelerate a single generation across multiple GPUs publicly available

there have been some experimental methods published recently, but nothing you can just set up and go as of yet

[wet locust]

@dim sonnet ok, thanks very much for the info, much appreciated.

[versed copper]

re: any modern browser, no, palemoon being an example. unreadable font colour.

[dim sonnet]

huh?

[versed copper]

black font on black background

for some reason

[dim sonnet]

wtf? why

does palemoon override the css?

lemme go install palemoon lol

[versed copper]

prompts not legible

[dim sonnet]

okay ow palemoon's design is... wowsa

[versed copper]

😄

reason for using: it's compatible with the old plugins, and has the old search field

[dim sonnet]

this is surprisingly laggy too

is this maybe just a case of, palemoon despite being updated isn't actually compatible with modern web standard

oh yeah it's a lot more than the text that's broken

js is all broken too

[versed copper]

i wasn't really complaining, just saying 🙂

PM also doesn't like civitai, i've learned to live with slight problems

[vivid geode]

Hello guys. I need some help. I have a comfyUI workflow that i want to use on swarmUI. My workflow has several nodes that i dont want the end user to interact with, is it possible to hide them from the swarmUI interface or not show them and show only a few specific nodes?

[potent coyote]

For some reason, every stableswarm generation looks different to comfyui. I'm not sure if this has always been the case or if it's a recent change 👀

[tame bronze]

SAM detector is not working for me when i try manually right click to detect parts of image but it works when i launch comfyui on its own can someone please check on your end if its a bug or just me? thanks

[dim sonnet]

if you add the new SwarmInput nodes, those will be listed as inputs in swarm and everything else will be hidden

[vivid geode]

Oh really

[dim sonnet]

You can literally import between swarm and comfy directly in the Comfy Workflow Editor tab, so you can see exactly what swarm generations do, and you can force whatever comfy gens you like to work in swarm interface, at will.

If you're just comparing the raw generate tab output to comfy's default workflow, it'll be a bit different, as the workflow swarm generates by default is slightly improved over the raw mindless default flow. Most notably with XL models it sets rescond higher on positive prompt and lower on negative prompt (ie bias the model to generate with reference towards higher quality images), or on SVD it sets a very high sigma-max (ie aggressively beg the video model to maintain coherence)

[dim sonnet]

should open an issue on their github, and link to this https://github.com/comfyanonymous/ComfyUI/pull/867 since it's the most likely reason

[placid island]

Hi @dim sonnet i'm new in stableswarmUI too, where or how can i download that SwarmInput nodes?

[dim sonnet]

they're just included out-of-the-box with any standard swarm install

(as long as you're updated, they're new)

[potent panther]

For this theme PR https://github.com/Stability-AI/StableSwarmUI/pull/288
I think it could be a good idea to take inspiration from these GTK themes:

• https://github.com/vinceliuice/WhiteSur-gtk-theme This is inspired by the MacOS theme (and has a great light them color palette)
• https://github.com/vinceliuice/vimix-gtk-themes Old Material inspired theme with a smilar light theme color palette to MacOS
• https://github.com/vinceliuice/Graphite-gtk-theme Interesting theme because it has an outline and uses both black and white
  These themes are made for Gnome DE but I think they look great and could be a baseline for a better ui design (especially light theme)

I wish pythongosssss was here or maybe he is but under another name

[ancient dirge]

[nimble belfry]

I was just testing out using my other swarm backend on my bot and either im doing something wrong or I found a bug. my code that uses the WS preview images: When the API call uses my local backend this is what I get returned. I wait to get all 0-3 batch index images (im doing a batch of 4 images) then I make them into a grid

Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":1,"preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":2,"preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":3,"preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}

But if I do a second api call while that backend is busy it sends it to my other swarm setup and this is what I get: The batch index never changes and only says 0.

Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}
Response JSON (excluding base64 data): {"gen_progress":{"batch_index":"0","preview":"data:image/jpeg;base64,[BASE64_DATA]","overall_percent":0.25,"current_percent":0.8214286}}

[dim sonnet]

uh... what?

oh shit it does stick to batch index = 0 wtf

[nimble belfry]

If I do a batch of 4 images the preview images are separated by batch index 0, 1, 2, 3. Thats how i normally tell them apart and how I know I received all 4 image previews so I can turn it into a 2x2 grid.

For some reason If I do it from another backend it stays at 0.

ye.

[dim sonnet]

oooh cause the remote backend is choosing that id and it's just always doing 0 cause it's only getting 1 request at a time

i needa intercept and overwrite it

.. wait but i do do that

[dim sonnet]

debug does but actual value shouldn't

OH, it doesn't update properly

i see

... actually it maybe does maybe doesn't i can't tell aa

[dim sonnet]

... wait are you doing a literal batch?

as opposed to a request of 4 images, you're using the internal "batchsize" parameter?

[nimble belfry]

yes swarm internal batch size

this is what i send through API {"images", 1}, {"batchsize", 4},

[dim sonnet]

yeah that's

awkward

uhm

[nimble belfry]

I could prolly make a second check for 4 0s in a row then process the preview batch. If you are not seeing an easy fix.

[dim sonnet]

i pushed a commit that corrects for batch size, mostly

[dim sonnet]

the correct solution is stop using the internal hack param and use the normal images param

but a git pull might sufficiently handle the hack internal param

[nimble belfry]

Then it wont look all sexy with 4 images loading at the same time.

[nimble belfry]

Hell yeah. all working now.

I mean come on, look how good that looks.

[lavish glade]

how to load locons in stable swarm?

[dim sonnet]

should work the same as any other lora

[lavish glade]

do i have to move them to the same folder? using stability matrix btw.

[dim sonnet]

yes, put them in the lora folder

---

btw, anyone who wants to try a thing -- update swarm, go to user settings, and swap to theme Modern Dark and lemme know what you think

[lavish glade]

getting a error with segment prompt(?)

20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:   File "H:\Data\Packages\ComfyUI\execution.py", line 151, in recursive_execute
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:     output_data, output_ui = get_output_data(obj, input_data_all)
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:   File "H:\Data\Packages\ComfyUI\execution.py", line 81, in get_output_data
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:     return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:   File "H:\Data\Packages\ComfyUI\execution.py", line 74, in map_node_over_list
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:     results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:   File "H:\Data\Packages\StableSwarmUI 2\src\BuiltinExtensions\ComfyUIBackend\ExtraNodes\SwarmComfyCommon\SwarmClipSeg.py", line 35, in seg
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:     mask = torch.nn.functional.interpolate(mask.unsqueeze(0).unsqueeze(0), size=(images.shape[1], images.shape[2]), mode="bilinear").squeeze(0).squeeze(0)
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:   File "H:\Data\Packages\ComfyUI\venv\lib\site-packages\torch\nn\functional.py", line 3916, in interpolate
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr:     raise ValueError(
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr: ValueError: Input and output must have the same number of spatial dimensions, but got input with spatial dimensions of [1, 352, 352] and output size of (2048, 2048). Please provide input tensor in (N, C, d1, d2, ...,dK) format and output size in (o1, o2, ...,oK) format.
20:56:19.762 [Warning] ComfyUI-0 on port 7821 stderr: 
20:56:19.918 [Info] No images were generated (all refused, or failed).```

[dim sonnet]

uh... huh? That doesn't make sense. That would imply ClipSeg returned entirely the wrong mask shape, which it really shouldn't??

idek how that could possibly happen

[lavish glade]

there's also this at start up

   at System.Diagnostics.Process.StartWithCreateProcess(ProcessStartInfo startInfo)
   at System.Diagnostics.Process.Start(ProcessStartInfo startInfo)
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUISelfStartBackend.<>c__DisplayClass11_0.<<EnsureNodeRepos>g__EnsureNodeRepo|0>d.MoveNext() in H:\Data\Packages\StableSwarmUI 2\src\BuiltinExtensions\ComfyUIBackend\ComfyUISelfStartBackend.cs:line 68
--- End of stack trace from previous location ---
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUISelfStartBackend.<>c__DisplayClass11_0.<<EnsureNodeRepos>b__1>d.MoveNext() in H:\Data\Packages\StableSwarmUI 2\src\BuiltinExtensions\ComfyUIBackend\ComfyUISelfStartBackend.cs:line 75
--- End of stack trace from previous location ---
   at StableSwarmUI.Builtin_ComfyUIBackend.ComfyUISelfStartBackend.EnsureNodeRepos() in H:\Data\Packages\StableSwarmUI 2\src\BuiltinExtensions\ComfyUIBackend\ComfyUISelfStartBackend.cs:line 76```

using stability matrix atm for install might try doing a solo install later to see if the same thing happens

@dim sonnet ^

[past rivet]

Hey @dim sonnet are their any good guides or good tutorials you know of the software without going in depth on comfyui or should I just use comfyui workflow if I really want some advanced stuff like consistent characters?

[nimble belfry]

any video by
Scott Detweiler

[potent panther]

It's very wonky

[potent panther]

Things that could be fixed:

1. Inconsistent rounded corners (prompt box, generate button, quick settings, etc)
2. The top bar text isn't centered inside the buttons
3. The new font feels more like a branding font (old Discord) than a UX font used in apps
4. Dropdowns are using the original style of Swarm
5. Hover color on button (and toggles) are a bright desaturated purple even if the buttons are grey
6. The top bar and bottom bar buttons are extremely big (likely a bit due to the font)
7. User Settings (and parameter configuration) aren't themed in a majority of the options
8. Comfy workflow bubble is not themed properly
9. Utilities aren't themed
10. Separators are almost invisible due to their dark color

There's a lot of inconsistencies basically

I like the buttons and toggles as well as the no outline style this has

[potent panther]

Could be nice to add icons to some of the buttons and tabs like the https://feathericons.com that are MIT licensed

[waxen moat]

I'm trying this rather simple workflow and I load it into the Comfy Workflow Editor, but then it does absolutely nothing when I click "Generate" in Swarm. I must be missing something very obvious?

[dim sonnet]

does it work directly in the comfy tab?

[waxen moat]

yes works fine

[gray geyser]

Doesn't it need a save image as the final node?

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Advanced Usage.md

[dim sonnet]

oh, yes -- SwarmSaveAnimationWS is what you'd want here

(it's just Video Combine but uses websockets instead of storing a file in the comfy folder)

[waxen moat]

clearly i am on the bleeding edge here 😎

[dim sonnet]

(I do have VHS_VideoCombine listed as a recognized final output node? so if that's VHS it should work as-is)

[waxen moat]

hmm

[dim sonnet]

does the "Use This Workflow in Generate Tab" work as expected? ie it swaps you to Generate tab and changes the params to match the workflow?

[waxen moat]

it does, but when i click generate nothing happens and nothing in the logs either

[dim sonnet]

check browser logs (F12) or server debug logs (Server -> Logs -> set view type to Debug)

[waxen moat]

i even added a new primitive input and that new primitive input showed up in the generate tab as a test

[dim sonnet]

huh

uhh

can you disable the comfy workflow from the generate tab, then re-UseThisWorkflowInGenerate again?

might've just glitched somewhere and the glitch is stuck

[waxen moat]

just so i understand better... how do i disable it from the generate tab?

[waxen moat]

this was in firefox with lots of extensions, so i just moved to incognito chrome window and restarted swarm entirely, but same behavior

with source scripts in the screenshot..

thanks for taking a look, i'll get some sleep and check in later

[dim sonnet]

button in the bottom left

or on comfy tab

[waxen moat]

ah, these buttons don't show up for me, which sounds like it indicates the disconnect

but it still works fine with the default workflow hmmm

[potent panther]

@dim sonnet I'm currently fixing the translation for the French language

for the authorship do I just put my username even though your tool already did a lot?

[potent panther]

what happened here "[FreeU] Skip Two": "[FreeU] Skip Two - Télécharger", bug in the translation tool?

[waxen moat]

still... models empty

so... i tried to put quotes around the ModelRoot path and that didn't work, so then I tried restarting StableSwam entirely, but now I completely can't launch it because those quotes are there in the config somewhere

15:01:14.555 [Error] Failed to create directory for SD models. You may need to check your ModelRoot and SDModelFolder settings. The filename, directory name, or volume label syntax is incorrect. : 'D:\Documents\stable-diffusion\stableswarmui\StableSwarmUI\"C:\machine-learning\models"\checkpoints'

[waxen moat]

opened a github issue 👍🏽

ah.. found the Settings.fds file and was able to correct my mistake there

[dim sonnet]

the AI that translated it had a silly idea I guess

offhand I believe only Chinese had human corrections, everything else is pure AI atm

(PRs to improve translations are very welcomed)

[dim sonnet]

ooh, ye -- the reset params button should resolve that if that happens. I'm not sure how/why that happens, so if you can replicate it again in the future please tell me the steps

[waxen moat]

here's another workflow... completely no connection between comfy and SS

and I just re-installed SS and let it install its own comfy and re-installed all the extensions from scratch

happy to do screen share if you want to dig into it further

[dim sonnet]

this workflow doesn't break for me

[waxen moat]

so weird.. do you have custom model paths?

the only other thing i changed was ExtraArgs had --output-directory set. i removed that

if i build it from scratch, then it works, but if I load the file with the Load button from the Comfy tab, it breaks the connection somehow

[obsidian jackal]

#🐝｜swarm-ui

[nimble belfry]

Is there a bug with SVD? I cannot seem to generate anything and get this in console.

11:06:08.355 [Info] User local requested 1 image with model 'starlightXLAnimated_v3.safetensors'...
11:06:46.553 [Warning] ComfyUI-0 on port 7821 stderr: Traceback (most recent call last):
11:06:46.553 [Warning] ComfyUI-0 on port 7821 stderr:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/execution.py", line 151, in recursive_execute
11:06:46.553 [Warning] ComfyUI-0 on port 7821 stderr:     output_data, output_ui = get_output_data(obj, input_data_all)
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/execution.py", line 81, in get_output_data
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:     return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/execution.py", line 74, in map_node_over_list
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:     results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/ExtraNodes/SwarmComfyCommon/SwarmKSampler.py", line 147, in sample
11:06:46.554 [Warning] ComfyUI-0 on port 7821 stderr:     sigmas = calculate_sigmas_scheduler(real_model, scheduler, steps, sigma_min, sigma_max, rho)
11:06:46.555 [Warning] ComfyUI-0 on port 7821 stderr:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/src/BuiltinExtensions/ComfyUIBackend/ExtraNodes/SwarmComfyCommon/SwarmKSampler.py", line 57, in calculate_sigmas_scheduler
11:06:46.555 [Warning] ComfyUI-0 on port 7821 stderr:     model_wrap = comfy.samplers.wrap_model(model)
11:06:46.555 [Warning] ComfyUI-0 on port 7821 stderr: AttributeError: module 'comfy.samplers' has no attribute 'wrap_model'
11:06:46.555 [Warning] ComfyUI-0 on port 7821 stderr: 
11:06:46.715 [Info] Generated an image in 20.10 (prep) and 11.68 (gen) seconds

[nimble belfry]

@dim sonnet I also tried this on a almost fresh install on a different machine and I get the same issue.

[dim sonnet]

oo, comfy changed some internal code, i fix https://github.com/Stability-AI/StableSwarmUI/commit/2b5ebabe5e4a5a56f4f5c16d6b78c19defd47da2

[nimble belfry]

Epic, as always. Thanks.

[potent panther]

@dim sonnet are there any reasons why some of the lines in the translation file don't have any spaces?

am I supposed to translate them or keep them as is?

[dim sonnet]

you mean like these spaces \n ?

if so: yeah the ones on the ends don't matter, and sequential spaces don't matter. That's just a software woopsie i haven't gotten around to fixing

leave them as-is exactly in the key, just don't worry about the translated value.

(a \n inside the text and not on the end should be kept where it is)

[waxen moat]

I reinstalled yet again with a fresh stableswarm and fresh internal comfyui that stableswarmUI installed. did zero customizations except installing comfyui-manager, ad evolved, and vhs video combine, but still I can't "use this workflow in generate tab" for any non-default workflow that I don't build from scratch 😦

incognito window, no vpn, absolutely nothing in the logs, even with verbose mode. just the usual 22:54:52.789 [Verbose] Comfy direct POST request to path manager/component/loads 22:54:52.791 [Verbose] Comfy Redir status code 200 from 200 and type application/json; charset=utf-8 for POST 'manager/component/loads'

[dim sonnet]

an error would most likely be in browser console, not server console

(hit F12)

[waxen moat]

thank you for the hint! it seems the VHS Video Combine node was breaking it

[dim sonnet]

oh wtf

seems like there are format values in your install of VHS that, uh, aren't?

lemme update mine, i haven't updated in a while

[waxen moat]

glad i wasn't completely insane at least 🙂

[dim sonnet]

what the hell

oh

oh my god

that's a new comfy feature they're using

oh that's so weird

okay

[waxen moat]

good luck 🫡

[dim sonnet]

run the update bat, should be fixed now

[waxen moat]

chef's kiss

[cold dirge]

does swarm work with amd ?

[versed copper]

is there a fix for this?

[dim sonnet]

yes, though there's some awkwardness in the install - see this thread https://github.com/Stability-AI/StableSwarmUI/issues/23 for details

[cold dirge]

thanks monkey

[dim sonnet]

Run the update bat and restart swarm, I added a thing to automatically handle that -- it's a new dependency comfy added recently, so older installs are missing it.

Alternate, manual install if you prefer: open E:\StableSwarmUI\StableSwarmUI\dlbackend\comfy and open a command line there, and enter .\python_embeded\python.exe -s -m pip install kornia

[versed copper]

hmm, update and restart didn't work

[dim sonnet]

did you run the update in the ~1 minute after I sent that message before I remembered to hit the push button

lolsorry

[versed copper]

thanks, updated, no more error 🙂

[potent panther]

No I mean the stuff like AllowedModels or BackendStartingPort

[dim sonnet]

that's a server setting

[potent panther]

I see

So no translation

[dim sonnet]

should translate it

[potent panther]

?

[dim sonnet]

?

[potent panther]

so I should translate these things?

[dim sonnet]

yes

[potent panther]

ok

[dim sonnet]

if it's in the translation file, it should be translated

translations only apply to display text

in the case of server settings, the display text in English is just the technical name copied directly

[potent panther]

What about the things that I can't really translate in French?

[dim sonnet]

like what?

[potent panther]

Backends

As far as I'm aware French devs use the English term

[dim sonnet]

if the English word is more fitting to use, then yeah use that

[nimble belfry]

Is there any way to choose what format the preview images are in? It looks like the video previews are in sequenced WebP but unless im doing something wrong I cant embed that format in a Discord embed.

[nimble belfry]

https://github.com/discord/discord-api-docs/discussions/5818

[dim sonnet]

video previews are in webp because it's basically insanity to do anything else

apng files would be disgustingly massive and webm/mp4 would be encoding hell

noting that previews are meant to be rapidly generated temporary displays

[nimble belfry]

yeah Discord is dumb. Making my life hard for no reason.

[dim sonnet]

funny thing is, discord fully supports webp natively

they just refuse webp embeds uploaded by users

[nimble belfry]

Its like youtube video . You embed them but wont let users do it...

[cold dirge]

is there a way to check your it/s in swarm ?

[nimble belfry]

Is it not just in the console? 20:46:07.696 [Debug] ComfyUI-0 on port 7821 stderr: 20%|\u2588\u2588 | 1/5 [00:03<00:13, 3.50s/it]

[cold dirge]

it does nots how up in my console

[nimble belfry]

o must only be there on debug

[cold dirge]

gotcha

do you know what folder name is for back end of comfy

it seem i deleted it when i moving folder around by mistask

for models

[nimble belfry]

I think its dlbackend

[cold dirge]

i have check point

but it not showing up in comfy so i assuming folder called somthing else

i remeber when i moved it in to swarm folder it was called somthing different xD

[nimble belfry]

huh? What are you looking for? Swarm models are meant to be in the Models folder

[cold dirge]

im looking for check point folder in model folder of back end of comfy

[cold dirge]

ah it has a - in it

[nimble belfry]

You can set it to whatever you want in settings

[cold dirge]

ah well that did not seem solve problem w/e i dont care XD thank anyways xD

i just wanted to compare auto1.9 to comfy speed in it xD

pretty sure auto is generating as fast as comfy now

[nimble belfry]

thats for sure not true lol

[cold dirge]

whatever they did it is going faster then before xD

i just prefer see real number to know difference

[nimble belfry]

Just turn it to debug. you can even see it in the UI server-> logs then set view to debug. generate an image.

[cold dirge]

where do i find debug ?

?

i figure out my answer xD that show me how long gen took

auto is 30% slower then comfy xD

er no 45% xD

it 6 second to gen in auto xD

if it a new gen it only 8% faster xD

[dim sonnet]

you can either set

or just go use

[cold dirge]

oh i see

ah yes this much more useful to see xD

new auto genning at 5/its

[nimble belfry]

wow they actually made it efficient?

[cold dirge]

yeah

[nimble belfry]

about time lol

[cold dirge]

i can pull it up on same prompt show you terminal xD

[nimble belfry]

I trust you lol

[cold dirge]

that why i wanted to check it

i love auto xD i was like dam this seem fast xD

well love auto for being able to edit lora weight value xD

if comfy could do it non annoying way i probably fully switch over to it xD

[nimble belfry]

what do you mean? you can do that in Swarm.

[cold dirge]

comfy said be alot of work to make it work in comfy

like being weight 0 on step 1 then liner going up to 1 by step 10 then fall back down to .75 by step 15

just nightmare to writei ncomfy xD

[nimble belfry]

Is that not what your talking about?

[cold dirge]

well that a single weight value what i want is to be able to write my own weight graph

[nimble belfry]

o I see you want to change it.

[cold dirge]

yeah alot of lora i make

dont work until 50% xD

if start at 0 just dont work xD

[nimble belfry]

sounds like you need to make better LoRA

[cold dirge]

so like in auto just write like 0 = weight value @ 0 = time for example if you wanted ramp from 0 to 1 to at 50% just write 0@0,1@0.5

that all you need have a custem graph xD in auto

for example i have a lora that rewrite people faces with flowers xD

so it generate person face until 50% then it throw a flower over top of there face xD

but yeah it could also just be a skill issue that i suck at making lora xD

[nimble belfry]

[dim sonnet]

yeee, changing over time is the big feature request people have for swarm/comfy that auto has better supported atm

[cold dirge]

yeah big request on forge too

an the guy who made add on was willing help forge person make it

might be able to ask him about it for comfy xD

[dim sonnet]

please do

[cold dirge]

i can grab the link again he explain how to do it

i dont understand any of it xD

[dim sonnet]

it's not hard in theory/concept to do it

it's just a matter of somebody taking the time to figure out how to fit the code in most properly

[cold dirge]

https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/68 here was reply he did to forge

yeah fair enough i think it one most core features of prompting make lora so free xD

like for example i able to turn my mimic lora off at 50% an add deflate lora on turn a monster in to a cushion for example

[nimble belfry]

neat

[cold dirge]

main use i find for it

is being able to control bad lora xD

one that just have way to much alpha value

an sdxl made it even beter since it not hard coded like 1.5 xD

[cold dirge]

like this impossible prompt for 1.5 because cant alter female features past 30% xD but sdxl has no problem rewriting full face

[cold dirge]

i am back again with more problems XD

i cant figure out how to fix tripo sr viewer

[waxen moat]

is there a way to get the comfy workflow editor tab to "activate" as if I was using it directly, but when it is triggered by the Generate tab? (just a vague feature request 😉

[dim sonnet]

triposr viewer dev seems to have died, i have a PR that fixes it here https://github.com/flowtyone/ComfyUI-Flowty-TripoSR/pull/35

[dim sonnet]

huh?

[waxen moat]

trying to use grid generator again with VHS but it's not showing the outputs in Swarm

2024-04-15 10:01:22.773 [Error] Invalid/null/empty image data from ComfyUI server for 'AD_latent-lifetimes-testing-DE-SATURATE-_00015.mp4', under {
    "gifs": [
        [{
            "filename": "AD_latent-lifetimes-testing-DE-SATURATE-_00015.mp4", "subfolder": "steerable-motion", "type": "output", "format": "video/h264-mp4"
        }
    ]
}
2024-04-15 10:01:22.774 [Info] No images were generated (all refused, or failed).
2024-04-15 10:01:22.775 [Error] Grid generator hit error: No images were generated (all refused, or failed - check server logs for details).

[cold dirge]

oh i see thank you

[dim sonnet]

should use SwarmSaveAnimationWS to ensure it works as intended

[waxen moat]

Can I combine SwarmSaveAnimationWS with VHS so that it works in swarm, plus takes advantage of VHS options like ProRes output? ... anyways will give it a try

Seems to work well with both outputs in the workflow 🫡

[cold dirge]

@dim sonnet maybe this is a dumb question
how do you install through swarm xD

[dim sonnet]

how do you install a comfy node in swarm? Just install it as regular, inside the dlbackend/comfy/ComfyUI/custom_nodes folder

[cold dirge]

oh alright just way i read it i thought like a feature in swarm putting it in i was just confused xD

[waxen moat]

this Pause Videos toggle does not seem to do anything, and no logs generated in browser console.

have lots of videos playing in the grid, but they never pause

[dim sonnet]

are those videos (mp4/webm)? as opposed to "animation images" (gif/webp)?

videos can be paused but "animation images" can't be

[waxen moat]

ah.. gotcha. just what the output is from SwarmSaveAnimationWS so I assumpe webp

[dim sonnet]

is selectable

[waxen moat]

ohhhh

gotcha gotcha. might be worth a tooltip for that

[dim sonnet]

literally can't do tooltips or anything like that on the comfy interface

only on the Swarm tabs

[waxen moat]

oh i meant on pause videos section of grid generator web page

[dim sonnet]

oo

ye okay

[waxen moat]

also another minor annoyance is that Swarm will complain if you put a trailing \ on the OutputPath path and you have to change it and restart the server again.

[dim sonnet]

yarr

[dim sonnet]

on which path?

in Server Settings?

[waxen moat]

yeah in server configuration

[dim sonnet]

tried just now, seems to work fine

OH i see

image history load gets upset

fixed

[trim orbit]

Hi all, i'm new on swarm-ui. I've just complete a fresh install for amd. I write a simple prompt a photo of a cat then generate. Nothing happen. On log i've an error about NVIDIA GPU. Someone can put me on right way ?

[dim sonnet]

AMD support is unfortunately a bit wonky, they do not have good drivers for windows -- this thread contains info about it https://github.com/Stability-AI/StableSwarmUI/issues/23

[trim orbit]

I'm on Ubuntu

[dim sonnet]

oh okay you'll have an easier time then, nice

ah see here: https://github.com/Stability-AI/StableSwarmUI/issues/217#issuecomment-1880434343

[trim orbit]

thx i'll do 😉

[potent panther]

Oh the scam is still here

Anyway I think I'll keep prompts as prompts in French

[heavy mauve]

how to export the logs? Do we have it automatically saved inside the directory?

[dim sonnet]

it's not currently automatically saved anywhere, no

reasonable feature request tho

[dim sonnet]

log saving to file added under server settings

check the checkbox and restart and it'll start autosaving

defaults to Swarm/Data/Logs/....

[heavy mauve]

Thank you! @dim sonnet

[potent panther]

@dim sonnet "[Close]\n ": why do some lines have the backslash n and a whole lot of spaces behind?

this is Line 15 btw

do I have to add it to the translation or not?

[nimble belfry]

https://media.discordapp.net/attachments/1105955639907328101/1229928555677352046/new_image.gif?ex=6631773a&is=661f023a&hm=5dbfaf17bd9eef5906169647c097db57ad48458eb517f1be4810a87020c2f6b3&=&width=1264&height=947

[potent panther]

neat

[dim sonnet]

cause i don't trim them atm and so they have fucky wucky html spaces on them. idk why that one does, it has no spaces

but say for example this one

that one has spaces because the code is too dumb to realize they're not real spaces they're just html tabulation

[nimble belfry]

This one came out good. Playing with making some logos.

[potent panther]

Okay another question

for refiner is it best to translate it or keep it as is?

[dim sonnet]

everything that's in the translation text file should be made clear in the relevant language

whether "refiner" or a related french word is clearer, is a question only a french speaker can answer

[potent panther]

Then I'll translate it

funnily enough the translation tool didn't translate the word itself but translated it when used in a phrase

[maiden shoal]

Is there a trick to getting custom nodes to work? I've cloned them to the custom nodes directory and restarted, but they are not available.

[dim sonnet]

For the most part, no, as far as swarm goes - if it works in comfy, 95% of the time it'll just work in swarm, and the only exception is an easy thing to fix when it happens

check Server -> Logs -> Debug and see if maybe the comfy logs include an error message whie loading

[dim sonnet]

if anyone wants to try SD3 API in swarm check https://github.com/Stability-AI/StableSwarmUI/issues/308#issuecomment-2063203524

[lavish glade]

anyway to get the machine to pump images out 1 image at a time without spamming the generate button? or is that what the generate forever does?

[barren moat]

Hey Alex I'm getting some errors trying to run SD3. Is there something that stands out to you that I did wrong?

I did load my SD3 api key

[barren moat]

There is a section to input the number of images you'd like.

[lavish glade]

will it do it in sequential order 1 at a time?

[barren moat]

yes

[lavish glade]

pog, guess i can set 1k then go to bed and wake up with something.

[barren moat]

There might be a limit but the generate forever button would be a better choice for that. Alex can let us know if there's a limit or try entering in 1000 and see if you get an error.

[nimble belfry]

@dim sonnet Did you change the way a lora is labeled in the API? yaml 18:44:53.806 [Debug] Refused to generate image for local: Invalid value for parameter LoRAs: Invalid value for param LoRAs - 'an0tha0ne.safetensors' - must be one of:

Looks like I now need to exclude the file ext?

[lavish glade]

mmm just took a nap and left the generate forever running for about 2 hours straight. not a prob. only thing i can think of a limiting factor would be space and cooling.

[potent panther]

.safetensors isn't saved in the metadata anymore I think

only the name of the file not the extension

[lavish glade]

i use stability matrix to keep my models organized.

only sad thing was a i had to add a symlink for locons/lycoris and point it towards the lora folder

[potent panther]

btw mcmonkey in the feature announcement

the mockup looks extremely old and non functional

like just the prompt box being green is already bad enough

it's not a good retro imitation and is just not the right choice

[potent panther]

I'll make a little concept later

[dim sonnet]

yeah generate forever does indeed go forever one at a time

there's a user setting for the time delay between when it hits

so you can set it a bit wide if you're worried about overheating

[dim sonnet]

The Key set on the User tab is used in the Generate or Simple tab, it is not autoinserted in the Comfy tab: If you want to use API directly from comfy tab, you have to set the key the way the comfy extension expects: either in a file labeled sai_platform_key.txt in the dir E:\StableSwarmUI\src\BuiltinExtensions\ComfyUIBackend\DLNodes\ComfyUI-SAI_API\, or in the api_key_override input node

[dim sonnet]

ooh, .safetensors was removed, but it's meant to still work if you include it -- try new commit, should hopefully actually work now (I'm a bit underequipped to test that one properly offhand)

[barren moat]

I did try it from the generate tab and that was the errors I got. I was only able to get it to work from the comfy tab by manually entering the api key directly in the node. BUT it did give me an error with the seed set to -1, it said it must be set between 0 and some crazy high number

I’ll try it again shortly when I make my way to my pc.

[dim sonnet]

oh, yeah: seed of -1 works for Swarm but not for Comfy

Swarm sees that and knows to apply randomization, comfy doesn't have that behavior

i wouldn't really recommend the workflow I gave for use directly in comfy tab, it's all set up to make for a nice import to swarm's other tabs

for comfy tab directly, just the SD3 node and attach a save image on the end

[barren moat]

I used your workflow for comfy and also for swarm. I’m assuming that’s how you intended it, right?

Import the workflow and then switch to the generate tab to generate the images from swarm, correct?

Or maybe I’m doing it wrong?

[versed copper]

when using lora:blabla

with lora chosen from dropdown

oops, might be a StabilityMatrix bug? (nope, same thing happens in standalone version)

[potent panther]

Likely an issue related to #🐝｜swarm-ui message

Did you try updating?

[versed copper]

totally up to date in both, i always update before starting

loras are in subfolders, work if chosen as loras

but not in the prompt

[dim sonnet]

delayed reply, but: alright well good we didn't go with it then lol

[dim sonnet]

the SD3 API reference workflow is just for swarm usage, not meant to be used in comfy directly

ye "use this workflow in generate tab" to use it

(or save it and enable the experimental new Simple tab and use it there)

[lavish glade]

kinda want a separate tab next to generate tab just for image history. especially when you use generate forever and produce 1 image a minute. (slept for 8 hours lol).

[potent panther]

btw mcmonkey I realized yesterday but is it intended that you can retract the left settings panel but not the batch panel on the right?

[potent panther]

that's at the bottom?

[lavish glade]

its is

can i move it?

[potent panther]

no

[lavish glade]

[potent panther]

could be interesting if mcmonkey could add that kind of modularity that is present in image editing softwares

[dim sonnet]

Oop, yeah bug from the commit for kaleb's issue, I fixed now

[dim sonnet]

Having a Gallery view at the top as its own main tab has been requested before

might have to add that

[lavish glade]

oooh

[dim sonnet]

it's intended but not for any particular reason, I just, made the choice to do it that way and was happy with it. Just figured it's rare that you'd actively want to hide that

I should probably just make it hideable ye

[dim sonnet]

that's what I want I want it so you just click and drag a tab and it shoves it somewhere else

that's hard to impl tho so i haven't even tried yet

[potent panther]

btw for Eric it could be good to take on some of the designs point of InvokeAI

[lavish glade]

ooof

[potent panther]

like making different tabs more compact by turning them into buttons with icons

that gives more space for other things

[dim sonnet]

it can be slammed shut now

no nice clicky atm, just drag it all the way over

[potent panther]

k

[versed copper]

i have never seen support happen this quickly, and being friendly.
thank you very much for the stuff you do

[gray geyser]

Which endpoint Api you'll are using for text-to-image

*api endpoint

[nimble belfry]

For swarm?

[gray geyser]

Swarm?

[nimble belfry]

You are in the #🐝｜swarm-ui channel

[gray geyser]

I'm facing errors

https://cdn.discordapp.com/attachments/1230799113197457461/1230821630331060284/image.png?ex=6624e4f8&is=66239378&hm=a9ac7fe14019b2e8e3c1fc1bd62dc184a7c0cda30c57c638e5593e13f3ccd645&

https://cdn.discordapp.com/attachments/1230799113197457461/1230805361909305344/image.png?ex=6634a7d1&is=662232d1&hm=379802da1e523c875304cffb974eee69ea69d9e6f17e2c33727a694568c22ed7&

[nimble belfry]

If your not using StableSwarmUI then you are in the wrong chanel.

[rustic river]

Hi I recently set this up after using nothing but A1111 for 15 months. It's a breath of fresh air over the A1111 in terms of codebase, installation process, resource usage and docs/clarity. Though the UI itself is so-so, I can see the massive potential it has to be extremely versatile and snappy.
I do have some issues though, the main thing being that when using the exact same settings as in A1111 (as far as I can see, Swarm seems to hide things that actually matter really well), the results are subpar, unacceptable basically. Is there something I'm missing?

[cold dirge]

🤔

what do you mean xD?

[rustic river]

I mean that when I copy my A1111 generation settings verbatim, the resulting image is different and looks much worse.

[cold dirge]

are you using fancy things or just text ?

🤔

[rustic river]

Nothing like ControlNet or something, if that's what you mean by fancy.
Just one LoRA

[cold dirge]

lora work different in comfy then auto

probably why it give you a different output

[rustic river]

Ah, that makes sense. Any way to bring them in line?

[cold dirge]

i mean lora should still work just same prompt probably wont work

i am not super amazing with comfy

[rustic river]

I mean, is there are known thing I need to configure, like a setting. Or, maybe the LoRA weight needs alteration?

I'll try without a LoRA though to see if I can get the same quality as I'm getting with A1111.

[cold dirge]

i am sure kalebbroo or monkey could give you more info when they have time i know i have not had much luck getting good results in comfy xD

atlest with lora

i dont think that swarm related issue i think a more of a i suck at comfy issue xD

[rustic river]

I haven't looked at Comfy yet, but I've been thinking for months about trying it. Swarm seems like a good way to get into it.

[cold dirge]

you can switch workflow at the top in swarm to comfy window

show you what lora is doing

[rustic river]

Alright, thanks for your help 👍
I'll do a few generations with both...

[cold dirge]

monkey is normally around you could alway ping him if you get stuck xD

[rustic river]

Oh, it's fine. I see him quite often in Oobabooga server as well.

Alright, I've made some comparisons. It's better without a LoRA, but the quality is still questionable.

Left = A1111, Right = StableSwarmUI

[cold dirge]

look about the same to me

[rustic river]

The ones on the right aren't as clean, the lines are just bad. The ones on the left are just softer, cleaner and overall higher quality, at least, to me.

[cold dirge]

i mostly just use comfy for ipadapt

then stick with auto for normal prompting xD

[potent panther]

The only difference between auto and comfy here should be the seeds

Are you using the exact same samplers and scheduler?

[rustic river]

Yeah, same sampler. Not sure about the scheduler, it's on Auto in A1111 and Off in Swarm. I can try to force a specific scheduler to see if it'll give the same result, but that'll be tomorrow.

[dim sonnet]

so, 3 key differences:

• 1: Auto webui does secret prompt reweighting (it normalizes so that the average weight across all tokens equals 1 iirc), comfy doesn't -- so that (anime style:0.75) is parsed differently. To replicate auto webui I think you'd need a like, (...:1.1) or so on everything else and then raise the anime style part to 0.8 or something (I'm guessing, I don't know the exact math offhand)
• 2: seeds different as noted
• 3: for XL 1.0 Base models, swarm emits a modified res cond by default - it tells the model to assume higher resolutions for positive prompt cond, and lower resolutions for negative prompt cond. On the Base model, this is basically just a guaranteed quality enhancement (some other UIs, eg Fooocus, do this out of the box. I imagine auto webui just flat doesn't). It's possible on animagine or the lora, rescond behaves differently. I didn't have a convenient toggle on this behavior (other than, yknow, marking the model as anything other than XL 1 Base, or going and editing the comfy workflow), so I added one now - hit the update-windows.bat and then under Display Advanced Parameters, find Swarm Internal and uncheck Model Specific Enhancements to see if that looks better to you. (If it does, you can create a Preset and name it Default and have that explicitly unchecked in the preset to make it always be unchecked for you).

Side note: seed-to-seed differences do weird things to extremely subjective measurements like quality comparison. Your comparison images to my eyes are... not very different on most sections of quality (aesthetic/style/cleanness of lines/etc) and the few things I can compare objectively, the left one imo is worse (eg bottom left has the cursed hand, and she's standing on an infinite blanket) (to be clear, I'm not saying "swarm looks better!" I'm saying there's seed-to-seed quality differences that are far larger than anything distinct to the two UIs)

[rustic river]

Thank you for looking into this! I know the ones in the left have some "cursed" parts, it's just that the swarm images look worse as a whole (even when leaving the seed out of it). The overall quality, and especially the lines and shading is just not on par with the images on the left. I understand it if swarm calculates things differently, and I'm all for it, considering it's speed and inner workings.
I'll dive into it tomorrow again and will post my findings here of course.
(Btw, it's impossible to see the difference between the images on mobile, or even or "desktop preview". If you open the original images on a pc monitor it'll be clear that the images on the right have poor lines, almost like bad scaling compared to the images on the left)

[dim sonnet]

OH LOL you probably just have a VAE applied in auto webui settings but you didn't do the same on swarm

that subtle rgb striation artifact is a known issue of the SDXL 1.0 VAE, that's not present in the SDXL 0.9 VAE nor the fp16 fix vae

i was wondering what you were talking about with quality differences but now that you mention to zoom way in on the lines, yeah, that's just the vae thing

[rustic river]

Hmm, it was on Auto, so, it should've use Animagine's internal VAE, I think, assuming it has one. I don't have a separate VAE weight for XL.

[dim sonnet]

if you don't already have a separate XL vae to use, probably grab this one https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

identical to the clean 0.9 VAE, but also as a bonus doesn't have the fp16 precision errors that the original did

[lavish glade]

hmm, generate forever is nice but it needs another thing besides a delay interval. might be nice to add a timer based stop. where say i can input the hours and minutes to do a full stop after current batch after say 6 hours of running. right now i got 20 image batch in core parameters with a 5 minute delay between each batch.

[potent panther]

@dim sonnet there’s this custom node to replicate auto’s behavior https://github.com/shiimizu/ComfyUI_smZNodes

Could be nice to add a toggle under the sampling section

[potent panther]

Also it has prompt editing from auto

[dim sonnet]

oh my god that's just full on chunks of auto webui source copy/pasted into comfy

[potent panther]

A

Forge

[rustic river]

@dim sonnet

• Disabling Model Specific Enhancements makes no difference (at least, for Animagine)
• The sdxl-vae-fp16-fix VAE makes no difference
• The two above together, and it looks 10 times better
  What?

[dim sonnet]

hwot

[rustic river]

Alright, alright, I'm gonna try again. Actually saving/naming the images now.

Uh, nvm, I'm stoopid

With MSE > without MSE

Yup, VAE makes no difference. So, the Model Specific Enhancements seem to break Animagine, huh.

[rustic river]

Reference and stuff

[dim sonnet]

that's. so weird

if anything that comparison should be reversed, as far as theory goes

[rustic river]

Shrugging...

I don't know enough about the inner workings of diffusion models to really know what's going on. Maybe you verify this quirk, or shed some light on it (but you already helped a lot! )

[maiden shoal]

I run Swarm on a Linux server and access it via a Windows browser window. When I try to save, load or browse workflows, it uses local directories (i.e. on the current PC). Is there a way I can change this to save the files on the server so that when I move between desktop, laptop, etc. all the same files are available?

[potent panther]

🤔

true is kinda better

[rustic river]

Well, odd.
I have to uncheck it for animagine.

[potent panther]

use the grid generator under tools

[nimble belfry]

i was just about to suggest that.

[lavish glade]

@dim sonnet how do i go about adding adetailer(?) and scoring extensions?

[dim sonnet]

use the buttons at the top

those are serverside

the comfy "Queue" sidebar menu is local files

[dim sonnet]

re adetailer, see here: https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7236821

re scoring: the builtin scoring extension is still a bit of a placeholdery hackup https://github.com/Stability-AI/StableSwarmUI/tree/master/src/BuiltinExtensions/Scorers but the readme for it explains install/setup

it's hidden on the main Generate tab as an Advanced param

depending on what you're doing you might prefer a comfy customnode for scoring

[lavish glade]

thx

so if i want to do 2 girl's faces i would do
<segment:girl's 1 face>
<segment:girl's 2 face>
correct?

[dim sonnet]

in theory yes however in practice i doubt clipseg will accurately detect that

probably needa do it a bit more manually

[maiden shoal]

Thank you for the reply. When I use the “browse workflows” button in your image, it opens local folders.

[dim sonnet]

it absolutely does not do that

[waxen moat]

i believe the FPS modifiers should all be int, not float

[dim sonnet]

it is technically valid, albeit weird, to do things like fps=0.5 to get one frame every two seconds

or fps=29.97 to do that weird math hack that some monitors do for some reason

[waxen moat]

fair 🙂

[dim sonnet]

oh for obscure broadcast reasons apparently https://filmpac.com/fps-battle-24fps-vs-29-97fps/

[lavish glade]

is there a option to add negative prompt word counting?

[dim sonnet]

was initially excluded as idk where to put it on the new alt-neg-box. But I added now as ... this awkward format:

needa find a better place to put it

[potent panther]

I was thinking it could be nice overlayed at the end of the prompt box

And like it slides away to let you prompt stuff

[lavish glade]

i mean i am kinda schizo with my prompts. postive prompt is at 215/225 and negative is from the looks of it 2-2.5 times that.

[lavish glade]

to keep it simple. ie "+215/225 -405/425"

or "Pos:5/75 Neg:3/75"

[dim sonnet]

Prompt editing! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-9206140

[rustic river]

Whoah, that's sick. I've never used alternating prompts in Auto WebUI...

I do use dynamic prompts (mostly "random", i.e. {cat|dog|mouse}), but it looks like Swarm already supported that 🙂

=====================
Btw, it seems that dots (.) don't work in OutPathBuilder.Format, can you confirm?

I entered raw/[year].[month].[day]_[hour].[minute].[second]_[seed]_[steps]_[width]x[height]_[model]
Resulting file name 20240424_011713_1378103283_80_768x1344_NewDigital_Life_-_v131_XLsafetensors

[dim sonnet]

yes it intentionally excludes . from the user setting there, as . has a special meaning in the context of filenames

(anything from file extensions to jump-up-a-folder to etc)

[rustic river]

So?

Never had an issue with dots in file names, except when programs didn't handle it right.

[dim sonnet]

here, update and go check this "AllowUnsafeOutpaths", then restart swarm and it should work as you intended from there

[rustic river]

Bruh

🏆

You're too freaking fast

[nimble belfry]

The top reason why StableSwarmUI is the best.

[rustic river]

Yeah, I'm liking it. Can't use it as my main yet, though...
LoRAs don't work as expected, no "Restart" sampler...

Though that probably has more to do with Comfy

Basically need this https://github.com/ssitu/ComfyUI_restart_sampling

Wait, if I plug that into ComfyUI, will the sampler show up in the Generate tab? Doesn't seem to do that, probably because it's an entire node.

[dim sonnet]

what's wrong with LoRAs? 0.o

[rustic river]

I don't know, I just haven't been able to get the same results as on Auto WebUI. I know... that's not the standard or anything...

Anyway, I got Restart working in Comfy, can't believe I managed that

[versed copper]

should [Property1:Property2:0.5] - i.e. switching after half the steps from 1 to 2 - work in stableswarmui?

[dim sonnet]

yes, as long as you're up to date - it was added literally two days ago https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-9206140
though note swarm's standard syntax for it is <fromto[0.5]:prop1, prop2> -- the auto webui syntax will also work the same tho yes

[versed copper]

oooh, nice.
0.7 would put the balance towards prop1?

or does that work the other way round

[dim sonnet]

yes. 0.7 would mean it does prop1 for 70% of steps, then prop2 for the latter 30%

[versed copper]

thanks a lot 🙂

[versed copper]

sidenote on this: StableSwarmUI actually translates <fromto[0.5]:prop1, prop2> in a wildcard to [prop1:prop2:0.5] in the picture properties
which confused the hell out of me 😄

[waxen hinge]

Hey Guys, Does stable swarm exposes an API?
I would like to configure additional backend nodes dynamically through API to the main server.

[nimble belfry]

Yes a very easy to use one. Search in this channel for where monkey posted his simple discord bot. If you can't find it or need a better example you can press F12 on your browser and generate something in the UI. You can see all the endpoints.

[waxen hinge]

thanks let me try

[nimble belfry]

https://github.com/mcmonkey4eva/SimpleDiscordAIBot

Feel free to ping me with any API questions I would like to think I know at least a little about the API.

[waxen hinge]

Sure thanks I am looking into it.

[waxen hinge]

I do see two three apis, but nothing related to add or deleting backend, do you guys have any documentation?

[nimble belfry]

O, I have only ever used it for image generations. I'm not even sure that's setup. You may have to wait for mcmonkey to be online.

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Advanced Usage.md you can check the docs but I don't remember anything about creating a backend with the API. I also have not had a use case where I ever needed to do that.

[waxen hinge]

I had checked docs already, couldn't find this info

[noble temple]

is there any way to use hires. fix on swarm?

[rustic river]

I think Refiner is where this is done

[dim sonnet]

Swarm uses its own API -- open a browser, and open the browser tools to show network calls, and do whatever actions you want, then copy the calls. For adding a backend you'll see /API/AddNewBackend to add one followed by /API/EditBackend to configure and enable it, and /API/DeleteBackend to delete it

going to publish formal API docs soon™️-ish

[dim sonnet]

yes as lounger said, "Refiner" in swarm does everything auto webui's hires fix does, alongside some related refinement techniques (that auto webui has separate or not at all), just enable Refiner, leave model as Use Base, leave method as Post Apply, set the Upscale to however much you want and whichever method you like

oh and set Refiner Control Percentage to same as in auto webui i think is called denoising or something

[noble temple]

Ah okay, I tried that but the result is a bit different from Forge but fine

[dim sonnet]

given exact same params, everything except seed should match

[noble temple]

Its a bit noisy and pixelated, while on forge it looks clean

[dim sonnet]

you might have a different percentage or upscale method then

[noble temple]

What method you use?

[dim sonnet]

the default percentage in swarm is lower than auto

just copy whichever one you're using in Forge

if you want a good option, lanczos is nice, but upscaler models are better (eg realesrgan)

[noble temple]

Yeah I'll try, now, another one, any of you guys ever had this when trying to use controlnet on it

tried reinstalling and all, same thing happens

[dim sonnet]

oo hol up

ah okay i see why -- matplotlib is used by controlnet_preprocessors but not actually listed as a requirement. Weird

[noble temple]

Yep the stacktrace shows it couldnt import it

[barren moat]

Is there any docs about inpainting for Swarm? When I click "edit" I can paint on the image and hit generate but I get the same image that has the "paint" on it. Like it's not sure the paint as a mask??

[dim sonnet]

the image editor is incomplete and a bit non-obvious atm -- by default yes you are in fact painting onto it. To use a mask you have to click the "+Mask" button

[barren moat]

Ok thanks, I'll give it a try

[noble temple]

Do you know if theres any way to fix it

[rustic river]

Same here. But playing with control percentage ("denoising") gives me some satisfactory results, though some percentages give very different result than A1111.

This could be due to how Swarm handles the upscaling weights I loaded in, dunno...

[noble temple]

Yeah, what I use to do in forge is gen in 512x512 and double it with hires, while this fucks up the image in swarm

[cold dirge]

Ollama - You need to install Ollama for this tool to work. Visit ollama.com for more information.
Optionally Set enviromnet variables for "ANTHROPIC_API_KEY" & "OPENAI_API_KEY" & "GROQ_API_KEY" with those names or otherwise it won't pick it up and the respective API keys``` how do i do this step ?

[noble temple]

and generating it at 1024 and doing refiner doesnt seem to do much, using xl btw

[dim sonnet]

update, I've pushed commits that (A) ensure matplotlib is not missing and (B) ensures controlnet preprocessor installs its requirements when installing

[cold dirge]

mainly api part

i have ollama running

[noble temple]

nice, thanks a lot

[dim sonnet]

shouldn't behave any different

[dim sonnet]

for XL you have to do 1024, you can't gen at 512, it'll just break

[dim sonnet]

again, the difference isn't the UI, it's the settings

[noble temple]

Hmmm it works on forge, looks awful at first but after refiner it cleans up

[dim sonnet]

you in the wrong channel?

[cold dirge]

should i ask in comfy ui

🤔

[dim sonnet]

what

ollama?

none of the channels here I'd say?

probably an ollama discord?

[cold dirge]

trying get ollama running in comfy

i will see if they have a discord

[waxen hinge]

Sure That should help. 😇

[rustic river]

Often when interrupting a batch, it just continues, unable to stop at all. Only seems to happen on my Linux system (comfy backend)

[drifting helm]

thats a comfy issue iirc, something to do with being unable to stop a certain step

[dim sonnet]

it shouldn't just continue without stopping -- in theory at worst it should finish the step it's stuck on or maybe another step past that, but then stop

this is annoying and unresponsive-feeling, but kinda an inevitable limitation of the underlying tech for right now

but if it stays continuing to go for a while longer / even returns images at the end? then something's broken

[rustic river]

It stops with the current image, then goes on to generate the next in the batch (and output it) and stops after that.

[dim sonnet]

oooh, that sounds like it might be overqueue not pushing interrupts forward properly -- I pushed a commit that should fix that if so