🐝｜swarm-ui

[hearty notch]

no swarm

thanks for the tip

[silver osprey]

or just load custom comfy workflow

idk if the integration actually uses the modules

[hearty notch]

to my understanding swarm is comfy when you are using as a backend, not sure if it will be deferent @dim sonnet ?

[silver osprey]

that's the thing, I'm not sure if the built in integration is working properly

you'd usually expect doubler the speed with AIT, heck, even triple if you have a 4090

[hearty notch]

and honestly, I am just having fun making cats XD @dim sonnet my cats vs your cats who win?

[silver osprey]

my cats =]

[hearty notch]

I will try it after i munch my burrito bowl hahah

i think your cats will be lawyers for Alex cats after the fight with my cats

[dim sonnet]

what?

it literally just adds the aitemplate loader module to a comfy workflow

there's no reason it would behave any different

ps speaking of comfy stuff i'm working on some special util nodes i need for somethin

idk if anybody has a use for this outside of what i plan for em but lol

[silver osprey]

idk then, what @hearty notch shared isn't the speed you'd expect from AIT

[dim sonnet]

the issue isn't swarm

idk what it the issue is, but not swarm

i've already shown you that my speeds double from the aitemplate checkbox in swarm

[silver osprey]

yeah, I saw

also, why is the latest ComfyUI commits mess up with AIT?

[dim sonnet]

i'd guess from the VRAM graph that you might be using up too much vram at high batch and that has negative side effects? idk. Try with batch size 1 either way

[dim sonnet]

that's a question for comfy or fizzledorf

[silver osprey]

not me tho ;], I just provide the modules haha

I think I know why AIT doesn't work correctly on new ComfyUI commits for my 4070ti. it's because new commits are aggressive with the VRAM, which screws with the entire 4000 series except for the 4090 and 4080ti

on commits from a month ago I get 3-4it/s on batch 4 with AIT, and with newer commits I get a few s/it also with AIT

[dim sonnet]

try launching with --disable-smart-memory?

Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.

[silver osprey]

already tried that..

it wasn't even close to older ComfyUI commits with AIT

[hearty notch]

@dim sonnet can you have a look at this?

you will see that 4 img batch is yeilding much better results timewise compared to 2 img batch

also there is an error when I try 3 img batch

it will not gen anything

only 2 and 4

[silver osprey]

you can't have odd numbers when doing batch

[hearty notch]

oh ok

I never knew that

well then I fold lol, i really do not see much step up in gen speed with AIT

[silver osprey]

idk then, it must be an issue on your end. other people report almost a triple speed boost with it on a 4090

[hearty notch]

can you connect me with somone from these people? maybe they can share

I can try webui if it's there also

[silver osprey]

they did exactly what you did I think, they used the commits I shared

[hearty notch]

ok, now I am keen to try again

I will try vanilla comfy

can you share a wf here?

so I can use it

[silver osprey]

just load the stuff I post, it uses a workflow I plan on adding to CivitAI

[hearty notch]

these?

[silver osprey]

yeah

[hearty notch]

ok, let me try and see

I will try first with up to date comfy and the nod alex shared,

then I will try the commit and the nod you shared

@silver osprey

do you have a simple one?

or just show me how to connect that nod

I will use basic SDXL WF from comfy

[silver osprey]

you just click the install missing nodes button

[hearty notch]

I think I need manager for that?

[silver osprey]

if you don't have that, just add the comfyUI manage to your extensions

[hearty notch]

wait let me install it

[dim sonnet]

don't bother with that, complexifies testing

use a default graph, add AITemplateLoader between model and ksampler

that, in theory, is all you should need to use AIT and test it accurately

[hearty notch]

yeah hold on a min

[dim sonnet]

and then just compare with vs without that added node

I did it. I created catdog.

[hearty notch]

only for base?

[dim sonnet]

use a default graph

definitely should not have a refiner in a default graph

a default graph as in

[hearty notch]

well i am using refinder with all the testing i am doing lol

[dim sonnet]

"I'm adding more variables and not thinking about them"

ain't how science gets done

if/when you have expected results on a default graph, then after that experiment with refiner

add variables only one at a time until you've found out what the issue was

[hearty notch]

this is 512x512

this is 1024x1024

only base

no refiner

this is from the nod and the commit of @silver osprey

now I will update and try again

now this is 1024x with all up to date CUI and AIT

basic graph

so it's true that people are reporting double, but with 512x

so there is 35% increase when using SDXL as intended

@dim sonnet is this what we also get in Swarm?

as I do not see the it speed anymore on the SUI console

[dim sonnet]

server -> server config -> set LogLevel to Debug

that'll give you all the comfyui outputs

[hearty notch]

ok let me try

[dim sonnet]

now That is a mildly diseased catdog

[hearty notch]

mad scientist

and yeah with swarm the 35% increase is there too

well 30 to 35%

[hearty notch]

wwait!! wat is that prompt!!!!

this is a new update??

[dim sonnet]

still in progress but yeah this is what i'm building atm

[hearty notch]

well here is a random blue cat for you then hahah

btw, i did notice that swrm now is not abusing the system recourses, I can see a very good use with VRAM, and GPU temp is more stable and now heating up too much

[dim sonnet]

from enabling AIT?

[hearty notch]

yeah this is what I notice today, I can not confirm if AIT is the reason tho

but my friend @barren moat also notice a very low vram use too

I usually get like 70+ degrees on the GPU

and now I am using the gen forever so it's non stop

[dim sonnet]

++ if you want to play with the experimental new regional prompting (NOT done, lot more to do before it's fully ready, eg having an interface to selection regions and all) I pushed an update, for now the way to control it is just type directly into prompt in a format like this:
a photo of a couple on a bench in a park <region:0.1,0.5,0.8,0.3,0.5> a bench <region:0.2,0.2,0.3,0.7> an old white man with white hair sitting on a bench <region:0.5,0.2,0.3,0.7> a middle-aged asian woman with dark hair sitting on a bench
where you can <region: ... > and that can be end to end regionalization and go back to global prompt (for preset compat), background to manually set a background region ie the region that consists of all areas not already selected, x,y,width,height as percentages (0.0 to 1.0) of the image space, or x,y,width,height,strength where strength is 0.0 to 1.0 (defaults 1.0) for how strongly to apply the region mask.
Global prompt (ie any text not inside a region) applies and merges everywhere, there's a slider under Advanced Options if you want to shift how strong.
If regions overlap they will merge - here's the masks used in the above example (woman, man, bench, background)

bear in mind it's not magic, eg in the image above the man decided to not have a head

[hearty notch]

Michael Jackson is squinting while reading all that!!

[cunning holly]

the slowness is a problem with the AIT nodes

[silver osprey]

they double the speed on commits from a month ago

with newer commits ComfyUI is too aggressive with the VRAM, so it screws with AIT

anyways, @cunning holly , are there plans to support BLiP diffusion in ComfyUI? it's just SD but with BLIP2 integrated with the image encoders allowing for things like this

it's so much better than IPA really, and it can also blend images at that level. they said they are also working on an SDXL version

not supported by AIT though. I'm wondering if it would be possible to make something like exLLaMa for diffusion? because exLLaMa almost squares the speed on LLMs while slightly improving quality. AIT can double the speed, but not THAT much. I tested exLLaMa today and it made LLaMa go from 8T/s to 62T/s

that's almost times 8 the speed.. I'm wondering if this is possible to achieve on diffusion models?

[dim sonnet]

not directly no

but the concepts used can be transferred

non-trivial task though -- diffusion models have way more internal complexity than LLMs

LLMs are simple software, big model. Diffusion is complex software, small model

[silver osprey]

SDXL isn't a small model

[dim sonnet]

it, uh, in fact is compared to any LLM lol

The smallest "standard" LLM is twice the size of SDXL

[silver osprey]

there are 3B llms

[dim sonnet]

3B LLMs are tiny

and ... SDXL is 3B

[silver osprey]

SDXL is 6b, isn't it?

[dim sonnet]

ehhhhhhhhhhhhhhhh

i mean they wrote that lol

it's only 6B if you add the base and refiner together

which is debatably valid

[silver osprey]

yeah, 6.6B base and 3.5B refiner

[dim sonnet]

no

3B base

and 3B refiner

3+3=6

except re the complexity of models thing, even that is debatably correct counting?

the base is 3 models in a trenchcoat, the refiner is 4 models in a trenchcoat

2 of those sub-models are the same between base and refiner too

(CLIP G and VAE)

[silver osprey]

I'm quoting: "6.6B parameter base model and a 3.5B parameter refiner model."

[dim sonnet]

hwot

quoting what

cause that's wrong

[silver osprey]

that's written in stability.ai somewhere

[dim sonnet]

https://stability.ai/blog/sdxl-09-stable-diffusion boasting a 3.5B parameter base model and a 6.6B parameter model ensemble pipeline (the final output is created by running on two models and aggregating the results)

i-

need to go poke somebody to fix https://stability.ai/blog/stable-diffusion-sdxl-1-announcement lol a 3.5B parameter base model and a 6.6B parameter refiner. is wrong

[silver osprey]

in some other places on the website they also mismatched the base and the refiner

[dim sonnet]

where?

[silver osprey]

the base is bigger than the refiner

hold up, I'll pull it up

idk, I can't find it atm.. but the base is bigger than the refiner, right?

[dim sonnet]

yes the base is slightly larger

different architecture & also has an entire additional textencoder

[silver osprey]

the architecture is the same, I studied that. when I provided the AIT modules, it didn't need for me to specify different architectures for the base and refiner

both are StableDiffusionXLPipeline

they are trained differntly though. the refiner is trained on SDXL's common mistakes I think?

[dim sonnet]

The architecture is different

like majorly different at that

refiner has significantly more weight dedicated to attention blocks at high layers, whereas base's weight is centralized at the mid blocks

[silver osprey]

they use the same pipeline to be compiled and loaded though

honestly, the refiner is kinda useless when it comes to modern XL models like JuggernautV2 and such

actually, the refiner is devolving the outputs of good base models

[dim sonnet]

maybe in diffusers? idk what diffusers does, but they are definitely different architectures

you can look at https://github.com/Stability-AI/generative-models/blob/main/configs/inference/sd_xl_base.yaml and the refiner config and notice they are explicitly different

[silver osprey]

no, in everything really. when I compiled the AIT modules it was loaded using the same pipeline

odd. when making the modules they both used the same pipeline..

[dim sonnet]

eg base has transformer_depth: [1, 2, 10] and channel_mult: [1, 2, 4] where refiner has channel_mult: [1, 2, 4, 4] and transformer_depth: 4

the AIT modules are like 5 MiB, I doubt they actually contain all that much

[silver osprey]

they just allow models that share the same architecture as the compiled ones to be loaded via MSVC engines, they weigh 5MiB because they don't contain weights

while TRT and OneFlow include the weights in their own engines

if TRT and/or OneFlow would not be checkpoint specific and be architecture specific, without a doubt they can be better than AIT

but that's not the case as far as I'm aware

the AIT modules have a range system instead of being specific to each settings. for instance, the modules I compiled can be used on anything between 64-2048px and 1-4BS and 0-150T

[dim sonnet]

they... clearly allow more than just the same architecture, if you built one for base and it works for refiner?

https://huggingface.co/Fizzledorf/AITemplateXL/tree/main/modules/windows/sm80/bs1/1024 there looks to be different ones for XL vs XL Refiner

[silver osprey]

that's the thing. they used the same script to be compiled, but the one compiled for base only work for base models, and the ones compiled for refiner only work on refiner models

[dim sonnet]

lol

[silver osprey]

yeah, I know.. I'm the one that provided those modules in the first place....

[dim sonnet]

is fixed now

[silver osprey]

anyways, is this still relevant? because I might be compiling new stuff for this soon

or perhaps by the time ComfyUI supports that there will be an equivalent of exLLaMa?

[barren moat]

@dim sonnet I'm having some strange issues when switching models. Almost like it gets stuck between models or something. What's weird is if I got back to the stock default sdxl base model my images come out like they should but when I try to swict to any custom model it's either ignoring my prompt or not loading any model.

But if I load an older image and tell it to "Reuse Parameters" it works fine untill I go to switch models.

That's with a custom model and now this is just switch back to the stock sdxl base...

I've tried resetting the UI back to default settings and still get the same issue. Here is a new screenshot switching to a different model...

Same prompts and settings just switched models.

If I go back to an earlier image and reuse the params it works again ...
.

My friend @hearty notch and I were trying to get it fixed but didn't come up with anything.

[thick moon]

O: is there a way to Reduce the displayed nodes or Sort them on the side when trying to Add a Custom Workflow in the Genrate Tab?

[dim sonnet]

by default when up-to-date, anything you made and titled a Primitive node for goes to the top, anything else goes under Advanced

so just add as many primitives as you want displayed parameters

(also default parameters, when matched, get retained)

[dim sonnet]

I don't follow what the issue here is?

without understanding your message but glancing at your images, is it just you're not expecting those images from those prompts? If so I think more likely the core issue is, uh, the models lol. A lot of models on civit are heavily tuned to produce pictures of girls more than anything else, and your prompt can be interpreted by a lazy CLIP as asking for (A) "girls", who are wearing (B) a t shirt with a vector design, which is what the images you got are

ie the base understands that your subject is itself a shirt design of a zombie man, but girl-focused models see the keyword "girl" and glom onto it as the primary subject, and reinterpret everything else as details to put onto the girl

[barren moat]

Yeah it’s difficult to explain. The models aren’t the issue, I’ve used the same models for a couple months without any problems.

If I load an image and reuse the parameters it works without a problem, if I switch the model and re try the same prompt i get those weird images. I’ve a few models (that I always use) and get those odd images. If I go back to the 1st model that was previously working fine, I still get weird images. I have to shut down the server and reload an old image (with a different model I want to use) and then reuse the parameters, then it works. This started last week and I reached out to my friend @hearty notch for help yesterday. It seemed to work fine until this morning it started this again.

Maybe I should delete it and start with a fresh install??? I haven’t changed anything (that I can remember) that would’ve caused it. It’s weird for sure so I wanted to let you know in case it comes up again. I’ll try the reinstall and see if that fixes it.

[barren moat]

Actually these images I showed for examples were better then what I normally get which is basically faded gobs of noise that doesn’t look like anything lol

[hearty notch]

Actually it’s a weird bug that’s starting randomly, what’s happening when you gen something on a new session, the results are weird and not following the prompt, I find a solution by loading (reusing) the parameter of a previous generated img then the ui will start to function again. The thing about this anomaly you don’t see any error it’s just like the ui parameter stuck on something no matter what you change on prompt or parameter of the ui

Hope this explain it more

[barren moat]

Fresh install underway now

[barren moat]

iirc this started happening after the update that seperated the 1.5 models from the sdxl models, not sure if that could be related?? With the fresh install I'm getting the same issues. CrystalClearXL is a pretty popular model and it's giving me issues with Swarm, but what's weird @hearty notch can use it just fine. TBH I can use in other UI's. JuggernautXL V5 is working fines and one other model I have seems to work fine, but the rest oif them aren't. Even the RunDiffusionXL model isn't working.

[dim sonnet]

if parameters are different when generating than they are on the sidebar, this will show in the metadata at the bottom (unless the metadata itself gets stuck on a prior gen)

metadata is processed independently

[dim sonnet]

those corrupt blobs are weird

do they happen without (A) refiner and/or (B) dynamic thresholding?

after the update that seperated the 1.5 models from the sdxl models, not sure if that could be related that's... odd, no reason for that to be related

that was a ui/model rec change

if you told me it started yesterday, i could see it potentially be a bug from the prompt region update, eg maybe it's not inputting prompts correctly

oh or from the SwarmKSampler update, maybe there's misbehavior somewhere there

[hearty notch]

This anomaly is few weeks back, not from recent updates, this is I can confirm

[barren moat]

It does it with and without the refiner

[dim sonnet]

can you give me a bare minimum like... step-by-step guide or video on how to replicate this?
Like from start when you launch swarm, what params do you set, what do you click, when/how do you change model, etc

[barren moat]

I can make a video tomorrow morning (it’s 11:25 pm here in San Diego) or you can log in my pc with team viewer (or something else) and check it out if you want.

Strange because I did a complete fresh install and still have the same issue. Seems related to the models but what’s strange is I’ve used the same models many times with swarm without issues.

[dim sonnet]

might be worth checking what the model type is specifically in case it's relevant

eg i downloaded RunDiffusion since you mentioned that and i immediately notice

it's a 0.9 model

this shouldn't cause any issues like that

but if it's always only relevant to 0.9 models then that's a clue at least

what in the fuck

okay good news i can replicate the bug lmao

[barren moat]

Yes 🥳🥳

[dim sonnet]

... lmao

[barren moat]

It’s some weird stuff

[dim sonnet]

are you sure this isn't new?

[barren moat]

CrystalClear also

[dim sonnet]

the above happened because of a bug with 0.9 models in the Prompt Regions update, tracked that'n down instantly

the blobnonsense I mean

the unexpected gens that are coherent but different than what you expect is separately from nonsense blobs

[barren moat]

Maybe that’s the issue and the old issue was related to my settings that we corrected the other day and this maybe new/not related to my original issues

[barren moat]

I also noticed CrystalClear is a .9 base model as well

[barren moat]

That's it! It's only happening with models based on 0.9

[dim sonnet]

the blob stuff is fixed

is the other issue also exclusive to 0.9?

[barren moat]

I'm only getting blobs from the 0.9 models

let me update it and try it again

did you push the change?

[dim sonnet]

yes

[barren moat]

ok updating now

[dim sonnet]

RunDiff seems to do more or less what i expect on the vector shirt prompt

[barren moat]

haha yeah it does

it's fixed!!

Thank you @dim sonnet !!

[dim sonnet]

yay

ps if you edit model metadata and set

it'll treat it as an SDXL 1.0 model

[barren moat]

ok I'll do that

[hearty notch]

This made my day! Thanks for sharing this golden nugget

[hearty notch]

I am looking at this and I want to have the option back to create such things

Actually that’s a lovely idea to have the option on the ui to make things go nuts, not sure if this is doable code wise, like a madness mode switch

Or this is something we can achieve with pushing the freeu settings maybe?

[dim sonnet]

you can create that by leaving the prompt box empty

[hearty notch]

I do that sometimes, but with SDXL, I am getting some normal results

I mean nothing super weird like that poor little sheep

[dim sonnet]

yeah SDXL base produces almost-coherent content when left unconditioned

finetunes are often liable to either create cursed things, or output corrupted approximations of what they were trained on

[dim sonnet]

^ eg the last output here being an asian woman who i think is not wearing a shirt, indicates that's probably a model trained on lots of nsfw

[thick moon]

Ah Okay Thank for the info

[hearty notch]

You think? Haha

Tbh, it’s really sad to see the community putting too much efforts on the NSFW, I really don’t see the value in there

Also, the other side is putting too much efforts or either anime or photorealistic, and they use these too styles as a benchmark on the model performance

Personally I think SDXL vanilla is quite powerful and capable

I really hope to see more efforts from stability on releasing periodically updated models

So we can have a proper competition with MJ, as the fidelity on their model (which I have a strong believe it’s a latent model like SD) but they are using Lora and presets on a more automated way,

Speaking of which, are we going to see a ui for training and fine tuning from stability soon?

[silver osprey]

I'm currently compiling AIT for CNET SDXL. is there a difference between unet_control controlnet?

[dim sonnet]

i mean i've been thinking about integrating Kohya as a backend for a training tab in swarm, that's a future project tho

[dim sonnet]

Iii do not understand the question, that sounds AIT specific?

for SDXL controlnets you'd want to implement ControlLora support ideally, i believe that'd probably require a code change not just compiling a module

[silver osprey]

no, the AIT module naming system works in a specific way. I believe I found a way to answer that question

nah, I'm getting this stupid error

can't even use the modules kek

using AIT on controlnet in general seems like an impossible task

guess we'll have to stick to pure pytorch for using CNET for now

[worldly oxide]

Hey does Qrcode monster for controlnet work in StableSwarm?

[dim sonnet]

iii don't see why it wouldn't

so probably

if it doesn't you can poke me and i'll fix it

[worldly oxide]

Thanks for the quick reply I'll give it a try

[barren moat]

Good morning @dim sonnet! Is there a way to write a prompt that will give you an image for each subject in your prompt? Kinda like the wildcards but instead of a picking a random subject you get an image for each?

Sample prompt.....

a {cat, dog, lion, bird} in a suite

so you get a generation for each?

[dim sonnet]

oo, interesting idea.
insofar as what you can right now, a cat in a suit (I think you mean "suit" not "suite" lol), and then select the Grid Generator tool and set axis to Prompt Replace and cat, dog, lion, bird

adding a prompt-level combinatoric autogeneration might be nice for convenience tho ye

[barren moat]

Yes suit, not suite lol. I haven't had coffee yet ☕ 🥱 😴

Ok great I will try that

[barren moat]

@dim sonnet worked perfectly!! Thank you!

I forgot I had my batch size set to 2 so it made 2 Identical images of each subject

[hoary basin]

Mid-autumn festival poster

[hearty notch]

🫨

[hearty notch]

Hi @dim sonnet , I am trying to install SSUI on a brand new installed windows, which exact python version you recommend, and what else to I need beside python and git before I hit the install-windows.bat

[dim sonnet]

you don't

lol

just run the installer, it does everything automatically

[hearty notch]

`Failed in attempting to update the source: winget
Failed when searching source: winget
An unexpected error occurred while executing the command:
0x8a15000f : Data required by the source is missing

No packages were found among the working sources.
Failed when opening source(s); try the 'source reset' command if the problem persists.
An unexpected error occurred while executing the command:
0x8a15000f : Data required by the source is missing
'git' is not recognized as an internal or external command,
operable program or batch file.
The system cannot find the path specified.
'.\make-shortcut.bat' is not recognized as an internal or external command,
operable program or batch file.
'.\launch-windows.bat' is not recognized as an internal or external command,
operable program or batch file.
Press any key to continue . . .`

this is what I am getting whem I run it

[dim sonnet]

this is a brand new windows install ye? Restart the computer probably

[hearty notch]

I did fiew times already

[dim sonnet]

winget is a part of windows that installs itself via microsoft store automatically shortly after the pc is installed, and seems to be the part that's misbehaving there

(I don't know why it installs after pc install vs during, but, idk microsoft logic)

[hearty notch]

ok I will restart one more time, and try one more time

hit me baby one more time

[dim sonnet]

if a restart doesn't suffice, might try the message's recommendation - open a terminal and try winget source reset

[hearty notch]

can i pm you?

[dim sonnet]

uhh if you wanna show something private or something sure

generally i prefer using normal channels when it's not explicitly private

[hearty notch]

thanks for the reply

appreciate the understanding on taking it to PM tho 😄

it's working now

[dim sonnet]

yay

[hearty notch]

after pulling last update

@dim sonnet am I missing something?

when AIT is enabled I ger this error

[dim sonnet]

comfy update broke ait :(

can disable auto-update in swarm and downdate the comfy backend to make it work again

[hearty notch]

oh

I do not know to to downdate tho

[dim sonnet]

i think 9546a798fba3c9fc9b6aee26cef46674a184727c should work

[dim sonnet]

open a terminal in swarm/dlbackends/comfy/comfyui and git checkout 9546a798fba3c9fc9b6aee26cef46674a184727c

[hearty notch]

let me try that

how to disable auto update?

[hearty notch]

work like a charm

[dim sonnet]

it's on option under Server -> Backends, just edit the backend and untick the checkbox for that

[hearty notch]

something very unique just happed check the seed on the gens

there is controlnet or init img

seed is set to random btw

is this possible?

I keep getting identical gens

all are dif seeds

did I find a bug?

[hearty notch]

looks like using AIT nod in a comfyUI wf to the gen tab is causing this bug

[dim sonnet]

0.o?

[dim sonnet]

do you have multiple samplers in your workflow? You might want to check if potentially some of the seeds are non-random

esp. if the primary ksampler's seed is static it's liable to make duplicates

if the seed attached to the swarmui Seed param isn't the important one and you use BatchSize or Images param it also might not increment properly

(if in doubt you can link seeds to a primitive and name that primitive SwarmUI: Seed to force it)

[hearty notch]

this is what I have in all my wf

i always convert the seed from the samplers into an input and connect to premitive nod

like so

[dim sonnet]

er... Seed Here is not the same as Seed

[hearty notch]

It was working fine for the last month

This happened only yesterday

[stable vine]

Will stableswarm ever get dynamic prompting via wildcards or similar?

[dim sonnet]

you can use <random:a, b, c> today, and a proper wildcards engine is planned

[gray geyser]

Does anyone know if I can find documentation on how the regional prompting of the ui works?

[hearty notch]

The wiki is a good place to start

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

[nimble belfry]

is there a code examples of the API specifically the ListModels? i am returning an empty list. Maybe im not adding the path properly?

        print("Getting model list...")
        url = f"{SWARM_URL}/API/ListModels"
        session_id = await self.bot.get_cog("APICalls").get_session()
        params = {"session_id": session_id, "path": "StableSwarmUI/Models/Stable-Diffusion", "depth": 2}
        async with self.session.post(url, json=params) as response:
            if response.status != 200:
                raise Exception(f"Failed to get model list. HTTP Status Code: {response.status}, Response Content: {await response.text()}")
            data = await response.json()
            print(f"Received data: {data}")
            return data.get("files", []) ```

[nimble belfry]

wait nevermind I got it. { "path": "", "depth": 2, "subtype": "Stable-Diffusion", "session_id": session_id }

[dim sonnet]

does it not properly autofill the subtype if you leave it off

[hearty notch]

anyone else is getting these faces ?

am I missing something new with the settings?

[nimble belfry]

I think it does the path was my issue.

[dim sonnet]

OH yeah yeah it was

[nimble belfry]

To choose the model with the API what key does it go by? title?
Payload being sent: {'session_id': 'A30DED9081586A502890AA7B1FE2', 'images': 4, 'donotsave': True, 'seed': -1, 'prompt': 'portrait of a frog wearing a crown', 'negativeprompt': 'low resolution, bad quality', 'model': '1. Stable Diffusion XL 1.0 Base', 'width': 1024, 'height': 1024, 'steps': '10'}
ERROR:root:Failed to generate image. Error: Invalid model value for param Model - are you sure that model name is correct?

I renamed the model title to 1. Stable Diffusion XL 1.0 Base so it should be correct.

[dim sonnet]

title is for UI display, you want the internal filename - name in the json for a model. I again encourage you to just do things in the UI and look in browser tools at what the API calls being made are

[silver osprey]

Hey guys. Will stableswarm have an LLM feature in the future? We won't stand closed source AI getting anywhere near open source imo, so isn't that something we must achieve?

[dim sonnet]

maybe?

I have a discord bot https://github.com/mcmonkey4eva/SimpleDiscordAIBot that combines LLMs via text-gen-webui with stableswarmui for image gen

can play with it in the r/SD discord or the text-gen-webui discord

[silver osprey]

I was thinking of future versions of stableswarm having a capability for oobabooga API as well as ComfyUI api, then you would specify the primitive with the positive prompt and you would chat with whatever LLM you're running with oobabooga and when you ask it to generate images, it will create a prompt and send it to the workflow then send you the outputs back

[silver osprey]

one issue with this will probably to fit both an LLM and SDXL in VRAM at the same time, idk if that's even possible without making them switch places in the VRAM.. @dim sonnet does that bot do smart mem management to always keep 1 model in VRAM and the other in CPU RAM and have them switch when being used?

[dim sonnet]

it's possible to do but my code for that bot doesn't do it currently

[silver osprey]

I see. I guess the way to do it would be to implement ComfyUI into Oobabooga itself then.. does Oobabooga have any plans to do so?

[worldly oxide]

Hey @dim sonnet I have just reinstalled Stableswarm on a new PC with a 4080 and I keep getting this error message, can you help?

[dim sonnet]

er... did you customize the output path in settings? that appears to indicate the save path has an unrecognized parameter in it

i'll fix it to handle that better rather than just dumping red error text

[worldly oxide]

Cheers I fixed it, in the Server Configuration I added "localhost" to Host

[ripe basin]

hey guys 👋🏻 Is it possible to use this UI so i can use all three of my nvidia titan cards at once? my PC has three GPU cards

[dim sonnet]

yes

[neon atlas]

How can i fix this pls. new workflow cant find the error works in comfyui though, not the generate tab

[dim sonnet]

pushed a fix for that, but it's erroring because your workflow doesn't have a standard prompt input

the fix will prevent the error but it might still be beneficial to add a prompt input since that's generally assumed by so many things to exist

[neon atlas]

***thank you ***

[hearty notch]

@dim sonnet hello, is there any way to (negative) interrogate images in the UI?

[nimble belfry]

When generating images with the API. I create an image grid that I send to discord and I wanted to be fancy and display the preview images in my grid as well. It seems like I need to send a keepalive ping of some kind?

ERROR:root:An error occurred: sent 1011 (unexpected error) keepalive ping timeout; no close frame received

This only happens when my images take a long time to generate. I do 4 images and sometimes they take 25 seconds each when I choose higher steps.

[nimble belfry]

@silent rose @uneven moth spam.

[uneven moth]

Ty ❤️

[nimble belfry]

[dim sonnet]

websockets have a keepalive ping as part of the protocol internals that are normally abstracted away from you

on swarm's side the timeout is i think 2 minutes?

oh nope i'm wrong it's 30 seconds

i could probably tweak that, but, yeah, your client should be handling the keepalive packets automatically

oh no 30 seconds isn't the timeout, it's how frequently it sends the packet

so i can't tweak that

(and it's set low to compensate for external APIs that expect it to be below a minute and get upset if it takes too long)

[nimble belfry]

Hmm, maybe I need to investigate it a bit more. It would really suck if I had to change how the grid looks. It is request heavy though.

It also takes a long time 25 seconds times 4 to complete if I do high steps.

[dim sonnet]

Should be fixable, worst case you just gotta add some manual keepalive packets i think

but probably whatever you're using as a ws api has an option you need to set somewhere

[nimble belfry]

Looking good right now.

[dim sonnet]

yo nice

[nimble belfry]

I like the boarder when it's generating

[dim sonnet]

ye that's neat

[dim sonnet]

clip/blip interrogator isn't a feature atm but is somewhere in the todolist

[nimble belfry]

I'm just using python websockets https://github.com/kalebbroo/bAIb_Ross/blob/main/cogs/api_calls.py#L96-L99

[red canyon]

hello there anyone using stableswarm with amd?

[dim sonnet]

according to google: Timings are configurable with the ping_interval and ping_timeout arguments of connect() and serve().

[dim sonnet]

AMD works with modern (ROCm-compatible) cards out of the box, and works with older (DirectML-compatible) cards if you manually install directml to the comfy backend: https://github.com/Stability-AI/StableSwarmUI/issues/23

[red canyon]

thank you, any exact answer in that threat I am looking for or should I go testing until I find one that works?

thread*

and it says to put an argument, where? I tried looking inside the windows bat file but nothing there

[dim sonnet]

https://github.com/Stability-AI/StableSwarmUI/issues/23#issuecomment-1667174967

[red canyon]

Sorry about the late answer, I tested this and it gives me this error, plus, may I ask where I can find the place to write the --directml, I checked on the launchers with no luck

nvm I found where the extra args is, thought it would be on an outside thing

still it says no version of torch-directional satisfied the requirement

should I look online for an specific one?

btw thank you for all of this

[dim sonnet]

does pytorch-directml work?

[red canyon]

unless I wrote something wrong

[dim sonnet]

uhhhhhhh

@cunning holly do you know how to get DirectML working?

[cunning holly]

yeah you need pytorch 2.0

[dim sonnet]

as opposed to 2.1? 0.o

(swarm's autoinstall for comfy atm is ComfyUI_windows_portable_nvidia_or_cpu_nightly_pytorch.7z)

[cunning holly]

yeah the latest torch-directml is from april for pytorch 2.0

but honestly it sucks

people with AMD cards are much better off with Linux

[nimble belfry]

@dim sonnet Thank you! 14:38:55.882 [Info] Creating new admin session for ::1 14:40:03.507 [Info] Generated an image in 12.18 (prep) and 52.79 (gen) seconds 14:40:54.552 [Info] Generated an image in 65.06 (prep) and 51.00 (gen) seconds 14:41:46.072 [Info] Generated an image in 51.04 (prep) and 51.48 (gen) seconds 14:42:37.389 [Info] Generated an image in 167.60 (prep) and 51.28 (gen) seconds I had to play with the keepalive time for a little but now even with settings jacked all the way up taking almost a minute per image. Its working without anymore timeouts.

Processing image data...
Processing image data...
Processing image data...
Connection closed OK``` Now to finish the buttons that should attach to the embed so people can upscale or do img2img from one of them. ![GuraSweat](https://cdn.discordapp.com/emojis/763994516281491476.webp?size=128 "GuraSweat")

[dim sonnet]

:D

[silver osprey]

hey guys, I think I fully mastered image blending on SDXL. would this be easy to implement into stableswarm?

also zero-shot generation

[hallow forge]

Can stable swarm run multiple gpu's in series?

[dim sonnet]

not sure about the specification of "in series" in that question, but yes swarm runs on multiple GPUs

[dim sonnet]

you can drag images into the prompt box to have them added as ReVision inputs if that's what you're doing?

---

Also! new feature announcement! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7236821
Automatic segmentation detailer. Useful and/or fun depending on what you do with it :D

[hallow forge]

Is that something that is possible/(talked about) in the future?

In reference to something like animatediff

Or any animation application for that matter

[dim sonnet]

is... what?

you're missing a lot of context to what you're asking here

[silver osprey]

that's not revision, revision is just a small fraction of that workflow

it's using IPAdapter

ReVision isn't actually able to blend the images, it's more like it mixes the styles.. but this is actual image blending achieved in ComfyUI

also if you try to do zero-shot with ReVision it won't keep the subject, but this does..

[dim sonnet]

well i definitely can't impl it in swarm if you don't show me the nodes that actually make it work lol

[hallow forge]

Apologies I'm not a tech guy getting use to the lingo. For instance in Animatediff you need to create several images in batches to create frames for animation. This takes a lot of vram especially when upscaling. I'm wondering if it's possible to have more than one gpu working together to render the images faster more efficiently instead of each gpu working independently with each queued image? like all of them contributing to a single queued generation

[silver osprey]

there are two workflows I made that are universal to all image inputs (it seems like that, at least..) I made a workflow for zero-shot, and the other is image blending. I've been holding back those workflows for quite a while now

[dim sonnet]

AnimateDiff relies on it being a shared batch per run, so, uh, not currently possible.
In theory it's possible to generate longer animations by running multiple separate animatediff runs in parallel and merging the borders

[silver osprey]

as seen here, I made it so the scales of imageA doesn't have to be identical to imageB; which is a known issue with IPAdapter

[dim sonnet]

(I say "not currently possible" as somebody might release a way to get around the batch limit)

[dim sonnet]

where applicable in swarm i usually just autoscale things to match

[silver osprey]

I'll try to implement the workflows on swarm, but it will require some additional nodes (maybe..?) just for swarm to be able to send image inputs to the right nodes in the workflow, then it should work

I'd say the image blending itself is close to perfected though

[hallow forge]

What do you mean by merging the borders like basic editing?

[dim sonnet]

no, as in:

say you want 200 frames, and are generating 24-frame standard chunks
you can split into 10 chunks, and generate in parallel every-other chunk, like, Gen/Wait/Gen/Wait/Gen/Wait/Gen/Wait/Gen/Wait
That's 5 parallel runs, using 5 GPUs
then for eg slot 2, you have a generated result before it, and a generated after it, so you can just do a 24 frame chunk where the first 2 frames are the last 2 frames of slot 1, and the last 2 frames are the first 2 frames of slot 3, and mask them off (to not replace them, like inpainting, but over the time dimension ie the batch), you're then generating 20 new frames that will naturally fit between the first and last, and so you can do all 5 waiting chunks in parallel, using all 5 GPUs again

[silver osprey]

@dim sonnet it did actually import, for some reason it lets me upload 1 image..?

just need it to send a specified image to a certain node, and another to a clone of that node

ohhh, is swarmloadimage the node I should be using?

nope, that didn't fix it =/

[dim sonnet]

loading images directly from workflow on gen page is a lil messy atm

why are your sliders so weird 0.o

[silver osprey]

damn, the workflow I made really masters image blending, I guess I'll just stick to normal comfy when I'm using zero-shot or image blending =[

[dim sonnet]

i mean if you don't want to share the workflow at all, then, yeah

if you let me get my grubby monkey paws on it though i can make it super shmoove

[silver osprey]

I'm planning to share it; installation is wacky though

also the zeroshot one is seamless

[hallow forge]

How would I set this up in comfy? is it possible to continue this conversation in dm's?

[silver osprey]

of course I can give it to you, it's just not really simple to setup, there are certain models you need and certain nodes.. but of course:

[dim sonnet]

i prefer against DMs when not needed
and: doing something like that today requires technical experience and the attitude to dig in and go figure it out the hard way

was more a thought for how it might be built in the future

[hallow forge]

that's why i asked lol, im interested in figuring this out! All good

[dim sonnet]

had to smash that workflow into pieces a lil to make it loadable, but

got it to load eventually

so essentially it's:

• feed first sampler with just unclip + ipadapter
• then do a refiner with just prompt
• the ipadapter needs a bit of a mess to load properly in comfy nodes

pretty sure i could make this down to just a checkbox that's available when using revision

the only 'special' part here is the addition of ipadapter

unless there's some magic to the prompt splitting

[silver osprey]

second stage is using the same base model but instead of IPAdapter it first goes through FreeU to make the diffusion on stage 2 more aggressive; IPAdapter on it's own then have the second stage use a normal model would make images with lower quality. if you're not using the IPAdapter with negative prompt and a general prompt template; the quality is destined to not be as good.

it's the little settings that allow it to blend stuff seamlessly

[silver osprey]

so not only that; also FreeU on the second stage, the use of a negative prompt, and the use of just a generalized prompt template such as SAI-digital-art or SAI-enhance

same goes for the zero-shot version

[dim sonnet]

prompt template part is irrelevant for the context here - swarm already has https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Presets.md for that

not sure what "the zero-shot version" is

I think you're mixing too many different concepts in one go here

[silver osprey]

it's the version with just one image input, with text:

[dim sonnet]

is there anything special about it other than only having one image input?

(relative to the workflow i'm already looking at)

[silver osprey]

it uses different FreeU settings, and no unCLiP

also the prompt isn't zero'd out before the first stage, allowing it to follow a prompt while keeping the subject in the image

essentially, zero-shot is: [IMAGE] as a.../doing..../wearing...../etc... while image blending is just [IMAGE]+[IMAGE]

[dim sonnet]

++ Pushed IP-Adapter support for when using ReVision

++ added FreeU Apply To so you can control whether freeu applies to base, refiner, or both

++ added ReVision Zero Prompt to aggressively zero prompt given to the base even if has text (so only the refiner stage gets it)

if you manually specify a refiner model that will cut it off from receiving the ipadapter

@silver osprey anything I'm missing to replicate your workflow purely in swarm?

[silver osprey]

you've got access to it, you tell me.

as long as FreeU is being used on the second stage with a prompt template, you should be good using IPAdapter_plus

[dim sonnet]

Well my answer is ip-adapter is all i needed

but you care about a lot of random details

[silver osprey]

that's about the entire capabilities of the workflow; it's the smallest settings that add up to make a pure blend of IMG_a and b

about zero-shot, I don't see any reason why it would be hard to import, having just one image input

this is probably one of the more important things you might've missed, but idk

[dim sonnet]

that's the slider at the bottom

[silver osprey]

and I'm assuming stage 2 uses the same settings demonstrated in the workflows?

[dim sonnet]

what

[dim sonnet]

that is a sampler

[silver osprey]

also stage 1 needs DPMPP_3M_SDE_GPU

[dim sonnet]

do they need different sampler methods

(no)

[silver osprey]

they don't but the goal was to just make a workflow that will work on any combination of 2 images, and the way to achieve that is to find the most suitable stuff for all use cases

[dim sonnet]

i can almost guarantee that using 3m_sde on base and 2m on refiner does not noticeably change how well it works on the average combination of any given 2 images

(vs just using 2m on both or 3m_sde on both or whatever)

[silver osprey]

oh, yeah, also; that's not the right IPAdapter model

[dim sonnet]

it's a dropdown you just pick what you want

that's the one i used cause the one you used errored out when i tried it

[silver osprey]

those 2 workflows were made for IPAdapter_SDXL_plus

[dim sonnet]

sir

I'm revisioning your status from "cares about details" to "problematically obsessed with details"

you can use the plus model if you want

i used the not-plus model, it worked well

[silver osprey]

it was meant to be the default settings, when new users use all that, they have no idea what do these settings mean; they just want to blend images. and the way to allow them to do that right away is probably to provide at least decent settings imo

[silver osprey]

I mean, you won't get blended results like that with just default settings and the normal IPAdapter, to achieve blending at that level you need to refine your workflow. when you skimp on the littlest details you loose some of that. you don't have to use any specific settings; but I can almost guarantee you that it'll make better results.

[dim sonnet]

as far as swarm development goes, I'm just building the toolkit and UI and all

You can set up all the details how you like em, then save a preset, and even share that preset so others can use the same one

If you can show me that using a different refiner sampler vs base sampler actually makes a notable improvement, I'll add an Advanced param for it

[silver osprey]

sure, hold up

[dim sonnet]

everything else you're using, can be configured in Swarm via selecting the settings you want I believe

[silver osprey]

oh yeah, also 30 steps with 9.0 CFG and SGM_uniform should be a default for the first stage, will provide examples for that in a sec

[dim sonnet]

you are free to configure your step count/cfg/scheduler at will already in swarm, easy params for em all

[silver osprey]

DPMPP_2m with normal scheduler for the first stage

DPMPP_3M_SDE_GPU with SGM_uniform

both are with same CFG and stem count

both did produce a blended image though

like; if something like Clipdrop will have this, of course it will need some kind of standard settings so users can just use it right away; simplicity is key when it comes to this- so it's always nice to have stuff that work with any input and produce near best results

[dim sonnet]

yeah, both worked fine

i wouldn't say either is better than the other

which is funny cause you used the wrong scheduler for the first one

(should be karras)

[nimble belfry]

I am adding logic now for the upscale button. Does the API calculate the width and height of the new image? I see there are some new parameters where you can enter in the previous ones.

[dim sonnet]

huh?

[nimble belfry]

The upscale button in the UI. I'm adding an upscale button to my bot. When I look at browser dev tools it's just a text2imgWS request with some extra perams. When I replicate that it does not upscale. So my question is do I need to calculate that increase in width and height myself or am I doing something wrong and it should be filling that in when I use the "previous" width height perams?

[dim sonnet]

oh, yeah, that just shoves the image to Init Image and sets width/height to 2x their original value. There is no "previous width" param

[nimble belfry]

I called it the wrong.thing This is what I saw . That must just be how the UI knows what value to x2?

lastparam_input_width:

[dim sonnet]

that's the browser storage/cookie name used to retain parameter values in the UI when you refresh the page, shouldn't be going over network at all

[strong steeple]

I was wondering if there is "highres fix" in stableswarnUI?

[dim sonnet]

yes, called the refiner stage in swarm, you just

enable refiner, use PostApply, set the upscale

also configure the others how you want

don't need to select a refiner model if you're not wanting that (which auto's highres fix doesn't do) but you can if you want

probably also want higher control percent, 0.3 or 0.4

[neon atlas]

will you be integrating an image editor like photo pea-embed or something else and when if you don't mind me asking (Exited for all you work thank you)

[barren moat]

Hey Alex are you going to put IPAdapter in the front end of Swarm? If it’s already there forgive my ignorance 😅

Something like Fooocus uses. If that’s even called IPAdapter. Basically using other images to manipulate your generation. Not sure if I’m explaining it correctly

[dim sonnet]

image editor eventually yes, that specific one idk

[dim sonnet]

yes just drag images to the Prompt box. By default it uses ReVision, but if you have https://github.com/laksjdjf/IPAdapter-ComfyUI installed to your comfy it will give you the option in swarm to use ipadapter in the ReVision settings

(atm not self-installing, might make it self-install or one-click-install in the future)

[barren moat]

That would be fantastic! I’ll have to ask @hearty notch for some help getting it installed. Thanks for the info 🙌

[nimble belfry]

im trying to install the UI on my server running linux mint.

17:34:32.400 [Info] Self-Start ComfyUI on port 7821 unexpectedly exited (if something failed, change setting `LogLevel` to `Debug` to see why!)
17:34:32.401 [Info] Self-Start ComfyUI on port 7821 had errors before shutdown:
ComfyUI error: Traceback (most recent call last):
ComfyUI error:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/main.py", line 69, in <module>
ComfyUI error:     import comfy.utils
ComfyUI error:   File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/comfy/utils.py", line 1, in <module>
ComfyUI error:     import torch
ComfyUI error: ModuleNotFoundError: No module named 'torch'

but I have torch installed...

Defaulting to user installation because normal site-packages is not writeable
Requirement already satisfied: torch in /home/homrlab_admin/.local/lib/python3.10/site-packages (2.0.1)

[dim sonnet]

don't need it installed globally, need it installed in the comfyui venv

the installer in theory should automatically install it there properly. It might've glitched from you having a pre-existing global install?

[nimble belfry]

I will try to uninstall it and reinstall the UI after dinner. That may work if that caused it.

[dim sonnet]

++ pushed a commit that should make the linux autoinstaller for comfy more aggressively dodge global package issues

[nimble belfry]

That did not solve the issue, however I think i got it working. It looks like I had to manually run apt install python3.10-venv

[nimble belfry]

now i need to stumble through setting up the API backend so i can use multiple GPUs

[dim sonnet]

0.o? you had python but not venv? weird

[dim sonnet]

really easy to do tbh

just install swarm on the other machine, and make sure you got a network address to it

then it's like, two buttons in your local swarm to add the remote swarm as backend

[nimble belfry]

Does it need to be the internal IP or just any arbitrary address set to it?

[nimble belfry]

Before I close the terminal, this is what it told me. ```homrlab_admin@HomeLab:~/Desktop/projects/StableSwarmUI/launchtools$ sudo ./comfy-install-linux.sh
[sudo] password for homrlab_admin:
mkdir: cannot create directory ‘dlbackend’: File exists
fatal: destination path 'ComfyUI' already exists and is not an empty directory.
The virtual environment was not created successfully because ensurepip is not
available. On Debian/Ubuntu systems, you need to install the python3-venv
package using the following command.

apt install python3.10-venv

You may need to use sudo with that command. After installing the python3-venv
package, recreate your virtual environment.

Failing command: /home/homrlab_admin/Desktop/projects/StableSwarmUI/launchtools/dlbackend/ComfyUI/venv/bin/python3

[nimble belfry]

@dim sonnet I think i have it setup correctly but the backend errors. 15:17:23.705 [Init] Initializing backend #2 - Swarm-API-Backend... 15:17:23.712 [Error] Final error (14) while initializing backend #2 - Swarm-API-Backend, giving up: System.ArgumentNullException: Value cannot be null. (Parameter 'value') However, I am able to access the UI using that url from the other pc. and the console on the pc that im trying to connect to creates a session from the other pc so it is reaching it. 15:17:22.412 [Info] Creating new admin session for ::ffff:192.168.0.160 not sure what im doing wrong.

[dim sonnet]

any valid address, but yeah if it's on a lan then do the internal one

[dim sonnet]

that's... weird. Even ensurepip isn't available? 0.o

ensurepip is the python package dedicated to being a pre-installed backup for if other shit is missing it can fix it

[dim sonnet]

surely there's more to that error message 0.o that's just the head of an exception where's the stack

... wonder if it's relevant that you appear to be hosting as ipv6

[nimble belfry]

   at Newtonsoft.Json.Utilities.ValidationUtils.ArgumentNotNull(Object value, String parameterName)
   at Newtonsoft.Json.Linq.Extensions.Value[T,U](IEnumerable`1 value)
   at Newtonsoft.Json.Linq.Extensions.Value[U](IEnumerable`1 value)
   at StableSwarmUI.Backends.SwarmSwarmBackend.<ReviseRemoteDataList>b__15_0() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 115
   at StableSwarmUI.Backends.SwarmSwarmBackend.RunWithSession(Func`1 run) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 156
   at StableSwarmUI.Backends.SwarmSwarmBackend.ReviseRemoteDataList() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 73
   at StableSwarmUI.Backends.SwarmSwarmBackend.ValidateAndBuild() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 68
   at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 224
   at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 258
   at StableSwarmUI.Backends.BackendHandler.LoadBackendDirect(T2IBackendData data) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/BackendHandler.cs:line 370

[dim sonnet]

is the remote swarm instance outdated?

[nimble belfry]

uhhh, shit maybe.

brb

[nimble belfry]

It was but that did not solve the issue ```17:49:24.828 [Error] Final error (17) while initializing backend #2 - Swarm-API-Backend, giving up: System.ArgumentNullException: Value cannot be null. (Parameter 'value')
at Newtonsoft.Json.Linq.Extensions.Value[T,U](IEnumerable1 value) at StableSwarmUI.Backends.SwarmSwarmBackend.<ReviseRemoteDataList>b__15_0() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 145 at StableSwarmUI.Backends.SwarmSwarmBackend.RunWithSession(Func1 run) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 164
at StableSwarmUI.Backends.SwarmSwarmBackend.ReviseRemoteDataList() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 146
at StableSwarmUI.Backends.SwarmSwarmBackend.ValidateAndBuild() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 69
at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 262
at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 262
at StableSwarmUI.Backends.BackendHandler.LoadBackendDirect(T2IBackendData data) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/BackendHandler.cs:line 370

let me also make sure the server one is up to date but it should be I git cloned it last night

[dim sonnet]

that's a different error

[nimble belfry]

what the hell.... I can connect just fine the other way around. My gaming PC can connect to the server. but the server side errors if i try to connect

[dim sonnet]

actually the new error... doesn't quite make sense, line number is in the middle of nowhere

actually i'm a little scared, every line number in that is a }

[nimble belfry]

17:59:31.873 [Init] Initializing backend #2 - Swarm-API-Backend...
17:59:31.880 [Error] Error #2 while initializing backend #2 - Swarm-API-Backend - will retry
17:59:32.881 [Init] Initializing backend #2 - Swarm-API-Backend...
17:59:32.886 [Error] Error #3 while initializing backend #2 - Swarm-API-Backend - will retry
17:59:33.887 [Init] Initializing backend #2 - Swarm-API-Backend...
17:59:33.918 [Error] Final error (4) while initializing backend #2 - Swarm-API-Backend, giving up: System.ArgumentNullException: Value cannot be null. (Parameter 'value')
   at Newtonsoft.Json.Utilities.ValidationUtils.ArgumentNotNull(Object value, String parameterName)
   at Newtonsoft.Json.Linq.Extensions.Value[T,U](IEnumerable`1 value)
   at Newtonsoft.Json.Linq.Extensions.Value[U](IEnumerable`1 value)
   at StableSwarmUI.Backends.SwarmSwarmBackend.<ReviseRemoteDataList>b__15_0() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 115
   at StableSwarmUI.Backends.SwarmSwarmBackend.RunWithSession(Func`1 run) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 156
   at StableSwarmUI.Backends.SwarmSwarmBackend.ReviseRemoteDataList() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 73
   at StableSwarmUI.Backends.SwarmSwarmBackend.ValidateAndBuild() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 68
   at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 224
   at StableSwarmUI.Backends.SwarmSwarmBackend.Init() in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/SwarmSwarmBackend.cs:line 258
   at StableSwarmUI.Backends.BackendHandler.LoadBackendDirect(T2IBackendData data) in /home/homrlab_admin/Desktop/projects/StableSwarmUI/src/Backends/BackendHandler.cs:line 342

``` I restarted it and checked it it was up to date. on server.

[dim sonnet]

okay the line numbers make sense again but back at the first error

[nimble belfry]

maybe i should nuke it and start over now that i have venv maybe something there messed it upo during install?

[dim sonnet]

that's on line data.Backend.MaxUsages = backend["max_usages"].Value<int>(); which is saying max_usages is missing from the remote server's data which implies this commit from 3 weeks ago isn't present on the remote backend: https://github.com/Stability-AI/StableSwarmUI/commit/2ce45f52459bf0757700b8d001e2483db9a77424

venv isn't relevant, this is in swarm's C# code entirely

you might nuke the src/binand src/obj folders to verify it isn't a compiler/cache issue?

[nimble belfry]

I ran this to check for update
root@HomeLab:/home/homrlab_admin/Desktop/projects/StableSwarmUI# ./update-linuxmac.sh Already up to date. root@HomeLab:/home/homrlab_admin/Desktop/projects/StableSwarmUI#

I am also getting some warnings but not sure if its relevant ```18:06:25.263 [Warning] ComfyUI stdout: Traceback (most recent call last):
18:06:25.263 [Warning] ComfyUI stdout: File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/nodes.py", line 1734, in load_custom_node
18:06:25.263 [Warning] ComfyUI stdout: module_spec.loader.exec_module(module)
18:06:25.264 [Warning] ComfyUI stdout: File "<frozen importlib._bootstrap_external>", line 883, in exec_module
18:06:25.264 [Warning] ComfyUI stdout: File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
18:06:25.264 [Warning] ComfyUI stdout: File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/comfy_extras/nodes_upscale_model.py", line 2, in <module>
18:06:25.264 [Warning] ComfyUI stdout: from comfy_extras.chainner_models import model_loading
18:06:25.264 [Warning] ComfyUI stdout: File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/comfy_extras/chainner_models/model_loading.py", line 8, in <module>
18:06:25.264 [Warning] ComfyUI stdout: from .architecture.LaMa import LaMa
18:06:25.264 [Warning] ComfyUI stdout: File "/home/homrlab_admin/Desktop/projects/StableSwarmUI/dlbackend/ComfyUI/comfy_extras/chainner_models/architecture/LaMa.py", line 15, in <module>
18:06:25.264 [Warning] ComfyUI stdout: from torchvision.transforms.functional import InterpolationMode, rotate
18:06:25.264 [Warning] ComfyUI stdout: ModuleNotFoundError: No module named 'torchvision'

I nuked it all and did a full reinstall here is the full console output
https://paste.denizenscript.com/View/116535

[dim sonnet]

oh, that error is very minor, comfy is missing torchvision from its requirements.txt for some reason, here's commit to ensure it's installed https://github.com/Stability-AI/StableSwarmUI/commit/5a23492e2fa2bf1ba2e17af75dfd8de5dea412ef (I don't think anybody uses LaMa though so doesn't make much difference)

[dim sonnet]

that error says invalid uri

[nimble belfry]

ooook now its working the way i want. I think maybe the update did not take fore some reason. because when i restarted the gaming PC and started Swarm again it updated again (the same update it showed me it did before) Did it not take because i kept the terminal open?

[dim sonnet]

0.o

you gotta restart swarm when updating

if you manually opened a terminal window that doesn't need a restart, just the running swarm process

[nimble belfry]

I had it running in terminal, ctrl+c, hit up on keyboard to run launch bat file, it showed it pulled from git and started. then later i closed everything and just ran the bat file by double clicking and it showed the same pull from git and started.

idk why.. BUT, thanks for helping me through it and sorry it was something dumb like that.

[barren moat]

Hey @dim sonnet can swarm create seamless patterns? 👀

[dim sonnet]

:3 it can now!
Under Advanced Sampling

generated this + checked via https://www.pycheung.com/checker/

[barren moat]

https://tenor.com/view/chirsfarley-shocked-what-huh-omg-gif-6963624868839649661

@dim sonnet that’s amazing!! Thank you very much!! 🙏

[tidal wyvern]

How would I use embeddings, it seems like nothing happens when I select one. Also is it possible to set up something like highres fix inside this UI (I suppose in Comfy I would need to do this?)

[toxic shale]

Invalid data: ComfyUI errored: {
  "error": {
    "type": "prompt_outputs_failed_validation",
    "message": "Prompt outputs failed validation",
    "details": "",
    "extra_info": {}
  },
  "node_errors": {
    "100": {
      "errors": [
        {
          "type": "value_not_in_list",
          "message": "Value not in list",
          "details": "lora_name: 'shadowheart_v3.safetensors' not in (list of length 385)",
          "extra_info": {
            "input_name": "lora_name",
            "input_config": null,
            "received_value": "shadowheart_v3.safetensors"
          }
        }
      ],
      "dependent_outputs": [
        "9"
      ],
      "class_type": "LoraLoader"
    }
  }
}

I'm doing something wrong. but I dont know what 😭

also. new UI gets a 10/10

love it

so... the issue was that the loras where on a different drive from model root folder :"D

placed them in the default lora folder, now it works just fine

[dim sonnet]

sorry i keep getting distracted from implementing the clicky part -- type <embed:filenamehere> into your prompt or negative prompt as relevant to use it

[dim sonnet]

more precisely, looks like Swarm & Comfy saw different filepaths? You should be able to put model files wherever you want, as long as the paths line up right

[dim sonnet]

update: went ahead and impl'd a basic version now actually

when you click it'll add it to your prompt box

it also tries to maintain a present vs not highlight but will be a lil shifty cause the syntax is quite open to variation

[hearty notch]

when using controlnet

[dim sonnet]

i think aitemplate doesn't work with controlnet

[hearty notch]

on or off, it will error out

[dim sonnet]

show me the error with it off

[hearty notch]

this

btw, even with comfy

sometimes it will error out

any controlnet preproccessor will error out even with review

[dim sonnet]

that's still AIT

[dim sonnet]

that, however, is ... something else entirely

code error in the lineart model code I guess?

[hearty notch]

i just reload the UI

and reset all settings

and it's off, but still getting the error

[dim sonnet]

0.o

is AIT overinjecting itself even worse now?

oh it got completely reworked apparently https://github.com/FizzleDorf/ComfyUI-AIT

oh and the rework doesn't work lol

for now the best recommendation i can give tbh is remove AIT from your custom_nodes folder until the AIT repos are less of a mess

[nimble belfry]

Whenever you work on the 2x2 image grid would it be possible to have on 4 images batch to create the grid with the preview images as well? So the UI would do all the image work for me and I could just send that grid to my bot? Instead of now where I get the previews and then add them to a grid.

[hearty notch]

any idea how to reverse comfy from an old commit to the most recent one?

i always forget the command XD

[barren moat]

git checkout (then the name of the current) usually master

git checkout master

[gray geyser]

Hello, what should i put here if i want to use google collab as backend?

[dim sonnet]

You'd have to launch the service on colab, then get a URL to it, and add it as an address, but do note google colab has cracked down hard against using UI stuff on their service so they may not be doable, even on paid accounts

[gray geyser]

Thank you.

[spiral plinth]

Hello. Can anyone tell what is the difference between this and stable studio? I have not download this yet because its a alpha but the community of stable studio seems to be dead.

[balmy orchid]

after not using the "save workflow" button for a few weeks (I was concentrating on animdiff in comfyui) I noticed that some of my recent comfyui workflows won't save at all. These same workflows will also not load parameters when the "Use this workflow in generate" button is pressed. Adding a save image node doesn't seem to make a difference.

[balmy orchid]

Ahh I found the issue - it was the Video Helper Suite "Video Combine" node (great node BTW). Its presence in ComfyUI a comfyui workflow prevents the workflow from saving or being used in the StableSwarmUI Generate tab.

Pro Tip: press ctrl+B in ComfyUI to bypass these nodes and you don't have to delete them to use the workflow with StableSwarm.

Just posting here in case anyone else comes across this issue. https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

[balmy orchid]

In a similar vein, I wanted to to make a grid of videos using Kosinkadink's VideoCombine node. It can turn a batch generation of images from ComfyUI into either a GIF, a h264 or h265 mp4, or a av1/webm file. Since it's not compatible, I have to render out image sequences for each grid iteration, rather than using this node to give me a tidy video file.

Unfortunately, I cant seem to render out batches of images, just the first frame. The grid tool definitely takes just as long as if it was generating all the frames, but it throws them into the ether or something.

That said, IF we had some capability to incorporate image sequences/video files into the grid tool, either by adding the video's output as thumbnails to the autogenerated grip webpage, or maybe generating GIFs of the frames so they playback in the StableSwarmUI main window, or perhaps some transport controls or a even just a "play" button to let us flipbook through a image sequence...

That would be very useful to those users that will be seeking video support.

[dim sonnet]

StableStudio was basically dreamstudio-but-open-source, and is broadly limited to being just that, StableSwarmUI is a full native local UI for stable diffusion, and is pushing to become the most feature-dense but still beginner friendly UI. It's already the easiest to install&use, and fastest to process (thanks to comfy backend).

[dim sonnet]

yeah there isn't really video support atm, but I can see about adding at least basic support (ie: make it not error, and able to render video files properly) soon

[unreal warren]

In the SwarmUI the segment:face,0.7,0.5 dont work ... its generate a picture (yellow bar) and then the blue bar to 50% and stop.
What do i wrong?
prompt: woman with a butiful face, segment:face,0.7,0.5

and a second question ... how i do OpenPose in SwarmUI?
I like the UI and that i can use all the graphic cards from my PCs!
Much better then all the other UI and i am a beginner (day 5 with Stable Diffuion)
Thanks for all the work!

[dim sonnet]

• need text within the segment part.
• first time it runs it'll download clipseg model which might take a minute
• if it continues getting stuck, send me a screenshot of the UI and console window

[dim sonnet]

isn't super pretty and automated to set up atm, but can be done via:

• open Swarm/dlbackend/comfy/ComfyUI/custom_nodes
• open a terminal and git clone https://github.com/Fannovel16/comfyui_controlnet_aux
• open Swarm/Models/controlnet
• add the controlnet model you want (if using XL there's some cnets for it @ https://huggingface.co/lllyasviel/sd_control_collection/tree/main)
• restart swarm after doing all that
• from now on you can use it at any time by selecting the relevant preprocess and model under this parameter grouping:

(the UI for that is gonna change a bit in the future, I have a plan for it)

[unreal warren]

thanks for answer ... i will test it and report

[cedar arrow]

Hi all, can anyone please guide me through stableswarm ui. Is it related to comfyUI and tiled vae to reduce VRAM?

[unreal warren]

segment doesnt work ... there is an error in the console ...
i downloaded as sayed and copy the diffusers_xl_canny_full.safetensors from the page there

i think i do something wrong ...

[unreal warren]

when i start then i get that error, i git clone it and copy diffusers_xl_canny_full.safetensors from the website.

[dim sonnet]

it's related to comfy yes, TiledVAE isn't overly quickly accessible (you gotta go through comfy workflow editor to put it in) but i can probably add that

https://github.com/Stability-AI/StableSwarmUI#installing-on-windows it's really easy to install

[dim sonnet]

well first if you want canny or depth use the stability models for that https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256

second 0.o why did it not autoinstall opencv

[unreal warren]

come they in Swarm/Models/controlnet?

[dim sonnet]

uhh if you open a terminal in Swarm/dlbackend/comfy, you can enter:
python_embeded/python.exe -s -m pip install -r ComfyUI/custom_nodes/comfyui_controlnet_aux/requirements.txt
idk why that's not automated, it should be

[dim sonnet]

yes there

[unreal warren]

he downloaded a few things but still the same error

[dim sonnet]

..?

[cedar arrow]

I have no experience in using comfyUI. I just use the automatic111 webui-user platform

[unreal warren]

maybe i must just install that open cv? Im on windows 10

[unreal warren]

maybe i can install the manager on comfyUI and then install that what he need with them? should be easier ... or?

[dim sonnet]

ye that should work too

[unreal warren]

what packages do i need? Manager is installed and i startet ComfyUI direct

[unreal warren]

i reinstalled it with the manager and after stop start it download many things and error is gone ...
Maybe its a good idea to integrade the Manager in the next version

[unreal warren]

controlnet with canny - diffusers_xl_canny_full work now.
has many other option now too but the get most an error, OpenPosePreprocessor is in the list too but it it came an error: something went wrong.
Hope you can fix it in the future ... for now its fine when its not work. I know its still alpha and error happend.
Or have you an idea what i must install for it with the manager?

[dim sonnet]

controlnet preprocessors

as far as swarm's involvement atm it just generates comfy workflows that make use of the controlnet addon and all

Swarm is a UI, it doesn't handle the internal processing, it lets comfy do the deep stuff

[dim sonnet]

@cedar arrow I added VAE Tile Size param under Advanced Sampling:

if that's toggled on at all it'll use tiling, and you can configure the size

if not toggled on it'll do regular vae decode

[cedar arrow]

Thanks I will try it tonight

[dim sonnet]

• The error when trying to use a custom workflow with VHS node is fixed
• added support for gif's throughout, including a fix to prevent the files breaking and the ability to play them in-UI
• added recognition that the VHS_VideoCombine node (and also animatediff's combine node) are valid final image outputs to the workflow validator so swarm will allow them and pick them up
• tested to confirm the grid generator renders gifs properly

(Note that it's gif only atm, no h264/etc. video files yet.)
(also yes if you yield a batch into gridgen it will toss out the extras as it's designed to only allow 1 output per square of the grid - if you output gif it works properly though)

[last rain]

I have a SwarmUI running locally with two A100s each running ComfyUI remotely

I can't figure out how to hand the models

I have SDXL on both A100s

getting errors like this using the comfy editor
10|comfy | ERROR:root:Failed to validate prompt for output 9:
10|comfy | ERROR:root:* CheckpointLoaderSimple 4:
10|comfy | ERROR:root: - Value not in list: ckpt_name: 'OfficialStableDiffusion/sd_xl_base_1.0.safetensors' not in ['sd_xl_base_1.0.safetensors']
10|comfy | ERROR:root:Output will be ignored

This makes sense since locally when I install swarmui it downloaded sdxl inside a /Models/Stable-Diffusion/OfficialStableDiffusion/sd_xl_base_1.0.safetensors
whereas on the servers it is in comfyui/models/checkpoints/sd_xl_base_1.0.safetensors

When I try to use the Generate it fails completely

I am really not sure how this is all meant to work. Is it mean to use my local model and upload it to each remote backend? That doesn't seem to make sense. If not why do I even need a local copy of the ckpt?

[soft spruce]

@dim sonnet What causes some models/loras to not sync with the model browser? I have quite a few that I know exist on Civitai, but aren't synced. For example, the ThinkDiffusionXL model. Just shows the placeholder thumbnail

[nimble belfry]

I have never had mine pull the meta from Civitai. I always added it manually.

I am interested in this as well.

[soft spruce]

It's 100% possible that I'm just an idiot. Thinking about this again, I realized that I have a civitai helper extension installed in a1111 that downloads some data and thumbnails, which is where I pointed my swarm models directory. So maybe swarm doesn't pull any data and instead it's just pulling from that

[nimble belfry]

That I do have as well. All the models I had from when I used 1111 have the meta. I'm particular in how I like my meta because my discord bot pulls that to use. I normally redo the images as well. But it is nice when they have a bit more starting with.

Alex, also is there a way to default the edit meta section to not use the current image by default.

One time I went in to edit a small description of a model and accidentally changed the image and had to remake it 😭.

[soft spruce]

Same here. I'm brand new to swarm. Just signed up this morning and I've already accidentally change the images of three models and had to remake them

[nimble belfry]

mcmonkey is the best dev I know. When swarm if fully released it's going to be the best UI out there.

[soft spruce]

So, I'm reading through the documentation now. There's mention of the possibility of using more than one GPU. My wife's gaming rig is sitting right next to my PC. It's got a 3080 in it and she never uses it. Am I reading this correctly? Can I somehow do some sort of black magic wizardry to use both my GPU and her GPU?

[nimble belfry]

Yes and it's super easy!! You just have to setup StableSwarm on that PC and set an IP address to it in settings. Then, go on your main PC and setup a backend and enter that IP you setup. Done.

Like this.

[soft spruce]

fascinating. thank you very much for that. i'm not very technically knowledgeable, but i'll be dag gone if that's not worth a try

[nimble belfry]

mcmonkey made it super easy to figure out. Even the API.

[soft spruce]

Is there a way to have more control over upscaling? I see the handy little upscale button, but I didn't see where I could choose upscaler models or define how much upscaling I want to do

[nimble belfry]

I have not really messed with that yet. Other than changing the init image creativity slider.

[dim sonnet]

so, first: I'd recommend running swarm remotely and linking swarm as the backend, rather than a remote comfy. A lot of things work better when swarm is present on the target machine (eg it feeds in Swarm's extra nodes, and can make comfy send images directly rather than storing to output folder, and etc)

[dim sonnet]

second: yeah Swarm assumes you have the exact same models between backends

[dim sonnet]

it uses local models to manage the list of models, but remote server is what actually has to load it

in the future i'll probably have it scan model list from the remote, at least from remote swarmbackend

[dim sonnet]

Swarm reads metadata that's either in a recognized format that other UIs have it in, or in the official standard ModelSpec https://github.com/Stability-AI/ModelSpec

most SDXL models have at least basic title/architecture stored in their modelspec, but a lot of them don't include thumbnail/description yet

that'll change over time as more people start noticing how helpful & convenient it is

[dim sonnet]

good point, fixed to now only autocheck if the model doesn't have an image yet

ps you can always pull back up prior images from the Image History tab

[dim sonnet]

ye that needs reworking, the "Upscale 2x" button was just meant as a quick convenience button, but it's kinda silly. The more proper way you'd upscale is to enable the Refiner parameter group and use the upscaling options there

or you can toss an image into the Init Image input and configure things that way

[nimble belfry]

there is currently no inpainting outpainting option yet right? I see mask image but I dont see how you can apply a mask. The info button on that does not really explain what that option is and it does not seem to be in the docs.

[dim sonnet]

the mask option that's currently there is a technical parameter - ie if you're hitting Swarm as an API you can use it to do what you need, but the average end user isn't really gonna use it directly - as far as UI usage there's not a clean image editor yet

[dim sonnet]

update @balmy orchid webm/mp4 are now (sorta) supported as well. They're a lil wonky but they'll save properly and render (mostly) properly in the ui

[golden drum]

Been checking out swarm a bit today. Looks promising. Keep up the awesome work! very excited to build custom backends in comfy and have inputs available in a more organized way.

Out of curiousity, Is there a roadmap of planned features somewhere?

[dim sonnet]

the main ones are in the readme: https://github.com/Stability-AI/StableSwarmUI#status

bunch more in my personal todolist

[golden drum]

Thanks, boss. Keeping settings locations organized and the workflows organized at the same time is impossible in the nodegraph. Especially when the workflows get huge.

[balmy orchid]

Wow! Thank so much for adding this!!!
I thought I should share a clip of how awesome you've made your GUI look!

This is invaluable to me, as doing x/y's in Comfyui is suboptimal for animation workflows, and really we have much a better presentation in StableSwarmUI.

minor quibbles - The generated grid webpage doesn't display the video images yet, but I guess that conflicts with the original design of that. And is there a way to sync the start of the preview videos so they play in parallel?

[dim sonnet]

grid page should work, just might have to poke it a lil (click an axis back n forth to make it reload)

sync button, uh - well on grid view snap an axis back n forth, main view... uh swap to Server tab or something and back might work? if not i'll have to add an explicit button for it

[balmy orchid]

Ahh, right you are - I just had to goose the grid page by clicking on one of the buttons. Sync is no big deal, just a nice to have. Thanks again for the quick work!

[dim sonnet]

Feature announcement: Bulk Pickle-To-Safetensors conversion utility https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7398005

honestly half the reason i built that was just to make embedding folders not-so-horrible

oh yeah also the announcement before that one, Docker files! @umbral siren docker works :D

[nimble belfry]

[umbral siren]

Wooooooo!!!! Woot woot

[dim sonnet]

Swarm will now show listings of remote models, so long as the remote is a SwarmSwarmBackend (ie you have installed Swarm on the remote machine and linked to that)

(be warned that not all features fully support this yet, still work-in-progress)

(lot of weird edge cases handling the concept of models that only exist on some backends but not all)

[woven torrent]

Hey, do Embeddings work the same as Loras do? When I select a model, it highlights it in blue and adds the filename to the line above. Same for Lora, they highlight blue and are added above. But for embeddings they don't seem to select for me, and nothing is added above... Is that how they're supposed to work, or are they not being applied, or?

[dim sonnet]

They do not work the same no. Loras are modifications to the model, whereas embeddings are special codes added to the prompt.
Loras have been fully supported for a long time, embeddings only became super clean a few days ago, so there might be lingering bugs

so, step 1, update swarm

step 2, if after reopening and CTRL+F5ing, you still have issues with embeddings, gimme screenshots or something to show what's going on

[woven torrent]

@dim sonnet Okay thanks I'll try that. Also that reminds me, when I see Lora's being put into prompts with <> tags, do I erase those from prompts in SwarmUI because I have the Lora applied, and just use whatever word the Lora says to use?

I feel like the UI could use an exclaimer on the Embeddings page, it seems intuitive to me that they'd be applied the same way Lora's are (at least from a UI perspective) so it feels like they're just not working or not being used right now

[dim sonnet]

The proper UI handling of loras is preferred, but when in prompts they will still work (for convenience of copying things between auto webui and swarm mostly)

embeddings should be visually clear that they're in use:

[woven torrent]

Ahhhh okay nice! Yeah mine doesn't do that. I just updated let me see.

There we go, now I see them highlighted... One perhaps small bug, in your example if I clicked purple box A, then B, then A, then B, it adds them to the prompt twice each. If I clicked A A, B B, it toggles the text from the prompt instead. For example I get: embed:bad-hands-5.pt embed:badhandv4.pt embed:bad-hands-5.pt embed:badhandv4.pt embed:bad-hands-5.pt embed:badhandv4.pt

[dim sonnet]

yes

it does do that

it's, uh, hard to have good behavior when you click them

they're not toggles

they're additive

having two copies of the embed is perfectly reasonable

two copies in two different places especially

[woven torrent]

Perfect! Look forward to playing with this 😉

Thanks

[golden drum]

is there a naming convention for primitive nodes to get them to be always visible in the grid generator dropdown list? Sometimes I have to rename things to get them to show up there.

[unreal warren]

is it possible to get the output in a UNC drive on Windows? Because i use it from network and want all picture on the same place.

instead C:\SwarmUI\StableSwarmUI\Output\local\raw \PC-Name\release\local\raw

[dim sonnet]

The only exception where primitives might be ignored is if the parameter is claimed by a default parameter (eg you have the primitive attached to your KSampler's Steps value, so its taken by the Steps parameter rather than left independent)
If you're seeing primitives go missing in any case where there's not a good reason like that, please gimme a workflow to test & replicate the issue

[dim sonnet]

all paths can be configured under Server -> Server Configuration tab at the top. I've never tested a network drive but should work same as any other

[unreal warren]

Just for Info: It work but only with drive Names like "T:\Netdrive\raw". UNC Path dont work. But its fine 👍

[balmy orchid]

Is there a way to set slider resolution on a parameter? I have some values that only work well in a range from 0.95 to 1.05, and the 0.1 slider steps make it so that I have to type any in-between values in.

[dim sonnet]

Er... on a technical level, can Swarm do that with parameters? Yes, that's always configured.
Can you customize it in the mainline parameters? Yes, go to user -> parameter configuration, you can select min/max/step/viewmax/viewtype for each parameter.
Can you customize it in comfy? Yes comfy's node registration in python configures those values, which are what swarm reuses to format custom parameters.
Can you customize it in a custom comfy workflow without editing the node python? Not currently. There's a plan to eventually have a UI for advanced node configuration unique to swarm custom workflow handling, but that hasn't been built yet

[limber trench]

Hi. I have 2 computers setup for swarm. Are the model directories supposed to synchronize automatically or do I have to manually copy files to insure that control net/loras/models are the same in both places? If this is the case, would it make sense to setup some automated process for this or use google drive with a symbolic link?

Error:
Invalid value for parameter Model: Invalid model value for param Model - '' - are you sure that model name is correct?

[dim sonnet]

you'll have to either manually sync or yeah use an automatic drive sync thingy

swarm doesn't try to sync them for you atm. Miiiiiight be an optional future feature but not rn

[limber trench]

Thanks

[limber trench]

Should just the Models folder be synchronized? I am trying out this tool. Worried that it may mess up metadata.

[dim sonnet]

models is the only relevant one you'd want synchd yeah

[limber trench]

How do I install custom nodes? I copied the folders to custom_nodes directory. How do I get stableswarm-ui to register them? I tried restarting.

I am getting some activity now. I think the issue was due to local comfyui server not starting.

Looks like it installed everything except for 2 nodes:

[loud bluff]

im getting a "Invalid operation: No backends match the settings of the request given!", but I see my Comfyui that I set up with API. Any idea on how to fix this?

[dim sonnet]

nodes get installed the same as in regular comfy -- if you're doing it manually, since you're grabbing Manager anyway, you might just only use Manager, and then use Manager to install other nodes, easier that way since it does the installation of python dependencies and all

[dim sonnet]

0.o? Does it fix itself if you restart swarm?

if not, go to server config and change LogLevel to Debug and try again, it'll give specific reasons why it failed to match

[loud bluff]

Thanks I'll try that.

[limber trench]

Should I be able to run a workflow designed on a local machine on a swam backend? I think that is the intention, but perhaps is not working yet.

[dim sonnet]

as long as the remote swarm install has a comfy backend, yes that should work

[hollow sphinx]

Hello, I try to use Stable Swarm with my ComfyUI directory, but it can't find my models for some reason

[dim sonnet]

is that ComfyUI\ComfyUI correct? 0.o

and either way when in doubt restart to be safe

(shouldn't really have to restart but, yknow, still in alpha, some things are a lil wonky)

[hollow sphinx]

yes it's the portable version

dammit I thought I restarded it already but it works now

I can use it with runpod or vast right ?

[dim sonnet]

ye

[spiral fossil]

is there a way in comfy to save the metadata in an automatic1111 format?

[dim sonnet]

Comfy metadata, Auto WebUI metadata, and Swarm metadata are 3 entirely separate formats

Swarm and Auto can be janked to be more or less intercompatible for basic gens if you push it

Comfy, not really, it has entirely its own thing

eg comfy metadata doesn't have "steps, cfgscale, seed, ..." it has here's a ksampler and all its properties. There might 1, or 2, or 20 of em..." or it might just not have a ksampler and have something else entirely

[gray geyser]

Hello. Any hints what could go wrong? I can't pass this step. Tried to turn off antyvirus, Windows Defender, firewall etc, but without success. It looks like a network problem, but everything runs locally.

[dim sonnet]

You seem to have enabled the scoring option by accident? Just uncheck it. It's under Advanced presuming you're updated

[gray geyser]

Perfect, thanks! It works. I was using Brave and then tried to use Firefox and Google Chrome where it worked. It seems this options was saved locally.

[narrow blade]

@dim sonnet is there any code for a bot where we can send those requests to stableswarm?

[dim sonnet]

as in a discord bot? i got this one: https://github.com/mcmonkey4eva/SimpleDiscordAIBot

[nimble belfry]

I have one written in python but it's not finished and definitely won't be as well coded as monkeys.

https://github.com/kalebbroo/bAIb_Ross

[unreal warren]

i cant get CLIP Sec to work. I try the example with
a photo of a cat walking in a park
<segment:cat's head,0.8,0.5>a photo of a dog's head
he make a picture (green bar) with a cat in the preview and then starts the blue bar and stops.
In console i get: [Info] No images were generated (all refused, or failed).
I use a SDXL Model, but tried it with an SD1.5 too.
What i do wrong?

[main egret]

Hi guys & gyals,
did anybody get to work Stable Swarm UI on SageMaker Studio Lab? Only saw a HowTo on Swarm UI on SageMaker which is not the free service version of the platform. Already got my request approved and set-up an account. Just trying to figure things out to get it working. Any idea?! 🤷🏼‍♂️

This is the HowTo I have found:

https://medium.com/@dminhk/running-stable-swarm-ui-on-amazon-sagemaker-notebook-23f7aa10c637

[dim sonnet]

well apparently someone has it working! Neat! Hadn't seen that before

[dim sonnet]

go to server -> server configuration -> change LogLevel to Debug, then try again - if will output a lot of debugging info to the terminal that should make it clearer why it failed

[unreal warren]

cant upload as plain text ... so here as txt file

[split swan]

finally found what I needed for merging

phew

[dim sonnet]

godspeed with the 27 sliders

[split swan]

oh Ive been at this in auto1111 for ages

I just needed comfy cause it hogs less ram than auto

so I can hopefully merge SDXL

[dim sonnet]

I figured out why -- apparently Comfy made a change to have masks in batches, and I needed to update the Swarm code to support that. I pushed the update, so if you run the update.bat and try again it should work now

[split swan]

I just need to figure out how to chain the nodes so it will save the model

[dim sonnet]

literally just connect the output to a CheckpointSave

iirc it goes into your comfy output folder not the actual checkpoints folder (unless that was changed recently)

[split swan]

but how do I trigger it?

It also demands a clip and vae connection

[dim sonnet]

oh, hook up the clip and VAE from whichever model you want to copy from

or use CLIPMergeSimple if you want to merge that too

the VAE you definitely don't want to merge

and you trigger it by just queuing in comfy as normal

[split swan]

hmmm it might have merged it but I just had no idea

there was no indication for it, no load bar or such

but there is a safetensor in there

@dim sonnet looks like this right now but if I toy around I'm sure I can make it look more convoluted

[unreal warren]

I confirm that it work after update perfect! Thank you so much ... thats a game changer.

[split swan]

hmmm how do I make sure meta isn't saved to my image in comfy?

[split swan]

well figured out how to keep it out of image but not how to save it to txt

[hearty notch]

wait wait wait!!!!! what is that nod! is that what i think it's?????

[split swan]

huh?

[split swan]

do you want the merge block node?

[hearty notch]

actually if you are comfroable sharing the WF, that will be awsome, I like to try this

[split swan]

sure let me find it

its basically set up as runetime block merge on auto

it doesnt merge them

just combines em per the blocks to generate images and test it

[split swan]

@hearty notch https://github.com/YinBailiang/MergeBlockWeighted_fo_ComfyUI https://github.com/azazeal04/ComfyUI-Styles

[hearty notch]

managerr nod will not install them automatically?

[split swan]

I don't know

[hearty notch]

many thanks for the share

that is so sweet of you

[split swan]

since I did it manually and figured out the wf

so I dunno

no problem

did a few of these with that to test it

[hearty notch]

report this emotional damage

[dim sonnet]

huh?

oh was there a spambot that's already deleted

[dim sonnet]

Swarm metadata is now in the same data format auto webui and comfy use (png text value, or for jpg usercomment exif), ie metadata-reader tools will now recognize it

[dim sonnet]

If you have multiple comfy backends, the workflow editor can now abuse all of them at once! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7555872

[dim sonnet]

More new features! i haven't slept in way too long all the features today! You can now set up generations how you want on the Generate tab with the easy simple UI there, and then click "Import" on the comfy tab and view the raw comfy nodegraph guts of it! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7558116

[nimble belfry]

That's awesome. I should probably learn how to use Comfy...

[neon atlas]

how can i fix this happened after update please thanks

[dim sonnet]

More new features! Wildcards! https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7568190

[dim sonnet]

oo, might be my fault - do you have metadata turned off? I accidentally had code to emit metadata even if it doesn't exist, which probably throws that error. Fixed now if you update

[neon atlas]

up and running with SaveMetadata off thanks

[neon atlas]

SaveMetadata not working for me maybe error on my part (getting same error)

[dim sonnet]

have you updated?

[neon atlas]

I have, I've been playing with it couple of hours now. Thanks

[dim sonnet]

er, well.. I'm currently testing with format=png and savemetadata=on, and it's fine for me D:

are you sure you restarted after updating?

[neon atlas]

jpg working fine though yes restarted after updating

[neon atlas]

reset again no go with png will go with jpg for now thank you for your help

[neon atlas]

Thanks

[worldly oxide]

I put my Wildcards with txt files in the Data folder, that I have been using with dynamic prompts but I don't see any lists show up!

[dim sonnet]

Have you updated to the new build? It should have autogenerated the folder for you, and you use them in the webUI at the bottom

[worldly oxide]

I'm on the latest build v0.5.8.0 and I don't see any wilcards show in StableSwarm, these are my settings am I missing something?

[dim sonnet]

eh... hit CTRL+F5?

looks like a page load fail probably from cache, so hardrefresh oughtta fix it

[wheat pawn]

I am having some trouble setting up SSwarm to access my comfy machine. Till now I have been accessing my comfy machine using a url and port. When I put that same url and port in sable swarms api access field, it won't connect. Are there other things I need to do to get things to work?

System.InvalidOperationException: An invalid request URI was provided. Either the request URI must be an absolute URI or BaseAddress must be set.

[nimble belfry]

By comfy machine do you mean you have SSwarm installed on both machines?

[dim sonnet]

^ best to have a swarm install on the remote machine and use SwarmAPIBackend. You can do a remote comfy but it has some limitations when you do

[dim sonnet]

ps here's your requisite LCM Quickpainting video but sped up. Testing swarm's new image editing canvas. (Video sped up 4x)

[outer garnet]

In stableswarm I am always getting the warning that my embeddings do not exist :O
Even when I select them right from the UI
To elaborate on that: It seems in the prompt they can be loaded, but not in the negative prompt :O

[dim sonnet]

Could you elaborate more? Ideally a screenshot of the error and the swarm window with the prompt/negative sent (also showing the Embeddings list with the model in it to be sure)

[outer garnet]

Sure ^-^

[outer garnet]

Note that the embeddings work in the not-negative prompt, and that while leaving away the bracets stops throwing error messages, but also still does not load the embeddings

[dim sonnet]

ooo

[dim sonnet]

that was a sneaky little bug - there was code that saw your prompt started with < and ended with > and so thought the entire line was just one big invalid tag, instead of recognizing it as being two tags in a row. I pushed a fix, should all work perfect now.

[outer garnet]

Yaii thanks ^-^

So I just update :O

[dim sonnet]

yep

[outer garnet]

Its working now thanks :O

[grim needle]

I'm not sure if this is the right place to ask, but I am trying to modify how the canvas sketching thing used by inpainting draws strokes. Would anyone here know how to go about this?

[dim sonnet]

huh?

[grim needle]

Hang on, I'm not sure this is the place to get help with modifying the UI used by Automatic1111...

[dim sonnet]

... please don't spam the entire server with a random question

pick one channel that sounds relevant and if you need to ask elsewhere somebody will tell you

The StableSwarmUI channel is definitely indeed not the one for an entirely different UI

[grim needle]

Alright, noted

Though, I suppose this UI might be a good alternative to hte UI I'm using now, might sidestep the need to solve the problem altogether

[dim sonnet]

ye :D

Key development focus this week is building out a thorough canvas based image editor, so if you want to start using Swarm and give some feedback while it's being built, that would be helpful

[grim needle]

Excellent! I don't know if you saw in developers/research, but I made a modification to SD's inpainting that allows "fuzzy" inpainting masks with gradual texture transitions. A proper canvas with brushes would be a good place to test it

[dim sonnet]

partial-opacity should work out-of-the-box already to my knowledge, if that's what you mean

(or at least it should in comfy)

[grim needle]

Think of it more like per-pixel denoising strength than something akin to alpha compositing.

[dim sonnet]

oo

yeah i've heard of people reinterpreting mask opacity as a denoise offset, never tried it but sounds like a useful concept

[grim needle]

Basically, before each step, it pulls back the latents to what they were originally before noise was applied by some amount according to the mask. The amount it gets pulled back is based on the sigma (step size) used for the step. I'll try to get a nice example image

[grim needle]

so in order: original, blood splatter added using soft mask, blood spatter added using hard mask

notice that the original flyers are preserved proportionally to how masked they were, and the blood texture gradually tapers out from the center

(sorry for the example, couldn't think of anything else haha)

(same seed and settings)

[dim sonnet]

oo yeah that's way better for preserving the base

[devout scroll]

termi

[nimble belfry]

What's a termi?

[edgy veldt]

Hi there, just curious if someone is building common wildcards library just like the official presets: https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Presets.md ?

[edgy veldt]

I have StabilityAPI setup by creating "./Data/sapi_key.data" putting my API key, but there is only SD15 option (not SDXL) in the StabilityAPI section, and it says "Invalid operation: No backends match the settings of the request given!" when I click "Generate". Is it not implemented yet?

[edgy veldt]

Is there any plan to incorporate SVD in short future?

[nimble belfry]

When they release a cool model why dont they get with monkey to have SW ready to use it? They could keep people in their ecosystem that way. Would bring a buzz to both projects 🐝

[dim sonnet]

there's plenty of wildcard libs out there already actually! Auto WebUI had a wildcards extension since last year, and Swarm uses the same format (mostly) so you can reuse anything made for there. Eg https://github.com/adieyal/sd-dynamic-prompts/tree/main/collections/jumbo
Civitai also has a wildcards section. Which... I would link, except, as usual their website is horribly broken so I can't even look at it atm much less link it. But it exists lol

[dim sonnet]

If/when it's implemented in Comfy, it will work in Swarm implicitly.

[wintry flume]

after making the move from A1111 to Comfy and loving it, i'm interested in checking out StableSwarm. i notice that the windows install bat isn't in the assets list for the two most recent releases; should i infer that those releases aren't considered release quality or not ready for windows or something otherwise?

[nocturne sierra]

How can i use multi gpu in swarm ui (comfyui)?

[leaden oyster]

If you already have it installed, on the server tab you have backends, you can set one per GPU, each GPU has a unique ID, you just select a different one for each added backend

but I'm sure more advanced users here can help guide you step by step if you need it : )

[wintry flume]

is this the right place to ask ComfyUI questions?

[leaden oyster]

we don't have a specifc comfyUI channel you can try here or in https://discord.com/channels/1002292111942635562/1089974139927920741 or if it is about installing it at https://discord.com/channels/1002292111942635562/1002602742667280404

[wintry flume]

of the three this seems the most appropriate then, thank you!

so i've got XY working for grids using XY Plot and KSampler (Efficient), but i'm stumped trying to figure out how to do a grid for the upscaled images

if i connect the XY Plot script to the first pass sampler i get a grid you'd expect. but if i connect the script to the post-upscale sampler, all images in the grid will use the first latent from the first pass sampler

i tried some batch nodes but they didn't seem to be the solution

[leaden oyster]

let me edit this since that's not correct, I'm going to try something

[wintry flume]

ah np, thanks for the edit i was trying to figure out what you meant 🙏

to clarify: what i'd like is for the grid to contain upscaled images (ksampler -> upscale latent -> ksampler)

[leaden oyster]

the upscaler sampler also has a script input

what you need yo plug into the X Y is either the sampler or the vae etc... to have a comparison

I don't think is possible to compare once rendered

since XY is working in the sampler

@wintry flume There seems to be a solution, the "CR Animation Nodes" can generate an XY plot from images/folders

So what you probably have to do is to save the upscaled images into a folder and then pull them to the CR node and do the plot

[wintry flume]

hmmm that's definitely interesting but i think that doesn't apply any labels? what i'm trying to do is put together a workflow to test various things like models vs. steps

eg

that's what i get from the script attached to the first pass sampler

what i'd like is to upscale those images before they go into the grid

[leaden oyster]

then plug it normally to the ksampler (upscaler) and put on X the steps and on Y whatever you like

it's the same for the original ksampler and the upscaler

[wintry flume]

so if i plug it into the upscaler it uses the same latent for each pass

i get a grid, but each image was generated using the same latent

[leaden oyster]

of you do "steps" doesn't work, gives you the same at every iteration?

[wintry flume]

it gives me this

i get the same result if i attach the script to both, i guess it's that the Upscale Latent node just stores the first latent sent to it from the first sampler

[leaden oyster]

yeah it upscales the latent but then you are resampling

try to inject a bit of noise and see what happens

you have it a 0

it's just passing thru

[wintry flume]

so maybe i should back up. my very basic workflow right now for an upscaled image is to do KSampler -> Upscale Latent -> KSampler

what i was hoping to do was generate a grid that shows me what images would look like at using different models and steps

[leaden oyster]

try putting 0.3 of noise at the upscaler sampler and plug the X/Y

on steps

model you are not changing (the checkpoint so that's useless, maybe steps and scheduler or sampler)

[wintry flume]

okay i have that generating

but i don't know how that would solve the issue. all the images will still use the same one latent

what i want is for each element in the grid to be one continuous pass of KSampler -> Upscale Latent -> KSampler , replicating my standard upscaling workflow

sorry, i feel like i'm doing a poor job of explaining what i'm trying to do

[leaden oyster]

i understand the concept, you generate an image and then upscale it but why do another sampler if you are not going to do anything to it, it will look the same, you can go straight from the latent upscaler to a VAE decoder

the idea (at least to me) is upscaling and bringing a little bit more detail and stuff

so I'd inject a bit of noise

[wintry flume]

because i'm generating images at 512 and then want to upscale to 1024

[leaden oyster]

and you should see changes on XY

[wintry flume]

it was my understanding that's the standard way to to do it

based on this https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/

gotta step away but i appreciate your time and patience. i'll keep plugging away 🙏

[leaden oyster]

exactly that one injects noise

it's at 0.5

which is what should happen, it'll make the sampler bring details and stuff, you can play with how much 0.3 - 0.5 is fine

but at 0, there's no change

[dim sonnet]

the install bat is in the readme https://github.com/Stability-AI/StableSwarmUI#installing-on-windows and it always automatically installs latest dev version atm

[dim sonnet]

yknow, a very convenient grid feature is built in to swarm - on Generate tab, under Tools at the bottom

[nocturne sierra]

How can I enter multiple GPU IDs in a single comfyui prompt?

[leaden oyster]

you can't, you use a GPU per prompt (inference)

you cannot split one inference into multiple GPUs, at least not that I know of

you can run a prompt in GPU_0 and another on GPU_1

[nocturne sierra]

thx

[carmine tulip]

Hi, does swarm UI have the same level of customability and speed as comfy UI

[leaden oyster]

if you use comfyUI as backend, yes, it's the same

the problem is installing it on your mac!

[nimble belfry]

I thought it had support for Mac?

[dim sonnet]

actually! there's a new feature that sometimes lets one "comfyui prompt" (ie: full workflow) use multiple GPUs if and only if that workflow has multiple inferences going on

[dim sonnet]

so ^ if you have multiple generations with multiple outputs, give each SaveImage/PreviewImage (your output nodes wherever they are) a different color (right click and select color), and then at the top-left there's a "Multi GPU" dropdown, select "Use All". Then, when you generate, the workflow will auto-split into multiple workflows (one for each unique color used) and split across your GPUs

[dim sonnet]

yep

[dim sonnet]

mac install guide if you need that: https://github.com/Stability-AI/StableSwarmUI#installing-on-mac

[barren moat]

Hey @dim sonnet is there a way to use Adetailer with Swarm? I'm doing some gens of football players to use for some mockup designs but the faces are kinda crazy. Not horrible considering they have a helmet but a little better would be great. Here's a few examples..

[dim sonnet]

tack <segment:face> a perfect hd football player's face yadda yadda onto the end of the prompt

maybe also <segment:football> a perfect hd correctly shaped handegg of a football lol

[barren moat]

Ok great let me try that, thank you!

[barren moat]

I'm getting some errors when I try segment commands. I did try to install a custom node a couple weeks ago but removed it when I was in over my head lol not sure if that is causing the warnings in the console??

I can run it just fine without the segment command, not sure what I did wrong.

[barren moat]

Maybe I should do a fresh install. I'm guessing it was that custom node I tried to install.

[hearty notch]

@dim sonnet any idea why this happned after a fresh install of ComfyUI?

I can not load the UI anymore as it will always error out,

[dim sonnet]

that error message isn't related to the segmentation, that is a problem tho (opencv is really annoying and doesn't like being installed, and the Remove Background parameter option requires it)