🧣｜comfy-ui

[steep marlin]

You could also just prerender the whole animation first, then using batching, layer diff the batch

[mild ivy]

you don't happen to know of a workflow for that do you?

[steep marlin]

Not off the top of my head, but you can probably whip one up ina few minutes. If you already know how to do animdiff and layered diff workflows, it's fairly straightforward from there. There are a ton of batch image load type nodes if you want to run layered diff on a fully rendered out sequence that you've already made

let me see if i can whip one up for you real fast, should only take a few minutes

got me kind curious about it myself lol, would be cool for sprite animations and stuff

[mild ivy]

yeah that's what I'm thinking of using it for, thanks

so I guess what you're talking about, there would be multiple 'ksampler' nodes then?

[steep marlin]

correct

[steep marlin]

meh, im running into a bunch of issues with the animdiff portion of the workflow. keep getting cuda errors bitching about parts being on the cpu and gpu at the same time

i might try svd or something else for animating, i know animdiff is pretty old by the standards now

[mild ivy]

thanks for trying at least

[steep marlin]

i think i've kind of got it working

[mild ivy]

could you share a workflow file?

[steep marlin]

yeah gimme a sec

the main issue i was having was that the motion loras were breaking the cuda stuff splitting it between gpu and cpu and it didnt like it

here's a test example of it, but for some reason the gif composites to a black background:

in my browser the pic on the right shows up as white background

[mild ivy]

well if it's totally black that should be fine, could use other tools to get the transparency

[steep marlin]

if you open the still image in your browser, you should be able to load it into comfy for the workflow

[mild ivy]

thanks I'll check it out

[steep marlin]

i suck at prompting 1.5, so it's a dawgwater quality example

"1woman, black hair, flowing hair, glowing eyes, photo, dslr" was the best i could do lol

[mild ivy]

it works alright as a prompt

[steep marlin]

anyways, it does appear they can work together within a single sampler

[mild ivy]

I'm thinking of using this for spell effects and stuff at first so motion and looping stuff isn't as much of an issue

[steep marlin]

yeah same actually, might be neat to play with

for some reason, i was originally thinking that layer diffusion could extract a subject from an already created image, i guess not though

and when i tried to do it in two stages, it broke badly

[mild ivy]

yeah that's why I wasn't sure how to do it, it connects back in a few places

well that and being relatively new to comfyui in general

got a working result, nice

thanks again

[steep marlin]

no problem, guess we both learned something new today

[carmine copper]

Comfyui customs workflow builder needed

[hoary zinc]

How can I get DWPose preprocessor? No, it isn't preinstalled with auxiliary node, there is only Pose Estimator. I've been trying to get it to work according to some videos, but even when I did it myself it didn't work

I found a mistake in my connection.

[carmine copper]

Need build the custom workflow for product image generator same as the refrance i can provide, so i can intgrated into my app

[silver nebula]

HELP! My ComfyUI just refuses to work ... and no-one has an answer!!! 😦

'SDXL' object has no attribute 'get_model_object'

[hoary zinc]

Can you show me your workflow?

[silver nebula]

You mean the nodes and noodles? Or the full ERROR?

[hoary zinc]

comfyui workflow

[silver nebula]

It was working fine, I updated, it broke!!! Seems that there is a rogue node I would need to delete, but I cannot see which one ... ?!

Advanced KSampler problem

[hoary zinc]

then downgrade

I also have a problem with DWPose. I got it working but I see artifacts in the image and some distorted parts. I use lighting model with 8 steps, euler sampler, cfg 6

and controlnet with openposeXL2 of strength 1

that's quite a complex workflow you have there, I can't help 😦

[lone bolt]

Is the CFG not too high for lightning models?

[hoary zinc]

no, I used this one with CFG of 6, yeah, but it'll be the same result when I use CFG 1 which was used for the generated images.

somehow openpose controlnet messes up the picture

it works

[silver nebula]

It was working perfectly; then I updated, and it "broke!"

[hoary zinc]

I don't know what I broke before. lower CFG to 1 worked.

[lone bolt]

So, it was the CFG?

[hoary zinc]

yes, it must have been CFG

I was toying with DWPose and nodes, is there a way to make image's posture from another image? So far I only see it's only working when I generate a new picture with prompt and combine it with the pose from another picture.

[silver nebula]

Some SDXL work in ComfyUI

[hoary zinc]

https://civitai.com/models/202108/bluepencil-xl-lcm-lightning?modelVersionId=369896 this model is affected heavily by steps and CFG. If I set anything higher than 8 steps and CFG 1 the image starts to mutate and looks like I watch it through multiple shards of glass

also only euler sampler works, when I tried dpmpp_2m_sde_gpu it was the same result as I mentioned

[hoary zinc]

Does anyone know how to make OpenPose node update it's skeleton picture? It happens randomly when I edit it and then change resolution and move with the skeleton a bit in the edit mode.

but it seems to be broken

it doesn't update automatically and then as a result isn't reflecteed in the image

When I open the editor it shows the pose I made but the picture doesn't reflect it until the preview changes.

[hoary zinc]

Does anyone have the same problem? It doesn't work through the manager nor manual installation.

[lone bolt]

Whats the console error at the start? Probably something with insightface

[hoary zinc]

I was missing that, yeah

I have all visual studios version including the one needed

[lone bolt]

Have you done a pip install Cython?

[hoary zinc]

nope

it wasn't in the requirements, so I'll install it, thanks

[lone bolt]

It was a pain to install ReActor. One error after another

[hoary zinc]

I'll see if I can get it done with my bat script.

I think it'll work. it downloaded cython

Will upgrading the pip installer break things? I still have version 22.2.1 and it's already 24.0

nope, it still doesn't work

I installed every requirement

[lone bolt]

Probably not. Btw, you can install python packages with the comfyui manager.

[hoary zinc]

Cython-3.0.10-cp310-cp310-win_amd64.whl this one was installed

and this message when I try to open "install.bat" as an administrator and without administratior rights it says the same message

[hardy tree]

i also remember reactor not working with another face replacement thingy, dont remember which one, basically you cant have both activated on comfy, but maybe they fixed it since then, im personally not using reactor these days anyway

[hoary zinc]

Is there an alternative? It really doesn't work even with all dependencies installed.

[hardy tree]

im using ipadapter faceid stuff as well as instantid, works very nice for me

[hoary zinc]

it wrote that the problem might be with a pip, I doubt it

[hardy tree]

reactor i think is using some deprecated thingy, or old technique, and the results dont look natural at all

[hoary zinc]

it cannot even be installed

[hardy tree]

well skip it then 🙂

there is also photomaker, but instantID is better than that

[eager escarp]

any particular ideas for applications for the new unetcrossattentionmultiply etc. nodes?

[hardy tree]

i didnt see those nodes

oh now i see it, latest comfy commit

maybe it improves the quality?

[eager escarp]

nah plays with the q v k

just wondering if there's a specific application for this

wondering if maybe something is coming up soon

[hardy tree]

they preparing for sd3 🙂

[eager escarp]

it's pretty good with ipadapter styles

[hardy tree]

huh

il try later

[steep marlin]

To anyone playing around with pixart sigma, the Nvidia align your steps custom sampler workflow seems to work pretty well and gives even better prompt cohesion. Granted I've only tested it a dozen or so times so far

[steep marlin]

i took it a step even further and am testing it with AYS+autocfg~~+PAG~~(pag doesn't work, should have known better since pag operates on the unet middle layer), appears to work well with them all. top image is just the default workflow. the prompt was:
"in the style of retrofuturism, a woman wearing a black outfit operating advanced instrumentation in a laboratory with pale orange walls, analog screens, retro tech, liminal, vintage film grain, retrofuturistic"

since the boost mode on autocfg is a patch to the whole sampling system, i found i have to turn off the boost function or sometimes the normal sampler will bug out, but that's a known issue with the node(it can affect the whole workflow, even if disconnected so long as it has been ran even once with boost enabled. the only way to fix it is to disable boost and rerun)

[pearl bronze]

hello

Is there anyone who can help me about faceswap?

[cedar mirage]

Well other then use the reactor node with a simple faceswap. Source image (e.g. generated), image with the face. Put both as input into the reactor face swap node and it should work. Maybe you need to download an additional faceswap model but you can find them with the comfy manager.

[pearl bronze]

Do you have any recommendations for an additional face swap model

[cedar mirage]

I use inswapper_128.onnx which is quite ok. Still wished they had released a version with 512 px resolution

[frozen wasp]

there is always network problem in China because of the gfw, I always got the error message when launching comfyui. is it possible to stop comfyui connect to github when starting

[mighty wave]

Hello, my ComfyUI stoped working after trying to install OOTDiffusion. It just doesn't start 😦 Any advice? Please... Thank you

[cedar mirage]

The quality of life suit addon is the one which is looking for an upgrade. If you move the directory to another place (from custom nodes to tmp or whatever) comfy could start again. Then select the option of the comfy manager to avoid checking for updates automatically.

[fickle remnant]

idk AYS seems to ignore the prompt a lot more, for example prompting for a black background with AYS makes a bright grey one at best

[eager escarp]

it's funny cuz they said they did that to avoid abuses... i have a comfyui workflow that blows it away

pretty pointless

[cedar mirage]

But you use another swap model or tile controlnet / ip adapter instead

[eager escarp]

...so?

point is, they aren't stopping ppl from doing high quality face swaps

[cedar mirage]

If other release models which are able to do it better, they keep there hands clean. On the other hand the better models are usable thru their discord bot as far as I remember. So their morale seems flexible 🙂

[eager escarp]

yeah

[steep marlin]

Depends a lot on the steps, cfg and sampler. And obviously the prompt and the model

[fickle remnant]

Talking about PixArt sigma in this case, I copied the settings posted and idk it varies so much

I still have to get the amazing outputs people have pulled out of it, my gens all looks like ass

[steep marlin]

ahh, i'll have to test it out some to see, might just be due to how pixart differs internally

[fickle remnant]

Testing another workflow now

[steep marlin]

yeah im doing a regular vs ays test right now to compare some seeds

[fickle remnant]

This setup seems to fare better with AYS

[steep marlin]

i'll test that out in a sec, but i'm definitely seeing what you're talking about when i say "black background" and it's making the background grayish

[fickle remnant]

Yeah, maybe its the combination of nodes, im not sure

[steep marlin]

i've stripped it down to just the basic ays setup vs regular karras setup and it's doing it on pretty much every sampler type that i've tested

[fickle remnant]

I first thought its a VAE thing cause colours, but this seems to be deeper

[steep marlin]

my guess is that it has something to do with how aggressive ays is with noise steps early on

ays is meant to be most useful for low step counts like 10-15

[fickle remnant]

Does it ignore other colours too

[steep marlin]

lemme play around

[fickle remnant]

Ive tried white and usually only get beige stuff

[steep marlin]

purple worked fine

lemme just make a big random batch {red|green|blue|etc etc etc} and run 20-30 of them to see

oh wait that doesnt worth with that special prompt encoder

ill have to manually run a bunch then lol

[fickle remnant]

You can use this for dynamic prompt/widlcards

[steep marlin]

oh im just sayin within the specialized t5 encoder thing, i know it doesn't account for weights correctly or something like that according to the git page. the {||||} thing i showed works fine for regular encoders though

[fickle remnant]

[steep marlin]

ays->normal->ays->normal and so on

black white pink red green

[fickle remnant]

So it desaturates at least

[steep marlin]

yep

the prompt said "a plain ____ background

[fickle remnant]

I mean its nothing bad per se, but if you want poppy colours maybe dont use it 😅

[steep marlin]

yeah, but other than that, it definitely helps shape things a little more reliably

oh and those were all at 15 steps

cfg 6

and with the 512 model

[fickle remnant]

what sampler?

I run the 1024 model

[steep marlin]

dpm++ 2m

karras for the regular

1024 was doing the same stuff

but 512 is hella fast for testing

the 512 model is still technically the "sdxl" version as well in terms of the rest of the settings

[fickle remnant]

Ill now try to fit an sdxl refiner in, with the workflow you showed I only got black image outputs, the one I posted at least works somewhat

still getting grainy bad quality outputs tho

[steep marlin]

which workflow? the alien one from a few minutes ago?

[fickle remnant]

No the initial one you posted a pic of

Im using this right now

throwing controlnet in there to see if it helps

[steep marlin]

oh i was using automaticcfg and perturbed, but perturbed does nothing for pixart. automatic seemed to help a little, played around with the advanced version of it a bit

[fickle remnant]

could also try running SUPIR over it

[steep marlin]

this is where ays really shines though: 10 steps

and i think the node is actually hardcoded for 10 steps, if you use more than that, it does some log scaling math on the hardcoded step table

[fickle remnant]

Yeah that seems really good

[steep marlin]

NOISE_LEVELS = {"SD1": [14.6146412293, 6.4745760956, 3.8636745985, 2.6946151520, 1.8841921177, 1.3943805092, 0.9642583904, 0.6523686016, 0.3977456272, 0.1515232662, 0.0291671582], "SDXL":[14.6146412293, 6.3184485287, 3.7681790315, 2.1811480769, 1.3405244945, 0.8620721141, 0.5550693289, 0.3798540708, 0.2332364134, 0.1114188177, 0.0291671582], "SVD": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.002]}

yeah from the comfyui git

ays probably just needs a custom set like this for pixart

id bet if there were an entry for pixart, it would probably start at some smaller number like 8-10 and then add a little more to the following steps to achieve roughly the same curve integral, since it ultimately decodes as an sdxl model and all

im going to try plotting these curves out and play around, maybe i can find a decent curve for it

[fickle remnant]

Have you tried mixing multiple sigmas. so basic one and AYS for example?

There is a merge sigmas node

Also AYS for img2img doesnt seem so good

[steep marlin]

not yet, but im about to test out the curve in red to see if a slightly softer start helps

gotta add in a new line to the code for another dropdown in the menu

[steep marlin]

middle one is my custom curve, i used your prompt

the "evil robot above a human brain, super advanced technology, 8k resolution, hyper realistic, day, 35mm film, editorial, high fashion, 2023, the details of its skin and texture, off white by virgil abloh" one

this is with the 512 model version for speed

still not doing great for black background though, maybe it needs to be heavier than the default curve, like extra strong in the first handful of steps, i'll test it out

[fickle remnant]

Right one looking the best is kinda rough 😅

[steep marlin]

well that's also at 40 steps, forgot to change it

[steep marlin]

i think im starting to see why it doesnt like AYS for the background coloring. looking at the code for pixart itself, it seems like it wants a linear noise schedule and obviously, AYS is following a curve

a bunch of other differences in it that i don't fully understand without doing a ton of reference digging. the real way i'd have to test for sure would be to run pixart from their source code and try to hackjob in AYS to it, but it's far too late in the night for that lol

[fickle remnant]

So its a rabbithole

[steep marlin]

yeah kinda, maybe someone already made sigma wrapper for comfy, would save me a ton of time

at any rate, thanks for bringing up the whole background color thing not working with it correctly, kinda made me scratch my brain for the night

[fickle remnant]

lol no problem

[steep marlin]

I'm just going to view ays+pixart as a combo you'd only use like a lightning model with 10 steps. Based on the 10 step tests earlier, it worked pretty phenomenally at that

[fickle remnant]

For me it still gives the best results for pixart so ill stick with it, until someone finds a better way

Uploaded my workflow to civitai now, maybe someone finds it useful

[steep marlin]

Yep, but overall, pixart is pretty fun to play with while we wait for sd3. Crazy how well it works with such a small model size

[fickle remnant]

Its doing good with SDXL as refiner

[steep marlin]

Hell yeah, that's obviously where it shines the most. Just use it for laying out the composition and resample with xl

You could probably even do a pixart->unsample to a few steps->resample with sdxl type workflow as well

[fickle remnant]

I never unsampled before I think, gotta try

[steep marlin]

I actually might rig that up before the kids wake up. And it's annoying, but when it works, it's amazing

[fickle remnant]

Dont you need sleep

[steep marlin]

Yeah, later on lol

[fickle remnant]

You should be able to get my tidied up workflow from my images posted here, if that helps

[forest birch]

Is there a lora/form of model out there designed to generate stuff as alpha? As in transparent, to not need to remove background afterwards, but it generates the image already with transparency?

[fickle remnant]

Yeah, layerdiffusion does that

But the results arent as great imo

[steep marlin]

i'll check it out

[forest birch]

Nice!

Doesn't have to be great, as greatness comes from trial and error, but amazing one such release exists nonetheless! :D

Does forge or comfy do it better though , or is the result gonna be the same there :P

[fickle remnant]

Comfy took me a while to get it to work properly, forge was easier to set up iirc

[forest birch]

Interesting. Due to strength needed of the first node to get it right?

Also, does it do just straight hard pixel removal? Or dithering, with for instance smoke?

[fickle remnant]

Nah I had weird issues, could also be my comfy install tho

I think its a straight mask, but you can apply dithering in comfy iirc

[fickle remnant]

Is there a prompt styler for pixart? Or do they use the SDXL one

[steep marlin]

here's an example, i'm not using the controlnet for the third one, only the same prompt. so this is an unsample to 3 steps->resample back up to 30 steps (27 steps in the ksampler)

from your workflow

had to drop the latent scale to 1.25 though because my gpu only has 8gb vram

the annoying part for an unsample workflow is that it's deterministic and the seed on the resample has no affect on the output, you have to perturb it elsewhere and it's hard to do without breaking

[mystic plaza]

can i pls get some help on how to fix this everything is installed properly

[fickle remnant]

Interesting

Gotta look deeper into it

[steep marlin]

like i said, it's annoying lol

and very fickle

[fickle remnant]

Did you try updating comfy?

[mystic plaza]

yes i did but still not working

[fickle remnant]

What is the console showing?

[mystic plaza]

console as in ?

as in the cmd console ?

[fickle remnant]

yeah

[steep marlin]

this will get you started. the unsampler needs to stay at 1cfg and the sampler types have to match afaik. you're basically dissolving an image back to the pure noise that generated the image. i've found that you'll rarely want to go above 10-20% on the "end at step" (also where the ksampler starts at). if you go too high, it just ends up replicating the original image too much like you're only denoising at 20% strength or something. so in this example, 3/30=10%. oh and make sure you don't add noise and don't return with noise. i mean you can return sometimes, but it gets more complicated because you're getting into a multistage workflow like sdxl+sdxlrefiner

[mystic plaza]

this is what my cmd console is saying @fickle remnant

[steep marlin]

you can also do some split sigma stuff where you rewind and resample as well, but it's even more fickle and annoying

[fickle remnant]

Anything above that regarding the core nodes?

Did you follow the instructions? https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

[mystic plaza]

yes

@fickle remnant

[fickle remnant]

Okay no clue, maybe #🤝｜tech-support can help

[fickle remnant]

Do you unsample with SDXL or PixArt?

[steep marlin]

Sdxl

Since you're wanting to remaster the image with higher quality models

The prompting for unsampling can get weird though, you can use no prompt or the original prompt. Then in the ksampler prompt, you can change things. Like if the original image composition had a woman in it, you could change it to a man while keeping most of the scene composition

And since it unsamples with 1cfg, the negative prompt going into the unsanpler is ignored

You can also mix in controlnets with the ksampler half as well, but it partially defeats the purpose

[maiden wigeon]

Hello, in my activity, I am frequently required to switch from an SDXL model to a stable cascade model or another SD model. The issue I encounter is the very long loading time of Comfyui when switching from one model to another (the time it takes to load the new model). Would you have any advice or an idea to give me to shorten this delay? Thank you.

[steep marlin]

put the models on an nvme drive, even ssds are relatively slow when you have to load 6-9gb models at 500mb/s

[maiden wigeon]

Okay, thank you. I don’t have an NVMe on this PC, indeed just one SSD. I will think about it. Is there possibly another software-based solution?

[steep marlin]

more ram would be the only other option, since comfy will offload models back to ram when not in use, but if the ram gets too full, it will remove them, which means they have to be reloaded from your drive again

but no, no real software solution

the models are huge, 10gb at 500mb/s is 20 seconds

[steep marlin]

here's a good example of how unsample can be used, the workflow should be in the second image

notice how a lot of the composition is still very similar, but i swapped the turtle for a frog and changed the style to watercolor just by changing the prompt

it's almost like a poorman's ipadapter/controlnet, but without having to load a bunch of other stuff

(uploaded the wrong images)

and also, just by upping the start/stop step from 1 to 2 or 3, it can drastically make it mimic the original image even more. but that's what i was talking about when i said it can be fickle

[fickle remnant]

Thats really cool

[fickle remnant]

I tried combining it with my workflow but it throws me errors

For the unsampler I get this :The solver failed because the input matrix is singular.

[steep marlin]

hmm, not sure about that error, never seen it

[fickle remnant]

Ah figured it out

Didnt use the same two samplers

[steep marlin]

ahh that would do it

oh and speaking of samplers, dont use ancestral ones with unsampler, they have to be ones that converge

[fickle remnant]

Im using the same as you do

I think im slowly understanding it now

[steep marlin]

you're basically just taking an image and reducing it to a "squint" of the image and then using that as a glorified seed to make a new image with

which is ironically kind of how cascade works under the hood

stage C makes the tiny image, then it gets upscaled and resampled by stage B

[fickle remnant]

For me it only turns my cartoon image in a photoreal one
Ill play around more

Second one is resampled, first is normal SDXL refine

I actually like it

Switched to end at step 5

[steep marlin]

yeah not bad. it's just another random tool to keep under the belt, it's not always worth it and it's not a workflow that has to be forced into everything either. but it has some serious perks if you want to drastically change elements in an image, while keeping most of the overall composition

[fickle remnant]

Which is good for refining images id say

[steep marlin]

like keeping the overall silhouettes and stuff. but that's why i was suggesting using it in a pixart->sdxl refine workflow. doing something like a 30% denoise on an image that wasn't all that great in quality to begin with, isn't going to yield great results. but if you pump that denoise up to like 55%, then you might completely change the whole scene too much. so that's where tricks like unsample come into play. obviously, you can still do a bunch of controlnet+ipadapter type workflows as well, it's just this one is quick and cheap sometimes

[fickle remnant]

I pinned it to around 0.4 denoise with canny controlnet

Thats how I keep the shape

[steep marlin]

yeah, but when pixart puts out some derpy face on something, 40% isn't going to be enough to change it enough

[fickle remnant]

But right now unsample looks more promising

Will have to try with more styles

[steep marlin]

i want to say some of the other apps like a1111 and forge and focus all have some kind of similar system under the hood for getting image variations

they are basically doing the same thing and maybe mixing in some split sigma stuff

like if you take a 30 step image, walk it back 15 steps, then resample it back up 15 steps with a different seed. you have to keep track of the sigmas so that you don't completely break the noise stepping

like the curves i was showing earlier, at certain steps, it expects certain amounts of noise and yada yada yada

point is, any other app that has a similar "variation" mode like that, probably is doing something like i just described

[fickle remnant]

I wonder how Adobe does their reference image AI stuff in Photoshop Beta now

[steep marlin]

no clue, haven't messed with it yet

[fickle remnant]

One problem with the unsampler seems to be weird image artifacts

Maybe I need a third pass with XL only

If you zoom in youll see what I mean

[steep marlin]

Shouldn't be getting those, probably some latent space mixup. Like in your workflow earlier, you were upscaling the image by 2x, that might screw things up. Also, if the sampler after the unsampler adds noise or the start/end/steps aren't exact, it can add noise incorrectly

Oh you might be talking about vae artifacts

Each time you encode>decode, you lose information. It's a lossy compression essentially

[fickle remnant]

Tried it with no prompt and quality prompt, both not great

[lilac kettle]

For those who use comfy ui, does anyone have a better workflow setup than the default?

I just use this basic setup

[fickle remnant]

Depends on how much more advanced you want it to be

I have this, but it uses a few custom nodes you can get with ComfyUI Manager

[hoary zinc]

How do people do straight lines in ComfyUI?

[fickle remnant]

In the settings of comfyUI

Not sure if its a custom node that makes it available

[hoary zinc]

thank you

at least I'll know which lines comes where

[dense berry]

😮 are we talking pixart?!

[fickle remnant]

maybe

[dense berry]

For pixart users: AYS needs a higher sigma value, multiply it for better results, I'm still working on a good curve, but the AYS curve is off for pixart sigma

[hoary zinc]

found my idea on fixing hands, but it's much more complciated workflow than I anticipated

with open pose to detect the pose first and then fixing it through depth mask or something similar

it's using 2 control nets

[hoary zinc]

What do these 2 nodes do?

[dense berry]

first is alignyoursteps -> a new scheduler by Nvidia, which is pretty damn good tbh

[hoary zinc]

I haven't got much into it. I want to first make a workflow to fix hands.

[dense berry]

if you update comfy, you have it by default

[hoary zinc]

When was the update?

[dense berry]

a week back or so

[hoary zinc]

I don't have multiply sigmas node

[dense berry]

And this scheduler is totally related -> having the right schedule (timing) makes all the difference between garbled hands and good hands

[hoary zinc]

only the schedule aligner

pretty

[dense berry]

i think it's from the https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler node

[hoary zinc]

Can anyone tell me why setting a noise mask doesn't work with tiled ksampler? Is it because it chunks down the image into smaller pieces? Otherwise it gives me an error: "module 'comfy.sample' has no attribute 'prepare_mask'"

[olive yoke]

has anyone authored a training script to optimize the sigmas for a given model, including potentially a model + loras?

i didn't really read the paper but is it automatic, or human feedback based?

i know it is per model

[dense berry]

for sigma, i have no idea -> it's a whole different architecture, as far as i know

that's about all i know

[dense berry]

Dunno if you can change the way the model works... i'm still looking into what's the best schedule/cfg combo, i noticed pixart is absurdly sensitive to it

[olive yoke]

are you using city96/ComfyUI_ExtraModels for pixart sigma?

[dense berry]

yes, there's no other loaders

however, using a default sampler/scheduler combo will not net you the best image quality pixart can produce

if the schedule is off -> images are just plain weird and have really bad proportions

the nvidia scheduler is pretty good, but the initial sigma might be a bit on the weak side.That being said, by boosting the cfg to absurd ratios and then using an ancestral sampler might also work

still figuring out what's the best way to go

[olive yoke]

at this point you should probably use the align your steps scheduler, because it is so much better than any of the other options

[dense berry]

this. Once you get used to that one, you're not gonna pick any other scheduler anymore

20 steps... each time i feel like -> hey it's getting close and then be like "nope, it fucks up somewhere else"

[brazen wigeon]

which model?

[dense berry]

that's my own pony blend

[brazen wigeon]

share?

[dense berry]

dm'd u

lemme know what u think of it

it should just behave like any other pony model tho

it's not that special -> and there's model that give better image quality by default

[brazen wigeon]

how do u even run pony models

i have never ran a pony based model

[knotty quest]

how do i achieve auto adetailer results on comfy? my facedetailer just doesn't get there

the eyes are ruining my gens pretty much

[brazen wigeon]

what prompt btw

[knotty quest]

nothing special, just 1girl, goth laying on her side w a few cats

no matter the prompt i use tho, the loras combo are causing that, but facedetailer can't fix it

now auto1111 works, but we all know at least 100 reasons why auto is worst

for me personally, it would slow down my process by a lot

[dense berry]

disable the lora's on the model used by adetailer?

[knotty quest]

adetailer (AUTO1111) is giving great results.

I need those results in comfyui.

[dense berry]

I've not used adetailer in comfy ever, i wouldn't be able to help there, but i know for pony models, and img2img, some lora's bite

dunno if adetailer node in comfy has a model input, but if it does, try to use an input model without loras

[knotty quest]

I suppose adetailer keeps the lora's, since it's using my style on the eyelashes.

[dense berry]

hmm

right

than it's just a matter of tinkering, i think the denoise should be higher in that case

[knotty quest]

even going to 1.0

it's like the image is always shit

changes the entire face, still shit image

[hoary zinc]

have you tried this?

I tried it once for my image and it worked quite well.

[knotty quest]

what exactly are you saying

[hoary zinc]

using face detailer

[knotty quest]

a refiner?

[hoary zinc]

yes

idk how adetailer works but I have it too

[knotty quest]

hmm i mean i could try it

don't feel very confident about it though

since adetailer uses the same model to gen

[hoary zinc]

before:

after:

and added "thin lips" into the prompt

[knotty quest]

i work with hundreds of images

i can't really prompt for each

could i send you one to give it a shot?

[hoary zinc]

well, if they are different

well, I'm still a newbie at it, but could try, but it'll take a bit of time since I have only GTX 1650 4GB VRAM on my notebook

[knotty quest]

no worries won't bother you with it then

i'll download and see what happens

[hoary zinc]

ok, but if you're interested in it I tried this one: https://www.youtube.com/watch?v=ekofgf9T-9c

but anything with mask doesn't or mostly doesn't work with tiled ksampler, but don't know why, maybe because it's tiled?

[knotty quest]

i've tried that, unfortunately i work with anime style, and this didn't yield the results i was looking for

[hoary zinc]

or you can try Krita

unfortunately what I tried for fixing hands didn't work properly

or at least in 512x512 resolution it doesn't work well with SD model

had to use krita

[stoic rock]

When I’m loading a Lora I always just set the clip strength to the same as the model strength. I don’t understand the difference. Should I be leaving clip strength 1.0?

[hoary zinc]

yes

clip is what encodes the text you write into tokens that are translated into language your PC knows and strength means factor

there was one technical video which explained all these terms, although I haven't watched the whole video it might be useful

[stoic rock]

Your explanation helped. Seems like for the most part I don’t want the Lora’s messing with the clip tokens.

[hoary zinc]

leave it at 1, default for everything is 1, also in the clip if you want to increase strength or how much impact that object or whatever you want e.g. you want a girl in a picture where is a boy you use (girl next to the boy:1.4) for example

default factor is 1, so for the girl it'll be 1.4x more so she's more likely to appear

hope this explanation helps

[stoic rock]

It does help… thank you!

[eager escarp]

doesn't it start at something nuts like 157.40724182128906?

with sdxl it's usually 14.614642

[dense berry]

correct. and it can handle it actually

starting at 14 is just a bit too low

[eager escarp]

oh what i mean is multiplying by 2 isn't nearly enough

[dense berry]

turns out, doesn't need to go that high:

you use an extremely high cfg to emulate it

just nudge it a bit over the top -> else the middle part of the generation gets wonky

[eager escarp]

huh, that's interesting

@stray hawk will probably be interested in this

i'm assuming those are outputs with those settings? looks pretty good

[dense berry]

yup

and these are song lyrics, actually, so i'm relying all on the base model actually being able to make something coherent out of garbage

[stoic rock]

Does your workflow nodes depend on AYS to tokenize the lyrics?

[dense berry]

it does use AYS yes, but i do add a little multiplication, because 14 as the first sigma, and the following ones aren't good enough

when actually prompting for something 😮

[stoic rock]

Beautiful image!

What is your preferred nodes for dealing with text prompt tokenization? Just standard ClipTextEncode?

[eager escarp]

ah ha, someone else that also likes to mess around with manual scheduling!!

[dense berry]

no, for pixart you need to use the t5 encoder

[dense berry]

i've tried so many scheduler formula's, and then nvidia came out with that, and the first images were kinda ok, but coherency was missing a bit, so when i was fucking around today, i had the idea of either splitting or multiplying their sigma schedule, and it worked 😮

[eager escarp]

cool

i got some extra nodes i added to that btw

allows some more manipulations, like creating iterative unsampling schedules without losing momentum

[dense berry]

yeah, i heard boto talk about it, but i wanted to not do that -> i wanted to get the pure result from pixart sigma

[eager escarp]

but yeah if you want to be able to manipulate sigmas more effectively... i should drop that here

added a few simple things that allow you to do stuff like actually concat sigmas, trim out consecutive duplicates, remove ones below a floor or set their value so you don't get NaNs/black images

here ya go

i'm lazy and just added them to the same file

also made it so you can use stuff like arccos in the manual scheduler

obv if you use that back up the original sigmas_merge.py in case i busted something i haven't noticed yet

[fickle remnant]

Yeah I have an unsampling workflow that uses SDXL as a refiner for pixart

I released one on civitai and working on an update later this week

Trying to fix my terrible pixart outputs

Multiplying the sigma already showed decent success

[eager escarp]

oh cool

[fickle remnant]

Got that idea from Eface tho

[eager escarp]

what i pasted there... pretty interesting

cool

there's all kinds of effects from the rho, the number of steps for the unsampling vs sampling side, etc

the number of iterations

doesn't need to be nearly that many just to get something period

i've found with res_momentumized you can often get away with generating images that way from an empty latent

so long as you boost that first sigma to 4 or so

what i've found is the sampler for that one needs to be something that adds noise... res, ancestral stuff

sde_cpu is also really pretty good

[fickle remnant]

My plan is still refining the output with SDXL, the unsampling idea was from RX808 here
PixArt by itself still lacks in details and SDXL can jump in there

[eager escarp]

cool

yeah unsampling rocks in the right context

[fickle remnant]

Unsampling works really well for refining

[eager escarp]

yep

[fickle remnant]

Im combining it with canny controlnet but not sure if thats even needed

[eager escarp]

i've found these flipped sigmas give the best results of anything if you can get away with using the same sampler for unsampling and sampling

and don't need to change any other parameters for each stage

ahh

yeah i've used the tile v2 one a bit sometimes as well

usually i don't use anything

[fickle remnant]

But unsampling also gives weird artifacts, so I run a 3rd pass with SDXL with low denoise

[eager escarp]

but i'm usually using the iterative approach to make a series of small, limited changes with a style ipadapter so that the composition only shifts slightly

what are you using for unsampling?

[fickle remnant]

Not sure from my head uhm

https://cdn.discordapp.com/attachments/1109185138383921182/1234881297579638805/240430165647_Elegant_Minimalistic_Photography_1990s_beautiful_Model_wearing_a_pink_jumper_Red__00003_.png?ex=66325854&is=663106d4&hm=7ab44a51bf1042e6aebd37c0bb21810f339c2fd5b64bf83760b1a03089714988&

I think this has my workflow?

Im not at my office PC so cant check rn

[eager escarp]

gotcha, yeah it's in there

[dense berry]

Scheduler only a bit off, i'll try more tomorrow

[eager escarp]

yeah you're using the regular unsampler node

[dense berry]

this one, i'm actually REALLY proud of -> theres actual pareidolia going on 😮

[eager escarp]

works great for a lot of stuff, but doesn't allow custom sigmas or samplers

[fickle remnant]

Its an older iteration of the workflow, I already did some changes

[fickle remnant]

True but I didnt feel the need

[eager escarp]

it took me a while to realize that you can just use samplercustom and flip the sigmas lol

[fickle remnant]

5head actually

Didnt think of that

[eager escarp]

if you look back i'm sure you can find me whining about the lack of an unsampler that takes anything but primitives as inputs

then one day i just tried the flipped sigmas and was like... duuude

yeah that's where that "clownsched" came from

unsamples, then samples, unsamples, then samples, over and over

[steep marlin]

Oh cool, yeah that's what I was playing with this morning, but I was just manually adding in new sigma tables to the node since all you need is 10 steps and it will stretch them out if you use more steps. But my experiments were a bust because I was trying to keep the overall average the same as the sdxl setting. I tried a stronger first few steps and a weaker first few steps, but was adjusting the rest of the curve to still average out to the same. I'll try this 2x multiplier later

[dense berry]

1.3 seems to be a sweet spot

I'm not sure if someone can read the pixart source code and reverse engineer what the actual formula should be?

or are they also just eyeballing it?

[eager escarp]

i have a feeling they don't really know exactly either

[steep marlin]

I kind of read it but they seem to do some weird conversation stuff like starting out discrete and then converting to continuous I think, idk, I was really tired so I might have just been trippin lol

Before sampling

[eager escarp]

the kinda stuff you can do with iterative unsampling with no controlnet or ipa

pretty fun with IPA too

workflow embedded

[pulsar drift]

hello,i was curious the difference between comfyui and webui,which one is better?can comfyui use deforum plugin?

when i download comfyui,should i care about its version?

[lone bolt]

Are you on windows?

[pulsar drift]

yes

win 11

[lone bolt]

https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_cu121_or_cpu.7z

Download this

Unzip it with 7zip

Go into Update -> update_comfyui.bat

And you should be good to go

[pulsar drift]

can comfyui do everything webui do?

[lone bolt]

You have even more freedom

[pulsar drift]

thanks

[lone bolt]

The possibilities are near endless. It takes a bit of learning. But don't fear. You will be rewarded

[pulsar drift]

😋 i am trying

[lone bolt]

One important thing! Get the manager:
https://github.com/ltdrdata/ComfyUI-Manager

[pulsar drift]

too many options to choose,different version of sd can have different effect

[steep marlin]

check out https://github.com/BlakeOne/ComfyUI-CustomScheduler, you can make your own table in one node instead of chaining together 5000 nodes. i saw the sigma graph earlier and it reminded me of this node that i didn't use this morning when i was manually adding table entries to the ays node

[pulsar drift]

i hope comfyui can save some time~~

[lone bolt]

There are a lot of workflows for that out in the wild.

[eager escarp]

that one is nice, but then you have to manually calculate the values

it's pretty easy to manipulate those giant tables i pasted once they're built

ultimately it's a lot more tedious to punch em in manually imo

[steep marlin]

yeah true, but what you're making reminds me a lot of doing lora training with cosine soft restarts. there is also another node like this that can take an input of an equation

[eager escarp]

the manual scheduler one? yeah i've used that too

i have equations that come very close to that iterative schedule, and are nice cuz it's clean math where doubling a value doubles the wavelength, etc, stuff like that

but... it doesn't play nice with these schedules

[steep marlin]

yeah that's the one

[eager escarp]

the numbers don't line up quite exactly with it being so quantized

so i abandoned that apporach

absolute best results were with making each segment individually, then concatting and flipping and cleaning up the duplicates and near null values at the end

[steep marlin]

which numbers? the ones calculated from the node?

[eager escarp]

yeah

i modified the manual scheduler node so you can use arccos, arctan, etc

allows you to do some nice neat stuff wit hthe math

[steep marlin]

yeah i was going to say, if anything it could just be a case where you need a double precision variable and it's pretty easy to throw in other functions like atan2 if needed(not that you're actually needing to draw circles)

[eager escarp]

it's not so much that

as it's that... for example, if you're having variable length wavelengths with each cycle

the bottom and the top of each won't necessarily coincide perfectly with the integer step count

in fact, it very rarely will

so you get unevenness in the schedule

[steep marlin]

oh that's just a scaling issue with the equation then, you need to have it scale with the steps

[eager escarp]

and the results are visibly worse, even though from a math standpoint, it's much more elegant

yeah

which means no nice neat pretty equations

gotta do each segment individually

hence the giant table approach. i could amke a node that does that internally, but why bother if that works and is easily modified

[steep marlin]

well tbh, it wouldn't be surprised if your issues with scaling were simply because of a -1 situation. like if you're doing 30 steps, the last step is 29 technically, since zero is a step. (i'd have to double check to be positive)

[eager escarp]

it's not that

[steep marlin]

the equation is an equation and all, and applying a scaling factor to handle it stretching correctly to N number of steps is what it is. it doesn't lose precision due to it, if the equation isn't discrete in the first place in the sense of having manually pecked out values. and simple float errors down in the bottom of the E-XXXX range aren't going to make any difference at all. so it just leads me to believe the math is off

[eager escarp]

well, i'm not enough of a mathematician to make an equation that can follow a shape as complex as the one i generated with the table

i'm trying to find the equation i came up with, seem to have misplaced it

[steep marlin]

ahh, i thought that table was based on a formula, my bad

[eager escarp]

oh, nah

it's basically a bunch of karras chunks

[pulsar drift]

can i ask what is this do?i already installed comfyui and some tutorial said to follow this stap?

[eager escarp]

maybe there is a fancyass equation that would do the trick very nicely... my math skills are limited

gone through calc III and lin alg etc years ago but lost a lot of it at this point

[steep marlin]

i see. basically, it looks like you're starting out with tight spikes, then slowly widening them out, while dropping the peaks. I'd bet if you took the integrals of each spike, they'd be relatively similar in area with some kind of slight falloff toward the end

[eager escarp]

exactly

[steep marlin]

i used to do a lot of signal conditioning and control circuit engineering. like tuning PID controllers and such, but it's been a while

[eager escarp]

oh nice

yeah the only things ive noticed definitively have been...

that peak shape is much better than a curved sine wave, unsurprisingly... spreading out each wave with each iteration, and having each be a slightly slower sigma value is also good... and you want to denoise all the way with each iteration

a steeper unsampling slope seems to change the image slightly more than with both having the same slope, but usually it results in a better image overall

it hasn't been that well tuned yet

tried maybe 20 versions of it max

[steep marlin]

well it's definitely a pretty cool experiment, makes me want to see if there are any chirp equations i could throw at it, since you're essentially making a reverse chirp signal, but that starts out loud and gets more quiet

[lone bolt]

Guys, i know you are having fun. But you scare away new comfyui users. Even i only understand half of those words.

[pure salmon]

I’m not even a comfy user and I’m always scared

[steep marlin]

lol, well if we crack some cool breakthrough, it can be made into a simple and friendly node for newbies

[lone bolt]

I coded a node counter node to count clowns Load Image nodes.

[steep marlin]

lol

[lone bolt]

It was over 70

[lone bolt]

The portable version i linked you should have everything on board

[pulsar drift]

ok

thanks

i got your point

this is Manual Install

i was silly~~

[lone bolt]

If you want pain, struggling and suffering. You go the manual route. Chill and easy mode: Portable

[steep marlin]

Yeah definitely go portable and use the vent setup. Global python installs are always a headache

Venv

Autocorrect

[pulsar drift]

i love you programer~~ always make things easy~

😋

[steep marlin]

https://www.mathworks.com/help/signal/ref/chirp.html#description I think the curve you're essentially making would be the convex quadratic chirp. As the frequency decreases, the intensity decreases. If I get some time later, I'll try to see if I can torture myself with making a python script to implement it where you can define how many humps you want with it and stuff like that

[eager escarp]

awesome, thanks

yeah the whole key is that the apex and low point of each segment can be precisely defined along both the x and y axis

especially so that the point on the x axis coincides precisely with an integer

[steep marlin]

well when you get into the continuous blending of going from one initial frequency to ending at another, that's far more difficult to pull off. might work really well for other types of sampling, like playground2.5 maybe?

but here's a rough sketch of it kind of pseudo working codewise, it would need cleanup and big changes to convert simple inputs and whatnot

you could also still code your setup as well, but it would more or less be similar to your node graph cluster, just in code instead

i might also look into bezier curves as well, that would be a really easy way to do it as well and have absolute control over vertex points

[steep marlin]

Correction, b-splines is what I meant, not bezier

B-splines let you edit a single point without affecting the other parts

[eager escarp]

can you get the apex of each to be pointy like in the graph i had?

i think that's what i needed the arccos for

but i forget

[steep marlin]

Yeah with filtering that I've forgot the damn integral math for

[eager escarp]

hmmmmmm

i had an equation that worked

[steep marlin]

It would be exponential

[eager escarp]

last two weeks have been so crazy with work my brain is fried and i have no idea where i saved all those notes

yeah, it had that in there too... memory is pretty hazy though

i had it so i could change the height, the rate of decay, the wavelength, etc pretty easily with the equation

[steep marlin]

Yep, but the pinch of the apex has to be done in the integral. My example was just populating a list of N points like 5000 or something

So you can't do the math on the baked points as easily

Without having to walk through it with a kernal of like 8 points at a time or something

[eager escarp]

gotcha

[steep marlin]

Like doing gaussian math

[eager escarp]

yeah, i found the shape was very important

just like with karras vs exponential vs normal etc and all the other custom ones

none of the default ones work well with res_momentumized for example

karras is okay-ish but once you've tried a few custom expressions, holy hell does it make a diff

(x/15)arccos(cos(2pi*(y**0.707)*8))**2

there we are

except i forgot discord fucks up equations

but yeah after playing around with that for a while, i found it was just better building each segment to have total control over the curve and length of each section

[steep marlin]

yeah i can see that working. that exponent makes sense. 0.707 is half of the sqrt of 2(probably hypotenuse related since you're applying the "sharpening" to one side of the peak at a time)

sqrt of 2 is pretty important in these types of calculations, and in signal conditioning, you use it for peak to RMS calculations

but anyways, i'm about done for the day torturing the poor souls in this channel with the math spam. if i remember to screw around with a b-spline setup, i'll let you know. i figure it could be made where you'd designate like N peaks, like 8 or 10, and then set their exact step widths and peak height manually for each and then apply a filter to it to pinch them or round them out.

[eager escarp]

please do

yeah, i used it cuz it's sqrt 2... i felt like i'd seen it somewhere and it sure feels better than 0.7 or 0.8

at least it looks like there's a divine elegance to it 😄

[steep marlin]

I'll try. The benefits of being able to experiment easily with sigmas is important enough to me that I can feel justified in spending the time on it

[eager escarp]

yeah it's one of the easiest things to screw around with, and very important

waaaaay understated in importance

it's even more important than the sampler itself in most cases imo

[steep marlin]

Yep

[pulsar drift]

can someone tells me what is the name of this plugin？it can seach nodes

[steep marlin]

Double click the canvas

[pulsar drift]

oh shit~~~that easy?

thanksi know

[stoic rock]

I wish there was some way to search civitai image posts filtered by ones that include the comfyui node workflow. I learn so much when I do happen across them,

Nice 26 node workflow

[eager escarp]

that would be a nice feature for sure

maybe suggest it on their discord

[opaque sentinel]

Hello Can someone help me with this?
im trying to run ComfyUI buy it doen't work

[pulsar drift]

hi~~~does this error means i download wrong model?

[fickle remnant]

You have selected an SDXL model with a 1.5 model as stated in the Red Text

[cedar mirage]

Check if you start your server with an argument -port as it is obvious required. Guess you set the ip to 0.0.0.0 to reach it from outside. Either you add :8080 or specify the port with the argument.

[opaque sentinel]

How to check? What I have to do?

[cedar mirage]

Which command do you use to start the comfyui gui

[opaque sentinel]

this one

[cedar mirage]

Try removing the mixlab custom node and see if the comfyui start up normal.

[opaque sentinel]

ok let me check. thank you

[stoic rock]

Good idea, I’ll do that. But for some reason I get the sense that’s not something they want to promote…

[leaden coyote]

if anyone wants help growing their datasets, I could do that in exchange for stock image subscriptions

and online GPU rentals

[loud crown]

Hello! Anyone know how I might install t2i-adapter-sdxl in comfyui?

[unreal reef]

Regarding the 'Latent Garbage Collector' node, is there any particular step in the workflow that it's best to place it in? Like, do you clear things out right at the start or at the end of a chain of events?

[languid summit]

I ran a vid2vid animatediff wf. However, I am not getting enough details and not the sort of look I want to achieve. Is my lora and checkpoint combination apt? Also it rendered only 1/3 of the video.

[stoic rock]

I always forget, do I place the CLIP Set Last Layer node right after the Checkpoint Loader but before LORA loaders, or do I put it after stack of LORAs?

[hoary zinc]

go in order: lora -> clip -> sampler

[hoary zinc]

.

[stoic rock]

So is this craziness?

[hoary zinc]

or you can try this

rgthree node

also your resolution is kinda big

[stoic rock]

The 4096?

[hoary zinc]

yes

the usual SDXL model uses 1024 px x 768 or something like this, it was trained on 1024 x 1024 px

[stoic rock]

I was wondering about that… I saw it set that way in another workflow, and blindly went with it

[hoary zinc]

I'm not that advanced to tinker with clip layers, but it'll surely do something.

[stoic rock]

Oh this is very interesting indeed… never came across that one till now. I like how it drops the Loras in. I wonder how it weighs and tokenizes the prompt out

[hoary zinc]

it was recommended to me, so I tried it

I sometimes use it instead of the normal clip because I can choose embeddings and don't have to manually write them out

copy them from excel sheet*

[stoic rock]

I have a foggy memory that there was a recommended way to encode prompts so that if they were longer than 77 tokens it would do a better job of handling it… but maybe I’m way off or was dreaming

[hoary zinc]

check this video, he goes into details about clip, tokens etc. https://www.youtube.com/watch?v=1h4e24Zn3fM&t

I think he talked about tokens having size of 77

[stoic rock]

Thank you, looks great

[hardy tree]

not sure if this would help https://github.com/SeaArtLab/ComfyUI-Long-CLIP

[hoary zinc]

btw has anyone noticed this error in cmd?

• "clip missing: ['clip_l.logit_scale', 'clip_l.transformer.text_projection.weight']"

it might be a node or something, the only thing I can think of is "power prompt - Simple (rgthree)" clip node

[hardy tree]

no that is ok

that is from rgthree but it doesnt matter :3

[stoic rock]

I’ll test that out 🙂

[opaque sentinel]

Good morning, how can i fix this problem? help me please

[lusty pewter]

Could be several things. The error alone isn't enough information. Using a LoRA? Maybe you've got an incompatible one for the model you've chosen. Could be a bad choice of ControlNet for the model you're using. Perhaps you've got incorrect node connections for the CLIP model. Basically, a bunch of possibilities. Maybe what I've said helps, but you have to go look through what you have in your workflow, which we don't see here, to find the issue.

[opaque sentinel]

😭😭😭 let me check

[pulsar drift]

i got similar error,i copy that notice to txt,and it is said something wrong with my ksampler node? and memory got some problem?but i am pretty sure my computer got enough memory~

how can i fix this problem? it is a simple workflow~

[opaque sentinel]

my Ksampler or showing the purple color too

[hardy tree]

i mean it says out of memory there

sometimes what you can do in comfy, even if it says out of memory, is go to the manager, click unload models, and then generate again,, it will continue and maybe pass the memory problem, but this doesnt always work

[hardy tree]

strange, i never had a problem with instant id...

https://github.com/cubiq/ComfyUI_InstantID/issues/139

[pulsar drift]

it is ok, i gonna try some other workflow~~

[opaque sentinel]

yes... because i installed model via ComfyUI manager?

[hardy tree]

no that shouldnt be a problem

well i mean im not sure tho

maybe double check your folders and what are you using for controlnet, etc

[opaque sentinel]

let me check, thank you so much sir

[hardy tree]

i mean it's been a while since i used instantid, maybe something changed and has to be fixed, il check tomorrow myself

[hoary zinc]

I guess you added ~~after_control: ~~random and it tinkers with the resolution it wasn't made for

something like that

Or do you use tiled ksampler or anything and set the tile size too high?

[opaque parcel]

does someone know how i can inpaint using comfyui/stability matrix?

[cedar mirage]

Just had the discussion yesterday.

#🏞｜general-with-images message

Within the image you will find a workflow for inspiring including masks and background removing

[opaque parcel]

1. Does that also work with stability matrix? (Based on comfy i think)
2. i didnt quite understand how that worked :,)

[cedar mirage]

I thought stability matrix is a package manager compatible with nearly all kind of UIs (comfyui, a1111, foocus, forge,…)

[opaque parcel]

hold up im trying sumthing, if it works im stoopi ash

omfg i wanted to change bc of nodes

but you just have to open website

[hoary zinc]

Is there anything wrong with Tiled Ksampler? It generates multiple people when I only stated 1 and that's at the very beginning of a prompt. When I try normal Ksampler it works as usual.

I'm using SDXL model 1024x1024

tiled ksampler and vae decode (tiled)

[hoary zinc]

output when I use "Load Default" workflow and replace "Ksampler" with "Tiled Ksampler"and VAE Decode with "VAE Decode (Tiled)"

nobody really uses tiled ksampler, but I stumbled upon this problem

[mystic plaza]

i guess i am using the wrong models or checkpoints

[mystic plaza]

pls help me to figure out what is going wrong here

[pure salmon]

could you post it once more so I could get a better look

[mystic plaza]

this is the error i am getting @pure salmon

[pure salmon]

👍 I’ll check it out

[mystic plaza]

i am doing something wrong or missing nodes ?

[brazen cosmos]

茜

A little girl is dancing

[mystic plaza]

how to fix out of memory error

[grand pike]

hello lovelies. could someone tell me why i keep getting shiny noise in people's eyes whenever I upscale? I've tried different denoising levels, different amounts of steps, 3 different upscale models (4xNomos8kHAT-L_otf, 4xNomosUniDAT_otf, 4xFaceUpSharpDAT)

Am I just not using a quality upscale model? or maybe it has something to do with the sampler or scheduler?

I'm using the workflow made by scott detweiller. IIRC I couldnt find the exact upscale model he used.

[hoary zinc]

I can take a look when I get home. On the other hand I got more colours when I upscaled one picture.

What are the other nodes you're using?

Also did you insert anything into the clip?

I'd increase steps, if you have a bad quality picture or don't want many changes, to decrease denoise level even further (each step generates less with more steps so it's more accurate)

dpmmp_2m was good

But depends on what you prefer

Try to use 4x_sharp mmodel or similar, there's a pack on huggingface

[hoary zinc]

Use tiled ksampler, but I've had a problem with generating pictures through tiled ksampler, what is your workflow?

Try to separate that workflow into a different parts. Like one workflow to generate a picture. 2nd workflow to fix hands, 3rd face and so on. It's annoying, ik, I'm struggling to do hand fixes and generating a picture in the same workflow but I run out of memory. One solution is to buy a better GPU with at least 8, preberrably 12GB+ VRAM

[mystic plaza]

i am trying the ipiv morph workflow so its a heavy workflow i am guessing

[hoary zinc]

I haven't tried animatediff, it wouldn't even run on my GPU, probably or took too long.

What are your PC specs?

[mystic plaza]

its a animatediff v3 thing

[mystic plaza]

its a p4000 gpu rest i dont know

[hoary zinc]

Forgot which one it is. What VRAM?

How much VRAM you have

[mystic plaza]

quadro p4000 with 8gb ram and 16gb shared ram

[hoary zinc]

And RAM

[mystic plaza]

32 gb

[hoary zinc]

Try to lower the resolution

512x512 or whatever you are using, use tiled sampler, if possible, I have never tried anything like this so this is one possibility

Another option would be to portion that worfklow

[mystic plaza]

i dont think portioning the workflow would help much

[grand pike]

@hoary zinc here's my workflow i'm using for the upscaling. I do have something inserted into the CLIP. Should my prompt be blank?

also i dont think i've tried low denoise levels yet. I just figured if the denoise is too low than it wouldn't upscale as much. like it would keep the blurriness

[drowsy compass]

what did i do wrong?

[hoary zinc]

you are changing width and height somewhere it shouldn't be done

[hoary zinc]

What is the resolution of images you're trying to scale up? THe last time I used it it took me 5 hours, because I had to use tiled ksampler which uses CPU instead of GPU. so I did from 1024x1024 to 4096x4096 otherwise I'd run out of VRAM.

I also used Scott's workflow as a reference

[grand pike]

most of them are 536 x 720

it takes about 11 minutes 4x upscale them

[drowsy compass]

here is my workflow

wait i cant see it lol

[hoary zinc]

I haven't tried Ksampler Efficient, what does it do?

[drowsy compass]

it just like all in one feature

i just don't know what is wrong

i tried to change checkpoint it just work fine

[hoary zinc]

what was the problem again, sorry, my memory is short and I was trying to help 2 to 3 people

it was with the width and height, right?

[drowsy compass]

the error show something about width and height, but i think there is also a problem with my checkpoint model

i can't neither use vae nor upscale

[hoary zinc]

idk if I'd use XY plot on something I want to upscale

this might be the problem

Which checkpoint you use?

[drowsy compass]

nah it's unused

[drowsy compass]

im using mix-pro-v4

from civitai

[hoary zinc]

hmmm nto using controlnet?

I'll show you mine

[drowsy compass]

not using it yet

[hoary zinc]

the part zou can't see is just a checkpoint, and intbinary (the same thing you have)

[drowsy compass]

can i have your workflow as my refrence? i still new with comfy-ui and ai art in general lmao

[hoary zinc]

sure

[drowsy compass]

i'm actually still using colab for now, i think i'll go local after i have an idea what i'm doing

[hoary zinc]

ok

[olive yoke]

why are you using all of these efficiency nodes

they are buggy

this is obscuring all the choices that matter

try the default sdxl workflow first, with the checkpoint you like. does it work well?

i guess people "just" want automatic1111, but faster

[drowsy compass]

i don't know man, i just started using it today

[hoary zinc]

I've noticed it just now. You have increment in primitive node.

change it to fixed

[stoic rock]

Do turbo or lightning checkpoints sometimes get majorly hampered if you add a Lora to the workflow?

I was testing out a lightning checkpoint… it was iterating through generations very efficiently. Then I tacked in one Lora… and it’s pretty much brought it to a halt.

[hoary zinc]

I haven't tried turbo models although it's said to be fast. What model do you use? Is it SDXL lora on SD checkpoint or the other way around?

Personally, does anyone have a workflow that can fix hands? I've tried meshgraphormer + bbox detector but it isn't performing well

in most cases the generated image has better hands and krita + AI does a much better job

[stoic rock]

I was trying out Boltning … I believe my Lora is SDXL, at least says that in the file name, though I noticed this Lora is greatly reducing the quality of the image…

[hoary zinc]

Can you show me the workflow with lora included?

[hoary zinc]

try to lower cfg

I use one similar lora, lighting LCM lora

and use dmp_2m or something similar, or euler for example, sde samplers are quite slow

also why does lora have lower strength

[olive yoke]

why are you using all of these odd custom nodes

use the vanilla nodes to do this

what are you trying to do?

[olive yoke]

this is configured wrong for a turbo workflow

use the example from the repository

[olive yoke]

why are you using 1016 as the resolution?

[stoic rock]

I get confused if I should make use of comfyui example turbo workflow for “turbo” models on civitai.

[stoic rock]

Im not even sure now. I think I thought I was making use of the wrong vanilla nodes for handling clip text encoding so i started reaching for others.

[hoary zinc]

try to understand the workflow you're doing

also check the requirements for the models, they might need different cfg, steps, sampler

[verbal sky]

Hi everyone, my newest video just went live (fun topics include Hyper, LORA power play, special effect overlays and more!) https://youtu.be/0BZi4I9LW-A Enjoy!

[opaque sentinel]

Good morning guys, how can i fix this problem?
Error occurred when executing BRIA_RMBG_ModelLoader_Zho
i already installed it. but everytime i run, it always show the error

[olive yoke]

what are you trying to do

why do you have all these custom nodes

you just have to download the model

nothing more to it

[hardy tree]

https://tenor.com/view/calculating-puzzled-math-confused-confused-look-gif-14677181

[opaque sentinel]

i already got the model and it still happen

[hardy tree]

im gonna ask you only one question: what are you trying to do?

[opaque sentinel]

im trying to do this one Commercial Product Background Replace

[cedar mirage]

Can you confirm that within the directory the model file is existing? Because the message is absolute clear. One Custom-node require an model and it needs to be placed in the according directory.

Rmbg 1-4 needs a model.pth file.

[stoic rock]

Use the Turbo workflow for a Lightning model?

[haughty orbit]

Hey Guys, I need to load json workflow before loading the editor screen, is there anyway to pass workflow dynamically.

For example: Instead of default workflow, i would like to load a custom workflow when i visit comfyui endpoint. On each visit workflow can be different.

[olive yoke]

what is your goal?

[haughty orbit]

just working on preset based session app. User can select and load any workflow in the editor area in comfyui

[spring jungle]

is there a way to combind all these output in to a node so i dont have reconnect them after making edit xD

like empty node xD

oh nvm i found node

[grim moon]

I'm getting this error while executing Instant-ID with insightface. It generates but I wonder if there is anything wrong with it or it could be improved.

[cedar mirage]

Did you installed a onnx runtime with cuda support?

[grim moon]

no, only pip install insightface

I have an AMD gpu, RX 6700, I think cuda thinks are not allowed (!?)

[lusty plinth]

Hello, how do I update ComfyUI and embedded python? Using portable version.

[silver nebula]

Edit .bat file, add git pull

[spring scarab]

I'm looking to create a workflow, and need a way to... per prompt a prompt. an example is i want to have a character, (with the prompt for it's creation, in a prompt.) and feed into it things like a smile or a frown, or other emotions that can be displayed visually. anyone know how i might accomplish this

[lusty plinth]

Does ComfyUI support newest Python aka Python 3.12?

[olive yoke]

is this posted somewhere?

[olive yoke]

what is your goal?

as far as i've tested it does work, but what are you trying to do?

[lusty plinth]

Just to have up-to-date my ComfyUI instance. Also hoping to gain a bit better speeds in generating.

[spring jungle]

question anyone know if there avd ksampler adetailer node xD

[haughty orbit]

not yet, we are still developing mvp. We have found the solution by changing some of the comfy code. We can discuss more over dm if needed about the project.

[quick birch]

Latent size :
For some reason it looks like to have the best results it is better to have a square for many models. Depending on the models and the size of the square sometimes you can have a full portrait or face portrait for the same request and sometimes for the same seed, is there a rule we can follow or is it a guess? I know SDXL have some reference value as SD but even following this reference the results are not always good.

[grand pike]

hi lovely people. I'm starting to get frustrated with my potato PCs long generation times. If i'm hooked up to the internet while running the exe locally, is there anyway to get some kind of speed boost? Or is it completely offline?

[olive yoke]

yes, always use square

you can request to generate images you can crop later

[olive yoke]

it is completely offline. how much would you pay to accelerate things remotely?

[grand pike]

Dang, even when connected to the internet? And I would hypothetically pay if I weren't such a broke bitch

[olive yoke]

yeah that's tough

[eager escarp]

well, some malware could help you out... if you installed enough viruses, maybe someone would steal your credit card and accidentally keep the billing address as the mailing address when buying a 4090

[graceful pewter]

Any got a good regional prompting or attention coupling workflow json??

[eager escarp]

i had a good regional one somewhere

[graceful pewter]

Keen 🙂

[wispy kelp]

where do i put these ELLA models?

If anyone could spell it out for me exactly whats the model i am missing and exctly where do i have to place it

[wispy kelp]

what does this mean? how does google fit in that structuire?

[wispy kelp]

ok figured it out but why only Euler works with ella

[grand pike]

@eager escarp I'll keep my fingers crossed lol

[wispy kelp]

Looks at Pixart Sigma

[silver nebula]

My Pixart-Sigma setup

[wispy kelp]

is it just me or comfyui is faster than even fooces dare i say

[timid plume]

Hello , every body. I am new to here.

Can I ask about how to fix such kind of hands ? I have tried MeshGraphormer Hand Refiner with ControlNet , it looks good for simple hands, but for such kinds of complecated hands, this solution doesn't work well.

this is the screenshot of part of my workflow.

[hoary zinc]

I've long given up on these kinds of solutions, they never work or the hand looks out of picture. I use Krita + AI plugin.

you just have to mark the area you want to change followed by a new layers you'll be painting a hand however you want it by picking a colour selector and taking the colour of the skin

then just paint it and AI will generate it, you can also help it with adding works such as "handshake" etc.

you can also choose a model which will be utilized by the AI plugin to correct the picture's marked area

[timid plume]

Thank you , but I don't quit understand, do you mean this ? https://github.com/Acly/krita-ai-diffusion

[quick birch]

It seems that comfyUI nodes have cache memory and this memory is not always flushed. Here is what happened
In the same workflow I copy two identical workflows from two different workflows. It is a kind of impainting (set latent to noise mask).
When I run comfyUI, the mask is only applied to one of them. After various tests, deleting the ksampler and replacing it with another one solves the problem. This is not the first time where I see node keeping old references in memory and doesn't give the right result.
Does anyone experiment with this kind of problem?

[hoary zinc]

https://www.youtube.com/watch?v=sCtRgc21CjA

yes

[timid plume]

Thank you so much !

[grand pike]

hey ya'll know how when you want to make a random list prompt like {thing|stuff|cookies}? Does anyone else have the problem where the prompt cursor defaults to the bottom of the text box, as soon as you type in a } symbol?

anyway to fix that? Makes it really difficult to edit prompts while the random brackets are activated

a lot of people in this thread reporting the same bug

https://github.com/comfyanonymous/ComfyUI/issues/3073

[steep marlin]

Sweet, he added a split sigma denoise node! Makes unsample/resample workflows easier when you're using custom sigma scheduling that doesn't have a handy denoise percentage built into them. Saves a few nodes that you'd need to calculate the percentage by hand

[wispy kelp]

how do i add a lora to a workflow

its so confusing, what do plug into what

in forge you just add a lora, bam... takes 2 seconds

[steep marlin]

Add in a lora model node between your model loader and ksampler. There are two versions: model only and the other one that gives separate control over model and clip strengths

If you use the latter, connect the clip from the load checkpoint node to it. Then connect the lora node's clip to your text encoders

Not all loras are trained to do both. A lot of sdxl loras are trained on model only, so the clip percentage slider won't do anything for those

[hoary zinc]

or you can use rgthree node called "power prompt simple" to automatically add lora into you prompt, whatever you prefer

[unreal reef]

Is that what causes this issue? I've been having the problem for a while now and have resorted to writing the prompt in Notepad so I can super-fast delete and paste.

[steep marlin]

Probably some kind of browser issue not liking something about the JavaScript. That or there's some incompatibility with another node. Someone else was having that same issue a while back with a plugin I was having them test, but I couldn't replicate the issue

[sly sundial]

I had the problem, but only when a specific node was in the workflow. I think it was something like "Superprompt"

[steep marlin]

It was you lol

[sly sundial]

It was! 😄

[unreal reef]

I recall you suggesting I remove that node and then I realized I don't have it. 😄

[unreal reef]

The weird thing is that it'll be just fine and then out of nowhere it'll start acting up without my having added any new nodes.

[sly sundial]

Strange. I've never had it again since I removed the node anyway.

[unreal reef]

My office just got a machine with a 4090 in it for doing AI stuff with our software... I wonder if they'd let me use it for gens in the off hours. 😄

[grand pike]

@unreal reef yeah it's a real pain in the dick. Hopefully they can fix it soon? Also I love your username lol

[unreal reef]

Haha, thanks - it's not original. It was from a comedy video (TikTok I think?) - "When your wife discovers your gamer tag." 😄

[grand pike]

A good wife would give you a high five I think

"Yo that shit is funny bro" - wife

[unreal reef]

Found it. 😄 https://www.tiktok.com/@marriedwithflaws/video/7055756213371407662

[sick badger]

same 🙂 like others...just thought it was my browser being stupid
SDXL Prompt Styler node does it

[midnight maple]

3-4x faster upscaling inside comfyui using tensorrt
https://github.com/yuvraj108c/ComfyUI-Upscaler-Tensorrt

[uneven stratus]

Heyy, Anyone having this issue with SUPIR Conditionner ? https://github.com/kijai/ComfyUI-SUPIR/issues/72

[lusty plinth]

How to update pytorch in ComfyUI?

[steep marlin]

Why do you need to update it? It's risky and can sometimes end up bricking an installation

[split gate]

using comfyUI, is there a way to download a workflow and also download the missing checkpoints and VAE's with it? i see the "install missing custom nodes" but that doesnt install the checkpoints.

[lusty pewter]

In the manager where you can install nodes, you could just search for the word "civitai". Doing that would net you some results like this, which might solve what you're looking for:
https://github.com/civitai/civitai_comfy_nodes

[lusty plinth]

To get newest and most update software to proooomt.
If I brick my install, it is my problem.
Tell me how to update it NOW.

[wispy kelp]

hooking up these nodes in creative ways and making it look clean gets addictive

it's like designing your circuit board

or something

[distant pendant]

if installed using pip - just nuke the env and start over, if mamba/conda do mamba update --all -c conda-forge -c pytorch -c nvidia, if portable release just download a new one

[sly sundial]

Too demanding, find it yourself!

[forest birch]

https://github.com/FizzleDorf/ComfyUI-AIT Is there another acceleration like this for comfy that is moreup to date? As it's almost been a year since it got updated, and with an older comfy, it works and is quite a speedup. Downside, you need a outdated comfy

[sly sundial]

New feature:

New shortcuts for canvas zoom in/out.

alt + + : zoom in
alt + - : zoom out
ctrl + shift + LMB + vertical drag: zoom in(up) + out(down)

[olive yoke]

i can investigate this to add into my LTS comfyui fork

i don't see where it hides the work it does though

it looks incomplete. it shoudl be using fx2ait to convert the model on demand, and cache it somehow

[olive yoke]

what workflow are you using?

[forest birch]

It basically accelerates generations. don't know how it does it in detail, but with it, my gens sped up by 46% higher IT's with a 3090

[lusty plinth]

We need more accelerators for ComfyUI!

[forest birch]

Aye. SD forge optimized the heck outta the webui, so we really need to get a few boosters for comfy as well

[worldly cliff]

Hi, How do i fix this issue, i am getting this error suddenly in last few days

Token indices sequence length is longer than the specified maximum sequence length for this model (81 > 77). Running this sequence through the model will result in indexing errors

using default sdxl model , latest comfyui only

[olive yoke]

this is kind of vague

[worldly cliff]

that is the only error, its running the output fine etc, but this always pops up

[olive yoke]

what are you trying to do? don't be vague

[worldly cliff]

i am running a prompt and got that error, let me send the full log . 1 min

[olive yoke]

is it a community custom workflow that uses a lot of community custom nodes?

[worldly cliff]

no, its my own workflow, using the smz Clip encoder, to replicate outputs similar to A111

[olive yoke]

why are you using a buggy clip encoder?

"replicate" and "similar" - what are you trying to say?

do you need the exact same outputs, or not?

and if so why?

[worldly cliff]

well, the left is from A1111, right with Comfy UI, using same exact settings

using img2img

i want the comfy output to be same as A1111, because quality is getting affected otherwise

model_type EPS
Using xformers attention in VAE
Using xformers attention in VAE
clip missing: ['clip_l.logit_scale', 'clip_l.transformer.text_projection.weight']
Requested to load SDXLClipModel
Loading 1 new model
Token indices sequence length is longer than the specified maximum sequence length for this model (85 > 77). Running this sequence through the model will result in indexing errors
Token indices sequence length is longer than the specified maximum sequence length for this model (85 > 77). Running this sequence through the model will result in indexing errors
leftover keys: dict_keys(['down_blocks.1.attentions.0.norm.bias', 'down_blocks.1.attentions.0.norm.weight', 'down_blocks.1.attentions.0.proj_in.bias', 'down_blocks.1.attentions.0.proj_in.weight', 'down_blocks.1.attentions.0.proj_out.bias', 'down_blocks.1.attentions.0.proj_out.weight', 'down_blocks.1.attentions.1.norm.bias', 'down_blocks.1.attentions.1.norm.weight', 'down_blocks.1.attentions.1.proj_in.bias', 'down_blocks.1.attentions.1.proj_in.weight', 'down_blocks.1.attentions.1.proj_out.bias', 'down_blocks.1.attentions.1.proj_out.weight'])
Requested to load AutoencoderKL
Loading 1 new model
Requested to load SDXL
Requested to load ControlNet
Loading 2 new models
loading in lowvram mode 3084.1030645370483

This is the full log btw

its generating the output but i am guessing its not getting the full prompt

[olive yoke]

okay. you have to carefully use hte same settings. are you following a guide? the right has a different sampler and CFG. the left doesn't look like it's all vanilla models.

CLIP is the least of your concerns

you are using a controlnet

are you asking how to get higher contrast output?

[worldly cliff]

no, i am wondering how to fix the issue:

Token indices sequence length is longer than the specified maximum sequence length for this model (85 > 77). Running this sequence through the model will result in indexing errors
Token indices sequence length is longer than the specified maximum sequence length for this model (85 > 77). Running this sequence through the model will result in indexing errors

i got it to generate the same output using the smz node

but the problem is its giving this error

[olive yoke]

i guess you will have to ticket the author of that node

on the appropriate github repo

[worldly cliff]

the reason being i tried other nodes that emulate A1111 type parsing but they arent really reproducing the output, but this custom encoder was able to do it, guess will ping the author then

[olive yoke]

i think this is a colossal waste of time though you should learn how to use comfy as is

and not really worry about a1111

or reproducing results. in 10 minutes of fiddling you will get an aesthetically pleasing result

i am telling you now that the error message isn't going to be related to your issue

[worldly cliff]

the reproducability issue is resolved, left image is the output with the SMZ node, its getting the texture, contrast, lighting everything perfectly ( the slight variation is because i am using Euler A, so it wont proceed exact same image)

But, this works only with the SMZ node. On the other hand, with efficient loader (set to A1111 mode), right side is the result

and this brings me back to that error related to the clip since SMZ node is crucial to this workflow

[cedar mirage]

The message is more an information then a real error. Did you tried a shortet Prompt? I guess the SMZ node is configured that the prompt should not exceed 77 Tokens. The inputs seems higher. As it still works i am sure it will simply cut if off

[worldly cliff]

yes it is getting cut off, because the error doesnt happen with other encoders like efficient loader or the default clip encoders

[olive yoke]

do you have some programming education?

[worldly cliff]

i am seeing how to contact the author to see if there is a resolution

[olive yoke]

i guess why aren't you just using a1111 then?

[worldly cliff]

yes, but not too advanced

[worldly cliff]

because, its leading to color change in output at high denoise, idk for what reason, that issue has no solution, while comfy doesnt have that problem

[olive yoke]

why is it important that the result is reproduced exactly?

what are you trying to make?

[worldly cliff]

need to process close to 1000 files for a studio

[olive yoke]

do you mean a game studio or?

[worldly cliff]

and they have lota aesthetic requirements

yes

[olive yoke]

what exactly are you doing? upscaling something?

when you say process

[worldly cliff]

no , changing the style

[olive yoke]

okay

let's try to be as not vague as possible

be really specific with what you are doing

it would take you like 15m to set this up correctly in comfyui with vanilla nodes, with a good aesthetic

stay focused on answering my question though

what are you trying to do? i am asking for your goals

[worldly cliff]

basically its taking input image and giving a 3d cartoon look, this workflow is all figured out nicely, except
A) in Auto 1111, there is issues with color at high denoise, see this post: #💬｜general-chat message

B) so recreated the workflow in comfyui and this issue is gone, but now i am facing the issue of prompts getting truncated because of the custom node

[olive yoke]

you authored 1,000 prompts?

what are they?

[worldly cliff]

its from a studio, i didnt author them