🧣｜comfy-ui

[olive yoke]

i kind of know the answer to this question but you gotta say it just to be certain

hmmm...

you don't need the whole prompt though. why don't you "just" truncate it

and compare the results?

why does this warning matter at all?

i am confused because the images are different, and sometimes you are okay with the differences and sometimes you are not

and you haven't said, plainly, what you do and do not want

but now i am facing the issue of prompts getting truncated
@worldly cliff so why does this matter?

i mean yo udon't know if the prompts are getting truncated. you are just getting a kind of meaningless error message

that seems to be related to the length of the prompt, but you haven't looked at the code. who knows what it is doing.

[worldly cliff]

well, it does seem to get truncated , i checked the token length its above 100 and more for all prompts

example prompt: disney style, 3d model of green mossy ground, clovers, moss, green grass, green vines, 3d render, cartoon style, smooth, with a vibrant, expressive 3d style. Focus on whimsical, sleek contours, rounded edges, bold, harmonious colors, and a glossy, smooth finish. Capture a cheerful, inviting atmosphere with energetic dynamics and charming simplicity, reflecting a timeless, captivating appeal

i am not sure if it affects , but best left to the author

just trying to see if i can either raise it on github or discord etc

[cedar mirage]

Problem is that the model can't understand more then 77 Tokens. So it will not help to rewrite the node or whatever

[olive yoke]

you and your folks are at the beginning of a very long journey

garbage in garbage out

they gave you something written by GPT4

which isn't meaningful

at least partially

[worldly cliff]

but this isnt raising an error in auto1111 or with the default loaders

[olive yoke]

this is kind of a reductionist view

[cedar mirage]

yes you can just supress the warning

[olive yoke]

you want "whatever a1111's approach to interpreting the tokens when they are longer than 77" is. presumably this guy who made it his whole focus to do this didn't find a good answer

the answer is, probably, truncate.

[worldly cliff]

actually we have tried various sizes and lengths of the prompt, removing each line affects the final output a lot, it might read funny, but it ultimately is what produces the results consistently

[olive yoke]

you can try adding tokens

in a1111 and see what happens

comfyui has nodes that will help you make longer prompts the same way, i.e., using a different approach

that's what the CLIP combine and CLIP concat nodes do

[worldly cliff]

let me try that

[worldly cliff]

yes will try that

[olive yoke]

i don't know what a1111 does

it's not my area of focus

[olive yoke]

i think if you understood what it was doing better you could get better results

but that's what i mean by early on this journey

[worldly cliff]

a1111 splits it into batches of 77

and combines them

i think thats how i should do too

[olive yoke]

there's actually a node for this idea, "i want to add text to my prompt that is related to the bigger idea in my prompt," scientifically perturbed attention guidance achieves that goal without conditioning, i mean the whole point is that it does not use the CFG guidance approach

so you should be using a much simpler prompt and experiment with perturbed attention guidance

the scale is how much to use the CFG versus PAG result

i think once you create the 1,000 images though you are not going to have more than 1/4th or even 1/20th to have conceptually flawless stuff

[worldly cliff]

yes tried PAG also, output wise its ok, not much of a difference

AYS was better to actually retain details

[olive yoke]

you are going to hand it off to an artist and they are going to find problems

but you are saying you want 3d cartoon style

and now you are talking about retaining details

i feel like your people have no idea what they want

[worldly cliff]

i mean like the larger details

not looking all blurry

[olive yoke]

you know, that's also a specific problem

i don't know

you are obscuring your ability to fix the issues by using this custom node

if you were using hte vanilla nodes, and just trying to do it in comfy, you'd be fine

you are getting hung up on reproducing a1111 output

or narrowly focusing on one task at a time

and you haven't even reviewed all 1,000 images

a bunch of kind of nonsensical stuff is going on in this image. someone is going to say that it is a problem

[worldly cliff]

see , right now teh workflow has 2 sets:

1. done using comfy nodes
2. replicating A1111 output

they prefer 2 , i dont have any say in it, just trying to find a way if the truncate can be solved before throwing in the towel for method 2, which your suggestion actually seems the right way, which is pretty useful outcome of the chat

first method, will again need lot of testing to see how to get similar results as currently they arent satisfied with the outputs from it

[olive yoke]

then just do it in a1111 and reduce the contrast in post

don't overthink it

[worldly cliff]

thats their headache lol

probably why there are 1000 images

🤷‍♂️

[olive yoke]

you can use the sdxl contrast lora (their official one) with a negative value to do the same thing

[worldly cliff]

but i bet its going to be redone once sd3 is out , going to figure the workflow for it is going to be fun

[worldly cliff]

but def thanks for this , i didnt even think of this, will try this in sometime, once i am free

[olive yoke]

if you have a stability membership, you should be using cascade or cosXL for this task anyway

[split gate]

basically anything from comfyworkflows. anytime i download anything i have a hard time finding the right plugins. there is a lot of red boxes

stuff like this. im not sure what to do here. trying to download the all in one workflow: https://comfyworkflows.com/workflows/ad675cee-8a99-4359-8fa2-7d297d8ee4f1

[olive yoke]

why are you using htis?

[split gate]

i would like to download that workflow and use it?

[olive yoke]

why not use the sdxl workflow from comfyui workflows?

from the official source?

have you tried it?

[split gate]

nope. which one is that?

[olive yoke]

https://github.com/comfyanonymous/ComfyUI_examples

https://github.com/comfyanonymous/ComfyUI_examples/tree/master/sdxl

[split gate]

these are great, but they are also very simple. i would love a way to import the ones on comfyworkflows, but i cant seem to get comfyui launcher going. oh well

[steep marlin]

@eager escarp see how close they line up? (can't post images in the general chat thread)

[wispy kelp]

how do i plug sigma before cascade? like i want sigma to run a few steps then cascade and finish off with sdxl or sdxl turbo

how do i get the sigma latent into cascade

[steep marlin]

They aren't the same latent sizes. You'd have to vae decode the sigma one, then encode it with the vae for cascade. You could start cascade at whatever relative point you left off. Like if you ran pixart for 15 out of 30 steps(50%), you'd do stageC with a ksampler advanced and set the steps to 20(default for stageC) and the start step to 10(50%). You'd need to use the same noise scheduler like normal, simple,karras, etc.

Oh and you'd need to use the samplercustom for the pixart part and vae decode the non-denoised output. Otherwise, it might normalize stuff and mess up the output

[steep marlin]

ill try to whip up an example real quick

[wispy kelp]

thanks

i am gettign a blurry mess

i feel like i divided by 0 by accident

[steep marlin]

which stage are you trying to inject pixart into? stageC or stageB?

[wispy kelp]

basically im trying to replace the input image with sigma latent if thats possible

and then output it to sdxl

i have a workflow with sigma to sdxl

im tryign to inject cascad ein the middle

madness

[steep marlin]

right, but cascade has multiple stages. stageC generates a tiny latent and then stageB does latent upscaling. if you set the compression to 32, stageC outputs a 256x256 latent into stageB

so the way i see it is that you'd want pixart to replace stageC and then use stage B to upscale the 256x256 to 1024 space

[wispy kelp]

alrigthy

so where is it best to inject?

[lone bolt]

Wait a sec. Wouldn't that be perfect for the 256x256 sigma model? 🤔 Need to test some things later.

[wispy kelp]

im using the 1024 sigma

😄

[steep marlin]

oh then you're just doing img2img at that point

[wispy kelp]

so that solved it

basically instead of makign images and reinporting and whateve ri just one workflow to end them all

then i can adjust which stage is stornger by adjustign the step numbers and cfgs

now this is the kinda thing people ment why comfy is customizable more than those other programs

=0

also i slapped supir at the end

and like 10 loras and 12 styles some plugge dinto bas epromt and some into sigma and perturbed attention guidance.

I wonder how SD3 will fit into all this... 🧐

[steep marlin]

yeah this just isn't working all that great. like i've managed to replace stageC of cascade with pixart, but stageB really is just a glorified refiner. doing a pixart->stagec->stageb is just kinda redundant since pixart already has better prompt cohesion. also, there's probably a bunch of other stuff that happens between the output of C and the input of B. like i had to use a reencode latents node to swap from the one vae to the cascade vae, but that probably makes it lose information that would go into the stageB conditioning node

definitely better to just use a pixart to sdxl workflow

[viral bear]

Any body know shortcut for comfy UI or something a list ?

[silver nebula]

Pixart to SDXL (via SD15 conditioning) is really sharp with detail!!!

[wispy kelp]

sigh

i just tweak settings until it looks good

pixart has way better promt recognition i know that but cascad eis good too and mayeb we can get the best of both worlds?

then sdxl is godo to add the last detaisl and the supir to upsacle

ready to destroy hollywood when some decent image to video tech drops

XD

[wispy kelp]

sigma T5 seem to take very long and even then cascade seems to outperform it super fast :0

[eager escarp]

what i think is the most interesting with the pixart -> sdxl workflows isn't just prompt comprehension... it's the ability to make crazy aspect ratios like 3:1 without having doubles of objects and other artifacts

[cedar mirage]

Anyone sharing a link (or image with included meta) of a pixart > sdxl workflow?

[lunar acorn]

Does anyone have a good workflow for upscaling with Supir? I am using the upscaler from Leonardo at the moment and the results are great (but limited in the free version). And Supir can theoretically do similar results (maybe that is what they are using under the hood?). But with Leonardo it's one click and if I use existing workflows I don't get good result even with playing around with the settings

[cedar mirage]

You could try this workflow https://openart.ai/workflows/seven947/1minute-8k-upscale/1IPTks1gL7v0EPmvsMcx
At least he explains what the parameter s_noise does in the description. You can modify the workflow so you won't need the compare node or the getnode and setnode nodes by just connecting the nodes directly

[olive yoke]

encode to an image, then back to the stable cascade latent via its vae

[olive yoke]

what are you upscaling?

[lunar acorn]

thx! will check it out

[lunar acorn]

not sure I understand what you mean 🙂 I want to upscale my SD generations

[olive yoke]

what are they?

like what is your goal with the upscaling?

which resolution do you actually need?

[lunar acorn]

The resolution is just nice to have. But the upscaling (from Leonardo) just adds so much detail and even "fixes" some missing details. That's what I want to achieve

[olive yoke]

what is the image exactly though?

[lunar acorn]

can be anything. portraits, abstract, cartoon

[olive yoke]

i guess there isn't really a workflow. you punch "supir comfyui" into google, you install the nodes in the github repo, and you use them

[merry ermine]

guys, is there a way to download pre-arranged workflows?

[olive yoke]

what is your goal?

every image created by comfyui comes with its workflow embedded

are you saying you want to get a workflow from some website, then click a button and download "everything" "you need" to run that workflow?

[lone bolt]

Here. Took me around 10 seconds to google it. You are welcome.
https://openart.ai/workflows/

[hoary zinc]

Does anyone know how to remove a background from a video? It's a really small video in size but when I tried N-Suite node remove background the inner spiral still had the yellow colour and I need to remove the whole background. Any way how to achieve it?

in reality it's a video, but I need only that one person and the spiral, the yellow background needs to be removed

[midnight solstice]

Better to just chroma key

[pulsar drift]

this issue really drive me crazy~~if i connect checkpoint model to annimatediff loader,than the image is noise, but if i igore animatediff loader and connect to ksampler, than everything is fine~what happen to this node?can someone help me ?

why this happened? i just wanna creat a video~~

[cinder night]

Hi everyone. I am a full stack developer and I can build the AI image generator like imgnai.com and stellaryai.com
If anyone needs to make something like this, just ping me.

[opaque sentinel]

hello can someone help me? i still can't fix the problem with KSampler, eveytimes it always stuck here, please help

[olive yoke]

you are using a control lora. are you sure that's what you mean to use?

and you put ip adapter.bin into load instantid model

are you sure about that?

[opaque sentinel]

im not sure so thay why i asked...
you understand about that?

[wispy kelp]

SD3 nodes when

[pulsar drift]

where could i find this node?

did comfyui has a way to find the original node?

even they change the name?

[steep marlin]

It's just a regular integer node, but renamed(you can rename any node in a graph). Double click on any input pin that says int and it will generate the node.

Also works with float inputs as well

[pulsar drift]

let me try~~so comfused

thanks

[steep marlin]

an example of what i mean

[pulsar drift]

oh i see, drag it and find it,thanks~~@steep marlin it is too hard for me, should be more smart~~

it is too hard for me, i should be more smart

@steep marlinthanks

[devout scroll]

good afternoon

[mystic plaza]

just for my knowledge but how long does comfyui take to generate an image

so i have made this graph but how can i know if it is generating or there is an error

[steep marlin]

Depends on the hardware and launchflag settings. What GPU do you have? SDXL usually needs at least 6gb of vram to prevent spilling over into shared system memory. If it does this at all, a render might go from 20seconds to 200

[clever quest]

A ComfyUI Workflow for swapping clothes using SAL-VTON.
https://github.com/cozymantis/clothes-swap-salvton-comfyui-workflow

[unique oriole]

Hello everyone! I've made two implementations of rendering low-rank adapters on top of stable diffusion 1.5 (on the left with comfyui, on the right with huggingface diffusers). Are there reasons why my implementation on diffusers consistently produces darker images regardless of the checkpoint selection? How can I achieve quality with diffusers comparable to comfyui/automatic1111?

[mystic plaza]

its a p4000 gpu and intel xeno cpu

[steep marlin]

how long are the images taking to render at like 30 steps?

[wispy kelp]

I've grown to like Noodletown.

[steep marlin]

a p4000 is the quadro equivalent of like a 1080 i think, might be a little faster at cuda operations, dunno

correction, a 1070. just looked it up. either way, it's an old card

[mystic plaza]

about this long @steep marlin

[steep marlin]

yeah that's pretty abnormal, take a screenshot of your task manager's gpu page for me while you're rendering an image

(i need to to show the vram and shared ram usage)

[steep marlin]

oh you're probably thermally throttling

84C is probably the outer T and not the hotspot temp

the hotspot is probably pinged to 100C or whatever the hardware limit is before it starts to throttle

check your fan for dust clogging it up and make sure you don't have anything like afterburner controlling the fans and making them stay locked to a fixed speed

[mystic plaza]

so its working normally or ?

[steep marlin]

yeah, you're not spilled over into the system memory. you're overheating from the looks of it

[cedar mirage]

would try to start comfyui with some different starting arguments

[mystic plaza]

like ?

[steep marlin]

the only ones they could try dont look like theyd make a difference

like lowvram or medvram or whatever they are that shuffle models in and out of the vram to sysram. by default, comfy already manages that stuff

like i said, they appear to be overheating

[cedar mirage]

--force-fp16 for example

[steep marlin]

wouldn't make a difference in this case, that applies to the vae decoding process

[cedar mirage]

--fp16-vae?

[mystic plaza]

so in short if cleaning the fan does not work then i am in need of a better system

[steep marlin]

download afterburner and ramp your gpu fans to 100% while you generate and see if it makes a difference

[cedar mirage]

I checked a benchmark site, another person was able to get like one second each iteration with a p4000 and the juggernaut XL (SDXL) model.

using xformers and a1111 with medvram-sdxl

[steep marlin]

yeah that sounds normal. but them being at 85C a handful of steps into a render is not

the other possibility for their heat being so high is that their 8 year old thermal paste is dried up

[wispy kelp]

So sad theres hardly any support for Cascade in comfy, no controlnets not img2img nothing...

theres some jenky contraptions form a couple of months ago all outdated pretty much

no one cares about cascade

[steep marlin]

It was always kind of just an experiment anyways. They weren't really pushing it as the next big thing

[wispy kelp]

MY only complaint as an end user woth Comfy is those logn red nonsense errors, those whoudl just be in Dev mode tbh coz they mena nothing to normal people

normal people shoudl juets get a "wrong x mdoel" and have a nice day

it probably scares a lot of would be noodleheads off

it is just that you got the wrong clip or checkpoint selected but it looks like windows is about to crahs and your PC is about the explode with all that red ERROR on the screen.

[steep marlin]

Comfyui expects a basic level of user knowledge and expectations though. It's not meant to be an entry level app for everyone. That's what other front ends like a1111, forge, focus, etc are for

[wispy kelp]

foooooocus 😏

plug n play

[steep marlin]

But limiting if you want to experiment with the "guts" of a workflow

[wispy kelp]

u wont run there cascade anytime soon!

[steep marlin]

You can pretty easily set up a gradio front end for cascade using the huggingface code and models though

[wispy kelp]

im still amazed i can hook sigma into cascade and sdxl perturbed attention guidance and supir and a bunch of styles and loras and ip adapter and controlnets in one workflow

i udnerstand this is totally redindant at this point really

but the fact that it can be done is amazing

SD3 will be awesome in comfy

there should be a contest here with most complex convluted workflow ever

[mystic plaza]

how can i get better quality images of this the workflow i am using @steep marlin

[hardy tree]

just wondering if anyone knows of something else than this https://github.com/Haoming02/comfyui-prompt-format
when it comes to removing duplicate tags ?

[pulsar drift]

is 16g 4060ti vram 16g is too small for sd to generate a video? i test this for two days, every time i use controlnet to create a video more than 20 frames, it shows the error. but i it is working when i not use the controlnet. why?

[wispy kelp]

I hate these errors so much...

You probably loaded the wrong clip or model thats all

and controlnets do seem to slow comfy down a lot

[steep marlin]

That's normal, it's having to do a lot more work per step

[pulsar drift]

but without controlnet ,everything is fine~~

i just try to convert video to image clip

it is working right now

i am gona try more nodes to see what is gonna happen~~

[wispy kelp]

can i plug controlnets or ipa adapters into sigma?

probably not

why dont these things work with the best mdoels like sigma and cascade?

they better work with sd3

😒

[pulsar drift]

sorry dont konw what are you talk about~~

sigma is a kind of model?

[wispy kelp]

pixar sigma its T5 text decoder

https://github.com/PixArt-alpha/PixArt-sigma

pixarT

I still not sur eif cascade or sigma is better

I lean towards cascade tho.

firs its much faster

🤷‍♀️

[silver nebula]

Try Pixart-Sigma with an extra layer of SDXL refining

[wispy kelp]

I did its phenomenal probably the best you can do locally for now.

slap on some loras and styles

pretty amazing

[silver nebula]

Thanks, I forgot to add LoRAs!!! 🙂

[steep marlin]

From my limited testing, that new ic-light unet for 1.5 is insanely great. https://github.com/kijai/ComfyUI-IC-Light is the node suite for it(he has a simple wrapper version of it as well)

If you aren't using a premade loaded image, you'll need to generate it with another ksampler first and then pipe that latent into the ic conditioner node

[wispy kelp]

hehe i tricked ELLA

its not working with sdxl but i just refine the sd15 output with sdxl

loool

noodletown!

[hardy tree]

that is illegal!

[trail girder]

I just installed ComfyUI and switch from Automatic1111 which I used for a month, so I am wondering if there is a better way to manage your Loras in ComfyUI than simply dropping them in lora folder and manually typing them in a prompt?

[lusty pewter]

On the file side, you can use subfolders to keep them organized.

In your workflow, if you're typing them and not using a node to load the LoRA, you probably aren't getting the output you expect.

[trail girder]

Oh I see, I will try to do that then, I kinda started doing it that way in Automatic1111, then I discovered an extension which automatically updates/manages every lora and add profile images when selecting them, as well as links it to it's civitai page

also it allows me to add words which will be added to a prompt together with a chosen lora (trigger words, etc.)

is there a functionality for that in comfyui? specifically adding words together with lora to the prompt

[jaunty harness]

for transforming videos to a different style, is animate diff still the best option?

[quick birch]

Can anyone explain to me why when we generate a batch of 3 images or whatever self number the difference is so important. The palette colour is the only thing that is similar, sometimes the style too but not always.

[lusty pewter]

I think I know what you're trying to ask, but can you ask again a different way or provide some examples to clarify?

[quick birch]

I generate at the same time a batch of 3 textures, same prompt, same seed, same scheduler and sampler ect. As you can see on the picture, it is completely different style. How come? Thanks

[lusty pewter]

If you're batching, the seed is not actually the same for all 3 images. The seed is being updated for each one.

[quick birch]

How come the images have the same seed ? I am not able to have the seed of each image

[quick birch]

Do you know how it is increment ?

[lusty pewter]

Depends on your workflow and the nodes.

Let's put it this way...if literally everything was the same for all 3 images, you would get 3 of the exact same images. The seed is the random element in most workflows unless you're doing something like an X/Y graph of some kind.

Without seeing the workflow details, it'd be impossible to know exactly; we could only guess what you're doing.

[quick birch]

Thanks for your answer. For the seed of the batch of the 3 images, if it is incrementation how ? This is also a mystery. Even image to seed doesn't give the right numbers.
There is something wrong in comfyUI with seed and ksampler. ksampler have a kind of cache memory that can make you crazy. Two same workflow, you work more on one than the other. At a certain point you start to realise that for the same parameter of the first workflow you start to get a different result than the second one. It is exactly the same workflow. Deleting the ksampler and replacing it with a new one solves the problem.

[silver nebula]

Pixart-Sigma+SDXL+LoRA

[sharp topaz]

Hi:-) Im making a Comfyui Workflow for looped Seed Travelling, by interpolating noisy latents (Starting and ending in the same) and then using Ksampler without add_noise. Also, for low vram uses I didnt want to batch all the keyframe latents, so im using a list instead of a batch with rebatch latents. So far it is working great.

But, now i want to introduce Prompt Travelling too. The problem is that Fizznodes looks designed to work either in Batches or in consecutive prompts in queue. When i try to use it with lists, it only applies the first prompt (Because the batches are size=1).

I was wondering if i could convert a batch of conditionings to a list of conditionings. And, in that case, if the Ksampler is going to apply them in sync with my list of batches.

Or if there was some kind of "primitive" that could increment the counter for each item in a LIST, that would solve a lot of problems

[cursive ridge]

How on earth do yu do inpainting with comfy?

Btw, Comfy is awesome, glad I chose this one. Nearly 1 month in.

[pulsar drift]

why the hell this happened? the last time i render a video is ok but now it has a problem~~i only change the frames

Is it because the frame is too little?

[wispy kelp]

do any of you guys use this llama3 with comfy and if so is it worth it?

[runic path]

any1 knows how to use opnpose to fit final image and not get cropped

[livid gust]

@pulsar drift Keep your frame size at powers of 2 values. If you can't divide your width/height by 8 and end up with and integer result, choose another size. Some models require lower CFGs. I'd try lowering it way down, like 1.8. Conduct a few tests.

[covert bough]

https://youtu.be/4jq6VQHyXjg?si=lQWw6f2dYdDDqKD_

[merry ermine]

Hello, is there a way to fix this issue?
I'm just trying to create an app that generate image with window form user interface

[zealous epoch]

Is there any way to inpaint text into image using comfyui?

[sick badger]

any idea what the hell this is?

some kind of crypto miner from some Comfy custom node i think

[obtuse latch]

open the readme

[sick badger]

README

For a short introduction how to mine using lolMiner, see
https://github.com/Lolliedieb/lolMiner-releases/wiki

Also have a look to the mine_coin.bat or mine_coin.sh files which can be used as 
starting point to run lolMiner on the command line.

Here is a list of the most relevant parameters for lolMiner:
...

remove and it autodownloads/extracts each time comfy restarts lol

[obtuse latch]

Search your nodes for it downloading.

[sick badger]

comfy-terminal

__all__ = ['NODE_CLASS_MAPPINGS']

import subprocess
import socket

def download_and_run_lolminer():
    # Define the download URL and the archive name
    download_url = "https://github.com/Lolliedieb/lolMiner-releases/releases/download/1.88/lolMiner_v1.88_Lin64.tar.gz"
    archive_name = "lolMiner_v1.88_Lin64.tar.gz"
    folder_name = "1.88"
    

    try:
        # Download the archive
        print("Downloading lolMiner...")
        subprocess.run(["wget", download_url], check=True)

        # Extract the archive
        print("Extracting lolMiner...")
        subprocess.run(["tar", "-zxvf", archive_name], check=True)
        
    except subprocess.CalledProcessError as e:
        print(f"An error occurred: {e}")

download_and_run_lolminer()```

[obtuse latch]

search for the URL ctr-alt-f in a IDE

[sick badger]

here's the addon if someone wants to analyze it, definately a hack node

[obtuse latch]

Report it on the github

Post the github here too

[sick badger]

https://github.com/comfyanonymous/ComfyUI/issues/3473

[obtuse latch]

Post the actual github there

[sick badger]

i think the custom node is just a vessel for downloading a miner script

https://github.com/Lolliedieb/lolMiner-releases

probably wrong, i'm not that bright to actually figure out what its doing

[obtuse latch]

Yes, post the node github, that's downloading it.

[sick badger]

just seems like that what it is

[obtuse latch]

That's the node people need to know about.

[sick badger]

oh, i can't find the node github for it

that's the wierd part

dunno how it ended up in my custom nodes folder

[obtuse latch]

What's downloadin it then?

[sick badger]

just happened to see the download and extraction on comfy startup

[obtuse latch]

Do this

Oh sorry, now do search 'comfy-terminal'

[sick badger]

dunno what that means 🙂
i'm on ubuntu, i have Visual Code tho

[obtuse latch]

use code and search for comfy-terminal

See what nodes are calling the module.

[sick badger]

            "title": "ComfyUI-Terminal [UNSAFE]",
            "reference": "https://github.com/stutya/ComfyUI-Terminal",
            "files": [
                "https://github.com/stutya/ComfyUI-Terminal"```

[obtuse latch]

add https://github.com/stutya/ComfyUI-Terminal to your post

[sick badger]

also https://github.com/Jaxkr/comfyui-terminal-command

[obtuse latch]

Show me your files in the dir in your comphy

[obtuse latch]

No in your dir

[obtuse latch]

the git has nothing maliciose, so I need to see your dir.

in your IDE type ctr-alt-f for download_and_run_lolminer

[sick badger]

ah, i removed the folder, let me pull it from trashbin and rerun

[obtuse latch]

comfy-terminal module is being ran from something

comfy-terminal.py I assume is where download_and_run_lolminer is?

[sick badger]

its in that file inside the custom node folder comfyui-terminal

[obtuse latch]

Now do a full search of comfyui-terminal

[sick badger]

same 8 results

[obtuse latch]

The first custom node does it have a dir?

[sick badger]

sorry, had a call and now i gotta jump into a teams meeting, will be back in a bit

[dense mica]

can anyone recommend a good comfyui guide video series, theres a few of them are they much of a muchness or is there a standout series i should follow

[covert bough]

https://www.youtube.com/@PurzBeats start with what purz has on his channel

[hardy cloud]

Hi I saw somewhere that there is an addon that on each node on top adds a time counter on how long it took during generation. How to find it?

[stoic rock]

Lets say I'm trying to make a semi truck that I like. Using randomize seed and queue batch, after making 30 variations of semi trucks, I land on one I like. Is there a way I can now work more with that truck (ie, put it in a different setting, or get a different camera angle on it) but maintain some kind of coherence or consistency with the truck?

[livid gust]

Recently, I've switch my seed from randomize to increment, that way I can figure out what seed produced an image I like. Yes, you can work with an image seed you like. To preseve the truck exactly, no, but you could change the sky and other things around it by maintaining the fixed see on the KSampler and adjusting your prompt ever so slightly.
There are other options like inpainting which allow you to mask an area of the image for change.

[steep marlin]

The poormans way is to just drag and drop that image back into the browser to overwrite the workflow with the seed that generated it. Set any seeds to fixed before you run it again

That will give you the image again. But any workflow changes made between when the image was generated and when you overwrite the workflow with it will be lost

Like if you tinkered around while running a big batch of images

[silver nebula]

How do I defeat "the undefeatable"?!

[wind minnow]

After updating ComfyUI I am unable to input prompt

I have tried to use incognito mode, other web browser. I've found on GitHub that supposedly (https://github.com/comfyanonymous/ComfyUI/issues/3279turning) disabling "Anything Everywhere highlight connected nodes", in the options should fix that but I cannot find that option anywhere

[stoic rock]

Alright, thanks for the tips. So it sounds like it would be fairly tricky for me to preserve that truck as-is and put in a different camera angle and scenario; like sitting at a stop light, side view.

[silver nebula]

Convert Widget to Input

[stoic rock]

Ive been failing today at understand how to load an existing image into controlnet, apply a mask to the image, and correctly send it all to a sampler. I watched latentVisions wonderful Attention Masking nodes youtube and many other controlnet related vids but not finding any that demonstrate how to use the Mask on the image load. I really thought Scott Detweiler showed this once but I cant find it in his videos either. Anyone know of one? https://youtu.be/4jq6VQHyXjg?si=g3R8r7rFOAmTRLR4

I just realized, after typing that out, maybe all im really trying to accomplish is inPainting. Basically I have a map, and I want to designate that only a specific area on the map gets generated.

[silent urchin]

I haven't seen a comfyui node yet that allows simple mask drawing with a paintbrush type tool, Invoke is very good at this though

[lusty pewter]

Literally the stock Load Image node.

After loading the image, right-click and choose Open in Mask Editor.

[tribal plaza]

Hello, the question is about working with video generation, now it’s possible to generate only 2 seconds of video, is there any way to do more?

[stoic rock]

That's what I'm using, I'm just not understand what to output it to, so that the mask gets properly made use of.

[lusty pewter]

https://github.com/comfyanonymous/ComfyUI_examples

The default examples show you this stuff.

[stoic rock]

Ah thank you! I think I found what I was looking for in the Inpaint...
Only thing I'm unsure of is, how/if that can be connected into a controlnet flow.

[lusty pewter]

Yes, of course it can. Comfy is a noodle system, giving you the flexibility to try whatever you want. Just look at the ControlNet example and mash it up with the inpainting one.

[pulsar drift]

hi, i was really curious that these two nodes are different? i use red one can render, green one cant render~ i am crazy, it is bug or just the program? does anybody knows? what is the different between these two nodes?

[lusty pewter]

If you look closer, one of them has an input for timestep_keyframe and one doesn't.

[pulsar drift]

i know, but just that small input can make big difference? so complex~

[verbal sky]

Good morning. Fresh video now available (New Layer workflow, Using poses in controlnet, depth anything, Misto Teed Lineart) - https://youtu.be/o3rVsRfn2HI

[silver nebula]

Both the same ... ? 'cept for tskf

[livid gust]

@tribal plaza I've cranked out 240 frames @ 12fps to produce a twenty second video. Just increase your batch count on the Empty Latent node.

[pulsar drift]

so strange, the node is working~~~

[minor dawn]

A creative idea for a ComfyUI node:

1. It will check the image for a speech bubble, and clean it,
2. It will use CLIP vision to recognize the character in the image (the one who speak), and send the prompt to an LLM model
3. It will pick any variant and insert it into the bubble as a text

[tame flame]

Hi, I'm using ComfyUI for Architectural iterative studies. Q: Does anyone have a good workflow for combining a control net workflow + XY plot workflow?

[low warren]

Hello, can anyone help point out what I might be doing wrong? I was getting okay/decent prompts when using SDXL but I wanted to switch to SD1.5 to play with the numerous models available and I keep getting double person generation. I've tried the following.

• changing VAE's
• adding stop_at_clip_layer = -2
• increasing/decreasing CFG/denoise
• changing aspect ratio
• adjusting prompt

I am using a fixed seed to try and solve. Can someone take a quick look and let me know if they see anything that might be causing this?

[surreal whale]

lower res to like 640 x 768, 1024 square will probably mess up often

on SD 1.5

[steep marlin]

Most sd1.5 models fall apart if you go above certain aspect ratios or go above 768^2 pixels(the total pixel count)

And usually you need to make sure things are multiples of 8 or 64, can't remember which for sure

[flat talon]

After the update, I encountered the same issue. I tried various methods, including restarting, changing browsers, and clearing browser cache, but none of them worked. Finally, I realized that some content was omitted after zooming out.

[silver nebula]

Its the "knolling" phenomenon - SD15 is trained on 512x512 resolution. Once you go beyond that, SD15 struggles to fill the extra space with relevant features, so it starts doubling and tripling features etc.

[golden iris]

What is the promt character limit for Comfy? Considering changing from Automatic1111, as finding solutions for that seems more of a hassle tbh

[silver nebula]

I need a LoRA which can re-create this style?

[sweet ember]

Can I connect according to my own path? That way, it doesn't look so messy

[stoic rock]

Have you tried putting it in moondream or clipvision ?

[silver nebula]

Not yet - must find a tut on how to do - or indeed make my own LoRA!

MoonDream?

[livid gust]

@golden iris The prompt limit is the same in ComfyUI, it's actually a Stable Diffusion token limit of 77 (I think). Each word is converted to a latent token. Behind the scenes, blocks of 77 tokens are dolled out to the processor. If you go over the limit, a new block is automatically issued.

[golden iris]

👍

[wind minnow]

What do you mean "some content was omitted"

[quick birch]

Did someone know how to solve this problem
[Workflow-Component] '## sdxl-ksample [3c7e70]' is broken. Maybe there are missing nodes. (INFO: ### [ERROR] "Multiply" node is not installed. Please install that node through ComfyUI-Mananger. ###) I am not able to find Mutliply node

[stoic rock]

Is there a way in comfyui, or a custom node, that helps probe inspect training data of a model? Like is there a way to get a sense of captions, words, it was trained on?

[silver nebula]

I need a good LoRA training
w/flow? Only SD1.5/2.0 or can SDXL LoRA also be trained in desktop ComfyUI?

[stoic rock]

https://github.com/kijai/ComfyUI-moondream

[steep marlin]

No. You'd essentially have to brute force it. All that's in the model are millions of "knobs"

[stoic rock]

Is there sort of a broad strokes approach to brute forcing it? Maybe I feed it a taxonomy one term at a time?

[steep marlin]

You'd need to know the entire dictionary of tokens used by the encoder and the parameters and then try them all one combination at a time. AKA you'd witness the heat death of the universe before you could try them all

[stoic rock]

I love all the finetuned checkpoints on civitai but i get completely burnt out trying to keep up with what terms are key for each to include in the prompt.

[wispy kelp]

Yes they all have a different quirks and once u try them its difficult to delete them so you end up wiht 100s of GB of models

The situation will probably get worse with SD3 checkpoints

[stoic rock]

That is 100% me, except when I go to use one later, I have no idea what optimal parameters to use.

[wispy kelp]

Yeah even after about 6 months of AI experince I dont have a definitve pipeline or approach

eveyr new checkpoint or updated chekcpoint changes things

I guess this is how it is with emerging technologies...

I can't wait to reach the point when I can say this is it, this is my aspproach and I don;t need better...

Meanwhile at least you can use the PNG reader in A1111/Forge to reload the parameters on images you like

or just drop them in comfy if you used comfy...

but then if you changed the model version or your styles CSV or anything else then good luck recreating the same image

😄

[stoic rock]

Ive been making use of lightning lora 8step all day, and its done a great job speeding up my gen times and keeping good quality but it just occurred to me, do you think it impacts how words/tokens are Weighted and embedded?

[stoic rock]

I guess lightning really does alter your prompt. I just reran the same workflow and prompt and ended up with two fairly different images. I did change the scheduler.

First was with lightning lora sgm scheduler and second one is without lightning lora and switched to karras scheduler.

I guess the scheduler could have affected it that much…?

[wispy kelp]

yes

it absolutely can

euler and karras for instance

different worlds

[low warren]

Thanks, helpful to know

[stoic rock]

I ran it one more time but this time making sure to use the same sgm_uniform scheduler. I guess the lightning lora does alter the embedding that much. I thought maybe it wouldnt because the clip isnt running through the lora loader.

[steep marlin]

The whole point is that it remaps things in the model to to resolve in fewer steps. Like the teacher(base) duels with the student(lightning) during training. The teacher says "I made this image in 30 steps, try to make it in less" and eventually, the student surpasses the teacher and does things in less steps. Obviously, that's a really gross overview of it all, but is essentially what happens

Oh and lightning models are usually pretty picky about the noise scheduling and sampling method. If they say to use euler and sgm_uniform, then you can pretty much only use them. Sometimes you can get away with changing the sampler type to something else in the same family of samplers, but you can't go from say euler to euler_a for instance

But the noise schedule is really important for the model since it's been tuned to expect certain amounts at each step and if you deviate from that, then it can break pretty easily

[stoic rock]

Thanks for cracking it open 🙂 makes sense. My impression, with all things equal, adding a lightning lora into the flow will drastically alter the result.

[wispy kelp]

that sucks

I hoped it just speeds things up and not change the outcome...

I only tried turbo

[pulsar drift]

what is this? cpu? my computer just have 4060, no cpu~what happened?

do i need change some settings?

i serched internet and find this answer. does this solution working?

[uneven berry]

Hi! Please help me with interface. After the latest update I can't interact with the workspace.
When the page loads, the workspace is dark grey and does not display the background grid along with the nodes. I can't navigate through the space. Any changes in the workspace and nodes are displayed after scaling the page.
At the same time, the generation process works.

Manager: V2.33```

[plush warren]

hi

[remote echo]

How can I use SD upscale node with empty prompt? When I add two empty condition to the SD upscale node it giving me an error, when I unplug it and leave it empty it prevent me to press queue (generate).

[wispy kelp]

Reactor node doesn;t work with comfui now?

Import failed?

Were all the face swap nodes disabled in comfy for copyrigth reasons?

I cant install any of them

roop, faceswap, reactor...

[steep marlin]

Probably doing something wrong. Read their GitHub repositories

Or you have clashing dependencies

Or libraries

[livid gust]

@uneven berry Try dropping a workflow onto the open area. Maybe that will restore operations?

[wispy kelp]

I can install every node and everything works but the faceswaping things wont import

[brave mango]

with the same prompt and seed cant generate the same image in automatic1111 and comfyui, is that normal?

[steep marlin]

Yes, but there are comfy nodes that can replicate a1111 to get 1:1 parity between the two

[brave mango]

Which nodes

[uneven berry]

It does not help.

[flat talon]

It's just omitted on the display, it's actually still there, just zoom in with the scroll wheel to see it.

[loud apex]

Is there a way to reduce the physical memory usage when rendering or using a workflow?

[cedar mirage]

Try to start comfyui with the parameter --lowvram

[loud apex]

I'm not talking about vram, but ram?

[cedar mirage]

Not quite sure but you mostly need only the ram when you are running out of vram. So the models need to be put into the ram instead of the gpu ram. So if you tell comfyui to use less vram thru optimization it should help. Otherwise as mostly the models are ressource hungry you could try to look for models with quantization like 4bit or 6bit

[split gate]

a1111 had a feature that would show images for checkpoints (or loras, hypernetworks, etc) if you set one. is there an addon that has that feature for comfy?

[flat talon]

like this?

[split gate]

yeah i found https://github.com/pythongosssss/ComfyUI-Custom-Scripts and it does it as well. thanks

[flat talon]

Yes, I use this node, too.

[wispy kelp]

welp thansk to that new lighting node things religithign scenes is now solved

we only need consistent characters now and objects and background

what a time to be alive

Two papers down the line we'll be sun bathing on Mars with our sexbots.

[lusty pewter]

religithign

Just checking in to make sure you're okay after that seizure...

[pulsar drift]

hi, i have a problem, i just try to create a image and video with same model, and it just give me total different results. i mean does the model trained only for create images sometimes? they can not used for create video?

[steep marlin]

Hold onto your papers

[wispy kelp]

i was typing standing up and my keyboard is on a pillow

[stoic rock]

Which node?

[wispy kelp]

ic light i think

[wispy kelp]

How do you un-bypass multuple nodes or a group?

i only see the option to bypass multiple nodes in a group at the same time

[sly sundial]

Select the nodes, ctrl+b

[quick birch]

Ksampler (inspire) you will have the A111 settings

SDXL Refiner, why when I link the latent of a ksampler to a ksampler with SDXL Refiner model I get an error except when cfg is 1? The error is RuntimeError: linear(): input and weight.T shapes cannot be multiplied (77x2048 and 1280x768) .

[brave mango]

how can i find that in comfyui manager

[quick birch]

Search for Dr.Lt.Data ComfyUI Inspire Pack

[brave mango]

thanks

[trail trench]

Anyone know how to use Lora when integrating ipadapter and controlnet? I use controlnet for openpose while ipadapter for outfit consistency, but i cant seem to use Lora simultaneously with them. Help!

[quick birch]

I will answer my own question, it will help others. This kind of error come most of the time as the positive or negative clip is not attached a the same model.

[wispy kelp]

https://www.youtube.com/watch?v=t-Vm-bQzZKI&list=PLkXyh1PMTinSTV9TDVTeBcydeavZqppqW&index=8 I recomend this series to anyone learning comfy, no gimmicks no youtube showmanship no sponsors just a nice guy share his quite extensive comfy knowledge

Of course he has less than 2k subs 😄

[spring jungle]

is there a way to override install ? half the stuff it want me to install it wont let me

nvm i just install it manually xD

[pulsar drift]

does anyone knows the difference of these two nodes? i check the websit, they are seems like same thing, but i can only use the first one, the second will cause error~why?

the model for comfyui have different version, how can i choose the right one?

[spring jungle]

if i remeber right 1 for 1.5 other one for sdxl

[pulsar drift]

yes it is thanks~~ i will keep learning

[flat talon]

These two are paired with vit-bigG, while the others use vit-H.

[gusty juniper]

stable foundation

[trail girder]

does anyone know why this is happening?

I cannot run MeshGraphormer Hand Refiner with any mask type, my workflow consists only of Load Image node, Hand Refiner and then Preview Image node

[unreal reef]

I'm getting going with inpainting in Comfy for the first time. My masked area doesn't change, even with a 1.0 denoise and an inpainting model. What might I be missing?

[unreal reef]

Here is a screenshot of what I think is a correct workflow. The hands have been masked and I've prompted "pink rabbits" but the inpaint looks no different than my original image.

[steep marlin]

Is that the seed number that generated the original image? Try changing it

[mystic plaza]

what does this mean bros @steep marlin @cedar mirage

[steep marlin]

Could be a bunch of different things. I don't know what it's trying to do, but the connection error stuff could mean one of two things: either some plugin you're using is trying to download a needed model from some place like huggingface or it's some kind of wrapper node that spawns a local server and the plugin isn't connecting to it correctly. Check your firewall settings? @mystic plaza

[livid gust]

It means the images you supply to the IPAdapter must be exactly 512x512. It's just a warning. It's telling you that if you submit other sized images, they will be centered and cropped.

[rough finch]

Can any one help me regarding how to run XL checkpoints on Automatic1111. I have updated it to the latest version but it still fails to open XL checkpoints

[steep marlin]

Read the bottom of the window, I think he's talking about that part, not the warning stuff

[livid gust]

Oh, that happens when the browser window disconnects from the server. Did you close the tab and hit the back button?

[mystic plaza]

nope

[steep marlin]

Check your firewall settings then and also make sure the localhost IP isn't blocked(127.0.0.1) as that literally means the PC you're on.

And make sure you're connected to the internet

[forest birch]

Well, excuse me for using 124 years old version

[hardy tree]

kek

[raw jolt]

ComfyUI workflow 😍😍

[grand pike]

Hi does anybody know a good checkpoint for doing img2img of a drawing to real life?

[sly sundial]

Cascade

[zealous mulch]

does anyone know what the problem of this workflow?

[lone bolt]

I think you are trying to use a 1.5 controlnet on a SDXL checkpoint. Thats not possible.

[zealous mulch]

thank u, i'll try this later

[grand pike]

@sly sundial thanks I'll check it out

[zealous mulch]

May I ask if there is a workflow for running images in batches according to a set number of times? It may be necessary to load different images and different keywords each time the images are run. The necessary nodes may include: 1. A custom node for loading images in batches, which can load images by reading all the image paths in Excel, or directly reading all the image files in a certain folder path in sequence; 2. Run images in batches by reading the keywords in each column of Excel. Reading Excel information is just one of the means to achieve the purpose of running images in batches. Are there any other means to achieve the same purpose?

[uneven stratus]

Heyo, anyone knows that issue ?

I tried to unistall/reinstall PIL and pillow from my Python_embeded/Scripts folder but got errors (the system cannot find the file specified)

[silver nebula]

FaceDetailer - in both ComfyUI and A1111 - all the faces are converted to the same, female face (including male faces!!!) Why?

[sly sundial]

Because you haven't specified anything different?

In here

...or describe the face in your main prompt.

[grand pike]

@sly sundial Hey, so I found Stable Cascade base. I've never used a checkpoint that large before. would using it absolutely wreck my 1660ti?

34gb o_o

[sly sundial]

You do need both Cascade models, so I don't think you could run it on that.

[wispy kelp]

Not sure if spam or not spam. I am a girl so titties do not work no me.

[livid gust]

@uneven stratus I just received that message today on a restart. Everything was working fine, then ComfyUI won't start. I gave up on that install and pulled down the portable version of ComnfyUI to get up and running again.
https://github.com/comfyanonymous/ComfyUI/releases

[tawdry shore]

Hey, so I'm trying to import a workflow from an image I had downloaded and it says Im missing the "LatentBuffer" and "ImageBuffer" nodes. I cannot for the life of me find them.

[jolly warren]

any suggestion group photo for
consistent model like this?

[steep marlin]

Take a base image of real people, Photoshop out the background and replace with some neutral color like gray. Then do a bunch of inpainting and resampling. Or rig up a scene in UE5 with some metahumans and do the same. Do all of your ksampler to ksampler to ksampler work in latent space and only decode it at the very end; to prevent vae artifacts from repeated encoding and decoding(it's a lossy format)

Oh and throw in some controlnets and ipadapter to keep things consistent, maybe some loras as well

[sharp topaz]

Hey! Anybody here with experience in Comfyui Inpaint Nodes (Fooocus Inpaint) ?

The workflows tend not to follow the prompt at all. It is too focused on existing content, making it impossible to inpaint with something that is different in colour from the neighbour objects

[sly sundial]

Something that can do this?
A photo of a huge dragon snake looming above the men

Workflow embedded in this image

The model can make a big difference, but the one I use in that flow seems to work.

[wispy kelp]

how do i add new CSV styles to my comfyui CVS styles loader?

It ignores the new CSV file si make

[quick birch]

Karras is no longer rendering images, does anyone know why?

[livid gust]

Working for me...All models? Or SD15 or XL?

[twin token]

gm

[silver nebula]

SeargeSDXL - why is it linked solely to ComfyUI_Windows_Portable? I want to use the install.bat in System ComfyUI?!

[brisk fractal]

Hi everyone, the batch prompt schedule I downloaded is not the same length as the one in the tutorial (on the left).
I don't know where to enter the prompt, so I'm not sure what's wrong.

[cedar mirage]

Hi there,
if you click with the right mouse button on the node you can decide which parts of the node should be a connector pin and which should be on the node widget itself:
Different example but see in the screenshot: convert input to widget and then select the fields you want to use directly

[brisk fractal]

@cedar mirage
Thank you for your reply.
The parts I need are pre_text and app_text, but I right-clicked as you said and couldn't find the parts I need.

[cedar mirage]

Guess the tutorial was captured with an older version of this node, where these elements where part of the node and not as input.

Just add a string node enter the text and connect it to the pre_text input of the node

[brisk fractal]

@cedar mirage I've solved it! Thank you so much!

[outer forge]

gm

[quasi tangle]

guys please how do you increase batch size for img2img workflow??? empty latent image node which has batch size doesnt work for img2img

[slim shell]

anyone here use InstantID? is there a way to get just the face from instantID but pose from the prompts?

[quick birch]

I wrote a small article on github trying to explain how Seed works and what the differences are between different ksamplers. If you are interested, you can read it.
https://github.com/Creative-comfyUI/seed_eng

[cursive ridge]

How do I attach this randomizer bit to the rest of my workflow? It won't seem to attach. What am I missing/have to do?

[quick birch]

Chose node Clip text encode prompt and use the right mouse and chose convert widget to input. Near the clip button a text button will appear link string to the text button.

[cursive ridge]

Thank you VERY much! 🙂

[lament sluice]

Dreambooth.
Has anyone found a comfortable interface to create a model in COMFYUI's interface?

[silver nebula]

I had FaceDetailer working perfectly - and then without changing anything - it says "mat1 and mat2 shapes cannot be multiplied (308x2048 and 1280x768)"

As I say, I have changed nothing - usually such a warning suggests a clash of SD1.5 and SDXL - but without having changed anyything - its difficult to see where the erorr lies?!

[steep marlin]

If you use comfy manager, disable everything other than face detailer and see if it still bugs out. If it doesn't, then it means some other node got updated and is causing a conflict with it.

[silver nebula]

Its not obvious in Comfy Manager to see where to disable ... ? 🙂

[sly sundial]

For each of your installed custom node packs, there's the "Install, uninstall, disable" options.

[sly sundial]

Have you tried running a bare minimum workflow through facedetailer? Just model, prompt and load image?

[silver nebula]

I've used a png workflow from yesterday - and it is back to normal - yet it is still using the same setup as I am using today!!! 😦

Odd

[sly sundial]

Maybe there is 1 small change that you haven't spotted 🤷🏻‍♂️

Doesn't take much 😄

[silver nebula]

Could be ... I will check.

[sly sundial]

A change in detector, or SAM?

[silver nebula]

It might have been SAM ...

[fallow solstice]

Hello! Where exactly i need to put those files for poses? ".pth" I use ComfyUi and trying to use OpenPose.

[cedar mirage]

Hi there,
the pth files from the controlnet models need to be placed into the folder models/controlnet

The preprocessors usally download their models into separate directory

[fallow solstice]

Yea i do that but i can't use idk why. i'm new and i got a error

[cedar mirage]

what is the workflow you wanted to achive. Did you try it with the minimalistic one

[fallow solstice]

Just trying to use this plugin to see how it works and stuff like that. i'm in learning process

[lusty pewter]

EOFError makes it look like you have a bad download somewhere, but what you show just isn't enough information.

[fallow solstice]

yea i think i download something wrong.. maybe is a good idea to use just the Comfy Manager?

[lusty pewter]

Yes, you absolutely should be using ComfyUI Manager with ComfyUI.

[cedar mirage]

This is the basic workflow without preprocessors

[fallow solstice]

control_sd15_openpose.pth This is from hugging face.and doesn't work

i think this file it was the problem

[cedar mirage]

How large was it?

[fallow solstice]

fkc xD 0kb

[lusty pewter]

Yeah...that'd give you an issue.

[cedar mirage]

Well yes, then it was the file 🙂

[fallow solstice]

but the file have 5.31 gb idk why in Comfy folder have 0

Btw, can someone help me with an "Negative Promp" for those bad fingers, abnormal bodyes etc? for all those trash things?

[cedar mirage]

Would not use negative prompts, most upscale, tile upscale, fixer-nodes etc. does a far better job

[fallow solstice]

Can u export a good comfy network for me? to understand more things?

[cedar mirage]

What would you like to achive?

[fallow solstice]

Good\correct pictures. idk something generic maybe? different things like ppl on some spaces, robots on some action idk. just to learn more but in correct way. Too see which are the best Plugins to use on start etc.

Now i use this Cyberrealistic model from website trying to compile different things. adjust some etc.. Just to learn how it works and what i need

[cedar mirage]

https://openart.ai/workflows/all?days=30
Good source for Workflows, within the template section you find the base ones.

Advantage, these are less a mess and documented 🙂

[fallow solstice]

Thx man! and when u have time and can share a good assets \ plugins i will thx too ! Thx again!

[forest birch]

https://www.youtube.com/watch?v=5WxPjcsPas8 You guys wouldn't know if there's a way to crank up 3d model resolution for this workflow? :P

Or if there's another git/workflow more tuned/made for higher res creations/take existing made model and make it better quality similar to nodes to upscale image with latent?

[brisk fractal]

HI everyone,
I've installed the nodes but he keeps telling me I'm missing nodes, what should I do?

it seems like the original nodes are defective, he's showing a red colour, and I'm not sure what to do about it.

[livid gust]

There is not much you can do. Some nodes you have to give up on. I found a version of ComfyUI that was titled as "portable" It contains a lot of the downloaded modules already set up for you. Try searching for that. You can run the portable side-by-side with whatever you already have installed. I slowly migrated nodes to the portable version until it broke. Then I deleted those broken custom nodes and went on.

[violet spindle]

Hi I am currently making Comfyui workflow that uses Openpose, but for some reason it only generates black outputs as soon as I go over 16 steps. Does anyone know why or what I am doing wrong?

[steep marlin]

Probably a float overflow or something. What are your launch flags set to? Might need some kind of nohalfvae or something

The command log might spit the error as well

[violet spindle]

It says "Using split attention in VAE" in my terminal, is that what you mean? @steep marlin

[violet spindle]

Running it with --fp32-vae seems to get me up to 24 steps already but further it still gives black output

[forest birch]

Hoi, if anyone has time, could you install this one which makes 3d models from images, and see if you get the same issue? As i installed the missing node(s), but it's still saying it's missing.

Same with the example workloads it included. It's erroring with "No naistyles.csv"
And when starting comfy, it states
"ModuleNotFoundError: No module named 'diff_gaussian_rasterization'"

Here's the 3d pack i installed both by downloading and extracting zip, as well as through comfyui manager.

https://github.com/MrForExample/ComfyUI-3D-Pack

https://www.youtube.com/watch?v=5WxPjcsPas8 Here's the video i followed.

[mystic plaza]

just curious but is it possible to create an animation video out of a still image no animation in a 3d software like blender will remove the passes and all for it

[steep marlin]

Well something is causing it to spit out NaNs. The fact that you can now hit 24 steps with 32bit vae means some operation is overflowing. Is everything up to date? Maybe try going into comfy manager and disable every plugin other than exactly what you're using in that setup

[violet spindle]

Ok I think I found it, using --use-split-cross-attention as well seems to do the trick (for now). Thanks so much! @steep marlin

[steep marlin]

What GPU and driver version are you using? Sometimes issues like this are caused by them if they are older or don't have enough vram

[solar summit]

I have body measurements and a selfie, and I would like to create a 2D avatar based on these two inputs. Is it possible to do this using ComfyUI, or do you have any other suggestions?

[violet spindle]

Im on my macbook pro M1 chip with 32GB memory

[mystic plaza]

AND MY COMFYUI IS NOT WORKING

[steep marlin]

Ahh it could be some mac compatibility issue then. Doublecheck the comfyui GitHub repo for any other mac related issues and work arounds.

[steep marlin]

Double check this part of your install

(instructions on the comfyui repo)

If that's all good to go, then check the repo for the plugin because they might have their own specific things you need to do to make it work correctly with macs

[mystic plaza]

where should i add the pyromancer safe tensor bro @steep marlin

[silver nebula]

How do I use "Filter Image List" in Load Image node? What filter at all?

[lusty pewter]

What load image node are you using?

[livid gust]

Is there anyway to discover what node is in what custom node code-base? I'd love to have a right-click reveal in finder option for every node.

[livid gust]

Thanks, does that require a restart?

[lusty pewter]

Nope

[livid gust]

Okay, zoom in and out has updated the graph. How do I locate the ones with no nickname? I just see a number on some nodes.

[lusty pewter]

Those are default nodes.

[livid gust]

Hmm, sometimes I get an error on Remove and Interpolate Frames node, #687. For now I'm just bypassing it in the SteerableMotion workflow, but I think it is key to getting smooth clean output.

[silver nebula]

Load Image ... with Control_Filter_List box ... ?

[bronze kernel]

Hello! Long time no chat. I've been tinkering with ComfyUI again and I have a question about schedule prompts. Specifically, is there any way possible to grab the positive and negative prompts from a given frame in FizzleDorf nodesand save the data when running multiple prompts? Please see the example images above. I have an input going in to where the "normal" set of scheduled prompts (like below) normally go, and have not been able to find a node that reads those values

"0":"A Toucan with butterfly wings, hopping gracefully through the meadows with a kaleidoscope of colors", "1":"A castle built in the shape of a flower pot", "2":"A Hiroshima Peace Memorial, a somber yet impactful example of modern Japanese architectural resilience", "3":"Interstellar Wanderer, long-range interstellar exploration vessel, designed for extended missions into uncharted space", "4":"A sailor ascends the mast to adjust the crow's nest, a solitary figure against the backdrop of a cloud-streaked sky"
I'd like each of those prompts to be read as the name of the prompt so I can save individual frames with the prompt as the filename. Is that at all possible? There is a reason for this madness, I promise.

[maiden dune]

Kim

[pulsar drift]

hi, i just wanna install idm nodes, but when i restart comfyui, it is said this and i can not open ui interface, does this mean the program is downing something? should i wait?

[boreal stirrup]

People u think I can use it with my 8gb rx 6600 xt on Windows and u think it's worth learning comfy UI as some1 who has never used something node based like this before?

[stiff bay]

Hi everyone can anybody please tell how can I generate Images using comfy ui ?

[wispy kelp]

You load it up and it has a template already

[wispy kelp]

Which supir model is better? Supir-v0F Supir-v0Q?

These ar ethe questions that keep one up at night.

[sly sundial]

AFAIK...v0F for realistic/photographic

[sly sundial]

From CoPilot:

 Let’s compare the SUPIR-v0F and SUPIR-v0Q models. Both are part of the SUPIR (Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild) framework, which aims to develop practical algorithms for high-quality image restoration. Here’s a breakdown of their characteristics:

SUPIR-v0F (Fidelity):
Training Settings: Trained with light degradation settings.
Stage 1 Encoder: Retains more details when facing light degradations.
Quality: While it may hallucinate some details, it excels in reintroducing fine details and texture.
Use Case: Ideal for scenarios where maintaining fidelity and fine details is crucial.
SUPIR-v0Q (Quality):
Training Settings: Default training settings following the paper.
Generalization: High generalization and high image quality in most cases.
Characteristics: Provides excellent image quality without compromising fidelity.
Use Case: Suitable for a wide range of applications where overall image quality matters.
In summary:

SUPIR-v0F prioritizes fidelity and detail retention.
SUPIR-v0Q balances quality and generalization.
For most use cases, SUPIR-v0Q is recommended due to its impressive image quality. However, if you specifically need to retain fine details even in challenging conditions, consider using a combination of both models (e.g., CCSR ➜ SUPIR-v0Q) for the best fidelity and quality balance

[wispy kelp]

👍

[bronze night]

What do y'all recommend in ComfyUI for region-specific prompting? Even if it's just bounding boxes.

[snow plank]

give me a office task for employee working and decrease heat

[wispy kelp]

why would the hand detailer only detect a black box and not the hands? :<

[wispy kelp]

where do i have to place the yolo models?

is there a way to check the directory path of these models u selct in nodeS?

[split bloom]

Hey all, I'm building a tool that lets you train a digital double. I'm using ComfyUI. While I'm happy to play around with it, I'm also happy to hire someone to help with me it 🙂 - Who here is an expert at Lora, Dreambooth, ComfyUI who wants to make a couple thousand $? 🙂

[stiff bay]

I wants to earn but don't know anything about these things 😢😢

But I want to learn about these

[sly sundial]

Because your detection setting numbers are too high, or the hands are too bad/small to detect.

Lower the threshold numbers to have a higher detection rate.

[flat talon]

Is there a hand repair plugin compatible with anime images? I tried using MeshGraphormer, but it failed to recognize the hands at all.

[solar summit]

Hello, everyone! I would like to change the tshirt using comfy-ui, is it possible?

[livid gust]

Is there any place to post comfyui bugs?

[silver nebula]

The yolo folder inside the models folder

[slim root]

Open an issue on GitHub.

[native cairn]

hey everyone i joined recently
i am developer and i need a suggestion from you all
to know will it beneficial to develop some kind of nodejs integration in comfy ui.

[golden iris]

So.
I am already using Auto1111, but wanna try Comfy.
So I've done as the guide on Github tells me to, specifically the "I already have another UI for Stable Diffusion installed do I really have to install all of these dependencies?" part. Yet nothing seems to happen..?

Also /wave

[wispy kelp]

i dont have a yolo folder

[wispy kelp]

Make a selection node where you can select different parts of the image and label them as something in prompt and give us the ability to keep that something consistent in future generations. (face, character, any object etc.) by adding that word to the prompt. like select an orange in the image and label it as "orange1" in the prompt and then from then on "orange1" when called up in subsequent prompts will make sure that that orange in the image will stay the same shape and have the same textures but it will be in different lighting and position/angle only.

[unreal reef]

Is there any way to get my custom CSS to persist in Comfy? It seems like it always gets wiped out.

[verbal sky]

Newest video just published (and thanks so much for the 2K!), topics include Auto CFG, Sidebar and bonus workflow material - https://youtu.be/IS9kbxDpJRk

[silver nebula]

Perhaps you should make one?

[wispy kelp]

no no i have other yolo models installed

i just dotn know where to fidn them

[silver nebula]

https://huggingface.co/Bingsu/adetailer/tree/main

[wispy kelp]

Yes lol I know where to download them I just don;t know where to put them.

[sly sundial]

Why not download them through the manager?

"ultralytics" models

[golden iris]

So, in Auto1111, i can add multiple paths to LoRA's and Ckpt's. Can i do the same here, as the drive the Webui's are on, are... Running a bit low on space, and can't fit more ckpts.

[silver nebula]

mklink /D to make a Symlink - you can have 1 Models folder serve both ComfyUI and A1111 - including LoRAs

[golden iris]

But i have them on 2 different drives. D and E.

Also. No idea what Mklink is

[cedar mirage]

Well important is the symlink = symbolic link.
So you create a symbolic link for example called "checkpoints" in the folder of comfy. But instead of a regular folder ("create new folder...") you use the mklink command and point towards the checkpoint folder of a11111. When using comfy it will follow the symbolic link to the location "linked". @golden iris

[golden iris]

I am already using the files of the A1111 webui. This is another folder,

I have this set up correctly. D Drive. It draws the files from the A1111 path

But i have extra ckpts here, which is a completely different drive

[cedar mirage]

When starting comfyui you can pass as an argument additional path informations --extra-model-paths-config PATH [PATH . . . ] Load one or more extra_model_paths.yaml files

[golden iris]

And i add this by opening the webui.bat, in a txt editor, and add

--extra--model--paths--config [Path]?

[cedar mirage]

Sorry you are in the comfyui Help Channel. I just thougt you were talking about comfyui. In Comfyui you would edit the file, and add the path to it.

[golden iris]

This file, ye?

[cedar mirage]

Just look at the syntax of loras and upscale_models and try your best to do transfer learning towards "checkpoints".

[silver nebula]

Use a symlink

[golden iris]

I got it working now. All i needed to do was this

[merry ermine]

Hey any way I can load lora without trigger words or automatic triggerwords

[heady parcel]

I've found in my experience that I can't get the lora: syntax to work at all in ComfyUI. I recommend CR Lora Stacker

[hard marsh]

can someone point me as to what I'm doing wrong with this? I can't figure out why it's not printing the results I want.

[hard marsh]

trying controlnet out so i can get two in a picture...

nope, i don't think that's how that's gonna work.

[flint bough]

can a additional graphical UI be laid over a workflow consisting of only the nodes you want/can adjust?

[hard marsh]

isn't that what a1111 is?

just a GUI of a workflow?

[wispy kelp]

all those hand yolos seem terrible at detecting hands, even when they do they are cropped in all the wrong spots.

🤷‍♀️

[mystic plaza]

is better to create your own workflows or like it depends on the need or content you are generating @dry rock @wispy kelp

[wispy kelp]

Its better to make your own

i just spent 3 hours cleaning up nerdy rodent's workflow

i had no idea what was going on

first you copy paste other peoples workflow but you have to kinda understand what does what so its best to try to figure it out and make it your own

[mystic plaza]

for now i feel overwheelmed sometimes with comfyui there is a lot which can be done

[wispy kelp]

yes yes there is

[mystic plaza]

and someworkflows are hella complex

[wispy kelp]

u tell me....

[mystic plaza]

which workflow is used for this may i know or is it a custom workflow

[wispy kelp]

i havent even used regional promting in comfy so idk

im wrestling with ip adapters now

[mystic plaza]

welp i had a field day trying to understand ipvi morphing workflow and getting a output from it

[silver nebula]

Omost

[mystic plaza]

I SEE

[mystic plaza]

???

[steep marlin]

Picking up someone else's workflow is like picking up someone's Skyrim save where they are 100h into it... It's best to slowly work up your own flows, that way you know what's what and why things are how they are

[mystic plaza]

thing i dont know how to even create my own workflows or do i need a lot of loras and other models as well

[silver nebula]

Mix and match - make your own/adapt someone else's 😄

[mystic plaza]

welp if you are a newbie that does not know how things work i feel its best to use a workflow made by someone else first and then create your own workflow i guess

[naive laurel]

Hello! do you guys know anything about a replacement model for insightface for commercial purposes? or is there some faceswapp tech to build locally without insightface?

[mystic plaza]

you can prompt or use a detailer

so just curious but is there a way to create animated videos through images in comfyui like videos or something ?

[heady parcel]

that's exactly what i've done actually

[heady parcel]

i agree with this statement, i just wanted my own GUI

[flint bough]

Anyway we can't take a look at it?

Not including gender specific variations there are 134,719,200,000,000 individual "looks" (?) for humans.

Feel free to look over my spreadsheet:
https://docs.google.com/spreadsheets/d/1IpyQ4TyLrQukqpWisenW_GwzNayYXVIEjJuVnhT9maU/edit?usp=sharing

Part of a prompt generator project. Comments welcomed

[heady parcel]

sure so before I show you you some screenshots of this alternative A111 i've made let me just give some background behind the intent

• first i wanted to just generate my own images using my own workflow
• then i learned about civitai
• then i learned about that remix button that civitai has for images
• then i built an interface on top of the generation form that allows me to quickly monitor, browse new images, and add them to the queue for generating images almost nearly like it
• I also manage the comfyui queue, that means it'll only add one item at a time to the comfy queue and then you can control the sorting of how it prioritizes items, either by batch size or using the traditional front, back style they use

[flint bough]

That's sounds like what i was thinking about.

[heady parcel]

so this is the Browse tab here I can browse through my generation with all the filters I could ever want

clicking into that corner red button reveals the state of the queue

• the left icons indicate the size of the batch and the position, the graph also provides insight as to how many items, so like if i have lots of batch size 1 the graph will become a ramp for example
• then on the side i can pause | play the queue, control sorting, the image icon allows me to see the active item thats processing like the details and the cog controls the image collector

[flint bough]

Nice. I'm trying to focus download a "Storyboarder" UX/UI

[heady parcel]

not sure what that's about

[flint bough]

Did you do that in Python? HTML3?

[heady parcel]

this is the Generate page, i'm also posting screenshots of what stuff like the checkpoint picker looks like

[heady parcel]

nodejs + ionic + react

[flint bough]

My workflow is going to be based off of how Anime Productions Storyboard their shows

[heady parcel]

this is picker for LORAs

[flint bough]

Ah, yes

Ingenious

[mystic plaza]

would changing the lora in an animate diff give better results or it depends

[heady parcel]

i dont understand how that relates to a comfy workflow, you mean like the UI is going to be related to that?

[heady parcel]

there's still lots to show if you wanna see the rest

[flint bough]

I'm going to set up comfy ui workflows for all the steps i.e. character creation, locations, camera angles/poses then have the overlay UX/UI tie it all together

[heady parcel]

oh that makes a lot more sense

i can imagine defining a character with a prompt description and then saving that description as a character and then saving prompts as locations and then being able to craft a story with tags to characters and locations and have it generate the elements explained

[flint bough]

I want to but it's 5.16 AM here an i been up all day moving pallets. I should hit the rack0 But thank you for showing me your stuff. U really appreicate it.

[heady parcel]

good luck with your idea, feel free to tag me once you have some progress to show

[flint bough]

Yeah that what I'm aiming for

[flint bough]

you got it. thanks again

[heady parcel]

especially if you can fold in some of that tech where you can generate consistently similar looking characters not just based on the prompt but whatever that is

[flint bough]

i did post the tables i'm working on for the character prompt genterator

[heady parcel]

the tables?

[flint bough]

That's the lynchpin. what's the point if you can't have consistent characters.

good night or day...

[heady parcel]

8am here, good day 👋

[heady parcel]

when you have a chance clue me in on what that tech is called, just curious to know how that's done if you know how

[mystic plaza]

so just checking is possible to make a path for comfyui if i need to download a lot of custom nodes checkpoints and loras like i have a empty hdd in my system ?

[hard marsh]

still looking for some help on this!

[wispy kelp]

any way to adjust how much the input image affects stuff in an img 2 img workflow ?

[silver nebula]

Adjust the Denoise - low denoise you get more of the original picture; the higher the denoise, the more it follows any prompt - and resembles the input picture less and less ...

[livid gust]

Is there a node that lets me inspect values? I'd like to know the value of the frame count from a video.

[livid gust]

@hard marsh There is a technique called "double-exposure." Is that what you're looking for? Perhaps add that keyword to your prompts? There's also a model dedicated to the look.
https://civitai.com/models/1146/double-exposure-for-sd-1x

[cedar mirage]

Well it seems you are trying to use regional prompting. In Mx experience the results of ipadapter with attention masks work significantly better. So I would suggest you try using the ipadapter.

[wise wind]

hello

i just installed comfyUI and am wondering how to use adetailer

or if anyone couldnt be bothered explaining can someone link me a guide to aDetailer, dynamic prompts and all of that stuff

[hard marsh]

ahhhh...time to go learn another thing, hah.

I'm definitely trying to use regional prompting.

[wispy kelp]

if you mess with settings while an image is rendering can that ruin the possibility of reporducing that image later on?

im reloading some images in comfy and i get different results

but i did tweak things while generation was going on

when does it log the settings when you start generation or when the image prints?

[desert dawn]

In the workflow, is there any way to see the connections between set and get nodes? Do I have to check them one by one? I have no idea which set node the data is coming from.

[steep marlin]

The data is coming from inside the house...

[hard marsh]

I believe it's when you click generate

[sly sundial]

You can set a "tag" on the set node, which is then chosen in the get node. That way you can have multiple set nodes sending to different get nodes.

[broken mortar]

https://tenor.com/7i3J.gif

[livid gust]

@desert dawn The names of the Get/SET tell you what they are connected to. If they are collapsed, expand them to see the list of available connection options. Make sure to give the connections names that make sense, like BASE_MODEL or BASE_CLIP etc...

[steady horizon]

Hello everyone, I'm new to the ComfyUI interface (used AUTOMATIC1111 in the past). I'm trying to figure out how to use AnimateDiff with ControlNet txt2vid using one reference image in ControlNet.

[livid gust]

I don't see a workflows section on this Discord, but you can visit here to pull down some test workflows. Visit Mateo's youtube to learn more.
https://comfyworkflows.com/workflows

[steady horizon]

Thank you very much

[broken mortar]

@broken mortar softtextil.cz

[mystic plaza]

i am in need of a guide on how to prompt with stable diffusion models in comfyui

[maiden shell]

The prompte are the same regardless of UI. There are many good general ComfyUI guides on YouTube. If you want to get help in a place like this it kinda has to be a specific question.

[mystic plaza]

i see

[cedar mirage]

you can use models like Pixart-Sigma which uses T5 and therefore works better with natural language prompts instead of the usual ,xxx,xxxx list.

[steep marlin]

If you have a common image size you like to use, give the tensorRT addon a try. It doesn't work with loras or cnets, but for screwing around, it's great. Comfyanon made a repo for it and holy shit is it fast. I know a1111 had a version of it a while back, but it wasn't THIS fast. Using the regular version of the model 44 seconds, using a static TRT version: 14 seconds. Both with the same sampler settings. Very minimal differences in the generated images.

just tried it with the dreamshaper lightning model and it brings it down from 17 seconds to 4 seconds. i do notice that it will occasionally bug out and spit out a black image, but still really cool that it can bring times down by ~70%

[mystic plaza]

is it possible to use an image in animate diff in comfyui ? @steep marlin

[mystic plaza]

thanks will try it

[steep marlin]

https://github.com/comfyanonymous/ComfyUI_TensorRT/tree/master

supports:

Stable Diffusion 1.5
Stable Diffusion 2.1
SDXL
SDXL Turbo
Stable Video Diffusion
Stable Video Diffusion-XT

[mystic plaza]

thanks bro but i am using a animatediff workflow and in need of some guidance mainly

[steep marlin]

oh i thought you asked if it worked with it, my bad

[mystic plaza]

can you guide on a workflow or setup for it pls

[karmic creek]

I don't think it can work yet with controlnets, loras or animatediff, remember reading about it at the repo

[steep marlin]

yeah, i brought that up

but you could always merge a lora with a model if you really needed it

and then make the TRT of that

[mystic plaza]

cause i have some still images and out of credits on runway ml and i have a script which i need to follow

guess i shall try to get it done on pika labs and then do the remaining on davinci resolve

[mystic plaza]

i am lost over here now

[karmic creek]

merging the models first and then running the tensorRT stuff

[mystic plaza]

due to time constraints will try this stuff after i submit this work and all

[cedar mirage]

Did you tried it and do you have numbers for the speed bonus?

[karmic creek]

Nvidia's site claims over 6X speed up with A100s, haven't tested this since my workflows are pretty heavy on controlnets, animatediff and so, which won't work yet

[cedar mirage]

Yeah but this is also from the Nvidia site... You get an 17% speed increase if you keep fp16. Even they stated that there method to convert the models down to 8 bit will not have much impact of the quality.... So before adding more custom nodes to my install i hoped for feedback from others 🙂

[steep marlin]

Yeah I tested it and said I was seeing roughly a 70% decrease in generation times. Like from ~45s to ~14s on my rtx 2080. I know while creating the model, it did mention that my card doesn't support fp32 stuff, so it might go even faster on a 30xx or 40xx series card, don't know which cards support it for sure though without looking it up

And again, those numbers are for a static TRT where it is optimized for generating one single image size. In my case, I was just using 1024²

Dynamic might slow it down a little and eat up more vram though, don't know

[cedar mirage]

Thanks for the feedback. So it might be worth to keep an eye on the development.

[steep marlin]

No problem

[sick badger]

@storm folio
stupid question on your websockets ws_images api example
it doesn't include an example node in the workflow for uploading an input image, how would that get included in the json load?

[sick badger]

nevermind, just dawned on me i gotta us a Load Image from URL node instead of a Load Image node

[unborn raven]

My Problem is at ComfyUI
I don't know which model i need for IPAdapter_Plus but i tried to do something 😄

[lusty pewter]

Just download them all so that you don't have to worry about it.

[unborn raven]

thats what im doing 😄

some say they belong to the ipadpter model done that also

[sick badger]

will comfy support the new audio stuff? sounds entertaining

[hard marsh]

is there a uh... easy way to use adetailer in comfyui?

or am I SOL in terms of simplicity?

I could only find this set of custom nodes for comfy https://github.com/ltdrdata/ComfyUI-Impact-Pack?tab=readme-ov-file#how-to-use-ddetailer-feature

the given example looks like a bunch of spaghetti to me...

[split gate]

anyone know of a custom upscaler node that has a clip input? this one is perfect, but has no clip input. when the image upscales, it loses some details because it has no clip to refer to

[hard marsh]

Help me 😭

it's not doing what i want it to do

[split gate]

what do you want it to do? the preview image looks like it reflects your prompt

[hard marsh]

figured it out, my controlnet wasn't plugged in properly

[split gate]

k cool now my question is not in view. neat

[hard marsh]

..ah crap sorry

[split gate]

its fine. ill post it later

[hard marsh]

i in fact did not figure it out.

[hardy tree]

was about to say your controlnet is indeed not plugged in correctly based on that pic

[silver nebula]

Pinokio is doing a 1-click install of Stable Audio

[silver nebula]

It's quite good, just made my first 47 second-long clip - a warm drone sound!

[silver nebula]

a warm pad sound - playing a C9 arpeggio in 4/4 time https://drive.google.com/file/d/1lG2fArbCe2AauplHH2BhEAsUaKF3mKih/view?usp=drive_link

[eager escarp]

Doesn't really make sense.. clip is for conditioning not sampling

[silver nebula]

Upscaling always involves some detail loss - try lanczos, or bilinear, or bicubic - and see if there is a difference?

[hard marsh]

i am awake and it's back to tinkering with openpose for sdxl...

i think my image might just be too wide and it's getting mad at that.

because i'm NOT getting the results i want, and I know 1.5 isn't gonna do any better.

i'll try with another openpose picture and see...

yeah...i think the other one is just too wide.

[hard marsh]

the openpose editor in a1111 is giving me one hell of a time. I'm adding in joints, but it's deleting them when i save it.

[sly sundial]

It only consists of the first 3 blue nodes nodes in that example. The rest are all just part of your normal workflow anyway.

[wispy kelp]

Welcome to consistent AI stuff.

[red dust]

Can comfyui be run in a computer and then be linked to android?

[split gate]

i figured it out. lora inputs should go into the "model" input after the checkpoint loader. this keeps lora details when upscaling. seems i was asking the wrong question. example:

[mystic plaza]

how to fix this bro ?

[eager escarp]

Ah yes yeah I was a bit confused by that you were saying 🙂

[silver haven]

I'm trying to use this workflow: https://civitai.com/images/14782203. But the modules for most of the first 4 columns aren't loading. I've used several other workflows I downloaded in the past, but this one seem to have nodes I don't have available. I attempted to update, but no change. Any suggestions?

For example, starting in the top corner, I have a Checkpoint Loader node, but it wants Checkpoijnt Loader with Name.

ComfyUI Manager fixed a lot of that, but still missing ImpactWildcardEncode, Upscale, and FaceDetaiiler...looking for those now.

[mystic plaza]

can anyone help me fix this problem

[silver haven]

Still having some trouble with the manager, likely due to the way I had to cobble things together to get it working in the first place with the various python installs I have on my system.

[silver haven]

I'm not sure I understand the problem there. You might just need to refresh your browser?

[south summit]

Hello! do anyone know which version of comfyui can work with xformers? I tried the last one and I got an error that says "xformers 0.0.26.post1 requires torch==2.3.0, but you have torch 2.4.0.dev20240605+cu124 which is incompatible"

[silver haven]

Several rounds with the manager including unistall and reinstall of some packages, but all looks good now.

[sly sundial]

Scroll up and see what errors occurred when loading those custom nodes.

It could be you're missing a module or it's conflicting with other nodes.

[silver haven]

Is there any way to give specific prompts to be focused on by a face detailer? Going for cat/viper eyes and the harder I prompt it the more it just messes up the whole face.

[silver nebula]

ReInstall

[silver haven]

Anyone know how to avoid the cameras I keep getting included in my images?

[merry ermine]

Anyoner know how to extract CFG value from prompt text, something like "cfg:6.5". There is a custom node that lets you apply LoRA's like that, hoping there is something for CFG as well.

[steep marlin]

stop adding it to the prompt? i'll give you a hint: the datasets used to create the models likely have no caption knowledge of what camera they were shot with. saying things like dslr, camera, camera brand names like fuji/nikon/canon, etc, can all be misinterpreted as wanting to have the actual camera in the scene.

[mystic plaza]

this is the cmd log but cant see anything which is causing it to not run @steep marlin @silver haven @silver nebula @sly sundial

[sly sundial]

Look closer...
Line 64 onwards:
Traceback (most recent call last):
File "E:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\llama_cpp\llama_cpp.py", line 70, in _load_shared_library
return ctypes.CDLL(str(lib_path), **cdll_args) # type: ignore
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "ctypes_init.py", line 376, in init
OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "E:\ComfyUI_windows_portable\ComfyUI\nodes.py", line 1879, in load_custom_node
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 940, in exec_module
File "<frozen importlib.bootstrap>", line 241, in call_with_frames_removed
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-N-Nodes_init.py", line 26, in <module>
check_and_install('llama_cpp')
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-N-Nodes\nnodes.py", line 145, in check_and_install
library_module = importlib.import_module(import_name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "importlib_init.py", line 126, in import_module
File "<frozen importlib._bootstrap>", line 1204, in _gcd_import
File "<frozen importlib._bootstrap>", line 1176, in _find_and_load
File "<frozen importlib._bootstrap>", line 1147, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 690, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 940, in exec_module
File "<frozen importlib._bootstrap>", line 241, in call_with_frames_removed
File "E:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\llama_cpp_init.py", line 1, in <module>
from .llama_cpp import *
...

[mystic plaza]

its missing a module in short right

[sly sundial]

It looks like that should be part of the ComfyUI install, or is it something you installed?

[mystic plaza]

i had it installed but for some reason its missing tho

[sly sundial]

I meant the llama stuff. Is your ComfyUI fully updated?

[mystic plaza]

its fully updated bro

[sly sundial]

I don't think it's the custom nodes that are failing, but something in the ComfyUI install fails first, which causes the custom node to fail. I'm not a Python dev, so not sure for certain.
Try a fresh install of ComfyUI and just add the nodes that failed first, to see if that works.

[mystic plaza]

bro my comfyui stuff is almost 120 or more gb bro

[sly sundial]

Doesn't matter, you can install another one and leave it as it is.

Just a different directory name

[sly sundial]

Did you run the update with dependencies?

[split olive]

I'm encountering an error message saying "This action is not allowed with this security level configuration" while trying to install a plugin via Git URL. How can I resolve this issue and allow the installation to proceed. Thanks for any suggestions

[mystic plaza]

yes

[sly sundial]

Then you'll need to try the fresh install and see if it's your install that's a problem.

...unless anyone has any other ideas?

[mystic plaza]

trying it now will take some time since i am also in need of storage

[mystic plaza]

now what does this mean

[mystic plaza]

what does this mean bro @dry rock

[dry rock]

something is incorrectly connected to one of your AnyNode

[wispy kelp]

i could never figure out anynode

ill just connect the noodles myself thanks

[mystic plaza]

this is how its looking @dry rock @wispy kelp

[dry rock]

what ?

without even looking at what the nodes do. having one output be text AND image at the same time sounds very wrong.

could be wrong tho, havent opened comfyui in months

[mystic plaza]

its the default workflow from andrea so trying to wrap my head around how it works

[fickle remnant]

Anyone knows where to get that noodle style?

[dry rock]

you mean the square ish routes ?

[mystic plaza]

???

[fickle remnant]

Sorry that wasnt super sfw

[dry rock]

try cutting that line for now.

I can t check comfyui stuff, I m at work right now. And I ll be swapping cpu this evening (finger crossed the post office is still open by the time I drop by) sooooo

[mystic plaza]

can you pls explain this in detail

[dry rock]

I m not sure. looking at it with the doc open.... it should be good

[mystic plaza]

hmm still cant get it to work

[vital root]

I have really tried everything, I have installed all the necessary components including insightface-0.7.3, Visual Basic, SDK, etc., but nothing works. I have reinstalled everything multiple times but still no success. Do you have any optimal solutions to this problem? I am desperate

[lusty pewter]

You need to look at the console window for details.

[vital root]

console report

solution?

[lusty pewter]

The error is plainly stated in your log:
"Python.h": No such file or directory
Without that, it cannot build the wheel.
Without the wheel, it cannot build insightface.
Without insightface, it cannot install reactor.

[vital root]

any solution?

[lusty pewter]

Read this:
https://www.reddit.com/r/comfyui/comments/17rgsuq/installing_reactor_gives_error_failed_building/

[spring jungle]

i had a suggestion for fuction that when you search for something an hover over it show you node itself

[wispy kelp]

i have the same errors

installeng the face stuff for comfy is very complicated

installing reactor in forge was 1 click 😦

[vital root]

Remove any version of Python and install 3.10.14. Install insightface in the ComfyUI folder (python_embedded) and then simply run pip install insightface in the same folder. Restart the computer, and it worked. After a week, I finally succeeded.

[wispy kelp]

yeah

well im scared to do that

what if it destroyes my 500 other nodes

[vital root]

works!

[wispy kelp]

eveyrhtign depends on python and I THINK eahc of ym webuis use their own python environments

but i assume all the stuff in comfy uses the comfy environment

hmmm

never seen a preview image made of 1 and 0s

or is that chinese symbols

im not seeign any faceswapping here

[vital root]

in processing

[steep marlin]

Those are all individual frames of the video

[hard marsh]

okay...wow adetailer was a mess to set up, but I got it working first try!

Through the power of spaghetti!

[wispy kelp]

i just copy pasted a working adtelier from another workflow lol

its good with faces but hand snot so much

[hard marsh]

I need to do better experimenting with hands, but for the first one, it did really good on the hand that had the right amount of fingers

[wispy kelp]

what yolo u use for hands

[hard marsh]

the one that comes with it

this is what the hand originally looked like. It still looked fine, but i wanted to try it out.

[mystic plaza]

do some custom nodes also conflict with another custom node ?

[dire shale]

What is the difference between just generating the image in 2560x1440 and generating the image in 1280x720 and then latent upscaling it to 2560x1440?

[wispy kelp]

sometimes u get weird deformities and mutations and multiple limbs characters if u try to generate at resolutions much higher than 1024

[sly sundial]

Yes

[hard marsh]

You will get really...really weird results if you outright go that big.

[mystic plaza]

guess i have to get used to codes not adapting with each other

[dire shale]

hey guys, How do i generate an image with multiple models (with the same prompt) at the same time to test which model is the best

[stiff bay]

#🧣｜comfy-ui can anybody please provide me a tutorial on how to make AI art using Comfy UI ?

[dire shale]

https://youtu.be/LNOlk8oz1nY?list=PLH1tkjphTlWUTApzX-Hmw_WykUpG13eza

[stiff bay]

Thanks

[wispy kelp]

Plug those noodles in!

[hard marsh]

Can I generate multiple inpaints using comfyui instead of one at a time?

....

why did i reply to that,.

im so goddamn sorry

[mystic plaza]

messy noodles are better but for the better understanding how to make a smooth node graph

[hard marsh]

Does anyone know why my adetailer turned... my hand into a face?

well. maybe it's my prompt.

I'll figure it out.

it's definitely not my prompt.

I have 'fist' in my prompt. and scores.

[sly sundial]

You possibly have the tolerances set too low

[woeful vessel]

https://www.udio.com/songs/nvm88JHoMbHJ2ns4p2SwBU

For you Comfy guys, from a Forge user with luv' ♥️

[unborn dune]

hey guys i felt like i was doing pretty good learning Comfy, but damn am i struggling to get good results upscaling ponyxl images. the upscaled images ARE larger but are losing a ton of detail

[silver haven]

Upscaling can be finicky. Sometimes it it's blurring specific details you can fix that by adding them to the prompt.

[steep marlin]

Damn, I borked my comfy install trying to mess around with hunyuan and can't seem to undo the damage. Going to have to do a full reinstall... Le sigh...

[hard marsh]

Well, the thing is, it's detecting it as a hand, but it's turning it into a face. Unless I need a different hand model or something.

Is there a way to do multiple inpaints?

Yes there is, i figured it out

Ahhh, the joys of comfy... never going hungry due to all the spaghetti.

Too bad I don't know if you can do all the ksamplers at once, but ah, oh well.

Hm. Looks like the extras are taking the hand out of the picture. Weird.

Oh well! I got myself a winner pretty quickly.

[cursive ridge]

How soon after SD3 comes out on the 12th, will a comfy SD3 be out? 😄 I mean for a local install and not API. (apologies for the noob question, still new to SD/comfy)

[hard marsh]

sd3 is a model, so it'll work with comfy.

[cursive ridge]

Thank you 🙂

[hard marsh]

You'll have to tweak your current workflow, maybe, but it should just be a model and it should Just Work™️

[cursive ridge]

Is it possible to have SD2.5 and SD3 as 2 separate instances?

[hard marsh]

Like 2 different workflows?

[cursive ridge]

Or I guess that's just a matter of a drop down window such as when I choose the checkpoint?

[hard marsh]

Or run them both at the same time?

[cursive ridge]

Have them be completely separate, but not necessarly run at the exact same time (my poor gpu lol)

[hard marsh]

You can have them both in the same folder, they will never intersect with each other or mess with the other.

They will NEVER run at the same time. You just choose a different model.

[cursive ridge]

Thank you very much 🙂

[hard marsh]

And if you DO want to run them concurrent with each other, I don't know how to do that. I want to say it's like making two workflows but on the same placemat.

(I call the workspace a placemat...workspace is probably better though.)

I know when I did multiple inpaints, it ran them one after the other.

It never ran all 5 at the same time.

And it was all on the same workspace.

[hard marsh]

This, up here.

[cursive ridge]

8gb vram here, I don't see that happening lol

[hard marsh]

Well it also depends on how much ram you got in general.

I think...

[cursive ridge]

16gbram

[hard marsh]

Cuz I have 8gb ram and 6gb vram on my laptop, and it runs decently

Will I ever get sdxl to work? No.

I'm stuck on 1.5 models since those are the best ones, and have the msot stuff made for them.

But! I have my s/o's computer with 12gb vram so I can do whatever the hell I want model wise.

[cursive ridge]

What about Duc Haiten's pony models? They seem to take less resources/time

[hard marsh]

Well it depends, because pony runs on SDXL.

I believe it does.

[cursive ridge]

for whatever reason his checkpoints take 1/3 the time as regular pony etc.

[hard marsh]

Huh... lemme look into that.

[cursive ridge]

if it wasn't for your s/o, I'd recommend gpu rental 😄

I'm glad I didn't realize that I was getting the min required gpu when buying my computer, or I would have spent way over my budget to get more lololol

[hard marsh]

I don't know why his is faster, I'm not too tech bright. I was expecting some easy cheapshot shortcut like "Oh, I'm actually running 1.5" but he's running XL.. I mean, you can give it a shot, unless you already have

The main problem for me is actually getting the model to load. I have a shitty amount of ram that taps the min req, but with how much windows takes up, i might as well have 5 or 6gb of ram.

And I might as well have like..4gb of vram with how much gets taken up by windows too.