🧣｜comfy-ui

[merry ermine]

looks ok 🙂

[wintry jackal]

I just learned that highres fix works with sd xl and it's amazing, but is there a way to use highres fix node with an image I already have??

[sly sundial]

You could try...Load image > vae encode > send to latent in and reduce the denoise

[wintry jackal]

Thanks I will try It tomorrow 🙏

[merry ermine]

ye most stuff like this is just a case of trying every denoise value between 10% and 90%

if you throw in one of the nodes that adds some noise to the latent in between samplers

you can get a much nicer result (better than A1111 highres fix)

[ashen valley]

i still didnt find a solution to fix this error even with chatgpt

can someone help me please?

The frustration is slowly creeping towards me 😅😭

[mystic plaza]

so what does this error mean bro ?

[distant lagoon]

It's complaining about the size of the input image it looks like. You probably need to crop or scale it with a scale node before feeding it in

[sly sundial]

Can you show the workflow?

[ashen valley]

yes one sec

[sly sundial]

https://tenor.com/view/were-out-of-time-times-up-over-no-more-the-end-gif-15996689

[ashen valley]

yeah sry i lost it

but now i found it again

i downloaded it from a website

and tried to install all dependencies

on startup its fine

[sly sundial]

Did you download all of the models and put them in the correct directories?

[ashen valley]

models?

i am new to comfyui

this is the console output

whats a .pth?

[sly sundial]

I think that's your problem. It's complaining that it can't find any of the listed models, and you've not added your own images, so it's looking in places that don't exist for you.

[sly sundial]

A file type

They're controlnet files/models

[ashen valley]

oh

[ashen valley]

you need controlnet models?

can you explain it a little bit to me?

if you have time

[sly sundial]

If you want to use controlnet, yes

[ashen valley]

thx for your help btw

🫂

[sly sundial]

I don't have time, but you'll need to locate them, download them into your models/controlnet folder.

There are different models depending on what you want to do. You need to work that out first, and you should probably learn how to start using it with simpler stuff.

[ashen valley]

ok

thank you very much

[sly sundial]

Places to look for models are huggingface and civit.ai

[ashen valley]

ok thx

i use civitai a lot

[wintry jackal]

Thank you very much I forgot about this, comfy has so many things that from time to image I forget stuff 😭

[candid tangle]

What's the best source for prompt syntax/tricks for ComfyUI? Things like alternating between dog/cat every step, changing weights, wildcards, etc

[merry ermine]

to be honest I am reasonably sure that the science says that heavy prompt engineering, on average, doesn't improve image quality that much

that's actually the reason why I have used the same prompt for the last 6 months

its notable though, that this is not the case for LLMs
where prompt engineering still has a very big effect

if you did want to try something new, this node is nice: https://github.com/Extraltodeus/Vector_Sculptor_ComfyUI

or if you are using regular text encode nodes then this: https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb

it gets you A1111 or Compel-style weighting

[candid tangle]

Just seems weird to me that there isn't a basic info page about prompt syntax for the regular prompt node 🤷‍♂️

[crisp trellis]

[silk aspen]

Hello everyone, I would like to know if could someone help me to understand what am I doing wrong..?!

I´m trying to upscale an image, and after that, I want to apply ReActor to make a faceswap, but I got the error shown at the image

The error appears in the ReActor node

[distant lagoon]

that's actually a really common error that you'll see from time to time when working with python environments. what it means is a function in the code of the packages did not understand the arguments it was being passed. like it was expecting one thing and got something else. why does this happen? because as versions of packages are upgraded the developers of those packages will change things. as far as your specific issue, I have no clue which python package is the issue, but maybe just try and upgrade insightface since it was referencing that toward the end and see if that helps

so like pip install -U insightface

it could also be wanting an older version, so another approach is to look at what package versions reactor wants, if there's documentation to that effect

[flat dirge]

I believe this is complaining that you are missing face embedding because you have nothing connected to face_model input.
https://github.com/Gourieff/comfyui-reactor-node?tab=readme-ov-file#main-node-inputs

[grand pike]

hey peoples. so i did an "update all" in comfyui manager and now its broken. the URL just loads to a white screen in my browser. this is the 2nd time this has happened this week. what do i do so this stops happening? can i fix it?

[shell sonnet]

I'm also curious about how to edit the command line args in comfyui, I'm using linux to run it so the answer of "oh check the windows.bat file" does not apply. I also checked in the main.py file and I could not find the args it's trying to pull from. Anybody have an idea about this?

[merry ermine]

here you go https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cli_args.py

[shell sonnet]

Thank you, I'll look into that.

Ah, I was over complicating it.

Literally just.... add em to the end of the main.py command,

I'll just make a bash/python script to do that for me.

[tough hornet]

Hello, I'm trying to reproduce this effect:https://www.instagram.com/p/C_La6hMoKsA/

I suppose I can do this with animationiff, but I don't understand how I can do it so that the background remains the same and that only the characters move?

Does anyone have a clue or a tutorial to give me? Thanks! 🙂

Instagram

[shell sonnet]

Ok, I have everything working and able to generate images as intended. However the workflow is spazzing out constly with huge varriance in fps

it's bad enough to the point where it's almost impossible to do anything to the workflows.

(however everything else runs perfectly fine in terminal.)

[livid gust]

@grand pike The white screen happened to me, using DuckDuckGo. Try visiting the same URL (http://127.0.0.1:8188/) using Chrome or Edge, they may still work. But I agree this latest update is a blunder. My fix was to pull down the portable version and rebuild from scratch. I had to abandon the update and use the old stuff.
Make sure your Manager's "Skip Update" checkbox is checked.

[merry ermine]

avoid Brave browser in particular

for comfy

[shell sonnet]

Using firefox on linux,

[merry ermine]

firefox could also be an issue

chrome or edge recommended

[livid gust]

I think it's a Comfy issue because the previous version worked fine on the same browser.

[merry ermine]

okay yeah it was just a suggestion

it might be the case that the method of adding flags was not right?

would ask on the comfy discord its 10x more active than this channel

[shell sonnet]

what's the comfy discord? I have not seen it.

[merry ermine]

will dm you the link

[livid gust]

I'm not too concerned at the moment. The old stuff works fine. I only have an 8Gb card, so the new stuff is out of reach anyway. Saving up for a new card, though...

[merry ermine]

8GB is absolutely fine for the majority of stuff

I personally use cloud so there is not a fixed VRAM size, and I did a lot on 8GB servers at first

[livid gust]

8Gb is severely limited for complex work using AnimateDiff alongside the IPAdapter. I agree 8Gb is okay for stills. I'm more of an animator.

[merry ermine]

oh I see

sorry yeah that's gonna be rough

[shell sonnet]

Thanks for the link Neon.

Edge is not an option for me, so I may try chrome?

[merry ermine]

yeah

in general chrome is the most reliable browser

I think its best to not focus too much on one browser

cos then if you need to switch it feels annoying

so I install like 20 browsers these days

sadly I have to switch a lot, stuff doesn't work

[shell sonnet]

I dont mind using extra browsers for specific things. I just like firefox for the extra security

[merry ermine]

yeah I use brave as my main for similar reasons

or Ungoogled Chromium

firefox didn't work with my monitor

oh except on apple products I use safari

not a big apple fan though

[silent cargo]

yoo so im using upscaler a lot on comfy ui but somehow the eyes of my character arent that great i wanted to ask if there isnt something like a highres fix inside comfy?

[lone finch]

can you share an example? does the eyes seem like lowres? like the pupil not fully circular?

i had that issue with quite a few of pony model.

[silk aspen]

thanks a lot for that info

[silk aspen]

thanks I will check it

[forest birch]

How do people achieve video with flux? Or are those just image to video with SDXL motion models?

[shell sonnet]

just in case the info does help, yes, firefox has an issue with comfyui ATM, Here's how I got it fixed. If you go to address bar and type in about:config

and then change this entry in particular from false to true.... standard default one. change this gfx.webrender.software false
to this one belowgfx.webrender.software true

It will turn off browser acceleration and will have minimal if noticable impacts, but it will fix the spasming/flickering on the firefox version of ComfyUI

Just posting this here just in case anybody else will have this same exact issue.

[silent cargo]

yes sure
maybe my standard is too high but on stable diffusion with highres fix i bring out better eyes than with upscaling on comfy

[silent cargo]

how high are you using denoise on the upscaler? that could be something im doing wrong, it just feels that everything over 0.15 is destroying every single hand at least xP

[merry ermine]

ah thanks, funnily enough I had turned it off already due to previous issues with other software

[shell sonnet]

np

[wispy kelp]

so the cogvideo imageencode node is just for show and only yet to be implemented? Coz it doesnt work

[merry ermine]

I was under the impression people had this model working in comfy

but possibly using a different node

this stuff is so bleeding edge that I think its best to just go in expecting things to not quite work without tinkering

I quite regularly find errors in node python code, its just worth baring in mind

it helps loads when they comment it saying they are not sure about that bit

its often not their fault cos Arxiv papers are so brief sometimes its tough to copy the methods

[livid gust]

@silent cargo Denoise is horribly misnamed. When denoise is 0.0, your output is a copy of the latents you supply. When it is 1.0, it is the prompt result. Denoise should be called Crossfade because it has nothing to do with noise. It allows you to blend between a supplied image as latents with the requested prompt. A value of 0.15 will result in mostly your input image (assuming you're not supplying empty-gray latents). To go higher than 0.15, I typically find 0.45-0.7 is a nice sweet spot, will introduce more prompt influence. What you can do is "fix" the seed, and then try different seed numbers until you find a "good" hand output produced by the prompt.

[steep marlin]

It has everything to do with noise and removing noise. When set to 1.0, it adds the maximum amount of noise via the sigmas, which means it's a completely new image. When set to a number like 0.5, it means it adds roughly half of the sigma noise schedule to things and when set to 0.0, it means no noise is added, therefore no changes happen. So if you're doing an img2img workflow, 0.0 means you get your image back, 0.5 means you get a good amount of changes to the image that follow your prompts and 1.0 means the base image has no real influence on the final image.

[livid gust]

I guess I never thought of the prompt image as noise, but I guess you're right, the prompt shapes the noise.

[steep marlin]

sort of, but it's a bit more complicated than that

point is, denoise is the correct word for it. is it elegant? no, but it does exactly what the name says.

[lone finch]

ngl, i think it just might be model issue. isaw from your workflow you use a bunch of lora. could be one of them. have you try using another model and see if it still have blurry eyes? i tried it on mistoonpony with no lora and such and imo the eye came out good enough.

[silent cargo]

wow that looks amazing, Thats a good idea maybe all those things happen due to that

[lone finch]

ngl, i recently try out this model and i fkn love it. the eyes come out so nicely.

try it out

[silent cargo]

thanks will try that 🙂

[silent cargo]

do you maybe also know why my Adetailer-settings sometimes giving overlaped faces? I'm abit overwhelmed with it atm

[steep marlin]

The image size is probably too big for the model version you're using. Seeing weird duplicate stuff like that is what happens. Like sd1.5 doesn't want to go too far beyond 768x768, so if you're making it work on some 2048x2048 image, you'll see stuff like this

At least that's my assumption on what's going on there based on the limited information

[silent cargo]

im using xl atm at 1280x720 that should be actually alright, right? (here would be an example if you would want to take your time)

• could it be that i need to increase the guide size on facedetailer?

[ivory warren]

Try to remove all the negative prompts and also try with 1024x1024 with at least 22 steps. Careful of the cfg as well. use something like 5.5

[silent cargo]

ty, will do

[silent cargo]

if i would try with 1024x1024 latent how would i get a 16:9 image as output, does the upscaler with an enough high denoise do that for me already?

[ivory warren]

Explicitly set the output dimensions to 1920x1080 (16:9 ratio) if it is an SDXL model. That way you remove the doubt.

Also, for your case I'd give these samplers: k_lms, ddim, ddpm and play with your CFG
` CFG 2 - 6: Let the AI take the wheel.

CFG 7 - 11: Let's collaborate, AI!

CFG 12 - 15: No, seriously, this is a good prompt. Just do what I say, AI.

CFG 16 - 20: DO WHAT I SAY OR ELSE, AI.

`

[silent cargo]

thanks for the help I will experiment on that one later on again with all of the tipps i've got

[livid gust]

Generally, you don't want to exceed half as much as the model was trained on. For SDXL, that's 1024. 1024+512 is 1536. So the highest you should request in resolution is 1536x864 to remain in 16:9 aspect ratio. Beyond that you may experience doubles in your output.

[silent cargo]

you mean in the empty latent image right?

[ashen agate]

Does ComfyUI manager work with swarm?

[distant lagoon]

definitely

I'd say that is largely its intended configuration

manager is always the first thing I install after comfy, hard to imagine life without it

[ashen agate]

Oh good, I'm brand new to Comfyui. Where do I find manager, if you don't mind pointing me the way

I mean to say... is this the best source? https://github.com/ltdrdata/ComfyUI-Manager

[silent cargo]

@livid gust @steep marlin @lone finch and @ivory warren thanks all of you for all the help time and tips, now I'm actually way happier with the outcome 🙂

[livid gust]

@ashen agate Yep, that's the only source I know of.

[sick badger]

That's the one. Once you get that up, you can do all almost everything else in the manager gui

[ashen agate]

Yeah, it's good, I got it installed. I won't be diving too deep into it right now, because I'm returning the PC I just got. The return has 4070 Super, the new has 4070 Ti Super. heh

Bestbuy has a good online PC deal.

[lone cape]

Finally I integrated Flux to my workflow with "concept selector" node. Several params for flux available, but as I tested different public workflows and settings, the negative prompt has no effect to the image if all other params (noise, size, seed) same. Lot of workflows trying to use neg. prompt, but no difference if changed or deleted. Using several unet models (official, gguf) have very similar results if the seed, sampler and noise freeze. Is it right that the negative prompt have no effect including if I use ksampler instead of "official" wf advanced?

[distant pendant]

both flux models are distilled and don't use negative prompts

[shell sonnet]

Is this a bad way to try to upscale an image?

If yes, what should I do to get better results?

[gilded jetty]

Does the red background not hurt your eyes?

[sly sundial]

Meanwhile, my nodes... 🤣

[gilded jetty]

Save your eyes before its too late

[sly sundial]

https://tenor.com/view/sunglasses-on-shades-look-stare-double-sunglasses-gif-17838198

[lone cape]

No problem with any colors 🙂

[lone cape]

As I see after long test, GGUF workflow working well with neg prompt using cfg 3, huge difference and relative good result if changed (or deleted) neg prompt only. Continue with F1.dev workflow + ksampler.

[ivory orbit]

Hi 👋
Does someone know how to fix anime full body pictures in ComfyUI (faces, hands, ...)?

[merry ermine]

impact pack face detailer and hand detailer are pretty much plug and play for this

[ashen valley]

Hey

im kinda having difficulties with flux

i have 64 gb of ram and an rtx 4070

it takes 13 years to generate 1 image

and that image looks.... interesting

do i have a mistake?

should i use a different sampler?

or is sth wrong with my clips?

[distant pendant]

set your cfg to 1 for starters, that will make it less baked

[ashen valley]

i got this now

in the diffusion loader

i forgot to change from default

to fp8

[distant pendant]

also probably wrong sampler for flux, but I mean there is an example workflow for flux, just use that

[ashen valley]

is fp16 better?

[ashen valley]

which one do you recommend

it generated this image (notext) in only 4 steps

[distant pendant]

the same one that the example workflow uses 😔

[ashen valley]

ok thx

[ashen valley]

this is the one

[wicked urchin]

Hi, im using the Krita diffusion plugin with ComfyUI backend. I try connecting to my local server but it says some nodes are missing (just updated btw)

tried installing thru github url but i get these errors

[wicked trench]

Hi there, anyone can help me to connect 2 workflows together I’m a bit noob

[thin crypt]

hey anyone can help me to solve to solve this error

[lusty pewter]

The message is pretty straightforward. You don't have the config.json in the directory it shows at the top.

[vestal topaz]

how to install flux, i mean i have its workflow but have no idea with this, can anyone help?

[vestal topaz]

how to let comfyUI show this to me?

[wicked trench]

Hi there, anyone know how I can automatically detect a part of a body then insert a text on it with a natural effect like an handwriting effect. With a1111 I use Adetailer + inpainting + Flux, it’s working but the area detected change too much, I would like to preserve the skin without modify it.

[thin crypt]

config.json file already have in folder but i did not recognize

[livid gust]

@vestal topaz That might be part of ComfyRoll node set. Try searching in the Manager and instal it.

[sly sundial]

It is from a custom node pack, but not ComfyRoll. I have it but no idea which pack it is from now 😦

[desert willow]

Hey is there an video or instuctions which can tell me how i can make videos using comfyui??

[humble harbor]

Hi. Does anyone know a node or a basic workflow so that when uploading an image or a photograph, its dimensions are transferred to the latent image?

[timber depot]

Hey comfy people, I published a little custom node set which might be interesting for you, they send a generated image directly into a specified Discord channel. https://civitai.com/models/717861

Soon also in manager 😉

[timber depot]

this is possible with get image size node and plugging the output into dimensions for upscale node, transfer them to input from widget.

While I'm here 😉

[merry ermine]

this is really fantastic

could be useful to fill a private discord with generations

[timber depot]

I use it when I am generating images with friends, or for my DND group, so that I don't need to always post them manually^^

[merry ermine]

would be fun in a group chat yeah

[wispy kelp]

Flux Openpose when?

[obtuse latch]

Hey lads, sorry for cross posting just looking for some suggestions on image blur: https://discordapp.com/channels/1002292111942635562/1089974139927920741/1280731128705192009

[humble harbor]

I'll try it right now. Lots of thanks 🙂

[silent cargo]

yoo guys whenever im using the ultimate sd upscaler on flux the upscaling restart after done rendering without giving out any image, its very weird every tip is highly appreciated

thats my workflow

[livid gust]

@humble harbor You can load images from a folder, connect that to get image size, then convert the WIDTH/HEIGHT from your latents into inputs to connect them. Then your latents will match your image size. When I'm doing this, I often abandon the "empty" latents and just use a VAE Encode to produce latents in the form of the source image.

[humble harbor]

Yes, that is what I have finally done, but applying an upscale latent between the aspect ratio node and the sampler. I have also applied a second image resize in series with the first one, to reduce the weights. Thank you very much @livid gust 👌

[verbal sky]

New video just dropped (including new intro) with advanced inpainting, compositing boosts, background removal and more! - https://youtu.be/9bwhoiYR920

[fading mural]

my comfyui suddenly doesnt wanna work

[vocal osprey]

https://github.com/crystian/ComfyUI-Crystools

[fading mural]

anyone else？？

or is it just me

at least tell me that

[flat dirge]

If you are using smZNodes you need to update it

[fading mural]

I am、I clicked 「UPDATE ALL」

still there

wait no、I don't use smZ nodes anymore

[flat dirge]

Okay I thought maybe you did because your error was the same as appears in two issues on ComfyUI_smZNodes. https://github.com/shiimizu/ComfyUI_smZNodes/commit/a7573b50970f0766dcc4cd5784cd15ccf4b085fe

[reef kite]

yo how to use different loras in 1 image is there an extension

https://civitai.com/images/27308396 like this they used 2 loras

so like i have only used 1 lora and 1 checkpoint cuz i read if I use 2 it wouldn't work is there like an extension?

[distant lagoon]

you can use like 10 loras

[reef kite]

yea but like how can i just put them all in the prompt ?

like i click on them and use the trigger words

?

[distant lagoon]

but sometimes you can burn out certain weights if the loras are doing similar things

so you might have to play with the lora strengths

so if you had a lora that preferred blonde hair and another one that preferred red,, what happens?

[reef kite]

yea but i want like actions like idk "smug face" theres a lora for that and the character itself idk could be anything i can just do that? like just add both without an extension?

[distant lagoon]

you can certainly do that

hold on, I got you...sec

[reef kite]

alright thanks man

https://civitai.com/images/858851 so heres like another example they used a lot

[distant lagoon]

grab that node

rgthree, load as many as you like

it works with any of the models as far as I know, though I've only personally done XL and flux

[reef kite]

alright ima look into it

cuz honestly looking at the image i dont quite understand it

is this like an extension?

i need to download that "rgthree"

[distant lagoon]

it's a custom node, you're in a comfy room so I'm assuming you know comfy

[reef kite]

yea

[distant lagoon]

so in comfy manager you would search rgthree

[reef kite]

oo alright bet

[scarlet rain]

Can I train a lora on Flux dev and then use it with a ComfyUI workflow using the flux nf4 model? I don't have a workflow for flux nf4 plus Lora yet.

[wicked trench]

hi, very nood question for comfyui, but how convert a captionning ouput ( clip text encore ) into text/string ?

[distant lagoon]

can you be more specific about what you're trying to do? where is the captioning output coming from and what are you planning to do with the string?

[scarlet rain]

Is it possible to use Flux nf4 with a lora in comfyui? Connecting CheckpointLoaderNF4 to LoraLoaderModelOnly gives me an error ".to() does not accept copy argument"

[scarlet rain]

Looks like the answer is not yet. I switched to GGUF and it's working, but slower than nf4.

[distant pendant]

you can use loras with nf4 in forge, comfy is a no go

[scarlet rain]

I haven't had much luck in Forge nf4 with Loras. My custom one doesn't work properly there for some reason. I've updated forge.

[distant pendant]

don't use it myself much but illyasviel boasted that loras work so idk

might be just some issue with key names

[scarlet rain]

I downloaded a random one off civit.ai and it seems to work better than the one I trained. I'm able to use my trained Lora on Flux GGUF in Comfy, but in ForgeUI, it just doesn't look like it's using the Lora at all until I pump it up to weight of 2.0 - 3.0 then it just starts messing everything up after that. I'm using a Lora weight of 0.9 - 1.4 in ComfyUI

[wicked trench]

Hello, anyone know how to resize + reshap a mask without touch the erosion settings ?

[sudden saffron]

Hey need some help.
Does anyone know how to create an XY plot for Flux ?

[flat dirge]

I know you can create plots for Flux with this, https://github.com/cubiq/ComfyUI_essentials

[sudden saffron]

Thank you will try that

[silent cargo]

does anyone probably know what checkpoint that is?

[sudden saffron]

LoRa for a "my hero academia" for the froggy girl + upscalers , if this is not cropped and it's not a horny picture, can't guess, if it is though, means pony

[silent cargo]

ty what pony would bring out such eyes, do you maybe have something in mind that could fit to that?

like those

[sudden saffron]

a-detailer for the face, maybe some lora-s for expressions.

If you'll check Civit you'll see there are 3-4 lora's used :
1 style
1 character
1 additional style additive
1 pose/expression/action

[silent cargo]

how did you saw that?

thanks for the information too

[wicked trench]

Hello, any expert with comfYUI ?

[raw kiln]

i would like to request assistance from someone to help set up nodes in comfy-ui

[sudden saffron]

On the right hand side of the image -

[silent cargo]

oh wow can you maybe send me the link?

[fallow quartz]

guys any idea why it is ignoring my prompt completly and generating random images???

[silent cargo]

try at the beginning of the prompt to put in "1male"

[fallow quartz]

its ignoring the prompt so puting 1male wont change anything, i cant put whatever, it just generating random images

[silent cargo]

for flux i always used this text encode:

try it out with that

[fallow quartz]

yeah i just when for the clip text encode classic promt then adding a guidance node and it works...

[silent cargo]

noice

[fallow quartz]

but its weird because swarm ui does use the ClipTextEncodefLUX in their worklfow

[silent cargo]

my computer is too laggy (atm) for flux so i really dont have a lot experience try to go to #🤝｜tech-support cs1o is a genius in all around ai

[sly sundial]

Copy your prompt into the t5xxl field as well

Your empty negative is a waste of compute time, use a conditioning zero out node and feed your positive to that and then connect that to your negative in ksampler.

[flat dirge]

If CFG is set to 1 it will already ignore anything sent to negative conditioning

[sly sundial]

It's not about ignoring it, but it will still process the node and take more time.

[fading trench]

anyone else get a cast_to() error from custom/Ksampler when using GGUF? I updated all nodes and UI :S

[merry ermine]

yes but it still calculates it

not 100% sure, as its possible there is a thing in comfy code to detect this issue and switch off neg prediction

[lone cape]

Finally I successfully modified KSampler for Flux. My main problem was, that flux was less "creative" than SD/SDXL, no "hallucinations", so for same (or little different prompts) I got very similar results while the seed changed. Now I play with noise and the variation levels now much higher, much more difference between images generated by same or similar prompts, but the bonus result that if seed has been frozen, the image can be variable by only noise. Long test running, result and node out soon, I have to check the code with very several prompts. The variation method will be useful for SD/SDXL models too.

[merry ermine]

sounds good

noise injection is great yeah

[wispy kelp]

Perkele.

[lone cape]

At second row I used. Same prompt.

[merry ermine]

yeah the noise helped

[obtuse latch]

I'm using a base checkpoint for generating then another checkpoing for upscaling. When using an upscaler, is it fine to use the clip from the inital checkpoint? Or is it best to have a seperate clip connected?

[distant lagoon]

you didnt mention which model, plus why ask...just try it and decide

looks like XL based on what I can see in the image

[merry ermine]

yes it will work

[obtuse latch]

I know it works, I'm asking if it's the best way to do it: Or is it best to have a seperate clip connected?

[merry ermine]

I think its better to have a seperate clip

[obtuse latch]

Thanks, will do that.

[merry ermine]

because if you add in a lora or something and forget

then it will be a problem possibly if it affects the wrong one

if you are very low on vram then it could be better to reuse

[obtuse latch]

Sounds good, the only reason I ask is because:

[merry ermine]

wow awesome

[obtuse latch]

Last question, I'm getting some pretty large differences in the upscale, any suggestions would be appreciated.

[merry ermine]

lower denoise is the main thing

also canny or depth control nets

[obtuse latch]

Hate to bug you more, is there anything specific you would reccomend to look into this?

Just going to read some docs

[merry ermine]

there's a really famous guide here: https://stable-diffusion-art.com/controlnet/

[wicked trench]

Hello there ! Anyone know how to blend a text into a skin with comfyUI for it's look smooth and natural ? Thanks

[wicked trench]

Hey there, anyone know how to center and fit a text into a mask ? My text also go outside the mask ( so is incomplete ) 😦

[bleak stump]

https://www.youtube.com/watch?v=9DbRJDitVhA

[lavish stone]

i can't seem to use loras in comfyui, i tried different lora loaders and they simply don't do anything, when the exact loras work on forgeui. I even downloaded the xlabs nodes for the lora and they give me "list index out of range error". Is there anyone who is using gguf flux dev and is able to use loras in comfy ui ?

[lone cape]

Automatic noise increaser, 9 steps, the very first pic the original without noise injection. Cascade, same prompt, and same seed.

[wicked trench]

Hi there, anyone how to modify the size of a text generated from flux and also, how center this text ( generated into a mask ). I created a rectangle mask that I place into a image then inpaint the content of this rectangle with a text but the issue is that 90% of the time the text or is too big or generated at a corner of the mask ( so is cropped )

[merry ermine]

Anyone else encountering high load on the GPU when having comfyUI idling?
18% with spikes to nearly 50% and perma bus traffic

Switching to another tab:

18% from a 4090 to render items on a screen?

[distant lagoon]

I do feel like I hear the fans kick on, but I haven't done the analysis like you have. If I remember I'll check later

[distant pendant]

it's probably because of canvas acceleration

when I switch to another tab my gens are faster

[lone cape]

I thinking how to exactly reproduce the original image if latent/noise/sampling modified internally. So I collected required generation data to tuple (the original is much larger, but not too interesting because contains prompt, sampler and scheduler, cfg, steps, model name, etc.... But here is the part what is important to read back and apply:

    "latent_data": {
        "noise_type": "white",
        "alpha_exponent": 1.146,
        "modulator": 1.36,
        "device": "cpu"
    },
    "sampler_settings": {
        "denoise": 1.0,
        "variation_extender_original": 0.0,
        "variation_batch_step_original": 0.0,
        "variation_level": false,
        "device": "DEFAULT",
        "align_your_steps": false,
        "noise_constant": 0.03,
        "variation_seed": 610662911819074
    },```

Non "standard" solution, meta reader needed and nodes to apply these settings.

[wicked trench]

I get cuda error memory on clip encoder node since last the update....

[merry ermine]

Hardware acceleration is disabled in the browser.
I uninstalled Comfy now, i don't need a software which spins fans for nothing.

[lone cape]

New ideas on my image saver node. The first if you use aesthetic scorer, you have trigger for save. If the scorer under 500 in the example, image will be ignored. The another is include model concept to path. for example Flux with upper settings save result like this: h:\AIPICS\Comfy\FLUX\FLUX1-DEV-Q4_1\Coffee\2024-09-Week-37\Image_20241009_120412_683550649516920_2400x3512_1.jpg (Coffee subdir come from the prompt input as preferred path instead of Dev)

[merry ermine]

this idea is great

using the score to decide

[lone cape]

Yes under 500 images no need to save, but a.scores not always correct for image quality. I already using scorer for my detailer to ignore if refined hands worst than the original 🙂

[merry ermine]

Good idea, in the past i used ollama vision

[lone cape]

I using this: https://huggingface.co/primerecomfydev/chadscorer/resolve/main/chadscorer.pth

[merry ermine]

I uninstalled all that comfy stuff for now, but thank you 🙂

[wicked urchin]

I'm getting an error at the Apply Contronet (Advanced) node. Here's the logs, I'm also using PonyXL checkpoint with controlnetxlCNXL_bdsqlszLineart

[thin moss]

people i become mad... i got currently on ksampler and also ultimate sd upscaler the error " ComfyUI Error Report
Error Details
Node Type: UltimateSDUpscale
Exception Type: AttributeError
Exception Message: 'NoneType' object has no attribute 'model_options'
Stack Trace" and i cant find anything about that error in the hole internet... can anybody help?

upscaler is upscaling ubt at the end there is this error

i try this workflow

https://www.youtube.com/watch?v=j-9m2hOcyiU

[livid gust]

None Type errors in general might mean you have a model specified in a node that doesn't actually exist on your system. This can happen when you open other people's workflows. Use the manager to browse for missing models, or type the model name into google to find it online.
Pay careful attention to names, because they have to exactly match.

[thin moss]

i will check out this

thx

can you see.. i can chose the models and it start rendering but then the error

[sly sundial]

Your tile width and height is way too high and you only have 1 step. Did you do this intentionally?
Can you select a checkpoint model (follow the purple line)?

Tile size depends on your GPU, but try 1024x1024 or lower.

[grizzled zinc]

Is there a way to use multiple images directly as prompts for Flux Dev? I understand the method of generating text from an image and using this as a prompt, I also understand connecting the image instead of the empty latent. But I'm after a more direct method of using multiple images to influence geometry/colours/style etc combined with either other images or a text prompt.

[merry ermine]

this is IP adapter

sadly there is not a very good IP adapter yet for flux

you could try this one https://huggingface.co/XLabs-AI/flux-ip-adapter

[grizzled zinc]

Thanks ! It’s all beginning to make sense now.

[maiden wigeon]

Hello, I am looking for advice. I have created a Discord bot that allows me to use SD models for different types of images. My issue is that when I switch from one model to another, ComfyUI restarts loading the new model from scratch, which is particularly limiting. I have already installed my models on an NVME and i'm running 32 gb of ram and 11gb of vram, but it still doesn't satisfy me. Is there no software solution to speed up model loading time between two switches? Perhaps a caching system or special nodes? Thank you!

[lusty pewter]

You are likely running out of VRAM. Keep in mind that you still need enough for your system to load the text model you're using and to do anything else required by VRAM. So when it doesn't have enough, it's gonna dump the model and load the next one.

[maiden wigeon]

hello and thanks for your answer, so with 24bg vram cards, comfyui doesn't dump models anf load the others ? if you haveenough vram you don't unload the model even if you change for another ,

[sharp raptor]

Hello, what causes this repeated tiling issue when upscaling?

workflow is embedded in image

[livid gust]

Couldn't find a WF in your image, but generally, tiles and repeats happen when you exceed the models width. You can't upscale latents past more than 50% of that value. Keep your latents in bounds, and upscale the final pixels instead.

[lusty pewter]

Why would it dump it if it's using it? Ideally, you'd want to keep as much in VRAM as you could, within reason.

[dry rock]

you re using a positive prompt describing the whole when upscaling. That s the problem

most likely you re upscaling with some sort of tiled method.

That s why you can see multiple "1girl walking a dog in forest"

one per tile of your upscaling process

Instead use no prompt or something very vague that could apply to every tile

Or decrease the denoiser so it deviates less from the original.

what do I mean by "tiled method". it cuts your input image in multiple sub images / tile, upscale each of them and then stitch everything back together.

[sharp raptor]

Yes this is what I thought was the problem. So I tried to replicate the issue but for just one tile.

this is what the result was with a img2img workflow and denoise at 0.35. Which is the same denoise I used in ultimate SD upscale

Size and amount of the tiles

@dry rock

[bleak stump]

New experiments + FLUX & SDXL LORA! - https://www.youtube.com/watch?v=acB5gA9WyPQ

[solid drift]

is there a way to iteratively load images in comfyui?

[distant lagoon]

yes

look up the was-node-suite-comfyui

[silent cargo]

Hey everyone! Is there a way to automate a list of character names through a specific part of a prompt in ComfyUI? To clarify, I have 20 characters, and I want to generate one pose for each. Since I'll be frequently changing the poses, I'm wondering if there's a way to automate inserting the character names into the text prompt node, rather than manually typing them each time. Any tips or workflows to streamline this process?

[winged hazel]

Hey guys, try this interesting FLUX model

https://civitai.com/models/741194/f1-everything-is-ice?modelVersionId=828874

[sharp raptor]

Something is wrong with my flux inpaint workflow
My lora is not being applied to masked area, even at denoise 1
low denoise produces a copy of my mask while it should just change the base image with a low denoise str
What is the problem?

[dry rock]

there's still 64 tiles and a positive prompt that is applied to each of them, run with an empty prompt and/or lower the denoiser even more

[livid gust]

@silent cargo Think of it as a batch with a length of twenty, then prompt schedule each frame with a different name...

"1": "picture of ellie",
"2": "picture of mike",

etc...

[silent cargo]

how does this work, I'm a bit irritated ?

[livid gust]

It's part of the FizzNodes. Try dropping down a Batch Prompt Schedule. Fill in the frame number and your prompts. Drop it in place of your current CLIP encode. You will end up with a batch of images at the end. You may have to save it as a video and pull out the frames after render. You can also increase the frame gap, which produces an interpolated version of your prompt, allowing the creation of a blend between "Mike and Ellie."

[pure siren]

gm

[mossy spire]

Hello, does anyone have any tips on smart ways to speed up the image generation or ways to reduce the VRAM a model either uses or settings in the system that might effect the vram?

[fading mural]

i use Automatic CFG, it speeds up generations by 30% on low end gpus

[mossy spire]

huh. that looks neat. i got somewhat limited vram with only 10gb not sure how much that will mess up stuff

is there any good place to learn about super technical stuff like CFG?

hmm interesting. it seems like comfy ui is using about 4GB of VRAM without having an active task running

[livid gust]

That seems high. If you have only one video card, then all open apps are using your vRAM. Make sure to close others down.
I often drop the GPU Clean node after a GPU Intensive step completes (AnimateDiff, IPAdapter, Controlnets, etc...). This can free up GPU memory.
Use a LCM model. It resolves fairly quick, within 5 steps. Other models can require 9-21 steps.
Don't overdrive the network with excessively large image inputs. Make sure their size matches your model's maximum resolution.

[ashen agate]

Can anyone help guide me to some outpainting and upscaling nodes?

Or simple workflows with that in them. I'm avoiding adding too many more nodes, certainly being careful now, as some of them keep breaking my install.

[lone cape]

Hi, I testing custom nodes (what is far from "standard" solution what compatible will all previously generated images and nodepack) to get back the image if play with noise/clipping/other settings. I have saved metadata like this:

    "latent_data": {
        "noise_type": "blue",
        "alpha_exponent": 0.5,
        "modulator": 1.25,
        "device": "cpu"
    },
    "prompt_encoder": {
        "clip_model": "ViT-L-14-BEST-smooth-GmP-TE-only-HF-format.safetensors",
        "longclip_model": "Long-ViT-L-14-GmP-ft.safetensors",
        "clip_mode": true,
        "last_layer": 0,
        "negative_strength": 1.2,
        "use_int_style": false,
        "adv_encode": false,
        "token_normalization": "mean",
        "weight_interpretation": "comfy++",
    },
    "sampler_settings": {
        "denoise": 1.0,
        "variation_extender_original": 0.0,
        "variation_batch_step_original": 0.0,
        "variation_level": true,
        "device": "DEFAULT",
        "align_your_steps": true,
        "noise_constant": 0.1,
        "variation_seed": 203043622053871,
        "batch_counter": 1
    },

what reproduce the exactly same image (and can be modify any saved settings for small difference). This is saved as meta, so have to write and read back with special nodes, but at first test working well. The bonus, that this meta contains network settings (I deleted because too long), so if forgot or changed, the process will use the original settings for Lora/Embedding/Lycoris/Hypernetworks reads from image and ignore workflow settings.

[forest birch]

Hoi, any idea why it errors like this when it's just one image i'm feeding it?

[livid gust]

You have to plug a VAE line into the decode. It's missing the input.

[forest birch]

Forgot to note i use this :P

As the second encode fails, but first gen does it just fine

Same error with manual wire

Here's the workflow.

Quite messy, but it's just to get the thing working. As this at least worked just fine for SD 1.5 to make literally minutes long videos by taking last frame of the first gen as target, first frame as reference, then shooting that over to the next sampler, and repeating over and over.

@livid gustFound the issue! It loaded from the wrong VAE :D

[forest birch]

Can not for the life of me find out why it just butchers the visuals on each further generation of the last video's last frame

[untold haven]

If i import an image made with automatic1111 in comfyui should i have the same output image?

[shell sonnet]

You may need to elaborate a bit more. I'm not sure what you're trying to ask.

[wicked lantern]

hello! Where is the best place to ask a question about ComfyUI and IPAdapterApply?

I"ve been trying to get this workflow going - apparently it's like Magnific: an upscaler with 'zing'. This one: https://github.com/fictions-ai/sharing-is-caring?tab=readme-ov-file
But after all is installed, I get the attached warning.
// Warning: Missing Node Types.
// When loading the graph, the following node types were not found:
// IPAdapterApply
// No selected item
// Nodes that have failed to load will show as red on the graph."

I've tried these links and fixes but I haven't got it right yet:
https://www.reddit.com/r/comfyui/comments/1dedkc9/ipadpterapply_missing/
https://comfyui-guides.runcomfy.com/ultimate-comfyui-how-tos-a-runcomfy-guide/how-to-fix-a-red-node-for-ipadapterapply
https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/123

I would really appreciate a little help. I'm not great at technical instructions although I do try following them step by step.

[livid gust]

@forest birch From your short image series, it looks like your CFG is increasing per-frame. Are you animating it? Try to make sure it's a constant.

[forest birch]

Interesting It's SVD image to video generation with the new SVD_xt_1_1. So CFG is the culprit? What should i do to make it a constant? Never had this happen before.

Cause so far, i've set it up to make 4 seconds long, but the CFG if the culprit, screws it over lol.

And do the seed have to be different for each continuation? AS i noticed if i use the same seed all over, first video will be fine, but second video will be a animated still. Where there's slight variation of colors/hue, but subject and background is a still image.

Here's the workflow i have it at. Basically like the multi gen sequence i made for animatediff for SD/SDXL

[forest birch]

Oh wait, maybe this one is the culprit

[sharp topaz]

I need your help :// I use ipadapter with att. masking for Renderings using a color map, and I have many (10+ masks). it works great except for this:

• If I use the nodes from Cubiq, I can chain them normally, and bypass the ipadapter related to the mask that I dont need to use. BUT the workflow gets exponentially heavier to run when I add more and more ipadapters.
• If I use the Inspire "Regional Ipadapter" nodes, it seems to be more optimized (it handles more regions easily). But there is no way to bypass or mute an independent region. If I do this, the "Apply Ipadapter Regions" doesn't run, since it is missing an input.

[dire sapphire]

Anybody have difficulties to install IPAdapter/Instant ID/face ID/ReActor to comfyUI ?

Some nodes are installed other not, I can't understand how this thing works

[distant lagoon]

if you're using comfy manager, sometimes the nodes are not known to manager and you have to install them by git url

so assuming you found a workflow containing unknown nodes, you'll need to figure out where those nodes came from

[bleak stump]

https://www.youtube.com/shorts/W6xmpX-fdmg

[lapis plank]

hello

[sharp topaz]

plsss :- )

[mortal ferry]

Hey guys, anybody know how to solve this ?

[livid gust]

Is there a comfy fix for this? I'm using SDXL, which I thought fixed some of these SD 1.5 problems, but it looks like it performs the same way. Even when pushing negatives, I can't get a decent embrace.

[merry ermine]

exactly this https://github.com/Danand/ComfyUI-ComfyCouple

[livid gust]

Thank you. Going to try it out.

[livid gust]

ComfyCouple didn't actually work for me. It does install, but that form of prompt blending didn't help much. I basically had to do away with all resampling and run a highly tuned prompt through a single KSampler.

[drifting linden]

ultraman

[rich agate]

Hey this is more about stable diffusion, but i tkink its suitable for comfyui.. I was wondering is there a way to figure out why and how SD determines if an image or parts of an image should be blurry or sharp and then hijack it and allow the user to specify the bliurry and sharpness? I tihnk IPAdapter does something similar to achieve what it does. I could ask the IPAdapter author if they are interested figuring it out.

[untold haven]

Someone does know the difference between the nodes Freeu and Freeu_v2 for comfyui?

[fading badge]

guys this errors looks massive

anybody now how to solve this?

[distant lagoon]

My best understanding, in the training of these models, there is some level of understanding of photography and the terms used, perhaps in the relationship between the tags used and original datasets trained. Terms, such as bokeh, cinematic, shallow depth of field,fish eye, dynamic angle, etc. At varying levels the models understand these terms and how to recreate these concepts. Now if you dont supply any photographic or lighting details in your prompt, I suppose it just chooses for you. I'm not sure if this is what you were getting at, but that's my perspective.

[rich agate]

Sounds like you are referring to text prompting, which is more on a higher abstract level. But what I am thinking of is more on doing away with that and specifying whether to have 'sharp/clear' details versus blurry details like a slider from -1 to 1. Then you can use inpainting to effect the area. It is clear that stable diffusion can generate very sharp or blurry details wherever. The problem is there doesn't seem to be a way to control it exactly except through text prompting which is a hit or miss.

[verbal sky]

New video just dropped for everyone to enjoy! Flux frenzy speeds, magical sdxl to flux outpainting, text fx and more 🙂 https://youtu.be/pGqes29RuhM

[distant lagoon]

given loras already exist that get rid of the bokeh effect, I'd say what you're suggesting is likely already possible

[wispy kelp]

So do these new flux controlnets like the all in one work with any old flux models on comfyui or only specific ones like GUF?
I am trying to make it happen with the schnell FP8 model but it aint happening

[distant lagoon]

so that wasnt the desired result?

[distant lagoon]

I hadnt played with it in a while, but at least with the xlabs sampler, I didnt have much luck with the FP8 models. the flux-dev model seems to work, and use lots of steps I was doing like 40

[wispy kelp]

I had to update comfyui... it works but its slow. I am back to sdxl speeds on my old card lol

[versed gorge]

Traceback (most recent call last):
File "D:\ComfyUI\main.py", line 90, in <module>
import execution
File "D:\ComfyUI\execution.py", line 13, in <module>
import nodes
File "D:\ComfyUI\nodes.py", line 21, in <module>
import comfy.diffusers_load
File "D:\ComfyUI\comfy\diffusers_load.py", line 3, in <module>
import comfy.sd
File "D:\ComfyUI\comfy\sd.py", line 17, in <module>
from . import model_detection
File "D:\ComfyUI\comfy\model_detection.py", line 1, in <module>
import comfy.supported_models
File "D:\ComfyUI\comfy\supported_models.py", line 5, in <module>
from . import sd1_clip

File "D:\ComfyUI\comfy\sd1_clip.py", line 3, in <module>
from transformers import CLIPTokenizer
File "C:\Users\mb_19\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\site-packages\transformers_init_.py", line 26, in <module>
from . import dependency_versions_check
File "C:\Users\mb_19\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\site-packages\transformers\dependency_versions_check.py", line 57, in <module>
require_version_core(deps[pkg])
File "C:\Users\mb_19\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\site-packages\transformers\utils\versions.py", line 117, in require_version_core
return require_version(requirement, hint)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\mb_19\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\site-packages\transformers\utils\versions.py", line 111, in require_version
_compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
File "C:\Users\mb_19\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\site-packages\transformers\utils\versions.py", line 44, in _compare_versions
raise ImportError(
ImportError: tokenizers>=0.20,<0.21 is required for a normal functioning of this module, but found tokenizers==0.19.1.
Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git main"

I encountered this problem after downloading the manager, does anyone know the solution?

[distant lagoon]

did you try those pip commands at the end?

unfortunately you'll have to start getting better at reading a stack trace

[versed gorge]

Yes I tried it and it doesn't fix my problem. I reinstalled the whole repository and it's good now

[versed gorge]

But now, I am facing with noise images :/

[sick roost]

https://youtu.be/GC_s3f4Nq04?si=zrNmQv9XKN2Ms0-m

[distant lagoon]

you're loading an fp8 model, have you tried it with normal flux-dev?

maybe the fp8 models work, I'm not sure, I've always used std dev

the fp8 clip model should be fine

[polar badger]

How do I automatically mask the face for imparting? Is there any custom node for this?

[alpine parrot]

@polar badger
Impact pack
https://github.com/ltdrdata/ComfyUI-Impact-Pack

It also has sam detector that I find really good at quick masking stuff

[polar badger]

im using setup like this isnt there any better way?

this is bad ik but doing the job for now

[alpine parrot]

its fine, if you need to bump details even more you`re gonna need to upscale base image

[polar badger]

i see, i want this for face swap with my flux lora.

[polar badger]

isnt there something i can do it automatically?

[alpine parrot]

What do you mean? like generate the base image and then run adetailer and save final image?

If so just chain the workflows.
t2i > ksample > decode vae > adetailer workflow >ksample > decode > save image

[polar badger]

alr it worked, kinda ty.

[alpine parrot]

Glad I could help mate

adetailer is fine for quick fix but upscaling to oblivion always best

[polar badger]

i see

[alpine parrot]

you can start with tiled diffusion, one of the easier ones with good results

[polar badger]

can i use flux with this?

[alpine parrot]

yes, it works fine

[polar badger]

alr can you share the workflow if you dont mind?

[polar badger]

i just wanted for simple task for masing the face to swap with my lora

[alpine parrot]

for just faceswap lora I think simple hiresfix would be ok, it wont be able to generate skin texture and stuff

unless your lora has this level of detail

[polar badger]

nah its alr im getting the very great details and face texture

[untold haven]

someone could help me to fix the faceid problem where onnx isn't loaded?

ComfyUI Error Report

Error Details

• Node Type: IPAdapterUnifiedLoaderFaceID
• Exception Type: Exception
• Exception Message: DLL load failed while importing onnx_cpp2py_export:

Stack Trace

  File "C:\ComfyUI\ComfyUI\execution.py", line 323, in execute
    output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

  File "C:\ComfyUI\ComfyUI\execution.py", line 198, in get_output_data
    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

  File "C:\ComfyUI\ComfyUI\execution.py", line 169, in _map_node_over_list
    process_inputs(input_dict, i)

  File "C:\ComfyUI\ComfyUI\execution.py", line 158, in process_inputs
    results.append(getattr(obj, func)(**inputs))
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

  File "C:\ComfyUI\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 610, in load_models
    self.insightface['model'] = insightface_loader(provider)
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

  File "C:\ComfyUI\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\utils.py", line 161, in insightface_loader
    raise Exception(e)

[distant lagoon]

what is imparting?

[livid gust]

@untold haven Have you tried using the Comfy Manager button? Click Update All. (and have a snack )

[untold haven]

yes

[livid gust]

🤷‍♂️

[lusty pewter]

Do you mean "inpainting"?

[distant lagoon]

no idea, this guy asked the question earlier

[versed gorge]

Flux-dev also doesn't work, is there a good workflow for flux model(f8, flux-dev, gguf)+flux lora+controlnet+ hiresfix(optional)?

[distant lagoon]

I have an almost identical workflow, it should work, maybe your model is corrupted? re-download?

[polar badger]

My bad, I mean inpainting.

[distant lagoon]

try segment:face:1 in your prompt

[polar badger]

I'll try. Ty

[distant lagoon]

I use swarm, so if that doesnt work maybe it's a swarm thing

it's part of a tool called segment-anything

[merry ermine]

clipseg is also suprisingly good

its has a higher failure rate but it mostly hits

[broken spoke]

Is there a workflow for https://huggingface.co/spaces/jasperai/Flux.1-dev-Controlnet-Upscaler

[polar badger]

why this dosent work?

[wispy kelp]

I am glad Comfy got rid of those intimidating red errors. XD

This looks friendlier.

[untold haven]

Someone can help me with insightface?
i don't know how to fix some problem
why the dll doesn't load?

[past acorn]

can some expert of comfy help me to set up my workflow

Hi Guys,

I hope you're doing well. When you have some time, I was wondering if you could kindly help me set up my workflow—I'd be incredibly grateful!

Here’s what I’m hoping to achieve:

1)Batch scan images from a folder.
2)Read the images and generate text prompts from them.
3)Use the generated prompts to create new images.
4)Set up a workflow with control net union for  SD/Flux/SDXL. Which one do you think is better?
5)Add an option to toggle seamless mode (horizontal, vertical, or both).
6)Option to remove background perhaps?
7)How do I add 2-3 loras or embeddings to it?
8)Add different stacking styles?

Thank you so much in advance for your help!

[distant lagoon]

those are all possible, but you should figure out where to begin because that's like hours of tutorials that nobody is going to write out in one post

[past acorn]

I think begin wtih saving the prompts

read the image generate prompts from image and saving it

any blogs or links for it would be nice

[distant lagoon]

I use a workflow like that, it feeds wd14 into internlm, I strip off the extension on the image name because I use kohya for training and it wants the caption be the same name as the image but with .txt

wd14 is really good at detecting all the things in the image, but flux likes natural language, so I feed those into internlm to get a natural language prompt. that's the thinking behind it, you can tailor it to suit what you're trying to do

[lilac patio]

good morningl i have a question regarding conditions or may be automated switches in comfyui: i have a workflow for archviz and for input im using a) a rendering with 3d placed people and autodetect them with sam2/florence2 and inpaints them with flux OR b) i make a maskpaint on the original input image and let flux generate a person where i masked it. i have put the respective nodes in a seperate group and by now i am controlling it with rgthree's "fast group bypasser" and an "image input switch". is it possible to automate it and make comfyui choose the right way, for example something like "if there a people found in the input, then take route img2img-flux, else take route txt2img-flux" and disable the other way?

[past acorn]

Thanks will try this

[stiff field]

Hey, what would be the correct way to connect the Eff LoRa Stacker node into a Flux workflow?

[sharp raptor]

I just trained lora and want to test different versions. Is there any node which can load multiple lora but only generate with one at the time to get a X-plot of the Lora versions?

Ik, there is efficiency sampler with XY plot but looking for something useable with flux

[surreal whale]

should be an apply lora stack node, drag out the noodle on the right side and drop it, should get a list of what connects to it

[agile panther]

i just updated my comfy-ui to ComfyUI: 2736d1cdf5 and when i open it this is what i get on any browser and am using Manager: V2.51.2 i updated my nvidia drives and rebooted with no differance

[north flint]

Please does anyone has experience using comfyui on google colab, I need some help. I am trying to install comfyui manager on comfyui in google colab but not successful.

[surreal temple]

figured it out. i'd rather Forge UI get a clean up of the UI.. 😉

though i would prolly learn a little more with ComfyUI.

forge tends to have a shit tonne of options with no real explination of what they are, and what they do. some have mosueover tooltips and that's awesome, but some don't.

anyway, thanx for tollerating my randomness. 😉

[placid moon]

Hello everyone! Could someone help me out with more details on how he’s doing this using Stable Diffusion/ComfyUI? I’m especially curious about this part of the video: https://youtu.be/SHmjC7t3fJA?si=vCqwT3uA-SrUmgSV&t=119. How is he making it happen? Where can I find more instructions, like installation, setup, and a quick guide for 'Blocking to Render'? #1292682865028501606

[surreal whale]

looks like this https://www.youtube.com/watch?v=Ckj8x4huv7E

[placid moon]

omg, thnks!!!

[lilac patio]

i have a workflow where i use a depth and a canny controlnet. when i load it up, everything working how intended, but when i change the strength value on one of the controlnet stackers for second generation after it worked, the output brakes and stays broken even i change to the exact value that worked in first run. the setting i changed loads when i restart comfy and then works. have anyone encountered such behaviour or may guide me where to fix this/get help? using pytorch 2.4.1+cu124

on startup i get these two errors:

removed**\ComfyUI_windows_portable\python_embeded\Lib\site-packages\xformers\ops\fmha\flash.py:211: FutureWarning: torch.library.impl_abstract was renamed to torch.library.register_fake. Please use that instead; we will remove torch.library.impl_abstract in a future version of PyTorch.
@torch.library.impl_abstract("xformers_flash::flash_fwd")

removed**\ComfyUI_windows_portable\python_embeded\Lib\site-packages\xformers\ops\fmha\flash.py:344: FutureWarning: torch.library.impl_abstract was renamed to torch.library.register_fake. Please use that instead; we will remove torch.library.impl_abstract in a future version of PyTorch.
@torch.library.impl_abstract("xformers_flash::flash_bwd")

updated comfyui + customnodes regularly, last time yesterday

[lilac patio]

this only occurs when i use a rendered image (zdepth) as direct image input for controlnet

fixed it

[lone cape]

Finished another preview node for saved prompt (.csv) file. Aesthetic score and category displayed on preview if available. Auto collected categories in buttons, keyword filter, sorting, randomized prompts from same category as selected one. Lora, Lycos, Embedding and Hypernetwork nodes left, but these will use same code but different path like the checkpoint selector.

[eager stream]

I had a question, with regard to Flux Comfy workflows - is it possible to combine two ControlNet types (say depth and canny)? I can't seem to find a way to do this, like with my old setup when I was using the Efficiency Nodes, which was easy to work with, and didn't need many nodes. Unfortunately those aren't updated for Flux as of yet

[past acorn]

which goes where btw

I think i connect the model to seamless

[past acorn]

i am using this set up atm

i think it will work with flux too maybe not sure

using cr multistack node

i need to fix my contrl net and latent image connection and its gonna be done

[wicked urchin]

when you save the currently used workflow, which folder is it saved in?

[sly sundial]

\ComfyUI_windows_portable\ComfyUI\user\default\workflows

[past acorn]

ok now its working but for some reason the depth says me to download control net when its already downloaded

[hexed igloo]

Try updating the kernel version

[fickle remnant]

Does anyone know where this menu comes from and how I disable it?

[merry ermine]

maybe right click inspect element

[fickle remnant]

Tried that but nothing in there gave me a clue

I just hid it through that for now, but id love to know where this came from suddenly

[merry ermine]

this is the problem with GUIs

[fickle remnant]

Nah more a problem of custom nodes adding random kinda useless stuff

like why do i need buttons to zoom

[merry ermine]

oh yeah if all GUIs were well behaved it would be fine

yeah I don't like it when a custom node pack adds scripts and extensions to the front end

[fickle remnant]

Some stuff makes sense, but a lot of things are just bloat

[sick badger]

settings/comfy/ under graph

[rocky sequoia]

is there a way to make a prompt list so every time a prompt is queued it goes to the next prompt? so it isnt a batch but a round robin?

[weak urchin]

https://civitai.com/models/546180/geeky-remb

[eager stream]

yeah, I used the CR multi-CN stack too, with SDXL, but I don't think the regular Ksampler works with Flux, but the CR Apply ControlNet node might just work for that

[past acorn]

for me the union depth and few other features not working when using union

the canny works but depth gives me an error

[eager stream]

yeah, it may not be compatible...

[past acorn]

operands could not be broadcast together with shapes (518,518,4) (3,)

m on my potato laptop now can't test

[eager stream]

I'm thinking that the Efficiency nodes should be updated soon for Flux - I believe the author is working on that.

[past acorn]

i thought my dektop with 24 gb ram and 16 gb gpu was good for flux but i was wrong lol

on my laptop i got 4 gb gpu xD

[eager stream]

I had to use the FP8 version of flux, myself, with my 4080 Super

[past acorn]

i am gonna try the q4 guff

but sdxl should work

can i run sdxl with 4 gb gpu?

is there optimized workflow for running sdxl? with low end device

[eager stream]

SDXL on a 4GB GPU? I doubt it

I think it needs 6-8 at a minimum?

[past acorn]

alright SD is my friend then 😄

on my laptop atleast

strangle i can que 100s of images with A111 but comfy says cuda ran out of memory

with sd 1.5

i am stll new with comfy man i hate nodes

but i got no other choice besides comfy atm

i am a programmer i like words xD

[eager stream]

I got sick of A1111 being slow as hell

[past acorn]

well the UI there is so straight forward

i put words i check mark and its done

no spiders with nodes and wires

but it is what it is i have to deal with darn wires and nodes i accepted it

[eager stream]

True

but I'm used to Blender's nodes so it didn't take too long to get used to

[past acorn]

but give me other option i go back asap

i love blender for making my low poly modelling

i made this scene in blender i was quite happy about it

[eager stream]

Yeah that's cool, I've never tried making stuff like that...

[past acorn]

oh its easy

if you do low poly its easy and fast

but those realistics a bummer

one thing takes one week

not ideal for hobbyists

[frail jasper]

Forge or SwarmUI

Swarm needs to be the mcmonkey repo

[cosmic rapids]

FLUX wold *\ComfyUI_windows_portable\python_embeded\Lib\site-packages\xformers\ops\fmha\flash.py:211: FutureWarning:

[ocean haven]

could someone confirm that recent comfyUI updates completely broke CORS so I can't setup comfyUI as a backend for a webservice anymore

[carmine copper]

comfyui_face_parsing
ComfyUI-Art-Venture
ComfyUI-SUPIR
ComfyUI_Custom_Nodes_AlekPet

after update the comfyui today those models cant be imported

[zenith quiver]

hi I am wondering if I can run flux 1.1 in comfy?

[distant pendant]

I'm pretty sure bfl didn't open source flux 1.1

so that's a no...

[flat dirge]

Only via API
https://github.com/black-forest-labs/bfl-comfy-nodes

[wicked urchin]

how to avoid getting this with Ultimate SD Upscale node?

workflow included in img

[ashen agate]

I'm using ReActor and wondering if there were a way of using occlusion around the eyes so if the image had glowing eyes, I could use that without inpainting

[alpine parrot]

tile controlnet, lower denoise

like this

text encoder should be blank

[wintry jackal]

Hi Im gonna delete this files so just before that I tought I could share it here, I don't often use discord because always forget it so here it is

[wicked urchin]

is it this model? https://huggingface.co/xinsir/controlnet-tile-sdxl-1.0

[alpine parrot]

This works

[torn moon]

hello guys

any of you use stability matrix?

[past acorn]

I can someone help me how to generate mutiple images from a text file in comfy?

i got 50 prompts written in text file

I want to generate 2 images each for every prompt

[merry ermine]

inspire pack load text from directory node

the name is slightly different but its in inspire pack

[lilac patio]

is it recommended to have a second comfyui install for testing and share e.g. modelfolder? something to backup before? any experiences?

[merry ermine]

would recommend docker

[wispy kelp]

I recomend waiting about 1 year.

By then all images/video (even local) will be more or less perfect).

[wispy kelp]

So anybody here celebrates Friday?

[lone cape]

Hi somebody have Python code example or workflow how to load SD3 and Flux Loras? SD3 result just noise, Flux getting error. https://huggingface.co/ByteDance/Hyper-SD/tree/main

[glass lagoon]

In ComfyUI, i can see you can use a mask to choose where ControlNet is applied to image generation. However, it seems to be either "completely on" (at the set strength) or "completely off" (at zero). Is there a way to say "masked area is strength X" and "unmasked area is strength Y"?

[glass lagoon]

omg I just discovered masks can be greyscale, and are not binary

[slender hedge]

ComfyUI FLUX-1Dev - UltraRealistic Upscale + Loras - fp-8
DownLoads Templates: https://www.patreon.com/posts/comfyui-flux-1-111180301?utm_medium=clipboard_copy&utm_source=copyLink&utm_campaign=postshare_creator&utm_content=join_link

[fresh vault]

that first one with the wolf looks nice!

oh, hmm.. was this spam?

well whatever, that wolf is nice

[mystic plaza]

how should i resize my image or what and i am using this image

[hasty brook]

Anyone have a ComfyUI workflow that uses both Flux schnell and ControlNet? Looking for examples

[mystic plaza]

@dry rock how to fix this error bro or do i need to resize my image into 3:5

[livid gust]

If you don't know this, the IPAdapter can only process 512x512 images. Also, start with the default Context Options until you get up and running. Don't change any of the values.

[brave gull]

im always pulling comfyui straight from latest update in main, works for me. i use this to communicate with it https://github.com/deimos-deimos/comfy_api_simplified

[brave gull]

why would a Lora trained w/ Flux work with SD3 ?

[brave gull]

if you download models with huggingface-cli, your models all end up in ~/.cache/ , and you can symlink the models to your comfyui models dir. or another comfyui models dir. as many as you want. i dont store actual copies of models in my comfui models directory

[verbal sky]

wow nice @sacred gyro - latest Colossus Flux Model is the current leader (only downside is longer render times of 35 steps) - https://docs.google.com/spreadsheets/d/1543rZ6hqXxtPwa2PufNVMhQzSxvMY55DMhQTH81P8iM/edit?usp=sharing

[sharp topaz]

HI 🙂 Do you know any node to directly download models from comfyui, or to load them through URL? Im trying to make a runpod template that is as easy to use for my students as possible. Ive tried https://github.com/ciri/comfyui-model-downloader, but i dont like that you have to input the model id, etc. An url should be enough

[hardy tree]

@storm folio with the latest comfy updates, can i move all the unet models (flux) to the new diffusion models folder (for the scaled flux), or is that new folder only for the scaled flux and I should keep the other flux models inside the unet folder?

hmm, there is also an error when i try to use the scaled flux:
Tensor for argument #5 'scale_a' is on CPU, but expected it to be on GPU (while checking arguments for _scaled_mm_out_cuda)

[lone cape]

No, these are only "speed" loras, half or quarter steps enough for same quality images, and hyper-flux useful only for flux, hyper-sd3 useful only for sd3.

Hi, when the LCM was released, I made a node that downloads the 2 LCM loras and automatically selects by safetensor is SD or SDXL. It was a pretty small node. :). But then Lightning and Hyper came out, and I did the same for these, had all 3 in 1 node, concept choose from a list, but Lora was automatically recognized and selected by checkpoint version and step. Then came Cascade, SD3, Hunyuan, Kolors, Flux, Playground, Pony, etc... I added these too, with selectable unet, text encoder or own vae if have, all with different sampler settings for each. Now this is my largest node, and it can do auto-detection by checkpoint, the list is not important, and a symlink trick useful to see unet files in the checkpoint selector. So Cascade Kolors and Flux files in my unet path, but selectable within checkpoint loader if symlinked. If somebody have downloaded files of several concepts, and wanna test my solution, pls dm me. The node not out yet in public repo, but developer can access.

[storm folio]

that diffusion models folder is one that replaces the unet folder

unet just doesn't really make sense when flux isn't a unet

[sly sundial]

Is there a method/workflow to merge loras into Flux diffusion model and then save it back as a diffusion model again?

I only see a node to save as a checkpoint.

[hardy tree]

ah ok ty 🙂

[storm folio]

ModelSave

[sly sundial]

Thank you!

[covert bough]

you should post the annoucements in here as well as on your discord

[lone cape]

New trick working with my custom model loader. Just symlink unet/diffuser models to the original model/checkpoint path, then not just listed, but working good with hacked loader.

[distant pendant]

...or you could just set all three to the same path in extra_model_paths.yaml

[merry ermine]

I think they are trying to make a node pack that does not require manually editing stuff like extra_model_paths.yaml

not 100% sure but that seems to be what they are doing

[onyx summit]

do you guys have issue to use the new model in comfyui? it tell me this

[shadow chasm]

how can i add img2img inside comfy?

[shadow chasm]

do you still need help

[onyx summit]

i need to rinstall it

[onyx summit]

i made it work

[hardy tree]

@storm folio so flux fp8 scaled was broken and needs to be uploaded again right?

[storm folio]

Yes

[hardy tree]

kk, cant wait 🙂 ty

[storm folio]

It was the test for the SD3 scaled model

So it had some issue that I fixed before doing the SD3

[hardy tree]

ah makes sense

[slim comet]

fp8 turbo model, anyone?

[cloud hollow]

cup of tea

[slender hedge]

https://www.patreon.com/posts/114567406?utm_campaign=postshare_creator&utm_content=android_share

[vocal vortex]

Any news on SD3.5 lokrs working soon in comfy? they give key errors where there are no key errors.

@storm folio not even lycoris loras work it has to be standard lora only.

[storm folio]

does it work now?

[sudden saffron]

Hello there, need an advice what the issue could be, please assist if it's the right place to ask..
Fresh comfyUI install.
No errors as what I see in CMD, attaching the browser log as well.

[livid gust]

Try another browser and or update your graphics drivers.

[remote sequoia]

hello, I using ComfyUI and i am trying to make my first lora (in hopes to make a realistic model) what a good number of images i should use to create the lora? Is it a general rule that the more photos the better? thank you

[vocal vortex]

I'll check as I see someone else opened a ticket.

[vocal vortex]

no.

[merry ermine]

using this example workflow:
https://github.com/Acly/comfyui-inpaint-nodes/blob/main/workflows/inpaint-refine.json

what causes weird Fooocus Inpainting, and how to solve it?

[brazen turret]

would someone experienced with comfyUI give me proper launch args for a 6GB card specifically the 1660 Super as I am having an issue where my id 0 gen will be super slow (90-160 s/it) and in console it says that it's requested to load SDXL and then loads it before the gen. But then on my id 1 gen it will gen super fast at around 7 s/it but it says unloading models for lowvram mode and then unloaded 0 models, but it's generating super fast so I don't have to wait 20 minutes or longer for a single generation. This is in lowvram mode, but in normalram mode it's slow every gen. So how do I get comfyui to a point where I can always gen at that 7 s/it on SDXL

[storm folio]

I merged a PR yesterday that should have added support

[vocal vortex]

it did not, or at least the one I just git pulled didn't

to be sure I am going to delete venv and start fresh

@storm folio YAY, it works. Thank you so much it must have been a venv issue

[merry ermine]

I got header errors with the fooocus inpainting
I recommend instead the inpainting controlnet from the Union Pro Max

if that isn't strong enough then Powerpaint V2 Brushnet is a strong alternative but you need to do that pass with SD 1.5 for it to work

[forest cove]

Anyone knows how to fix when I save a workflow then I load it it won't generate any images anymore?

[brazen turret]

@storm folio since you would be pretty experienced could you help me with this?

[proud void]

Not an expert, but may need to run with --lowvram to split the Unet into parts, possibly also --disable-smart-memory to force aggressive unloading from vram into ram. If those are insufficient you may need to experiment with the arguments for loading unet and text encoders in fp8 instead of fp16. Not sure what other arguments would really help with this card in particular. I remember it having some quirks compared to other 1000 series and later cards.

[covert bough]

start by looking at the console log and seeing what the errors are - as long as you've shut the machine down, turned it off, and let it stay off for at least 10 seconds first

[manic moat]

Hello please can anyone help me face swap tool option I tried lot nit not getting wat m expecting
Need same face in other images without change expression n jaw line

[remote sequoia]

hello, i ran a loRA training and the output_dir is models/loras but i can not find it, any suggestions? the datapath (text files for the pictures) i can find and are in the right folder but am lost finding the actual Lora model. Thanks I am running ComfyUI btw

[toxic gate]

is it possible to take anime photos and turn them into realistic with FLUX yet?

[past acorn]

i like doing opposite turn real into anime 😄

[hollow vault]

Installed this recently, does flux work with it?

[verbal sky]

Good morning all - new video is now live, so much fun in store and I apologize in advance for the puns 🙂 Topics include a new loader, compositor, flux controlnet and several tips and tricks including a memory booster! Enjoy! https://youtu.be/WSx74Uep590

[silver nebula]

https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/tree/main

[thorn gazelle]

I'm on the latest comfyui but I don't see the new skip layer guidance node. any idea how to get that to show up?

[sly sundial]

Did you update in the last hour or two?

[thorn gazelle]

yes. updated right before posting that, and I keep checking to see if there's anything more

[sly sundial]

It's possible your updates aren't working as expected then. Are you using git pull?

...or the update .bat file?

[thorn gazelle]

git pull

I can see the latest changes referencing the node in the git log and everything

figured it out. comfyUI manager took over the listing of nodes and defaults to making it a remote source. changing the "channel" setting to local fixes it.

[vocal vortex]

@storm folio Same issue we had with SD3.5 Large we now have with med, and I just git pull too.

[crimson minnow]

HI GANG

am i doing something wrong

i click "install missing nodes" and get this

clearly theres more than 1 missing but they dont show up

[vocal vortex]

It means, normally, that the missing nodes are not a part of the db held by the manager. IOW, you will have to seek them out and their githubs.

@crimson minnow Sometimes I run into nodes that are private too.

[crimson minnow]

Okay noted, yea can still see their names, just have to manually find them, thanks for confirming 💚

[vocal vortex]

Good luck as sometimes I can't. I wish Comfy showed the names of the missing nodes so I can sort of get an idea what they did for those cases I just can't find them.

[storm folio]

should be fixed now

[vocal vortex]

Just saw. Thank you, and it does work

[proud void]

when running sdxl in fp8 mode in comfyui, is it worth having both the unet and textencoder in fp8?

[sharp raptor]

Is there any comfy Node which can save image with prompt data visible on the image itself?

As i Edited in the example image above

[livid gust]

There is a font node you can composite over top your final.

[sharp raptor]

Cool! What is the name of it?

[livid gust]

Double-Click and type "Image Text" in the search box.

[forest birch]

Damn, making tensorrt of SVD_XT 512x1024 be heavy on the memory lol.

[grave knoll]

Does someone know where I should connect this please ?

[lusty pewter]

I mean...you can do whatever you want with it, but the nodes you've posted there are just going to be text, so I assume you want to condition that text into an actual prompt?

You don't have much workflow there to give us context as to your goal.

[grave knoll]

oh sorry @lusty pewter well basically, those nodes are supposed to help the prompt generation. Now I wonder if I would have to connect it to the rest of the workflow

Im testing flux

so the setup is basic

right now Im just coping / pasting the text into the text encode

[lusty pewter]

Essentially, you'd convert the CLIP Text Encode text widget to an input and connect it there.

[grave knoll]

This one ?

[lusty pewter]

yes

[grave knoll]

ohhhhhh shit !

Thank you !

[lusty pewter]

As you may have just discovered, widgets can be converted to inputs to allow for you to pass something in from another node.

[grave knoll]

yup !

[forest birch]

Is there a way to split up model conversion with tensorRT for comfy? Tried to do 1024 flux, but comfy just broke and cmd window closed xD

Need way WAY more than 24GB vram and 64GB ram for that lol

[vocal vortex]

is there a save to lora or a save to checkpoint we can extract a lora from for SD3.5 yet? the current save lora node isn't for 3.5.

[carmine copper]

i need help

[alpine parrot]

reinstall python dependencies, run as admin

[forest birch]

Do you guys know of a inpaint node to brush onto the image in comfy like the inpainting features in forge can?

[alpine parrot]

load image > left click load image, open mask editor, brush > hit save

[kind karma]

Is there a image to text node that allows for custom instructions?

[livid gust]

I've used Blip and Forence2. You can prompt the node like "describe this image to a child." Or "What is in the background of this image?"

[wintry jackal]

anyone knows a workflow for inpaint sketch?

xl or 1.5 doesn't matter

[sharp topaz]

When applying multiple Ipadapters with attention masks, ive found that the least efficient way (in vram) is to chain Ipadapter Advanced nodes, the second least efficient is using Apply Regional Ipadapter from Inspire, and the best is to use Ipadapter Regional Conditioning and Ipadapter From Params

The problem is that this last technique doesnt allow to combine embeds... Ipadapter Regional Conditioning doesnt have a combine embeds parameter

[frail heron]

Im just getting into comfy, and its pretty powerful...but also a weee bit confusing. So by default, it doesnt seem to have a "negative" prompt like SD does. How do people work around that?

Its generating some attributes I dont care for, but... not sure how to get them out of the image.

[covert bough]

are you talking about flux?

[little anchor]

Hello guys is there is anyone here that's good with comfyui and have some time to help me to setup some models

For example I've CogVideoX5b but it's really slow and i heared there is some other tools that could speed up the process

[frail heron]

Yes, i am. My apologies.

[covert bough]

okay. flux doesn't use negative prompts, because in order to use flux correctly, you have to set cfg to 0 - and with it at 0, negative prompts are ignored - so what you have to do is be careful about how you craft your prompt - use tight, concise terms. dont' ramble. tell the AI what you want it to draw, and be specific and detailed.

[frail heron]

So how do you handle removing things your not wanting in an image?

[steep marlin]

By writing more descriptive positive prompts. If you want a photo of a dog and want to make sure it doesn't come out as a charcoal sketch, you'd prompt something like "a photo of a dog" instead of just saying "a dog"

If you wanted to make sure a woman is wearing a red dress and not a blue dress, you'd say "a woman wearing a red dress" and therefore you wouldn't need a negative prompt including things like blue and blue dress

If you're having troubles with fleshing prompts out, use an LLM to prompt expand

[bleak stump]

New concept videos I made for The Voidz! - https://www.youtube.com/watch?v=pyNTS4oYGFU

[frail heron]

So, lets hypothetically say it keeps generating armpit hair... youd say "shaved armpits"? lol

[livid gust]

Don't prompt what you don't want, prompt what you do want. It will never make armpit hair.

[fresh vault]

Pretty interesting. Clean images, looks like Flux to me at that quality. Hopefully we can get some tech to produce longer video without all the jump cuts. 🙂

[steep marlin]

Yeah basically, but people just need to think about prompting with that style of logic in general. Be descriptive in your prompts.

[low spire]

I forget...how do I circumvent this to update Comfy?

[little anchor]

Hello

I need help running Moochi with multiple gpu's on ComfyUI please

[sly sundial]

I think you can run the model on the first GPU and choose another for the clip with this node

[alpine parrot]

git config --global --add safe.directory

[frail heron]

I had a ton of issues with comfy, node installation, etc, I moved to Linux and it couldnt be better.

[wispy kelp]

Omnigen on comfyui?

When, or has it happened?

😏

[little anchor]

4 days ago

[placid flower]

?

[fickle remnant]

Seems like youre missing something for CogVideo

[placid flower]

Its solved.. thanks for the concern

[verbal sky]

Hi everyone, my new video is now LIVE 🙂 High resolution face swaps, model configuration optimization, finger fixes and power tips and tricks. Enjoy! https://youtu.be/vozQbuAMYN0

[restive escarp]

possible to use workflow with AMD card? most i try use some sort of cuda only process like reactor

[verbal sky]

I don't see why it wouldn't work since it's not using reactor 🙂 If you're able to normally render stuff in Flux, you should be good to go

[restive escarp]

thanks! ill check it out

[primal dagger]

As I'm sure you can tell I'm very new to ComfyUI

[white pawn]

Does anyone have an up-to-date workflow for creating realistic people (faces)? I'd be very grateful

[livid gust]

Click the Default button and type your prompt. You don't need any special workflow to make realistic people.

[wind minnow]

What would be the best way to animate this in ComfyUI?

I just want like slight movement of these symbols. I've tried some AnimateDiff workflow but the animation did complete change the input

[lilac patio]

mhh, i had art-ventures custom nodes up and running, then installed mixlabs and since then getting import failed on art-venture.. anyone encountered something like this and know how to fix?

[prisma rose]

Nope , same for me.

[prisma rose]

You can try SVD XT, or cogvideox fun perhaps

[lilac patio]

it must be possible to have these two nodepacks running together, it worked fine two days ago

[sly sundial]

What's the error on startup?

[lilac patio]

fresh install (windows_portable), installed manager, installed art-venture (until here all fine) then installing mixlabs -> error occurs

i get the error described here https://github.com/sipherxyz/comfyui-art-venture/issues/13 but when i try to install segment-anything and omegaconf with pip as suggested there, it trys to use a python i do not have, the path is different.. it looks at
PS C:\Program Files\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages> pip install segment_anything
Fatal error in launcher: Unable to create process using '"D:\a\ComfyUI\python_embeded\python.exe" "C:\Programme\AI\ComfyUI_windows_portable\python_embeded\Scripts\pip.exe" install segment_anything': ...

should look at "C:\Programme\AI\ComfyUI_windows_portable\python_embeded\python.exe

how do i f ix that

(i tried everything suggested there, including renaming the folder)

reinstalled pip, requirements already satisfied now after pip install segment-anything and omegaconf, replaced segment_anything from original zip to site-packages

even after uninstalling mixlabs, the error on art-venture (IMPORT FAILED) stays

[sly sundial]

Is your install in "C:\Program Files\AI\ComfyUI_windows_portable"?

[lilac patio]

yes

[sly sundial]

Are you running this from the cli?: C:\Program Files\AI\ComfyUI_windows_portable\python_embeded\python.exe -m pip install -U {package to install}

[lilac patio]

yes

wait

cli? from inside comfyui client you mean?

[sly sundial]

No, command line interface (terminal)

[lilac patio]

yes

[sly sundial]

Did you have it installed in D:\a\ComfyUI at one time?

Not sure why it's coming up with that path in the error

[lilac patio]

no

yeah i dont know too

i have d d: drive but there is nothing comfy related

[sly sundial]

Have you tried uninstalling both custom node packs and reinstalling each again? Restart between each pack.

[lilac patio]

yes

in every possible order

as soon as i install mixlabs art-venture stops working

[sly sundial]

Does one work without the other after installing it fresh?

[lilac patio]

art-venture works when i install it first

[sly sundial]

Does mixlabs, on its own?

[lilac patio]

yes

[sly sundial]

Sounds like they're just too incompatible to run together.

[lilac patio]

they did a few days ago 😦

everything just worked as intended

i made a huge workflow releasing it for free and already recieved lots of love for it.. now i cant use it myself anymore

[sly sundial]

Perhaps the nodes were updated and it broke. You could try installing an older version.

[sly sundial]

mixlabs update 3 weeks ago, and art0venture last week. When did you last update them?

[lilac patio]

around 5 days ago, i cant remember exactly but i always update on a regular base

[sly sundial]

Do you know how to change to an older commit?

[lilac patio]

unfortunately not

i have downloaded mixlabs 0.44 and replaced it custom_nodes

that worked, mixlabs showing me that there is 0.46 and i should update

but art-venture still import fails

im 100% sure that 0.44 worked on my thing

not sure about art-venture

[sly sundial]

Locate the commit hash for where you want to go back to, then go into the custom-nodes art-venture directory and run git checkout {commit hash}.
Give that a try and re-run the requirements install manually.

Commit hashes are listed here
https://github.com/sipherxyz/comfyui-art-venture/commits/main/

So, to go back to the Oct 31st version, is hash 8d538c9678dd5fad21e35681a5c0f0202b247c36

[lilac patio]

rund update comfy and dependencies?

PS C:\Program Files\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-art-venture> git checkout 8d538c9678dd5fad21e35681a5c0f0202b247c36
Note: switching to '8d538c9678dd5fad21e35681a5c0f0202b247c36'.

You are in 'detached HEAD' state. You can look around, make experimental
changes and commit them, and you can discard any commits you make in this
state without impacting any branches by switching back to a branch.

If you want to create a new branch to retain commits you create, you may
do so (now or later) by using -c with the switch command. Example:

git switch -c <new-branch-name>

Or undo this operation with:

git switch -

Turn off this advice by setting config variable advice.detachedHead to false

HEAD is now at 8d538c9 Merge pull request #56 from sipherxyz/develop
PS C:\Program Files\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-art-venture>

[sly sundial]

From the same location run: ../../../python_embeded/python.exe -s -m pip install -r requirements.txt
If you want to change back to the latest version again, run git checkout main form the correct nodes directory

Each nodes repo has a list of packages that need to be installed and it's in requirements.txt. You can compare them on each and see if they try to install different versions of the same thing.

Conflicts can also be caused by same name nodes. This is common because there are so many that do similar things.

[lilac patio]

mhh, no success

[sly sundial]

Try rolling the other pack to an older commit too. Same procedure.

I just give up with some packs because they break too many things. They're just not usable without too many headaches.

[lilac patio]

i see, what makes me mad is that really everything worked just a few days ago

i even made a showcase video with it

frustrating.. but really thanks for your help on this, much appreciated

[sly sundial]

Have you update ComfyUI as well?

[lilac patio]

sure

its a fresh install too

install comfy -> install manager -> install custom nodes -> everything works until -> install mixlabs

update before install manager though

[sly sundial]

Did you download all the models for mixlab? I don't think that's the problem, but the Readme instructions do mention it.
https://github.com/shadowcz007/comfyui-mixlab-nodes?tab=readme-ov-file#models

I really don't know what else to suggest, sorry.

[lilac patio]

i do not see required models there :x

[sly sundial]

Near the bottom

That link takes you to the section

[lilac patio]

aaah

i will try

i think i did not do this before

[sly sundial]

This sounds like your issue too https://github.com/shadowcz007/comfyui-mixlab-nodes/issues/363

No response to it though.

[lilac patio]

yes, sounds same

[sly sundial]

Then follow:
https://github.com/sipherxyz/comfyui-art-venture/issues/13#issuecomment-1972001803
and
https://github.com/sipherxyz/comfyui-art-venture/issues/13#issuecomment-2296780269

Good luck!

[lilac patio]

thanks, thats what i found too.. tried everything mentioned there already in the first place

[merry ermine]

lol yeah this is common
there is a problem with mixlabs

[lilac patio]

it has been around for a long time as far as i know, but it never occured to me

i dont knwo what i did right before or wrong now

[merry ermine]

mixlabs has been breaking loads of people's installs recently

its just one of those potholes to remember

[lilac patio]

for my work i cant replace its functionality.. so its just trash now?

[merry ermine]

quite possibly yeah

I haven't seen much of an explanation around as to how the repo broke

[lilac patio]

thats just sad news to me, i really spent a lot of time on a potential money saver to many people and gave it away for free

recieving already lots of love from around the globe and now have to tell people that its not working anymore

[lone cape]

Hi, new idea for my "visual nodes". Previews not just images, but contains lot of information. The model name, architecture (unet, diffusion), type (I call concept like SD1, SDXL, Flux, etc...), average aesthetic score with relative percent, and average sampling time. Not just display previews by modal, but previews sortable by several values, and 1 button filters by subdirectory and concept. Supported concepts: SD1, SD2, SDXL, SD3, StableCascade, Turbo, Flux, KwaiKolors, Hunyuan, Playground, Pony, LCM, Lightning, Hyper, PixartSigma. Different architectures like Flux or Kolors can be symlinked to the standard checkpoints folder, so these type of models can be selected by same loader. Not too easy to setup, the basic workflow required because custom loader and sampler. Please read the manual before use: https://github.com/CosmicLaca/ComfyUI_Primere_Nodes/blob/master/Workflow/Manual/nodes/basic_workflow.md

[lilac patio]

does anyone by chance know an existing nodepack that gives a floatslider like this?

ive tried to create one with claude but failed

[sly sundial]

All of the nodes work like that 🙂

[lilac patio]

i know :p i would love to have limits to it.. a visual thing would be nice too but not necessary

some sliders would be a great benefit i think.. i'd love to have some at least

maybe with two inputs für range values

and something like steps

or how ever its defined wich numbers will be applied

the mixlabs thing was awesome for that.. but i'd rather like to work with my workflow again 😄

replaced everything from mixlabs now

thanks for help ❤️

[sly sundial]

You're welcome, well done!

[merry ermine]

mixlabs's end took out a lot of people's workflows yeah

[lilac patio]

i used to have xformers installed, why is pytorch attention so much faster now?

on my workflow everything works 3times faster now

and less ram is consumed

[sly sundial]

Progress? 🤷🏻‍♂️

[lilac patio]

thank you comfy 🙂

indeed

[uneven stratus]

When i run this, i got this result

[sly sundial]

Are you running the windows portable version of ComfyUI?

[uneven stratus]

Yep!

Got this issue every months, but this time i seems not being able to solve it..

[sly sundial]

Go into the directory it's installed in and run .\python_embeded\python.exe -m pip install -U torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu124

Run in your E:\ComfyUI_windows_portable\, or wherever you have it

[uneven stratus]

yep

Working you're the goat @sly sundial 🙌

[sick badger]

what custom node can i use to add extend an image downward from say 1024x1024 to 1024x1344 with the image moved to the top and just a blank area at the bottom

[surreal whale]

search for how to outpaint

[steep marlin]

Depending on the model, there are controlnets for it. I think shakker labs is one of the better ones? Can't remember the major names of the top of my head. They usually have some all-in-one "pro" controlnet that can handle outpainting. Just make sure to use the node that lets you pick the mode, since those pro models do canny/depth/normal/in paint/out paint/etc

Yeah it was shakker labs. The models are called controlnet union pro

[sick badger]

wanting to add some text to the bottom of the image like a caption but rather it not be on the the image itself
was able to extend with the outpaint padding node

[steep marlin]

Ahh I see

That's just simple Photoshop work then to slap the text on top afterwards

[sick badger]

trying to create a character card thing, using a local ollama to generate some fun text elements but if i do a long sentence i'd rather the text not be on the image

looks dumb on top, so wanna move it to an extended part like a bottom caption area where i can use wrap text so i can add a backstory with longer text

but all of it needs to be inside the workflow to work on my web app

got the padding to work just not sure that's how i want it yet

[surreal whale]

https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes/wiki/Layout-Nodes

that and other nodes in there might be helpfull for you, i suggested outpainting to you but it sounds like you dont really need to do outpainting

[sick badger]

using comfyrolls text overlay, however it doesn't have a word wrap option
will figure out eventually, appreciate the helpful tips as always

[sick badger]

fixed my old custom node, all the old outdate PIL versions
think this will get me started

[forest birch]

@storm folio Hoi, do you know if comfy/python would be able to offload data it loads into vram onto for example a second gpu to hold instead of ram? If i have a 3090, but 24GB isn't enough, to "expand" vram to hold 12 additional GB? Or would for instance x4 lane bifurcation make it slower than ram? :P

[dry rock]

Natively no. Otherwise https://github.com/city96/ComfyUI_NetDist. Have fun.

Regarding multi gpus, most solutions out there are just "run 1 instance per gpu"'.

(at least regarding inference, there are a bunch of other solutions for training that I don't know much about)

You can also try to use "accelerate" https://huggingface.co/docs/accelerate/usage_guides/distributed_inference but that s outside of comfyUI scope.

[forest birch]

Aye. Performance would be only 1 card, it's more to use a second gpu only as "extended vram A.I data holder" in it's vram, but not performance itself

[dry rock]

problem is kinda the same it would require some kind of unified memory across multiple gpus.

Having 1 gpu access another gpu's vram requires the cpu to pass along the information. Slowing everything down.

[north flint]

Hello all, please does anyone know how to fix this?

[lusty pewter]

Well, first off, the node tells you not to use it. Use the node that it's telling you to.

[marble venture]

is anyone else finding that the new comfyUI update from last night is causing vae encode to stop working and crash when using ae.safetensors?

[sharp topaz]

Hi! Is there anything similar to the mask bounding box crop components, BUT instead of cropping a bounding box that includes the whole mask, it crops the biggest square/rectangular box that is inside of the mask.

¿?

What I want is to get a piece of the masked region, but this piece has to include only information the is inside the mask

[crimson snow]

Is it normal for a 896x1152 image at 30 steps to take 4 minutes to generate on an 6700 xt?

[livid gust]

Oh, you mean AMD? Could be. Are you using the Load Default button setup along the side?

[fallen moat]

What are the rest of your settings and what model?

That's definitely within the realm of possibility with directml on Windows.

[digital mortar]

Hello, not sure if this is the best place to ask this. But i built a basic diffusers pipeline and its working with our model and the Lora we are using. But we are having a hard time matching up parameters from A1111 to get similar results. Does anyone have any experience with this stuff?

[sharp topaz]

So basically a Controlnet-Tile should be called controlnet detail/deblur? The tile part is part of the workflow not the model right?

Or does it also retain coherence with the rest of the tiles?

[merry ermine]

tile and deblur are different

the main thing tile does is that when global semantics clash with local semantics, it uses the global semantics in those areas

this is actually enormously more valuable than deblur

however both can be done at once

[lilac patio]

hey, when using flux1-depth-dev-lora from the example workflow i get some errors:
lora key not loaded: vector_in.in_layer.lora_B.weight
lora key not loaded: vector_in.out_layer.lora_A.weight
lora key not loaded: vector_in.out_layer.lora_B.bias
lora key not loaded: vector_in.out_layer.lora_B.weight
.....
comfy etc. is uptodate, any hints?

[primal dagger]

Can someone help me with adding a lora picker node to this workflow? I want to try using loras in this but I'm not sure where to put it.

[obtuse latch]

So are the new versions of comfy breaking the json workflows?

I exported a workflow on one machine and opened it in another and all the nodes are wrong.

[sly sundial]

Just add it straight after the model and clip loader nodes, then connect those to it.

Does anyone know how to drag over an area and use that to add a group?

[livid gust]

The new Comfy trashed my workflow. The Get/Set nodes no longer work. I can't find the old tools to search for missing nodes. The funny thing is that none of my nodes error out during pre-flight, so it looks like a bug to me.

How can you seach for missing nodes using the new Comfy?

A manual install of KJnodes fixed it, whew.
And manually updated the manager.

Is there anyway to get rid of this toolbar? I have no use for it.

[livid gust]

Thx, that worked.

[forest turret]

Hello, I'm looking for help, I tried many times to fix hands in my workflow but the results are pretty low.
I'm looking for an anwser for "how to correct hands".
It's the image I'm trying to fix, it was generated with Pony Diffusion 6 XL (SDXL) and know technique I tried :

• inpainting (with differents node variations)
• controlnet (don't really works)
• detailer with SEGS (don't works)

Best results was basic inpainting (for 20 generations, 8 was good).
Do you know others techniques ?

[void shale]

get a model that can constantly generate hands correctly, circle the hands area and give it denoise at ~50 to 90% and let it work on it

[forest turret]

Does that model must be an anime style ?
Do you have an example ?

[void shale]

if you want it for anime then i suggest these options

1.find a model that can constantly generate hands correctly with prompts to optimize it, use inpaint and continue(just like above)

2.find a non anime model and use it to inpaint, the inpaint then goes into an inpaint with the same area into an img2img with the anime model at ~25 -50% noise
lower will lower the anime, higher might make it ruin the hands

sometimes you will need to make a large batch and then choose one of the options, use specifically img2img models or inpaint models

you can also use embeddings for hands

for better results i'd suggest talking in the anime channel
https://discord.com/channels/1002292111942635562/1091193032273043516
and bring a cross before you enter it, might need to recite a few bible chapters beforehand

[sly sundial]

That middle, square icon can be useful. Select any node and click on it to zoom in and focus on that node.

[verbal sky]

New video tutorial is now live! Tons of efficiency tips in using the new Comfy Interface including Flux Inpainting and other tips and tricks. Enjoy! https://youtu.be/d0_6pKLskLk

[obtuse latch]

The new update is dogshit.