🧣｜comfy-ui

[bleak crypt]

Gotcha.

[lone cape]

Hi! If I have an image with exact mask, for example a product or person, can I put this image to the new background using IP adapter nodes but without changing the original masked image, while consistency needed for light and shadows by the new background? I already have good results but the originally masked image always changed a bit, while the final consistency is really good.

[karmic creek]

https://www.comfy.org/

[fading badge]

guys do you know why my generated pics like this?

[craggy basin]

a broader pic showing the other nodes and settings might help

But offhand, know that not all the options such as dpmpp_3_SDE etc will work, nor Karras

on SD3

[karmic creek]

model, resolution, ksampler settings, etc... just with the output is difficult to see what you have going on

[fading badge]

the whole thing just simple setup

i am kinda new

i didnt found the problem

[karmic creek]

maybe the vae, try connecting the vae from the checkpoint loader straight to the vae decode node

[fading badge]

ok

[karmic creek]

the rest seem to be fine

maybe lower a bit the LoRA strenght

the neg prompt might be a bit too much too

try first with nothing there and if you find problems start neg prompting them one by one

[fading badge]

i will try another workflow

[craggy basin]

What VAE are you using?

[steep marlin]

It's your vae. You're using a taesd vae which is normally for previews and such

Either use the red pin from the model loader or find the appropriate vae to use in the vae loader

Most common models include their vae with the model checkpoint

[still tartan]

Change your clip to -2. IIRC the pony models require that setting to work properly.

[scarlet kite]

where do I place this in my Comfy folder because this is the last thing that I need and I am getting an error /t5-v1_1-xxl-encoder-bf16
I dont have a t5 folder in my models folder can I just make one and place the t5 encoder there?

[stoic gulch]

thats a clip, its goes in the clip folder, and you can further organize clips with folders

so you gotta search forwherever you put it in the clip folder, within th UI

[forest birch]

Hoi, whenever i use more vram than gpu got stored, it starts to chug on the nvme for pagefile as "shared ram", is there a way to make it use ram for that instead?

[distant pendant]

it should default to cpu ram and to pagefile only after that's exhausted

you could disable pagefile for nvme but I don't think this should be happening in the first place

[slim seal]

is there a way in comfy to know where a custom load model node is looking for the model?
I downloaded a workflow, and I have all the model files it is looking for but I have no idea where to put them

[round summit]

is there a node or set of nodes that would do this?
switch to another node depending on text input
like If text = tall then use this node else use this other node

i would like to switch nodes depending on the text input.. im not sure im saying it correctly

[feral knot]

Hello all, I'm training a model in kohya_ss and I have a prompt in the "samples" collapsible section that outputs an image at the end of every epoch. When a checkpoint is saved at the end of each epoch, a sample image is made, but when I use the same prompt and checkpoint in ComfyUI, the results aren't as pretty. How can I achieve the same results?

[heady parcel]

hey guys i just want to understand a technical aspect of comfyui, i have 8gb of vram on my nvidia card and most of the time i can process most items, but every now and then i get an out of memory error on t he ksampler.

1. is the out of memory error caused by a combination of batch sized and loras used?
2. for example i never go over a batch size of 10, does this mean if it's a batchsize of 10 and 5x loras used it's likely the cause of the out of memory error?
3. would reducing the batch size to lets say half at 5 while keeping the same settings help with processing that item?

[round summit]

a node that wold do this?

[peak surge]

question, is there a watermark placer node that allow me to place a transparent PNG over the generated image?

i want to avoid the photoshop passage

[spring jungle]

https://github.com/shiimizu/ComfyUI_smZNodes/issues/83 comfy please add support for this ! smznode is my fav node in comfy

if he want to give up i understand xD

please support running auto parser ! xD

[steep marlin]

You could probably code one fairly easily. Just parse the input text for the words(I'd make a little dictionary of all the various synonyms you might use) and branch the output based on that. Probably put an override toggle on the node in case your prompt has a word like wide in it to describe something in the actual scene.

[spring jungle]

@storm folio if you can atlest xD

[round summit]

ah yea, i know limited coding, very limited. was hoping there was already something there. however i may have figured a rough work around with text to numbers numbers to something. something to something.. im already confusing myself here

[steep marlin]

Should just be: if 'wide' in text_input:
Pick some wide resolution
Elif 'tall' in text_input:
Pick some tall resolution
Else:
Square resolution

If I get some time later, I might whip up an example. I'll tag you if I do

[round summit]

that would be amazing. no hurry though. thanks 😄

[sick badger]

I do something similar when I want my users to be able to choose between a twitter banner and a mobile wallpaper using python and the comfy web socket API. Dynamic swap of the resolution with if then statements

I use the comfy API almost exclusively now

[round summit]

oh yea! thats interesting. yea i set up a family thing but i would like to give them more options than they have now*

[sick badger]

I can send a py file example I use for just a quick web interface that connects to your comfy server later tonight.

[round summit]

thanks, but i use a private bot 🙂

[steep marlin]

had a few minutes to try to whip something up, you're looking for something like this, right?

if so, here it is. just edit the file to change the triggers and resolutions

if you need more triggers add them like this:

just remember that if multiple are mentioned in the prompt, it will pick the first one

edit the list before opening comfyui

[round summit]

woah! lol ok i'll be looking this over 😄 thanks a ton

[steep marlin]

np. i kind of liked the idea personally. seemed potentially useful. if i never end up using it, oh well, i wasted 10 minutes lol

[round summit]

lol looking over the examples, this looks perfect 😄 again thanks 😄

[steep marlin]

again, np. good luck

[steep marlin]

Oh and one last thing: I'm not sure how it handles letter cases because I hate python, so probably make sure to use lowercase for the triggers and the actual trigger word in the prompt text

[round summit]

sounds good 🙂

[round summit]

works perfectly 😄 pretty simple code. i can learn from this \o/

[merry ermine]

so im trying to apply faceDetailer in comfyui while using SD3, but the facedetailer preview image always looks noisy and sort of distorted. anyone can help with what i might be missing in the setting?

[round summit]

in WAS Text Find and Replace how do i add multiple words to find?

[sly sundial]

Have you tried "Any-Node"?

[round summit]

yea i did, but still needed a way to switch what one turned on. but someone helped me out 🙂 thanks

needed to be switched by using prompt

[round summit]

is there a node like Any Switch, but will allow my to pick which one to use. (not a boolean). and it needs to select between more then 2 input nodes.

[mystic plaza]

what does this mean bro

[sly sundial]

What is it you want to switch between?

[round summit]

wow, would that take a model input?

[sly sundial]

No, but there's no point because models can be selected in the loader anyway.

[round summit]

im running a discord bot for family and would like them to be able to switch between 3 or 4 sd 1.5 models

without haing to reload the models

i can run up to 4 models seperate (same time), but no way i can let them switch

[sly sundial]

I think having multiple loaders would cause all models to be loaded into memory anyway. Do you really want the overhead of that?

[round summit]

4 works fine, i just dont know how to switch between them from a bot

i can do 2 easy with model switch

[sly sundial]

Does it use comfy backend?

[round summit]

model input switch

yes

from my pc

it runs great, just need a selector with input

[sly sundial]

🙂

https://github.com/JPS-GER/ComfyUI_JPS-Nodes

[round summit]

lol yea LOL

wow

[round summit]

thanks a bunch! this might be it \o/

been here(ai art) from around what was is.. Craiyon? something like that. before disco D. when it was 15m for a bad image XD. i still learn new stuff!

[round summit]

works perfect 😄

[sly sundial]

Awesome!

[mystic plaza]

any idea on what this means bro ? @sly sundial

[sly sundial]

No, not enough info.

[mystic plaza]

this is the command log error text @raven parcel

[scarlet kite]

Anyone have a Img to Img workflow for Pixart Sigma?? I am kinda new to Comfy and got Pixart working perfectly but now I want to use Img to Img

[lapis girder]

Hi everyone new to this forum dont know if this the right place to ask questions please let me know if its not.Basically i trained a lora model with a bunch of specific car images (27) with following parameters- batch size 1,epoch 1,repeats 60,learning rate and text encoder learning rate 0.0000004,unet learning rate as 0.0001,max resolution 1024x1024 on sdxl base model. I have a simple comfy flow( just added the lora between the checkpoint and the clip text encode in the default workflow) . So for 512 x512 mpty latent i am getting okayish images with poor quality for the car but on 1024 x 1024 latent i am getting really good images but all the images have tepeated cars stacked on top of each other. Any idea what might be the issue here( i tried positive , negative prompt for single but not working)

[thorn olive]

Which folder should I put the adapators in?
I need to install 1.5.
I downloaded it automatically through the manager, but it doesn't see it.

[sly sundial]

In the models/clip_vision folder

Did you refresh or restart after downloading them?

[thorn olive]

I'll try again

[fading mural]

what does the X mean?

how do i fix?

[sly sundial]

You can ignore it. I have the same, but it works

[craggy basin]

It says there is no 'auto' value option for VAE in Ksampler Efficient

[fading mural]

I put --normalvram into my .bat but it still launches stuff in low vram, anything to fix that?

[heavy silo]

Everyone, I have encountered a big problem
I am on the cloud server and it cannot find my model
I've been crazy for hours now, and I still can't solve it. Who knows what I should do

[steep marlin]

By getting a GPU with more vram or changing what all you're loading. Comfyui will automatically switch modes to prevent bugs and crashing.

[fading badge]

when i am downloading some nodes i encountering some errors

how can i solve exit code 128

anybody knows?

[stoic gulch]

https://www.youtube.com/watch?v=p1jKqV9IV8I

no more spam sorry i love you all lmk if you need help

[mystic plaza]

can you pls link me to where i can find apply ipadapter script ?

[hoary lodge]

hi, i have a problem, when I try to generate an image after 15 seconds the interface says "reconnecting", when I try to generate it again, there is a "TypeError: Failed to fetch" error. thank you for your help

[proud pawn]

Hi community, I was wondering if anyone is using ComfyUI ControlNet Video Builder with Masking, specifically the node "AnyBus" is not being load for me. Is anyone in a similar situation? Thanks

I mean MarasitAnyBusNode

[alpine panther]

I need help

[mystic plaza]

maybe you need to adjust the model strength and see if that works

[ebon heron]

hi

[wraith scaffold]

anyone knows any nodes for define accurate age or some like ?

[fading mural]

what is text l and text g?

[gentle karma]

https://tenor.com/dgT0f2UMC5l.gif

[spring jungle]

from my understanding G never used in any training only base SDXL model make use of it in covering art style xD

[covert bough]

on the ClipTextEncoderSD3 node - what is the empty_padding value for?

[rustic jackal]

I want to use a node that does a lens blur, however it has only one image input. How would I use the ZDepth to control it? Is there a way to use the zdepth as a mask? Or composite it somehow? Any help would be great ❤️

[dire shale]

hey guys

[grave panther]

Hi

[dire shale]

can anyone tell me how to install comfy ui 3d pack with comfy ui manager? i installed it but i am getting import failed error

[fading badge]

did you downloaded zip file of this?

[livid gust]

Has anyone coded a conditional STRING input for a node? I can't get any other node to drive my conditional input? All wire connections are rejected. I've reviewed the code inside the Show Any node, but I'm no closer to getting this to work.

[frail ocean]

is there a way to use the previously generated depth map image into comfyui?

without having to use a depth map model inside the ui

[fading badge]

guys do you know how to animate gacha type animation in comfyui

can anybody tell me please?

[hoary lodge]

hi, i have a problem, when I try to generate an image after 15 seconds the interface says "reconnecting", when I try to generate it again, there is a "TypeError: Failed to fetch" error. thank you for your help

[fading badge]

bro show the error code

[hoary lodge]

thx

[fading badge]

did you tried restart pc

?

[hoary lodge]

yes

[fading badge]

try decrese width and height size

check internet connection

maybe your internet is slow

and search for antiviruses

i am 80% sure you have some viruses

[hoary lodge]

I use windows defender

[fading badge]

try side softwares for antivirus search

win def is not strong enough

[hoary lodge]

ok I text u later!

[fading badge]

ok👍

[trim anvil]

Your CFG is waaaay to low

Try 8-9

[digital quest]

same

[scenic flower]

/upsacle

[livid gust]

@frail ocean You mean a Load Image node? Save the depth map as an image.

[round summit]

trying to update Comfyui but it seems to be stuck at Install: pip packages for the last 5m. i update every 2-3 days and its neverr taken this long.
Is it safe to close it out and try again?

[hoary lodge]

hi, i check my pc with malwarebytes and he found nothing, and my internet speed is 90MB download and upload

[hoary lodge]

minimum

[fading badge]

so

something in your computer

dosesnt allow to you generate

turn off your microsoft def

and generate

[covert bough]

FoFr just dropped his new node https://github.com/fofr/comfyui-prompter-fofrai

[strong mango]

Good day everyone, I'm looking for a workflow, I think it should be possible, but I don't know which custom_nodes to use
I tried hi-res fixing a picture by dividing the picture into several blocks and calculating them one by one using a method similar to inpaint to avoid exceeding the hardware limit.
I know I can do this manually, I'm just wondering if there is a way to simplify it so that it can be done independently within comfy ui

[surreal whale]

maybe this https://github.com/ssitu/ComfyUI_UltimateSDUpscale

[mystic plaza]

@dry rock so just wondering in comfyui where do we add out animatediff motion models bro in the animate diff folder or ?

!!! Exception during processing!!! Directory 'F:#Projects\1) SD Animation\16) Loka Loka LCM\Passes\HED_Fixed cannot be found.'
Traceback (most recent call last):
File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Advanced-ControlNet\adv_control\nodes_deprecated.py", line 31, in load_images
raise FileNotFoundError(f"Directory '{directory} cannot be found.'")
FileNotFoundError: Directory 'F:#Projects\1) SD Animation\16) Loka Loka LCM\Passes\HED_Fixed cannot be found.'

[mystic plaza]

what does this mean is the lcm missing or ?

[sly sundial]

Does anyone know if the Impact pack facedetailer works using [SEP]? Like adetailer does, as per this https://github.com/Bing-su/adetailer/wiki/Advanced

[strong mango]

Ah yes, that's exactly what I wanted, thank you very much!

[mystic plaza]

how to fix this video setting error bro @sly sundial @dry rock

[dry rock]

Seems pretty explicit to me, "F:\#Projects" is not a valid path

so either it's the wrong folder, or it doesn't like funky non latin characters in the path

never use non latin character in path names regardless of your os, programs, etc if you want to avoid weird issues

[mystic plaza]

but i am not using any latin characters tho bro

[dry rock]

Ok. Is it fixed now? Was the problem caused by the path?

[mystic plaza]

not fixed it yet

its still saying path

[dry rock]

what s the video path ?

[mystic plaza]

how do to check that now ?

[dry rock]

I m guessing it should be the content of the "video path" node linked to the "video" input. I can see it half hidden in your screenshot behind "Video Settings" node.

[mystic plaza]

this what its set to bro @dry rock

[dry rock]

so yup ... there s still some # ) and spaces in there

[verbal sky]

Yippee latest vid's out, enjoy! (Flo2, BG Swaps, String fun, Character color correction & more) 🙂 https://youtu.be/weJhnFWu1Rw

[dry rock]

So does it work after moving the files to full latin/simplier path ?

[loud prawn]

Is there an openPose version working with SD3 yet?

[median berry]

only for diffusers code bases. nobody has converted them to a format that comfyui can use

i don't get it but that's how it is. the cnet models only work with diffusers code

[fading badge]

bro why is the zuckerburg quote in my cmd 💀

[dire shale]

Hey guys I have an AMD Ryzen 9 7950x3d so i was wondering if i can run ComfyUI fast on this very powerful CPU and i also have 32GB ram but i don't have a GPU i am looking forward to buy one but its going to take a while so till then i want to run it on CPU.

[quasi tangle]

pls anyone have a workflow with prompt injection node? pls can you share it with me?

[mystic plaza]

nope

[dry rock]

ok what s the error now and what are the paths ?

[dry rock]

That s because of some node extension (cant remember which one) probably WAS nodes

[novel arch]

how can I add some python lib to do one specific thing to a node to be able to use it in comfy ? I am trying to use "pyautogui" to simulate keypresses, normally it works. We don't even need any output or input , we can add it to the workspace as easyuse uses for cleancpu or to the menu with a new button

[livid gust]

@novel arch Feel free to give my node a try. It's on revision 2, but it is my first Comfy node. It allows you to run python code natively inside the ComfyUI. The code generates output via a print() statement. It also accepts a maximum of three optional inputs, so you can fetch values in the workflow to process them and generate a new output. I haven't tested it with importing external libs, but as long as you can target your import with a valid path, it may work.
https://github.com/AtomicPerception/ap_Nodes/blob/main/ComfyUI_Python_1.2.py

[cursive ridge]

Anyone have a comfy workflow for yiffymix v43? OR one which just includes a custom Vae and vpred???

[fading mural]

is there a way to force force comfyui to use —normalvram? it ran fine on a1111 and i heard low vram drastically slow downs generation

[steep marlin]

Normal is the default

[steep marlin]

And if you need to use a different option, you use the flags --lowvram or --highvram. If you don't have enough vram to handle your workflow, it will automatically drop to lowvram mode. It doesn't really slow down generation in terms of it/s but it will slow things down between steps of a workflow where it has to swap models out like ip adapters or controlnets and so on

If you run out of vram, it will go into the shared memory pool which will be far more detrimental to your generation speeds, or it will cuda OOM error depending on your Nvidia cuda sysmem fallback setting

[covert bough]

@storm folio how about creating a way for models to be stored on a different drive, in something like a models folder, and allow comfy to use them from there? (without having to also install auto1111)

[storm folio]

there's a config file

extra_model_paths.yaml

[covert bough]

i tried that. comfy refuses to find the models where i told it to look, and all i can assume is that since i don't hve auto1111 installed, it's not happy for some reason

[still tartan]

because you have to list out the folders for comfy or it will default to the normal models folder: comfyui:
base_path: D:\AI_Resources
checkpoints: models/checkpoints/
clip: models/clip/
clip_vision: models/clip_vision/
clipseg: models/clipseg/
configs: models/configs/
controlnet: models/controlnet/
diffusers: models/diffusers/
embeddings: models/embeddings/
gligen: models/gligen/
hypernetworks: models/hypernetworks/
inpaint: models/inpaint/
insightface: models/insightface/
ipadapter: models/ipadapter/
loras: models/loras/
mmdets: models/mmdets/
onnx: models/onnx/
photomaker: models/photomaker/
prompt_expansion: models/prompt_expansion/
sams: models/sams/
style_models: models/style_models/
ultralytics: models/ultralytics/
unet: models/unet/
upscale_models: models/upscale_models/
vae: models/vae/
vae_approx: models/vae_approx/

[eager escarp]

paste your config here and ping me and i'll help ya out

i have lots of folders for things

[hardy tree]

lots of nodes, lots of folders, man you have a lot of stuff indeed 🙂

[covert bough]

i figured it out. i had, for some totally unknown reason, a > on the end of the path

[vagrant bough]

How can I solve this problem? I have re-cloned comfyui_controlnet_aux into the ComfyUI\custom_nodes folder.

[mystic plaza]

https://github.com/lllyasviel/IC-Light/tree/main

[mystic plaza]

in which folder do i add this safetensor bro and thanks for the help

[fluid cobalt]

Hey 🙌 is it possible to generate a lightmap images of a reference image? Like these.. any idea? Could not find something helpful in the short tie I had to do a research

[steep marlin]

Filtering with some basic photoshop/krita skills. There is a decent comfyui addon with almost every common tool, so you can pretty much do 100% of the work right in comfy. You'd use levels to crush the blacks darker, along with some other adjustments and then run a fat gaussian blur on it

https://github.com/chrisfreilich/virtuoso-nodes is the pack that has some good photoshop-like tools in it (i use it a lot for minor tweaks that im too lazy to do the photoshop shuffle with)

[steep marlin]

here's an example of what i mean, if you're looking to do it in comfy

i exaggerated it to really show the effect, you wouldn't want to go this heavy handed with the crushing though(forgot to throw in a gaussian blur to smooth it out more)

in Krita, it would look something like this:

photoshop equivalent would be pretty much identical

point is, you'd pinch in the black and white input ranges(leave the outputs levels alone), shift the midpoint gamma toward the whites, and then heavily gaussian blur it

[ocean haven]

I made this free webapp for changing facial expressions of your characters https://fyrean.itch.io/emote-engine

powered by this workflow https://openart.ai/workflows/DZqWIDYghMS7czb5NTGB

[mystic plaza]

in which folder should i add this in bro @steep marlin

[loud prawn]

Thanks for the answer. then I will have to wait until somone converts it

[silver nebula]

ComfyUI asks for torch 2.0.O (compiled with CUDA enabled) to be installed. I use pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118 to install. I start ComfyUI.bat and the next thing is - it decides that it needs torch 2.3.0 - so instantly d/loads and installs that version - whilst uninstalling torch 2.0.0!!! Help!!!

[mystic plaza]

so just checking but is there a way to control the camera when we want to animate stuff in comfyui @dry rock or do i have to create a code of sorts for it ?

[sly sundial]

You were way behind! What help do you need? Looks like you're where you should be now.

[mystic plaza]

@sly sundial any inputs bro

[sly sundial]

Your question is a bit vague. Depending on what type of animation you're doing, there are motion loras. I think they're for animatediff.

[mystic plaza]

its like for a product shot like camera zoom in and then orbits around the product and alll

[sly sundial]

Vague about how you're animating. My answer is the same.

Search Civit for motion loras.

[silver nebula]

Not yet - it is quite confusing! ComfyUI demands I install torch 2.0.0 - and once I have done so - installs torch 2.3.0 over the top!!! Why?

[sly sundial]

You probably had such an old version it thought the latest was 2.0.0, until it updated and realised that 2.3.0 is the latest.

That's what I'm using, stop worrying about it 😄

[mystic plaza]

but will i be able to control the camera movements in it bro

[sly sundial]

Did you look at them?

You need to do a little research yourself sometimes 🙂
I found motion is sometimes a bit vague but there are many loras for different types:
https://civitai.com/search/models?sortBy=models_v9&query=motion lora
Specifically for directions:
https://civitai.com/models/153022/animatediff-motion-loras

[sly sundial]

This may suit what you're after: https://civitai.com/models/541515/fpv-drone-motion-lora

[mystic plaza]

thanks will have a look at these bro

[silver nebula]

I am getting stuck at Huggingface Hub - no matter how many times I uninstall version 0.23.0, and install 0.20.3 - the 0.23.0 version is still installed?!

[mystic plaza]

have you tried installing what its asking for bro ?

[sly sundial]

You're not reading the message properly. It requires at least 0.23.0 to work, but it finds 0.20.3 so it fails and tells you to install a newer version.

[silver nebula]

Errrmmmm - problem solved! I deleted ComfyUI and reinstalled completely!!! Thanks for all the help 🙂

[verbal sky]

🚀 Brand new video tutorial just launched, enjoy and please share! (New sampler speeding image gen time by 70%+, Face detailing using masks, cinema fx bonus) ⭐ https://youtu.be/igO7Ax0BRII

[silver nebula]

I agree, and when I install the newer version - ComfyUI immediately lets me know that the old version is still in place!!! Bah 😦

[sly sundial]

Are you sure you're installing it in the correct environment? If you're using the portable version you need to add the path to the embedded python.

[silver nebula]

OK, let me know what files to add to Environmental Variables (PATH)?

Done - comfyui_portable\python_embeded added to PATH

Will try again ...

[sly sundial]

You don't need to add the variable, just add the path in front of your python command, to make sure it uses the embedded python and not one you have in your OS

[silver nebula]

I am beyond calm 🙂 Please spell out for me what that means - python.exe pip -s -m python embeded ... etc?

I simply cannot concentrate - it is quite stressing for all 3 of my ComfyUI installs to fail together - System ComfyUI, Portable ComfyUI and SwarmUI ComfyUI!!!

Mebbe I should leave it for a while? 😄

[Sorry for my moanin']

[sly sundial]

DM'd you

[hardy tree]

@storm folio hey awesome work on the latest audio support stuff, like load audio (upload), as well as save audio preview and also the empty latent seconds selection. however, just one thing perhaps to fix/add is related to the save audio part: since the audio widget player is just playing the music, like a preview anyway, is there a way to make a simple preview audio node as well, because i can right click and save from within the widget if i want to save the audio, but the way it is now, it is always forcing a save on disk, and i dont necessarily want to always save everything. but awesome stuff ❤️

[runic mirage]

Anyone know a good workflow to tests multiple lots with the same prompts?

[steep marlin]

What do you mean lots? Is that an autocorrect of loras? If so, use the efficiency node pack, they have a really easy to setup x/y plotter that you can use with loras

[shell bobcat]

hello everyone, I was wondering if there was a way to combine two images using comfy ui. If there is a way, what do I need to do it?

[sly sundial]

Like this? https://civitai.com/articles/3099/combine-multiple-images-with-ipadapter-a-workflow-for-comfyui

[shell bobcat]

It's a bit like that, but what I want exactly to do is combine or mix two character images into one single character that has elements of both characters. For example I want to combine a fox and a turtle . The output I want is something like a fox with turtle shell.

[distant pendant]

I think clip vision in cascade was good at that but not sure how that extends to other models

[shell bobcat]

Sounds good, i'll take a look at it

Thank you for the advice guys @distant pendant and @sly sundial

[forest birch]

I recorded a concert, but i managed to get a finger in front of the cam a few times, so i want to remove those few frames that has that, but have either comfyui detect the missing video, or i mask the frametime the removed clip start and ends, to then read from the last few frames to then interpolate at that video's framerate and resolution. Is there a workflow for this?

[severe peak]

Hi, Someone know why this node keep being processed even with no output?

[lusty pewter]

Is that a group node? If you right-click on it, does it have a "Convert to nodes" entry in its right-click menu?

[severe peak]

No it's just a Ksampler Efficient renamed to HiresFix

[lusty pewter]

lol...as I say frequently, people should get away from Efficient nodes. They do things that are otherwise unexplained. You can demonstrate that by recreating a workflow that Efficient nodes are trying to replicate, choose the same settings and feed it the same seed and sometimes get different outputs.

But I'm going to guess that what they're doing inside the node requires intermediary work before the outputs and they're not correctly checking for outputs before performing the work.

[severe peak]

I need to feed a Ksampler with a script, and AFAIK this is the only one that can do that, Also I have a previous Ksampler Efficient before this one, and it's not reprocessing so it's kinda weird

[severe peak]

On the script I use the same seed

I tried to unlink the script from the Ksampler, and its not computing anymore

[lusty pewter]

Then it's as I figured. It's set to process the script but not checking for no output before doing that part.

[severe peak]

makes things so hard, just like when using switches, all the nodes will be processed even if false

[forest birch]

Do you guys know of a neat workflow that can make configurable vectors, or SVG files from images? As i will use lora's to simplify images for easy vectorizing for a few projects

[hardy tree]

@storm folio hey thanks for updating with the PreviewAudio node, works wonderful! also, the new samplers like deis, _pp, etc are also giving some cool results too, always nice to play with new samplers or schedulers. 🙂

[storm folio]

glad you enjoy them

[steep marlin]

See how it has an image pin on it? That means it's likely viewed as a live node because you have vae decode set to true. Set it to false and it will stop doing that

Internally, it's vae decoding the image, even if it doesn't show it on the node

[severe peak]

But If I set it to false and use a latent VAE decode that doesnt have image output connection, this unconnected branch will still compute? I have to try

[carmine copper]

Anyone can help here im trying to add llm using gimini api

[steep marlin]

It shouldn't. As long as the preview and vae decode are disabled within the node, it shouldn't count as an output. You should then be able to connect a separate vae decode node to the latent pin(like you would a regular ksampler), and connect an image preview to that new decode node. If that separate image preview node is bypassed/muted, that efficiency node shouldn't execute any of that branch

[severe peak]

No unfortunaltely the node itself is still computing for no reasons

[steep marlin]

I see, I thought for some reason that was a node from the efficiency pack. Efficiency has a ksampler and a script pin where you can plug in a hires fix node

I should note that I also use rgthree

Which supposedly helps with execution order stuff

[turbid bolt]

are there any nodes other than cg-random that'll randomize models and loras? cg-random doesn't output prompts for checkpoints or loras so doesn't work with extended saves

[cursive ridge]

I'm having to start over, a bit confused! Is this standard for github?

[steep marlin]

no, github is usually a lot faster than that. either their servers are overloaded, repos have some kind of throttling limits to them or there's something wrong with your route to the datacenter

[cursive ridge]

My civitae download of a model took the same size only took a couple of minutes. Though it is a holiday Monday here. Perhaps more people are realizing that comfy is just the best 😄

I don't really want to install, nor learn a new UI for SD.
Is this possible within Comfy instead? Has anyone done it?
"To merge two models using AUTOMATIC1111 GUI, go to the Checkpoint Merger tab and select the two models you want to merge in Primary model (A) and Secondary model (B). Adjust the multiplier (M) to adjust the relative weight of the two models. Setting it to 0.5 would merge the two models with equal importance."

[steep marlin]

yeah you can easily do it in comfy

[cursive ridge]

Would you by any chance have a workflow for that? I'm pretty happy, did not want to start downloading a new UI for each project.

[steep marlin]

you dont even have to save it if you dont want

stock nodes btw, dont even need a custom addon to do it

oh and i'd assume the ratio of 0=model1 and 1=model2

so 0.5 should be 50:50

[cursive ridge]

Thank you very much 🙂

I was using some insanely complex SD3 ones, and they were too much for my GPU, and my newbie comfy knowlege as well lol

[steep marlin]

np, have fun with it. comfyui really is a powerhouse all in one app

[cursive ridge]

I lucked out and started with comfy. So many say it's difficult to learn, but I tried to learn swarm and was completely lost!

[wispy kelp]

any idea why the face detector yolo adteliare cuts off half a hand?

hand detection is all over the place. why?

[modern forge]

  0.0 seconds (IMPORT FAILED): C:\ComfyUI\custom_nodes\efficiency-nodes-comfyui
7.3 seconds (IMPORT FAILED): C:\ComfyUI\custom_nodes\ComfyUI-Impact-Pack```

why am i getting this?

[sly sundial]

You need to scroll up through the startup messages and see why each one failed there.

[modern forge]

ok i fixed each thanks

[mystic plaza]

so i just have a small doubt how can i match or generate images based on the image i have been given @sly sundial

[sly sundial]

By using an image2image workflow, like this one: https://www.patreon.com/posts/stepped-upscaler-107112755?utm_medium=clipboard_copy&utm_source=copyLink&utm_campaign=postshare_creator&utm_content=join_link

That creates a much larger output, and so requires a lot of memory, but there are others around.

[mystic plaza]

welp the image has a green screen and i also have a alphas mask

[mystic plaza]

like i need to match the perspective of the image

[cursive ridge]

All was good until I installed comfyui manager, now: "OSError: [Errno 10048] error while attempting to bind on address ('127.0.0.1', 8188): only one usage of each socket addr" Help?

[mystic plaza]

this is what i am running but also need to generate images as well as match the perspective like i heard there is a way to do it in comfyui but cant figure out how tho @sly sundial also the attaching the images i have here so pls help me out and thanks

[sly sundial]

You've already got an instance of it running and trying to start another one.

[cursive ridge]

I don't though

[sly sundial]

Well, something is running on port 8188

If you can't find it, try a reboot

[cursive ridge]

Why did I forget, what always fixes any problems in windows? 😄

[mystic plaza]

any inputs you can provide me with

[sly sundial]

No

[cursive ridge]

I had to reinstall comfy etc. etc. and I have to say how happy I am that all my old images have my previous workflows embedded right into them!

Some people are extra motivated and keep prompt and workflow spreadsheets; I sort of prefer this way though 😄

[cursive ridge]

Where/how do I connect the FreeU to this?

[sly sundial]

Between the model loader and the ksampler

If you're not sure, then drag some spaghetti out of each end of the node, and it normally tells you what you need there.

[cursive ridge]

Thank you very much. I'm going to stop lisenening to Gemini advanced now!!!!

[silent turret]

Anyone here using Cuda 12.4? Did anyone manage to build a wheel for xformers and flash_attn2, on Win 11 Comfyui

[sly sundial]

Before and after upscale

[lofty stream]

Has anyone tried / been able to get ComfyUI running on AWS or another cloud platform? I have projects I’d like to build but my computer can’t handle it, and want to try something else before building a PC

[steep marlin]

There isn't a prebuilt one and windows is not currently supported for flash attention 2. Though it might work if you go the WSL route, but I haven't tried it.

[distant pendant]

flash attn compiles on windows in general, not sure about cu124 specifically

there are also wheels https://github.com/bdashore3/flash-attention/releases/latest

[fading mural]

is ipndm_v faster than euler_a?

[mellow tundra]

hey team, quick question: in my custom node, what difference does it make if I return my output inside a ui object?

return {"ui": {"text": (output,)}, "result": (output,)}

I've seen some doing it and it doesn't seem to affect anything.

[merry ermine]

Hi - we're looking for someone to help us build an AI image generation workflow that can create a large number of consistent, high-quality images of AI influencers. This workflow will need to main character consistency across multiple settings and poses while producing a high volume of images efficiently.

If you're interested in working on this project - please reach out to me directly

Will compensate for your time

[steep marlin]

He hasn't had any for the newer torch branches in a while. And no, 12.4 doesn't usually compile for windows correctly at the moment.

they're all still for 12.2, but then again, 12.4 is still a nightly build. all i know is it's a problem for those that want to use lumina t2i

because it requires flashattention2. i've wasted hours compiling it over and over(takes over an hour to compile FA on a 13600kf btw) to no avail

[distant pendant]

what's the issue with just using stable torch with cu 12.1 and flash attn wheels for that

it works with exllama

[steep marlin]

lumina uses conda install python=3.11 pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=12.1 -c pytorch -c nvidia -y

along with FA2

but there isn't a combination of precompiled windows version for it

all i know is you'll see tons of posts about it over on the FA github and the lumina github, for now, you have to use linux or WSL on windows

[distant pendant]

I don't know much about lumina but it will likely work with different torch version too, you can also make a github action following the ones in the flash attn wheels repo with whatever combination you need

[steep marlin]

from https://github.com/Dao-AILab/flash-attention:

"Linux. Might work for Windows starting v2.3.2 (we've seen a few positive reports) but Windows compilation still requires more testing. If you have ideas on how to set up prebuilt CUDA wheels for Windows, please reach out via Github issue."

"FlashAttention-2 currently supports:

Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing GPUs for now.
Datatype fp16 and bf16 (bf16 requires Ampere, Ada, or Hopper GPUs).
All head dimensions up to 256. Head dim > 192 backward requires A100/A800 or H100/H800. Head dim 256 backward now works on consumer GPUs (if there's no dropout) as of flash-attn 2.5.5."

the big problem for me is that i'm on a 2080 :/

oh i guess someone finally managed to compile it for 12.4:
https://github.com/Dao-AILab/flash-attention/issues/595#issuecomment-2192690209

at any rate, it's still not fully supported for windows and therefore anything that needs FA2 support will be hit and miss. for me, i won't be able to use FA2 at all until they support this 2080

[distant pendant]

I guess 2080 has cc lower than 8 for some reason? strange

[steep marlin]

i dont know the specifics, but it's also an early gen of the cores, so maybe it doesn't support all the features on the 30xx/40xx cards. kind of like cpus that support avx, but not avx512 or something

[distant pendant]

seems like whole 20xx series is 7.5 (and flash attn 2 needs 8.0), nvidia done fucked up

[steep marlin]

the 2080 came out in 2018

[steep marlin]

so one of those features in red is likely why

[distant pendant]

no idea but 8.0 has been a requirement for quite a while now so wouldn't count on flash attn suddenly adding support

[steep marlin]

exactly, anything they might use as a workaround would be like in gaming, using the CPU or non-RT cores for doing raytracing

aka exponentially slower lol

but fallbacks like that exist for a ton of other things, just to enable compatibility to wider variety of platforms

[sly sundial]

Is anyone else having a problem with nodes not appearing in the list to add? I'm trying to search for the RGTHREE bookmark node, which I have installed and have used in the past, but it's not appearing in the list. I have a workflow with it in, and it works, so the only way I can use it is to cut-n-paste from that workflow.

@storm folio Is this going to be an issue with the custom nodes (rgthree) or the UI?

[storm folio]

probably the custom node

if you check in the browser console do you see errors?

[storm folio]

looks like a problem with the custom node

[sly sundial]

I've tried the workflow link fixer, reinstalled the nodes

Deleted ttN

[mellow tundra]

@storm folio I'm implementing a custom node, using other nodes as a reference. Would you be able to explain why some nodes return the output like this...

return {"ui": {"text": (output,)}, "result": (output,)}

... instead of the standard way?

return (output,)

[storm folio]

that one above is to return stuff to the user interface

"ui": will send that to the frontend

[mellow tundra]

oh got it, so for instance they can implement a js that will receive the value and display it within the node, like pythongosssss' ShowText

[steep marlin]

@storm folio Hey, you don't have to waste your time with this if you don't want, but I was wondering what I'm doing wrong with my image previews for my standalone diffuser. I've tried using several latent2rgb methods but they both start from pure noise and slowly resolve into the image. The last few steps of an image generation look pretty normal and close to the final image.

This is the function I'm using in with my callback(based off your comfyui code):
`def latent_to_rgb_comfyui(latents):
#scale_factor = 0.13025 not sure if it's actually used or needed for this type of decoding
latent_rgb_factors = [
# R G B
[ 0.3920, 0.4054, 0.4549],
[-0.2634, -0.0196, 0.0653],
[ 0.0568, 0.1687, -0.0755],
[-0.3112, -0.2359, -0.2076]
]
latent_rgb_factors = torch.tensor(latent_rgb_factors, device="cpu")
latent_rgb_factors = latent_rgb_factors.to(dtype=latents.dtype, device=latents.device)
latent_image = latents[0].permute(1, 2, 0) @ latent_rgb_factors
latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1).mul(0xFF)).to(device="cpu", dtype=torch.uint8, non_blocking=True)

return Image.fromarray(latents_ubyte.numpy())`

I also tried one that involved, same deal:
rgb_tensor = torch.einsum("...lxy,lr -> ...rxy", latents, weights_tensor) + biases_tensor.unsqueeze(-1).unsqueeze(-1)

I was wondering if there's anything I'm missing that would make my previews more like yours? I've tried different things thinking it's some kind of -1/1 vs 0/1 range thing, tried mixing in the multiplier, etc etc. Oh and this is all based on diffuser library code

[storm folio]

Make sure the scale is correct

[steep marlin]

Where do I apply it? It's for pixart sigma, which uses the sdxl vae

[old jay]

hi, could someone help me with setting up multiple loras for pony in comfyui?

[livid gust]

I quit using rgthree. It takes over the system and lowers my workflow FPS.

[old jay]

i can't connect them both at the same time

[livid gust]

They pass through one another. Try chaining them together and grabbing the output at the end of the chain.

[old jay]

like, connect the first lora's clip to the second's?

i'll try, thank you

oh my god it works

thanks a lot

[sly sundial]

Use a "lora stacker" node.

[tidal tulip]

How make art like this ?

[steep marlin]

i think diffusers library just handles noise scheduling or sampling differently than you do or something. the preview code works fine, i tried it on a saved latent partial diffusion from comfy and get pretty much 1:1 with your preview port. ughhhh, i guess i'll have to go down the rabbithole of code to see where the big difference is. i like the learning experience though

[storm folio]

why are you using diffusers?

[steep marlin]

oh because i was making a boring little lightweight diffuser for the family to mess around with and because 90% of it is way over my head, so i like to experiment to try to learn lol

[near haven]

I've been doing comfyui for a while and the constant wall that I'm finding myself come up against is niche scenarios that end up using basic principles that I just haven't been exposed to yet and I'll often ask community memebers what the solution is. Many thanks to them for helping, but I just know that I can somehow use a local llm to get me answers faster and streamline my skillset. My issue is that I've never tried to use or train an llm. I'm really hoping that there's one out there that is just trained on comfyui, but I have a feeling that there isn't. I know that it's possible to do what I want/need to do, but I don't know where I would start... ><

Anyone have any advice on how to solve this problem?

[stiff tendon]

Use cartoon style checkpoint model it would not be same cuz no one have trained model or lora on indian cartoon style

[wispy kelp]

does supir only work with juggernaut models coz i am having bad results with other sdxl models. 😦

[surreal whale]

work with others

[daring tundra]

Ever wanted to create your own ComfyUI node but don't know how to code? No problem! Check this tutorial:
https://youtu.be/VM3FXXEl5Z4

[nocturne pumice]

does comfyai work with zluda on amd gpus?

[brazen turret]

Any input swithes that only load the selected inputs instead of all inputs? Was trying to use ComfyRoll Suites input switches to switch between models and clips and such. But it's loading everything instead of just what's selected which is taking up a lot of memory. I don't want to have to redo a bunch of sections of my workflow everytime I want to switch some things around.

[steep marlin]

Try looking into the rgthree addon, it might help or have what you're looking for. I there's a switch anything node, along with global settings for recursive execution optimization I think

[brazen turret]

How about something for sub groups or stacking groups inside groups

[winter estuary]

hi

how would you get taesd

working

with a checkpoint that isn't sdxl or sd3

like has a custom name

[steep marlin]

Find out which model the custom checkpoint is based on, then pick the taesd vae version. Or just trial and error it. Try with the 1.5 taesd, if it makes a garbled image, try the sdxl one

It will either work or won't

[winter estuary]

Got it

[merry ermine]

anyone here trying out the new controlnet union sdxl model?

im kinda stuck at selecting controlnet preprocessor, i only see none / tile / tile

opposed to what the video here is showing https://youtu.be/UBFEw1IUX_I

[covert bough]

anyone know how to get the comfy manager button to show back up on the toolbar if it vanishes?

[steep marlin]

use a different preprocessor then like AIO aux from https://github.com/Fannovel16/comfyui_controlnet_aux

[covert bough]

@thorny lily is there a place other than here to report bugs with ComfyUI? i don't really want to open a github issue for this

[thorny lily]

GitHub is the correct place, there's also a Comfy Org Discord linked at the top of https://www.comfy.org/

[brazen turret]

isn't there a prompt library node that allows you to save and select saved prompts and prompt parts to inject into a prompt node? I swear I saw that somewhere but I can't find it

[livid gust]

@covert bough It might be related to the setting of "remember" position. You may have loaded a WF authored on dual screens. What you can do is open Comfy with an empty workflow, go to the gear icon and turn off that remember feature. Then load the troublesome workflow.

[brazen turret]

PFFT Image Resize NUKED my computer XD

[sly sundial]

Did you check that the it started up without any errors?

[brazen turret]

Anyone know why Image Resize is sooo intensive on your hardware? I can run basically everything else in my workflow just fine without lagging my computer but when it hits image resize my entire computer starts to lag extremely hard to the point it has locked up my computer twice. I am simply using it to downscale an image to 1440x2520 from 3072x5376 before running it through my second ksampler for hi-res fix.

[steep marlin]

It's CPU based, so you likely have a weak CPU. If you're just downscaling, use bicubic, lanczos is kind of overkill

[brazen turret]

I have a 5800X tho.....

I wouldn't call that weak

[steep marlin]

Then something else is wrong or your OOM

You're*

[steep marlin]

Oh and afaik, it's a single thread operation, so only single core performance will matter for it

[brazen turret]

that doesn't explain how it's locking up my computer tho...

I've just switched to bicubic and I'm about to hit the image resize so we'll see if that helps

Seems to be a memory intensive operation as I go from 12GB ram usage to the full 32GB

but my cpu is barely touched so I assume I am running out of memory in my RAM

[steep marlin]

Yeah that will do it. It's probably thrashing the hell out of your page file when it does, which will lock the PC up usually where things like even the mouse will hitch

[brazen turret]

any other nodes I can use to resize the image that might use GPU instead or be less memory intensive as scaling down an image shouldn't use this rediculous amount of ram

hmm and my second pass ksampler is running out of memory even tho it's running it in low vram mode, which it wasn't doing before

[steep marlin]

A lot of custom nodes that use models aren't scripted correctly to let comfy manage their memory usage, so any settings like lowvram won't work

[brazen turret]

Yea but this is the standard ksampler not a custom one

[steep marlin]

But what models are leading from start to that ksamplet

[brazen turret]

Like 3 models. PonyXL, one for control net, and one for upscale model cause all others don't get loaded thanks to rgthree's Muter nodes

[steep marlin]

Well something is screwy, it also doesn't help you're trying to sample a 3-4mp image

[brazen turret]

I use image resize to resize back to 1440p

[steep marlin]

Yes, 2520x1440=3.6 megapixels, which is 3x larger than the standard 1024² size. Make sure you're using tiles vae encode/decode and a tiled sampler

Most systems aren't going to be able to sample an image at that full size without tiling

[brazen turret]

What has tiled sampler

I am using tiled decode and encode

[steep marlin]

Yeah that helps at those steps. Use ultimatesd upscale for doing tiled sampling

Google for it

[brazen turret]

I have this node but no idea how to actually use it:

these are the only nodes that come up for a ksampler input

[steep marlin]

https://github.com/ssitu/ComfyUI_UltimateSDUpscale

Use this, it's kind of the gold standard for upscaling

But you can also use it with no upscale

For things like resampling

The node version will say (no upscaling) in it

Other than that, you gotta Google and YouTube for more guides on it.

[surreal whale]

use microsoft paint to downscale

[brazen turret]

so does that work like a Hi-Res Fix cause that's what i was using the second ksampler for

to basically add more details in

[brazen turret]

@steep marlin

[brazen turret]

uhh.. the tile sizes seem to just not match up at all rn and idk why.

[covert bough]

yes, i did check that, there weren't any errors. did get it fixed however, turned out to be browser cache corruption

[covert bough]

i turned that on and off, didn't help. what did, though, was clearing browser cache. got it fixed late last night.

[brazen turret]

Figured it out, wasn't using the image resize node correctly so my image was still like 5K px in height

but the SD Upscale is just making my image look dirtier. like less precise than before. like some of the scales on the skin are deformed and stuff, I'll post some snips

left is original (upscaled using model), right is SD Upscaler output

it just looks worse. instead of enhancing the details it's smudging them

[brazen turret]

Yea I can't seem to get any more detail using Ultimate SD Upscaler. idk if I am doing something wrong or not.

[brazen turret]

@fickle remnant ban this kid

[nocturne ice]

How do i run LORA's in confyUI?

[brazen turret]

I recommend using rgthree's power lora loader or the power prompt which allows you to load loras in the same way you load them in A1111 by using lora:name:strength

honestly rgthree's custom nodes is a great find has a lot of reall nice stuff

especially the muter. with that I have 1 workflow for like everything, just found it yesterday thanks to someone on this server and now I have both SD1.5 and SDXL models setup properly in the same workflow which I can switch between super easily

@nocturne ice

[wispy kelp]

Is Kolors supported? XD

of course it is...

[wispy kelp]

cant get kolors to work. ideas?

its not downloading 15gb model the stuff when pip the requirements txt, i downlaode dit to try to make it look like the tree it shows there. config.json has different names tho

[steep marlin]

Sometimes when you manually download the json files, it will append some bullshit before the normal name. Like instead of it being config.json, it will name it something text_encoder_config.json or prefixed with w/e the folder name was

Double check all the various json files in the kolors folders to make sure it didn't happen

[wispy kelp]

i just deleted the mess I made and apparently the workflow automatically downloads all the files...

[steep marlin]

You might want to swap out the tenc though with one of his premade quants like q8 or 4, there's a second encoder node for loading them

They go in the models/llm/checkpoints folder, make them if they don't already exist

damn, i'm actually kind of impressed with it so far though

[proud void]

is there any official documentation anywhere for comfyui command arguments?

[surreal whale]

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

[sage fractal]

where do i get this?

[sly sundial]

From the manager, "MileHighStyler"

Looks like this one has had a lot of additional styles added to the defaults.

[sage fractal]

it appears i already have this one installed

doesn't work

[void shale]

i need some help, anyone here has a good workflow to make an img2img that is low quality and make it look much better?

[sly sundial]

I have a couple on my Patreon, or I can run an image through it for you if you like.

[void shale]

Its an image with my significant other so im kinda worried posting it on the web
Do you have a workflow? I can share my current one(which yield okayish results, nothing great)

[sly sundial]

Only what's on my Patreon. I ran it on this image to improve it

[void shale]

much better than what mine can do

this is mine, it can barely do its work

[brazen turret]

I really wish comfyui had group stacking.

[void shale]

you mean like batching?

[brazen turret]

so I could have multiple groups contained in a single group. yea like batching ig

[brazen turret]

not that batching

[void shale]

you mean this?

[brazen turret]

the blue box you have is a group I wish I could contain multiple of those inside of another group

[void shale]

so like making a batch of batches of different creations?

[brazen turret]

no

so I can move mutliple groups around at the same time and mute them as well etc

it would be really useful for my workflow

since my workflow contains mutliple types of models and such and I would like to add multiple types of generation to the workflow like img2img and SVD etc all in 1 workflow

[void shale]

mmm im sure that if you could make those nodes yourself it is possible

maybe weeks of research to do so?

[brazen turret]

anyone got tips on using conditioning concat and combine? Not getting the results I expect

rn I am starting simple to kind of get a feel for it but I plan on using a third prompt as well where the first prompt will describe the scene and the second and third prompts will describe 2 characters, altho I'm not sure how well I'll be able to get the character prompts to interact with each other for posing and such. but I figured if I seperated my prompt into multiple I would be able to preserve more of each prompt considering before I was losing much of the prompt in the generation, say no cherry blossom trees, or the poses being entirely incorrect, or the details of the character design being wrong and mixed between the 2.

[sly sundial]

Has copy/paste of nodes stopped working for anyone else? I was able to do it up until today, but now it ignores anything I copy and always pastes the last node I copied yesterday.

[steep marlin]

It's been doing that to me for ages, probably a browser caching issue or something

Or some rogue addon bugging it out

[sly sundial]

Maybe 😦

[sharp raptor]

What is the node that looks at a image and then describes it in text for the clip conditioner?

[steep marlin]

Unclip, but you can also use the small vlms like moondream2 or florence2 for a more natural description

[sharp raptor]

thank you

[sly sundial]

There is also a WD14 Tagger from Pythongosss

[steep marlin]

yeah it's not bad if your looking for DB style tagging

[sly sundial]

The best I've found for natural language is Florence2. Not tried Moondream2 though.

[steep marlin]

if you want captions for sdxl/sd3/other LLM based models, florence2 is the absolute best

yeah lol

moondream2 is pretty good, but it's quite a bit bulkier now vs moondream

florence2 is like the nextgen of captioners though. it's even on par with big chunky models like cogvlm, but even the large version of florence2 is absolutely tiny

i've found next to no reason to use any other for captioning, unless you want a full VLM like llava that can write out pages of descriptions

[verbal sky]

Good morning everyone! New tutorial video now available (fun caricatures and some bonus hacks!). Enjoy! https://youtu.be/0w_DTdc2V4A

[steep marlin]

seems to even work well with artistic stuff (been screwing around with kolors, it's been pretty awesome so far)

[sharp raptor]

had import failed with florence2

[steep marlin]

youre using https://github.com/kijai/ComfyUI-Florence2 right?

[sly sundial]

Just install the nodes from the Manager and they will automatically install everything you need.

[steep marlin]

and make sure to use either the base or the large models, don't bother with the finetune (ft) versions, they are trash for regular use. a lot of model creators make specialized finetune versions of models like these solely for passing certain types of benchmarks

[sharp raptor]

yes I installed this one via manager

[steep marlin]

its probably clashing with some other node in that list, there are other import fails

[sly sundial]

You need to scroll up to see what the errors were on loading.

...or look in the log files.

[brazen turret]

since this chat seems somewhat active rn ima reply to this in hopes of getting some input:

[sharp raptor]

I guess its this but when i run pip install ml_dtypes it says requirements already satisfied at 0.3.2

[steep marlin]

how are you running it? you might just be updating your global python install

[sharp raptor]

Like this

[carmine copper]

How the blending in this result

[steep marlin]

yeah that's wrong, it's just going to update your global install. to install for comfy it has to look like this: G:\ComfyUI\python_embeded\python.exe -m pip install ml_dtypes change the path to your comfy folder

comfy uses a venv, not the global install

[sharp raptor]

still not 0.4.0

[steep marlin]

no that's still wrong

you have to use the absolute path to your python.exe within the comfyui folder

hang on, ill make a copy paste for you since i can see your path

C:\AI third\ComfyUI_windows_portable\python_embeded\python.exe -m pip install ml_dtypes

[steep marlin]

try that

might need " " at the start and end since there's a space in the folder name

so "C:\AI third\ComfyUI_windows_portable\python_embeded\python.exe" -m pip install ml_dtypes

or just manually navigate to your python_embeded folder and run the python.exe -m pip install ml_dtypes from within that folder (you can type cmd in the file path in the folder browser to open a cmd with the current folder)

[sly sundial]

You should start with a .\ if you do it from the folder though.

[steep marlin]

don't need to, it takes precedent over global if you type the full python.exe name out within the folder

[sly sundial]

Best to be sure, not knowing the PATH 🙂

[fading badge]

bro it doesnt matter how much time i uptade comgyui manager, it always says try uptade, why this happening and i cant import ultimate upscale node too

[steep marlin]

thats the folder, now run python.exe -m pip install --upgrade ml_dtypes

you might need the upgrade tag

[sharp raptor]

Thank you!!

[steep marlin]

did it work?

[sharp raptor]

yes have 0.4.0 now

[steep marlin]

alright cool, go check and see if the node loads now

[sly sundial]

It's already up to date. That button is always there.

[sharp raptor]

yes the node is there 😄

[fading badge]

it is always shows this?

at everyone?

[sly sundial]

Yes

[steep marlin]

grats, now go pick up your "i survived python venv library updating" tshirt lol

[fading badge]

i have one problem too

i cant import ultimate upscale node

[sly sundial]

You need to look at the errors when it tries to load the node.

[fading badge]

explain please

it just says import failed

[sly sundial]

Look in the console, or the log files, for errors when the upscaler tries to start.

[fading badge]

where i can see lof files?

i am new i dont know so much things

[sly sundial]

In the comfyui directory

The error will be in the text window that opens when you start Comfy. You'll need to scroll up to see it.

[fading badge]

ModuleNotFoundError: No module named 'numba'

maybe it is the problem

i dont know what is this

[sly sundial]

Yes, very likely. Like Platinum had to do above, you'll need to install the missing module.

Can you send a screenshot of the full error?

[fading badge]

yes of curse

[sly sundial]

You have several custom nodes that are failing to load. The numba error is for a different one.

That's for WAS node suite

[fading badge]

yeah you are right

i recognized right now

but how can i download a module?

why this is not downloadin automaticly?

[sly sundial]

It's the "fun" of using Python 🤷🏻‍♂️

Try deleting Upscaler and reinstall it through the manager, restarting between each stage. It should install requirements on startup.

[fading badge]

it didnt work

i tried so much time

manually and from manager

[sly sundial]

The full error isn't in your capture, but it says something about missing file/directories

That's not always what it means though 😄

Is you ComfyUI updated?

[fading badge]

i am uptaded several times in these days and wished some day maybe the missing node will be import

[sly sundial]

Can you share your "C:\ComfyUI_windows_portable\comfyui.log" file?

[steep marlin]

C:\ComfyUI_windows_portable\python_embeded\python.exe -m pip install rembg

to handle the ModuleNotFoundError: No module named 'rembg' error

just copy paste that exact command in cmd, should execute it without having to do a bunch of navigation

[fading badge]

good?

[steep marlin]

looks it, try seeing if the node imports now

[sly sundial]

If you see errors like this: ModuleNotFoundError: No module named 'rembg' then you just need to run that pip install and change the module name to anything that is missing.

[fading badge]

still it give errors

again ultimateupscale is missing

[brazen turret]

it has support already, I read up on it just the other day, drop them in lora folder

[sly sundial]

I don't know why you have two installs of it, but that won't help.

   0.0 seconds (IMPORT FAILED): C:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_UltimateSDUpscale
   0.0 seconds (IMPORT FAILED): C:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_UltimateSDUpscale-main

[brazen turret]

like I said it should have support already in the latest release. They pushed a bunch of silent updates for it in the last 2 months from what I have read

[sly sundial]

It has been available for a long time. Just treat them the same as any lora

[fading badge]

i forget unistall first one before manual install

i must delete and reinstall again?

[brazen turret]

you must be trolling

there is no way you aren't

[sly sundial]

Probably best to stop Comfy server, delete them both, restart and install from manager...then restart again

2 weeks

[brazen turret]

don't humor him, he is clearly trolling, he's asking when DoRA support will be added

when I and someone else have already told him it's already supported

[sly sundial]

I know, I answered him. 😄

#🧣｜comfy-ui message

[brazen turret]

now... if only I could get my conditioning concat/combine nodes to do WHAT I ACTUALLY WANT

AHHHHH

[fading badge]

bro new enemy boss showed up

exit code

so can anybody help me about this?

[sly sundial]

With what? There's no error.

It was downloading and you stopped it....or it couldn't download. I'm not sure.

It looked good until then anyway.

Have you tried Fooocus? 🤣

[fading badge]

bro dont you see it says failed😭

why are you messing with me

i just need this one node bro

why is this so hard

[sly sundial]

It failed because it had an exit code, which means something interrupted the download. I'm out now so can't help any further. You just need to restart and let it finish downloading.

[fading badge]

i will unistall this and download again

because comfyui manager just shows try to fix option

and when i try to fix it shows again same thing failed

[sly sundial]

Further up...looks like you have network issues.

[fading badge]

exit code1

[fading badge]

i dont have it bro

just when i try to download something from github

my download speed slowing

i dont know why

[livid gust]

You can trust to manually download that one. Just place the folder into custom_nodes. Remember to remove that "-main" on the end of the folder. That's how you ended up with two copies. If you remove that, the manager should be able to treat your manual install as if it was a pulled version.

[fading badge]

i am doing right now

yeah import failed again

[ebon jasper]

hi, does anyone here know if/how its possible to run the krita ai plugin without installing their required models?

i dont plan to use any of them and installing them would just unnecessarily bloat my system :/

[arctic night]

Hey, everybody. If anyone knows how to help, please help.
Yesterday after loading Liveportrait I had everything working fine, but was missing one node. I was advised to do this (git fetch --all , followed by git switch develop)- in the Сomfy shared folder. In the beginning I did it via cmd in the "ComfyUI_windows_portable\ComfyUI\custom_nodes" folder. Then it turned out to be a mistake and I had to do it in the "ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-LivePortraitKJ" folder.
Liveportrait started working fine, but other nodes stopped working. I tried updating everything, removed the "ComfyUI-LivePortraitKJ" folder from "Custom Node". Restart computer, remove the failed nodes, and try loading the nodes through Manager. But it is useless, these nodes are loaded but give an error.
And it's not just one package, it's several. These are very important nodes to me. Who knows how I can fix the error without deleting Comfy- portable?
In the screenshot you can see which nodes are not loading.

I also have a warning.
Could you please tell me in which folder it should be done via cmd? Otherwise I will break everything again:))
[07/10/24 20:31:44] WARNING Your inference package version 0.9.13 is out of date! Please upgrade to version 0.13.0 of inference for the latest features and bug fixes by running pip install --upgrade inference.

[merry ermine]

i've had issues with portable version, i suggest using a proper installation of comfyui that git clones the repo... here are the two links you'll need

1. comfyui https://github.com/comfyanonymous/ComfyUI
2. comfyui manager https://github.com/ltdrdata/ComfyUI-Manager

[arctic night]

Thank you. I'm thinking of reinstalling too. It's also possible that Windows 11 was updated last night. Maybe that's why, I don't know.

[merry ermine]

no, can't be, i use windows 11 too with all the recent updates

also when you install nodes after comfyui is up and running, make sure to check if there are notes in red on those nodes that can conflict with other nodes

[arctic night]

Yes, thank you. I try to upload nodes without any problems.

[merry ermine]

i mean downloading and installing custom nodes via comfy manager, cause some of them are deprecated and can conflict with the newer ones

[analog forge]

Do I need a seperate Python install to create Lora with ComfyUI? I am getting this error when Itry to do the last step with the <Lora Training in ComfyUI> custom node.

[sick badger]

Did you try "pip install accelerate" command from inside your environment?

If you run venv's you'll have to pip it into that venv. If you run a global python environment, you can pip it from anywhere as long as python is set in your env variables

[steep marlin]

By default, comfyui uses a venv. To add or change installed packages for comfyui, opena command prompt from within the python_embedded folder and run python.exe -m pip whatevercommandshere

That will ensure anything you add/remove will only affect the virtual environment that comfyui uses

Otherwise, it will default to whatever global PATH version of python you have installed on your system and comfyui doesn't use that

[sharp raptor]

What is going on..? 🤔

Normal comfy installation stopped working after I installed a custom node. So I ditched that for portable. This morning I tried starting Portable and it gave me same Cuda error. I have not done anything and it worked yesterday 😂

What should I do?

[steep marlin]

https://pytorch.org/ and find the right version to use for comfy

think it's on 12.1 now or at least works with it

[sharp raptor]

Btw fooocus UI works fine

[steep marlin]

yeah focus is probably using its own venv

these types of apps are meant to be ran with a virtual environment(venv) and if you try to use a global python install, it gets screwy real fast

[sharp raptor]

Ill check the pytorch site

have to brb atm

[steep marlin]

yeah thats the base cpu version of torch

youll needpip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

[sharp raptor]

everything already satisfied

[sly sundial]

Did you run it in the Comfy venv?

Same as you were shown yesterday... C:\AI third\ComfyUI_windows_portable\python_embeded\python.exe -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121, or run it from within the embed directory.

[sharp raptor]

Oh so the embeded folder is the venv?

[sly sundial]

If you're using the windows portable version, you always need to run pip that way.

It just makes it use the embedded python instead of your locally installed version.

[sharp raptor]

okay!

[wispy kelp]

is there a way to put the same promt in multiple boxes at the same time? like relay it somehow

coz often i cant reuse the same text input coz of the CLIP conflict form different models...

[steep marlin]

yes, on the text encoder node, right click it and convert the text to an input pin, then doubleclick the pin

it will make a new node that you can put text into and plug into many other nodes at the same time

[steep marlin]

like this, like i said though, make sure to doubleclick on the new pin after you convert it. it will create a text node for you

[steep marlin]

oh and it also works on pretty much anything for any node. like you can do this with sampler names, seeds, cfg, steps, etc etc. pretty much any value on a node can be converted into a pin like this

[wispy kelp]

comfy never fails to amaze

a better question would be is there anything comfy cant do? XD

[mystic plaza]

i having this kind of error where i cannot run it as a adminstrator as it is saying system is not compatible

[devout scroll]

Hello everyone

[analog forge]

I need help setting up a workflow to train a Lora. I can get the text description files to generate but It will not generate the actual Lora file at the end. Is anyone familiar with doing this?

[sly sundial]

If you don't get any joy here, try in #🔧｜finetune

[analog forge]

Thank you!

[carmine copper]

What the best way or fixed ipadapter weight value to input?

I set mine at 0.90

Im building a webapp and i need find a solution for a fixed weight ipadapter or a solution to do automatically based on the template refrance

[livid gust]

Does anyone have a python example demonstrating how to fetch information about the current node? Most notable, I'd like to fetch my own ID and know when my node is bypassed.

[bleak stump]

https://www.youtube.com/watch?v=6VVZk_J5vDk

[livid gust]

@carmine copper You could cheat if you're in control of the Comfy distribution. You could alter the code, line #214 inside IPAdpaterPlus.py found inside folder custom_nodes\ComfyUI_IPAdapter_plus. Change the default end_at value for the node from 1.0 to 0.9, and resave the python file. Then all new instances of the node will have your default value. Note: this will not change values of nodes already in workflows.

[steep marlin]

What do you mean, locally? Or some kind of telemetry BS to track people that might use some node you made?

[livid gust]

I'm trying to make my node more aware of Comfy system features. If you've used the system for a while, you've probably experienced a bypassed node still executing. This is quite annoying, SD Ultimate Upscaler, I'm looking at you. I thought the Comfy UI would handle bypassing, but I wonder if I might need to poll some variable to determine what state my node is in to detect if I'm bypassed.

[surreal whale]

im not advanced in any of this but ive used this to fix a thing https://github.com/Trung0246/ComfyUI-0246

[steep marlin]

No, I haven't experienced bypassed nodes executing and I'm notorious for have a huge messy "desk" of various projects where half of the node clusters will be muted or bypassed. Do you have a specific example showcasing this happening?

Also include your comfyui log, since there could be plugin or library conflicts

Oh and I use ultimate a lot, 99% of the time, that stage of my pipe is bypassed until I find what I like.

[brazen turret]

Can someone with more advanced knowledge of Comfy give me some tips on what else I can add to my workflow? This workflow is supposed to be my ultimate all in one workflow that should have everything but not sure what to add now.

I would love to have a better Hi-Res fix added in but in all my experimenting it usually just degrades the quality and doesn't add more detail to the image like I want. Assisting it with LineArt ControlNet seemed to help it nuking the image but alas that's the best I was able to achieve.

oop seems I missed a few of the NSFW models in the workflow when I was trying to clean it up to post here for help. Please don't get mad at me staff I did try to clean it up beforehand

also don't get mad that it's a GIANT WALL OF TEXT, I didn't realise it would all be 1 line of json causing the huge wall.

[livid gust]

@steep marlin My log is a mess. I just noticed Comfy is loading every custom node three times.

It looks like the log also triple processes everything. It's like there are three copies of Comfy running on top of one another. Which might explain why a bypassed node might fire off an update event.

[merry ermine]

lol that looks weird and funny, are you running portable version?

[livid gust]

Yes, I am.

[merry ermine]

the way it looks in your console its messed up

i would recommend git cloning the repo

[livid gust]

Thanks for the instructions. Comfy already runs. I've removed all custom nodes and the triple problem went away. So there are indeed conflicting nodes somewhere. I'm copying back in each custom node set and conducting a restart to verify what I actually need.
Why does everyone feel like they have to make an Integer node? So many duplicate functionality in most node sets.

[merry ermine]

integer node? you mean for seed?

[round summit]

in Comfyui does [car|robot:5] work like in a1111 where it starts car then 5 steps in chages to robot?

[merry ermine]

your framed the question nicely, so i asked chatgpt

[sly sundial]

I would recommend you don't. This requires you to install and configure a lot more than using the portable version. Not ideal for anyone who doesn't know what they're doing.

[merry ermine]

oh? didnt think you would need to install useless nodes at all, only the essenetials that you have control over

[round summit]

nice, thank you 🙂

[sly sundial]

You need to install the python environment and everything that goes with it.

[merry ermine]

installing python is pre requisite whether you run portable, or direct repo

also if you wanted a1111 instead

[sly sundial]

No it isn't

[merry ermine]

you run comfyui w/o python?

[sly sundial]

Portable comes with its own python embed.

[merry ermine]

bad idea

installing python takes a few seconds btw, and you can safely run python version python-3.10.6-amd64

i suggest intalling git and python separately

Git-2.45.1-64-bit (all default settings)
python-3.10.6-amd64 (make sure to tick add path to environment check box)

[livid gust]

I managed to slim down my custom_nodes folder quite a bit, and the triple log-issue has not resurfaced, yet. I like the portable and its embedded python. It gets me up and running quicker.

[hardy tree]

@storm folio hey just curious, what is that latest AuraFlow stuff? is that some new model that was released somewhere? 😮

[storm folio]

it's a model that's releasing soon, the authors gave me early access and I got permission to push the code before the release

[hardy tree]

ah awesome 🙂

[merry ermine]

good to know. but installing python and git separately is nothing complicated in case you were concerened with the installation process.

also doesn't take more than a few seconds to get the comfyui open in browser

[devout scroll]

Does anyone know how to drive this and how to recognize animal faces?

[storm folio]

https://huggingface.co/fal/AuraFlow
They released it

[limber gull]

For some reason everything I produce is low rez and nasty. Any ideas on how to fix this?

[sly sundial]

First try 🙂

[sly sundial]

Settings are different for turbo models, so you need to adjust your steps and CFG accordingly.

Workflow should be embedded, but it's same as yours with adjusted steps and CFG...and different model.

[limber gull]

Image cued. Waiting for the turtle to catch up.

I'm using SDXL_Turbo 1.0

[sly sundial]

Isn't that just the base model? You may want to download a finetune

[true onyx]

I'm a new user trying to lean edge detection. I put my control net from start 0 to end 1 and while the lines look good on the anyline preprocessor the quality is much worse on the generated picture. How can I fix this?

[steep marlin]

lower the strength on the cnet node

you will rarely ever want to use anything above 0.9

depth is a lot more forgiving though, so it can be used up to 1.0 easily, but canny will usually fry the hell out of things if you have it at one.

also, maybe lower the cfg a little as well, i see it's at 8.0, maybe try it at 6-7

[true onyx]

put strength at 0.9 and cfg at 6, still looks fried to me

[steep marlin]

Lower it even more then, you'll have to experiment. Mistoline is pretty potent stuff, so you might even need to go down to like 0.5-0.65. Basically, lower it until it reaches the point of losing its tight fit

Also, it could just be a bad seed

But when it comes to values for settings, there is rarely a one size fits all, so it's all down to experimentation

[sly sundial]

😄

[true onyx]

how??

[sly sundial]

Using my own workflow, which is on my Patreon ("Creators Lounge" tier only)

[true onyx]

I do notice that for both your images the number 1928 (right under the emblem) is not generated correctly. Is that asking too much or is it possible? here is my mistoline output if someone wants to give it a go

[sly sundial]

The details are not clear enough. I cut the image from your screen shot and used that

[true onyx]

here's the original image i downloaded if you want to give it a shot

[sly sundial]

Yep, that's much better 🙂

[true onyx]

that's a huge improvement. the only flaw I still see is the 'EBU 28" on the top right but I feel like i'm nitpicking

[sly sundial]

Yes, you are 😛

If it's required for advertising, photoshop it

[true onyx]

I pm'ed you btw to not flood the entire channel with 1 specific topic

[stray siren]

Just swapped to comfyui from a1111, my generations with the exact same prompting, loras, checkpoints etc are giving me... issues

[hardy tree]

that's big boi, 16gb 😮

[stray siren]

nsfw warning: https://imgur.com/a/3bbX4gS 1st one is what im trying to replicate, in a1111 it was pretty close other than expression and i checked the author's properties, it was using comfyui

but every single one im getting now is like the 2nd picture not sure what is causing that (im trying to get a human lol)

but its all mlp?!?

[fading oyster]

are there any common optimizations I can do when running comfyui to speed things up with minimal quality impact?

[turbid bolt]

to speed up my gens i use unipc and low steps

unipc works well at low step counts i find

[wanton kestrel]

https://tenor.com/beh9p.gif

[distant lagoon]

does anyone have luck running this on rocm in a stable fashion? when I run this on cuda in runpod vs local on rocm, it's night and day different

doesnt crash, no OOMs etc

[daring tundra]

hello

Ever heard of the Strawberry People?
They're an incredible hidden tribe deep in the Amazon rainforest, and they need our help!
Join me in my latest ComfyUI tutorial

👉 https://youtu.be/SWPKJIg--8w

[true onyx]

can someone help me out with this error? I googled it but only found posts that say it was resolved about a year ago?

[livid gust]

@true onyx Looks like a mis-match between CPU and GPU. Have you altered the command line which launches Comfy?

[true onyx]

No, I have not. I'm going to do a clean install and see if that helps

[true onyx]

I just updated the NVIDIA studio driver and that seems to have fixed it

[true onyx]

it actually returned after not having it for about 5 image generations, no clue what's going on

[steep marlin]

youre running out of vram, it's moving stuff to sysmem

[steep marlin]

try this and hit apply, run some tests and see if comfy spits a big OOM error eventually. if it does, that would have likely been when you would have seen the Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0 error

[true onyx]

I did this, I'm getting OOM errors now

[steep marlin]

Yeah you need to lighten your load then if you're ooming. What mode is comfy in, normal or high vram mode?

Also, depending on your workflow, if there are any big text encoders like t5, use lower quants of them

Like use fp8 instead of 16

[true onyx]

I don't know what to tell you, but even with no system fallback i'm getting the same error

[steep marlin]

well it's probably something in a custom node then that's requiring the tensors to all be on the gpu then, but since you're ooming, comfy is moving stuff around. do any of the nodes say bnb or bitsandbytes? like check any model loaders for it

also, maybe try adding --disable-smart-memory to your comfy launcher and maybe add in --lowvram to it

[sly sundial]

It starts in normal_ram but switches to low_vram when it loads his models. I think it was just doing too much for 6GB VRAM.

[steep marlin]

oh damn, didnt know he was on 6gb vram

[true onyx]

According to my task manager i have 8 GB dedicated GPU memory but 16 GB shared memory usage

[steep marlin]

the ideal amount of shared memory usage is zero btw, you never want cuda operations to spill over into it

what model are you using and what kind of controlnets, etc etc, we need more details

[true onyx]

I paid for @sly sundial s workflow so I'd need his ok to share this info

[steep marlin]

dont need the workflow, just tell me what kind of models are being used. is it sd3? sigma? sdxl? are there depth cnets? ipadapters?

[true onyx]

sdxl, there is an ipadapter, 2 control nets and a depth estimator

i just run out of memory with depth anything, it's the latt model to be loaded before the crash

[steep marlin]

change some settings then. lower the depth anything resolution down or something or make it use a smaller version of it

like if it's depthanythingV2, use the base size model

for the actual cnet, if it's using a stock sdxl cnet, there are 128 and 256 versions of it, use the 128

[livid gust]

I can't get IPAdapter to work with SDXL. Try using a SD1.5 model instead. Control Nets are also limited to 512x512 image sizes, right?

[zenith quiver]

Is live portrait only in comfy?

or is it in auto?

[mystic plaza]

how to fix this error

[ebon jasper]

opinion on drltdatas workflow components?

[mystic plaza]

pls help me out with this bro @dry rock

[covert bough]

https://christian-byrne.github.io/custom-nodes-security-scan/

[zenith quiver]

Hi I am having trouble downloading insightface can someone help?

ERROR: Could not build wheels for insightface, which is required to install pyproject.toml-based projects

[sly sundial]

This has some tips https://github.com/deepinsight/insightface/issues/2430

[fading mural]

what did wang do

five nights at monkey

[maiden wigeon]

Hello everyone,

I have been searching for weeks but I can't find how to do it. I would like to automate the creation of images on SD from a Python code(without webui). I don't know where to start. I can code but I don't know how to create the bridge between my code and Stable Diffusion. Do you have any resources to recommend?

---

[merry ermine]

most of the UIs can run in API mode

[maiden wigeon]

Comfyui Can? How Can i create my own api hosted on my computer?

[steep marlin]

https://github.com/comfyanonymous/ComfyUI/blob/master/script_examples/basic_api_example.py and comfy has to be set to apimode or whatever it's called. It's a launch flag. Probably --api and --share, but I'd have to double-check

You make a workflow and save it with dev mode

And then you can call it and fill in the blanks or change settings on the nodes of the workflow

I've used it with openwebui before(an all in one LLM chat interface but can also be linked to a1111 or comfyui to generate images in the chat window like chatgpt does)

[ebon jasper]

do any of u know a node that can do controlnet preprocessing?

[storm folio]

There's a canny one in the base install but for the others you have to use custom nodes

[ebon jasper]

somee else told me that theres a node in the aux preprocessors thats an AIO node, i missed that wrhen looking at the nodes

btw, where are the node templates saved?

mine sometimes disappear or reappear seemingly randomly

[cursive ridge]

The commands menu being on the bottom is AWESOME! It no longer (usually) dissapears all the time now!!!!!

[maiden wigeon]

Thanks a lot i'll try this

[copper grail]

I have some issue with stable diffusion i can't figure it out how to generate good images. i follow allot of tutorials and everytime i get very weard images. like the sky, clouds etc or so weard. What is the isseus is it because i need to generate more to train it or what is the matter that i get such bad images generated

[sick badger]

Way too many weights in your prompt probably burning it

Also increase to 768x768 maybe

Of those don't help, turn off that Lora

[glad wind]

I'm learning all this too and have been having similar issues. From what I can tell, the models are very picky when it comes to CFG and sampler step settings. Taking some time to find the proper settings through experimentation helped me get some good renders. I would try simplifying your prompts according to your models recommended prompt formatting (read the fine print for your model on CivitAI), and adjusting your sampler steps, CFG, and Denoise strength to match your specific model (again, reference CivitAI for your particular model). Different models give drasticly different results. Let me know if you are able to find a solution that works for you.

[copper grail]

Tanks allot for your help ill try those out and let you know if still have issue

i found out that the issue is lcm-lora

[livid gust]

Your in the wrong channel, but I'd guess your CFG is too high. Lower it to 4.8, then try 1.0. It depends upon what model drives your workflow. For LCM workflows always start with CFG=1.0.

[ebon jasper]

https://github.com/ltdrdata/ComfyUI-Workflow-Component
does anyone know why this isnt in comfyui manager? should i avoid using it? i want to use the image bridge node, are there any other extensions that provide such a node?
GitHub
GitHub - ltdrdata/ComfyUI-Workflow-Component: This is a side projec...
This is a side project to experiment with using workflows as components. - ltdrdata/ComfyUI-Workflow-Component
other parts of it are kinda buggy for me so im not sure if i should use it (the image refiner)

[half crater]

Is there a workflow that can handle these complex compositions?I couldn't do any of the openposes I tested

Especially these complex compositions, there is no way

[ebon jasper]

in comfyui when using the set latent noise mask, how do i control how much influence the stuff below the mask has on the generation?

[steep marlin]

The denoise amount on the ksampler. In terms of how much of the base you'll see: 1.0=zero influence on the inpainted area, something like 0.3-0.5 will mostly retain the base image, anything lower and very little will change

[ebon jasper]

wasnt there another setting in A1111 besides the denoise strength?

i thought i remembered a setting of how much of the original its able to "see"

[steep marlin]

Idk, I don't use a1111, but it's the denoise strength

[cursive ridge]

Help? I"m trying to get Reactor faceswap to run. I've searched for reactor within the ui manager and nothing, I clicked on an open space and searched, and nothing. I installed reactor faceswap from github clone and nothing. I also gitcloned and installed the Reactor node, and still nothing. Did I install them into the wrong directories or something?

[river monolith]

https://tenor.com/bzbmt.gif

[queen temple]

maybe that node is not supported any more

[sly sundial]

Most likely a missing package, but you'll need to look at your terminal window or logs to see the error on loading.

[cursive ridge]

It does currently work for other people, so I've definitely done something wrong...

[hollow bay]

Hey guys, can anybody tell me, how i can get greater resolution on my ouput?
I really dont know how to use emptysd3latent image in combination with my vae encode for inpainint ( due to my masking )

[livid gust]

Try inserting an Upscale on the blue-pixels line before you VAE Encode, delivering Latents to the sampler.

[merry ermine]

do you need more resolution, as in pixel count?
or do you need more details?

resolution is about upscaling as atom.p said

details can be boosted by many methods

one of the strongest is noise injection

stop the sampler part way and add something like velvet noise to the partly-finished latent

then resume with a second sampler node

there are some pre-made samplers that inject noise automatically, but even once can add a lot of detail

if you want a good example noise, add these two together

left is pink and right is velvet

if you want to make your own then Power Noise Suite with Power-Law Noise and Blend Latents are good

the way you blend the noise is crucial, because these two noise colours want to make a VERY different image

the pink noise wants to make a nice hazy artistic image

the velvet noise wants to make a super sharp contrasty realistic image

so you can get the two noises to fight each other and get a good middle ground

a huge benefit is that this does not require changing your CFG setting

[west jacinth]

Could you tell me how to disable animation looping? Closed loop false, what else ?

[bleak spoke]

help me pls, Why do I generate such art?

[queen temple]

modern art, cool

[bleak spoke]

noo

[queen temple]

looks like an xl model on a 1.5 vae

[bleak spoke]

I want it like in the first art

[queen temple]

i suggest to embrace the change

[bleak spoke]

?

[queen temple]

yep, a vae problem

[bleak spoke]

What is the vae for AutismMIX SDXL?

[queen temple]

yep

[bleak spoke]

?

What kind of vae

Can I have a link?

[queen temple]

i suggest this https://civitai.com/models/140686/fix-fp16-errors-sdxl-lower-memory-use-sdxl-vae-fp16-fix-by-madebyollin

[bleak spoke]

how install?

Just put it in the vae folder?

[queen temple]

y

[bleak spoke]

😦

[queen temple]

what's the cpf?

are you using something like lcm?

try to use the vae already inside the model

[merry ermine]

pls post workflow

[merry ermine]

maybe try getting rid of Model Sampling Discrete

[golden iris]

Is there a way to reinstall Comfy, without having to delete it all? Having some issues with an upscaler math expression. The execution just... Stops.

[proud void]

Does comfyui support prompt comments?

[merry ermine]

don't think so

but you could add a node

I am guessing you mean adding this // to the start of a line and then it ignores that line

[proud void]

something like that or equivalents

[merry ermine]

fairly sure it doesn't exist

there are so many custom nodes that are to do with prompts though

one of them might have it

would be easier to just make a new node for this

[proud void]

I am coming from webui that supports this functionality, but I can go without it or shop around a node

[sly sundial]

You could add a node to recognise and strip comments before passing the rest to the prompt node.

[livid gust]

Can anyone recommend a node set that will extend an animation. Say I have a 16 frame animated diff sequence. I'd like to extend that to a few seconds.

[bleak spoke]

Please help me how to install ComfyUI for AMD 6600 video card

[chrome ether]

Is there a constraint on upscaling when using --disable-cuda-malloc?
I'm runing ComfyUI on my old PC with a gtx970 and I can't get it to upscale an image, and it actually generates an uglier version.

[prime portal]

does anyone know where samloader models are located, say for instance "sam_b.pit" ?

[proud vale]

I am totally new to comfy-ui , are there some tutorials describe what the different nodes work for ?

[sly sundial]

There are far too many nodes to know what they all do. The best thing to do is try and build your own basic workflow and/or look at existing workflows and try to understand what each part does. You juts learn the nodes as you go along, and new ones appear all the time.

[hollow bay]

hey galaxy, do you know whats the best method to inpaint, but keep the original style (in comfyui) @sly sundial

[sly sundial]

Inpainting is something I almost never use, so probably not the best person to ask.

[mystic plaza]

so anyone tried messing around with the blender 3d rendering side of things do we need to add materials texture in our scene when rendering along with the depth pass line and cryptomatte pass would like to understand how this works

[mystic plaza]

@dry rock

[dry rock]

never used sd with blender sorry

[hollow bay]

alrighty, thanks for the quick reply tho!

Does anybody know, how i can improve the edge transition of my inpaint? (i marked the inpaint area with red lines)

[prime tide]

comfyui fills my ram until my pc crashes

And my ram usage goes up to about 16 gb total

[merry ermine]

yeah thats fine

[prime tide]

at the usage keeps growing until my pc crashes

[merry ermine]

can you get more ram

[prime tide]

I've got 32 gb

[merry ermine]

it went up to 32gb ram?

sometimes a custom node can have a memory leak

[prime tide]

I wonder which one it is

[sly sundial]

Have you added a blur to the mask edges?

[hollow bay]

No

[sly sundial]

You should

The flow is embedded in this image

[hollow bay]

bro, youre a legend, always helping me out

I will asap try this

[hollow bay]

Hey man, just wanted to say thank you, worked like a charm.
Ill just have to play a bit with the settings! @sly sundial

[mystic plaza]

cartoon style or a 3d render just curious

[hollow bay]

https://civitai.com/models/340599/zavys-cute-isometric-tiles-sdxl

[mystic plaza]

i see

[hollow bay]

That was my prompt: big black pug face temple, zavy-ctsmtrc, isometric

[mystic plaza]

nice work

this is what i am working on and then trying the comfyui blender rendering process

[mystic plaza]

just not sure if i need to add the textures materials and lighting when rendering this out in comfyui ?

[hollow bay]

unfortunately, i dont know but id be happy to see the final result!

[merry ermine]

https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#amd-gpus-linux-only

[hollow bay]

How can i place a hexagon mask without drawing it myself? Its always on the same position like here:

[merry ermine]

Detailer (SEGS) node from https://github.com/ltdrdata/ComfyUI-Impact-Pack

and then a Yolo model for the detection

you may have to fine tune Yolo

[hollow bay]

The input resolution and hexagons will always be on the same position. Can I just use SEGS to place the mask shape?

[hollow bay]

Like before getting an input image i want to put a mask with hexagon shape in the middle of the input image

[merry ermine]

SEGM Detector (combined) node should do it

it will just put a mask based on SEGS

the issue is this might require you to fine tune Yolo

which is apparently not too hard

but a bit annoying

you could try a Yolo-World model

Yolo-World are detectors which can take prompts

for example you can prompt Yolo-World with this person, bicycle, car, motorcycle, airplane, bus, train, truck, boat, traffic light, fire hydrant, stop sign, parking meter, bench, bird, cat, dog, horse, sheep, cow, elephant, bear, zebra, giraffe, backpack, umbrella, handbag, tie, suitcase, frisbee, skis, snowboard, sports ball, kite, baseball bat, baseball glove, skateboard, surfboard, tennis racket, bottle, wine glass, cup, fork, knife, spoon, bowl, banana, apple, sandwich, orange, broccoli, carrot, hot dog, pizza, donut, cake, chair, couch, potted plant, bed, dining table, toilet, tv, laptop, mouse, remote, keyboard, cell phone, microwave, oven, toaster, sink, refrigerator, book, clock, vase, scissors, teddy bear, hair drier, toothbrush

and this was the effect on my image

it kinda got the answers wrong but it is not too bad