🧣｜comfy-ui

[median furnace]

Anyone have tips on optimiziation of SVD for apple silicon macs? I have a maxed out M2 Max and 25 frame SVD runs extremely slowly via Comfy (regular SDXL is tolerable so was hoping for SVD to work). I know it has something to do with compatibility with pytorch tapping into metal, but was hoping someone already played around with it.

[olive yoke]

moderate your expectations because the apple hardware is quite weak in terms of potential performance

it's not due to the runtime being buggy

it is comparable to a 4050-4060

and maybe worse

[plush steppe]

Hi, I wanted to ask a question to see if someone could help me.
I connected in Windows my plugin streamdiffusion inside touchdesigner to be able to create images in real time with stable turbo in a performance I'm doing, but the status of the plugin is not stable, sometimes it disconnects by itself and sometimes more than 1 minute is maintained, but still unstable and I have to press several times the re init button, anyone has any idea what could be the reason for this?

[plush steppe]

I have a slight suspicion that it may be due to the connection with the api, but why is it not stable?

[eager escarp]

is it just me or did the order of the schedulers invert after the latest update?

[visual roost]

In the dense forest, the sunlight casts colorful light and shadow through the dense leaves. Xiao Wang and Xiao Li walked through the intricate woods, and every step was accompanied by the soft sound of fallen leaves being trampled.

[sly sundial]

...then one of them farted

[sly sundial]

It's just you 😄

[south widget]

how does tome node work? I haven't found much in the docs

[sly sundial]

There are docs?! 🤣

[south widget]

somehow 😂

[eager escarp]

Great 😄

Lifes small irritations lol

[unreal reef]

Is there a functional difference between using the Preview Image and just right-click/saving ones I like and using a Save Image node?

[sly sundial]

It's safer to save them all automatically and just delete the ones you don't want. At least you don't forget or miss something that way.

Like everything, it's just personal choice.

[olive yoke]

what's your objective?

what is the application?

[plush steppe]

I am creating a real time performance with dance and live visuals, but the status of the api seems to be not stable.

[olive yoke]

do you have a software engineering background?

[plush steppe]

I have software skills, I am a dancer and I like coding.

[olive yoke]

okay, i think this is going to be pretty hard to solve

are you following a guide?

[plush steppe]

Is it going to be difficult just because I'm a dancer? lol I followed the guide and managed to run the stream, but it is not stable that is my problem

[olive yoke]

can you share the guide?

[unreal reef]

I thought that too, but then I realized that everything is getting put into the /temp/ folder even if I don't save it. I wasn't sure what the "Save Image" node did exactly if the images all save automatically.

[plush steppe]

https://github.com/olegchomp/TDComfyUI

[olive yoke]

unfortunately this touches a lot of things that can be pretty fragile

[olive yoke]

you can try to get everything in the exact versions it was tested with

[lapis marsh]

is anyone here a comfyui dev? Id like to have a small chat about how it works under the hood

[olive yoke]

what are you trying to do?

[lapis marsh]

building a tool similar like comfyui but for text

i wanted to ask a few questions on how litegraph is interacting to the backend

[olive yoke]

litegraph helps build a json object encoding the graph that is submitted via an ordinary http REST API to the backend

the backend isn't aware of litegraph at all. it could be JSON from anywhere

[olive yoke]

what is the concrete question?

[lapis marsh]

do you guys send the raw serialized json or do you format it into a specific structure?

[halcyon thorn]

pure json

[lapis marsh]

the pure serialized json has all the nodes, inputs outputs etc

[olive yoke]

a functiont transforms the litegraph data model of a graph into a graph encoded in JSON that is very similar.

[olive yoke]

no, not all the nodes that appear in litegraph's model appear in the graph that is sent to the backend

for example, primitive nodes

"this was a mistake" in some sense

it has its benefits too

[lapis marsh]

so nodes like math nodes etc all run in the backend with corresponding backend logic?

[olive yoke]

what math nodes?

no i'm saying the opposite. not all nodes in the UI have corresponding backend nodes

what's your goal

[lapis marsh]

but does that not mean as the graph executes that the results are sent back and forth

[olive yoke]

what do you want to do?

[olive yoke]

no, there's no back and forth

[lapis marsh]

our goal is to create something like comfyui for text llms, that it acts as a kind of middleware where you can select and use loras, middlewares etc

[olive yoke]

any nodes that exist only in the UI are computed entirely in the UI before being submitted

[olive yoke]

are you saying you want to create a graph based programming tool for building llm pipelines?

[lapis marsh]

exactly

[olive yoke]

okay, and you're asking what is important about graph based UIs? like how are they architected?

[lapis marsh]

yeah, I have a few ideas in my head but they're not fully clear yet

[olive yoke]

okay let's say there's a triangle.

in one corner, there's "pure product"
in another corner, there's "pure engineering"
and in the last corner, there's "pure mania"

where do you want to be inside this triangle?

like what is most important to you? that whatever you make be as useful to as many people as possible? that it be as well written as possible? or that you personally feel the most motivated writing it at all hours of the day?

[lapis marsh]

I am unsure tbh, If its me alone I'd embrace the mania, but the backend is written in go and I am building a tool for it, so they need to modify the backend a bit (I dont know go)

[olive yoke]

this is a pre-existing backend?

[lapis marsh]

yeah

aphrodite to be presize

its basically an llm tool to run multiple llm workers at once

[olive yoke]

hmm

https://github.com/PygmalionAI/aphrodite-engine this?

[lapis marsh]

yes

[olive yoke]

okay

[lapis marsh]

https://waifu.pygmalion.chat/redoc#operation/show_samplers_v1_samplers_get

[olive yoke]

well unfortunately i can't really be helpful for mania lol

[halcyon thorn]

oh pyg is still alive?

[lapis marsh]

alive?

[halcyon thorn]

people still use pyg models?

[lapis marsh]

this is a koboldcpp replacement for mass deployment

theyre working on the model

[halcyon thorn]

ohh

[olive yoke]

this looks like it's a wrapper around vllm

which is a good but buggy project

[halcyon thorn]

please be neox based

[lapis marsh]

the idea is to be able to run concurrent requests configure sampler orders etc

[olive yoke]

actually i can't generalize

[lapis marsh]

my concern was that if we run it in the backend with concurrent request, transforming stuff etc it would be a lot of work for them, so i was thinking about a dual backend

[olive yoke]

i can't speak to the quality of an LLM backend, other than to say that (looking at the product corner here) people will want to use the best model, which means you only really need to support 1 architecture

so anything that supports many architectures is kind of a waste of limited programmer resources

[lapis marsh]

so a tought of mine was that either
A. We compose a json of executions and be handles it all and then returns a single res
B. litegraph is the execution engine where we send multiple req and build the "result" together in the FE.
C. We have a nodejs server which handles the execution we did in step B so it can run like na api as well

[olive yoke]

litegraph isn't an execution engine. it's just a UI

you don't want to program in a graph

[halcyon thorn]

if you generalize enough then different architectures shouldnt be an issue

[olive yoke]

yeah but that generalization is basically worthless

as soon as the model is outperformed by something else, everything about that model isn't going to be used by anyone

the interfaces between them are totally fungible

so it would make the most sense to focus on the one model that currently performs the best today, which there are indeed open source models that outperform proprietary stuff, but it's like cogagent/cogvlm or whatever

[lapis marsh]

the engine is not model specific

[olive yoke]

anyway i digress

[lapis marsh]

it can host any llm model

[olive yoke]

well

i'm telling you that's a problem lol

that's a curse, not a feature

comfyui doesn't support every model

you think it's a feature and it's the opposite of a feature

it's a minus sign

[lapis marsh]

okay not any, it supports mainly gptq formatted models

wait

[olive yoke]

it's something that takes 0.5 points off your physics AP exam problem, but in reality, if you get the minus sign wrong, your spaceship goes into the ground

[lapis marsh]

if lets say SD3 comes out, wont be comfy auto compatible with it?

[olive yoke]

well our dear developer had to go and write code to support it

and maybe it borrows some stuff here or there

but things must be written to support it anew

[lapis marsh]

yeha, but is it different to run SD1 then 2 or 3?

[halcyon thorn]

yep

[lapis marsh]

im not that knowledgable in that area

[olive yoke]

they have their similarities and their differences

[halcyon thorn]

comfy public has 0 sd3 code afaik

[olive yoke]

it's actually really unusual that SD1.5 is still "used"

i mean, nobody uses GPT2

[halcyon thorn]

nai leak

[olive yoke]

fooocus does (mania corner)

do you see what i am saying?

pick a corner

i can't really help you with mania

are you asking what is the right way to engineer this?

or how to make something that is actually valuable to people?

pick a point inside this triangle

[lapis marsh]

how to engineer this with actual value to people

[olive yoke]

okay

so actually this is the opposite of the mania corner lol

that's the opposite of what you said

[lapis marsh]

im still really uncertain

[olive yoke]

you said if it were up to you it would be mania, which is a very manic thing to say!

lol

[lapis marsh]

yeha, but im not the only person here

personally mania, but im not alone in that team

[olive yoke]

i mean you instantly had the insight that the open source projects are ordered by the amount of mania of their maintainers, which is a very not manic thing to know

[halcyon thorn]

just make comfy but text and not image

[olive yoke]

i didn't even have to explain to you what mania meant

you knew right away

which is very not manic

usually the really bonafide manic people aren't aware

[lapis marsh]

ah, i get that, i burned out

[olive yoke]

lol

[lapis marsh]

when i worked on a sutpid complex project

i was down for 1 year

[olive yoke]

okay so

we're getting to something juicy then. i think you are being really sincere, and you definitely want to work with your colleagues here to make something good

this is really good

[lapis marsh]

yeah, I am not as knowledgable on how llms work in a low level, but I wanna prepare a basic setup, and introduce more nodes with their guidance

my concern was do ppl use the aphrodite engine to get "step" results and compute the rest like concat etc locally, or does the aphrodite engine handle everything in a custom middleware

[halcyon thorn]

load model, load tokenizer, load lora sample etc

[lapis marsh]

yeah

[olive yoke]

in my opinion, there are a lot of things that have already been "Decided" in the LLM space:

• people will use the best models for the task. this means they will happily use GPT4, Claude, Gemini, alongside CogAgent, Deepseek-Coder, etc.
• the API is the OpenAI chat completions API with tools.
• there is only single question QA (the LLM is a library for another system) and chat UX (the LLM is the UX)

so the most productful and well engineered thing does these things the best

[olive yoke]

suppose your frontend supported "both" aphrodisiac and openai's models. you know, if openai's models work better, what would people "use"?

[lapis marsh]

Oh, the ui we build is not aphrod specific, its just related to the aphrodite engine

[olive yoke]

yes

[lapis marsh]

ohh

i get what u mean

mb

tbh i did not think about that

[olive yoke]

that's good. maybe this is giving you an idea for how you can build this "middleware" you're imagining

[lapis marsh]

if we do it in the "middleware" way on the engine itself then it would be strictly for use with our engine

[olive yoke]

your instinct for "it is a facade to aphrofuturism" but really "a facade to useful stuff"

that's all i'm really saying. and then, what is the UX it supports

it is kind of clunky to do chat UX with node graphs

and then it's like, what is someone going to be more productive in doing for a QA style programming task? writing code might be easier

[lapis marsh]

The idea was that you can use a frontend like SD or whatever, and send your "workflow.json" to the backend, it uses it as given and returns ur message

(if we do the its a frontend only and specific to aphrodite engine)

@olive yoke can I dm you a document?

[olive yoke]

i will take a look shortly

[lapis marsh]

thank you

[olive yoke]

just working on other stuff in the backgorund but

i like coordinating all the efforts together

[lapis marsh]

@olive yoke that document is what the main dev sent me, Im a really good frontend dev who basically wanted to help out with this

[hardy pebble]

hi, I am trying to use a comfy ui workflow which connects photoshop to comfy ui and generates real time preview. so I have installed everything but i get this error.

do any of you guys use it or know?

https://www.reddit.com/r/comfyui/comments/18jygtn/new_ai_news_photoshop_to_comfyui_v1_is_finally/

[olive yoke]

this is just saying you are trying to use a workflow with a controlnet checkpoint you do not have

[hardy pebble]

oh

i have all 14 checkpoints

but i am sharing the ones from auto1111,,is that a issue

[olive yoke]

the exception is saying that the filename referenced in the workflow does not exist

i don't know why it's so arcane, but that's what it is saying

[hardy pebble]

in video he uses these

in my files i have these

in comfyui it shows this

so do i have to select my tile model in this comfyui panel?

Ohhh yess

it worked damn

thanks soo much

i would have never known

ty

i hope it does not run into issue again, it is doing something. atleast its working now

it works but one question, how can i pop up the preview?
it is like this

but in video its this

comfyui does have a node for this, it is there in my workflow too, but how do it pop it up

i am very noob : (

any of these options?

double click does nothing.

[hardy cloud]

Hi! Can you suggest a node that can get a filename from Load Image? I want to use it in promt. It seems so simple, but I couldn't find it quickly

[steep marlin]

It would take like two lines of code to change the default load image node within the nodes.py file. If I get some time later on, I might whip a custom node up because it is actually pretty handy if you want to save a text file along side an image with the same name (for variables)

[sly sundial]

Why not just embed the prompt in the image and use that?

You could also save your prompts directly into a file

[steep marlin]

Long story short, I use some custom made LLM nodes for captioning within comfy and save the tags to a file along side the images

[steep marlin]

#✨｜sdxl message like this

click it to see an image i posted of it in another channel here

@hardy cloud

put it in your custom nodes folder, let me know if it bugs out ever

[hardy cloud]

I'm trying to do exactly the same thing but I'm a newbie and I couldn't make it work in batch + 🔤 CR Save Text To File for some reason won't accept string as input for file_name 😭

[steep marlin]

Did you try the node in the zip I posted? If the string format is an issue, you can probably change the output within the custom node. I'll have to double check the exact format word for it, but it will be text instead of string. There might also be some string to text node within the million nodes in common packages

[hardy cloud]

No no now it works with your node! Thank you!

[steep marlin]

Awesome, glad to hear

[hardy cloud]

It worked one or more times and then stopped too. Strange. Similar bugs started popping up in the section where I clear the file name using TextSplit (Chibi Nodes) and CR Text Replace (Comfyroll Studio). And even if the cleanup passes, at the end CR Save Text To File says that my string is not a string. It worked before with another node where I found the file name and stopped. Then I ran yours and it worked too, but then it stopped with no any changes. I don't understand. It seems to be some kind of global error with my comfy (restarted) or browser...

[wispy kelp]

the red errors are always scary and everytime i try a new workflow it has to dl more and more nodes, so confusing, this is why i am not comfortable with comfy

a plain promt/neg promt image genartion looks more confusing even if its an absolute breeze no brainer in forge or foooocus

it is very customizable tho i get that but so far i was able to do everything i wanted with forge so idk.... you stack a few contronlets and stuff and its almost the same customizability

[olive yoke]

do you have some experience programming?

[hardy cloud]

Rather no than yes. Only with chatgpt, I can open vs code and correct something according to the guide/tutorial/comments from somewhere

[hardy cloud]

See, I have two steps in my workflow to clean the filename. In the first step, I clean the name for the .txt file. In the second stage I clean it for the image caption text (the file name contains a concept, style, there are several of them in the dataset).

So when I select a file in Load Image for the first time and its name is displayed normally, I get an error (screenshot 2), but if I select another file and then the first one again, the name in the Load Image node shows name + (1), (2), (3)... and then there is no more error (screenshot 3). WTF is going on? 🙈

But the most interesting thing is that if I change CR Text Replace (Comfyroll Studio) to similar Find and Replace nodes from WAS Suite or YMC Suite, exactly the same error occurs (screenshot 4). It does not matter what exactly is entered in the Find field. Could you please check my workflow? Add any path in the last node and select any image in load image. Seems to be a problem in my Comfy or Chrome.

[lone bolt]

I think the text split node returns a list and not a string. Didn't read the whole text though

[lone bolt]

No, wait. I looked at the code and it should return a string. But in your case it returns a list

[hardy cloud]

Exactly! But why? And why does it happen with all Find and Replace nodes (Comfyroll, WAS, YMC) for me? And why does it change to string if a file is loaded the second and subsequent times it has a (1) in its name in the Load Image? (any Load Image node)

I honestly don't even know who exactly which team to report this bug to

[lone bolt]

Ok, so the untouched load image node works, and if you change the image it stops working?

[hardy cloud]

Untouched node with any previous already loaded file, when "(1)" in the name -> yes. But with new file -> error.

[lone bolt]

Ahhh, ok. @steep marlin I think your code is missing an update somewhere.

[hardy cloud]

I'm not sure I understand what it means when Load Image adds a (1) to the file name? Does that mean it's loading it from somewhere else because it already "knows" the file?

[lone bolt]

In the loader node there is a method called IS_CHANGED. I think the bug lays in this part.

Let me try something.

Ok, i got something

Well, i thought i have it.

[hardy cloud]

Oh this is even more interesting. Just before @steep marlin made his "Load Image with File Name" node, I found Chibi Nodes where there is LoadImageExtended and it also has the "filename" output I need (but there along with the full path) and I ran into this problem first on that node. Now I rechecked and the behavior is exactly the same! If the file is loaded the first time then there is an error, if it is loaded the second time with (1) in the name then there is no error and everything works. How is this possible?

[lone bolt]

Question, could you bypass the Show text nodes and try it?

Not that they somehow return lists for whatever reason.

@hardy cloud I am getting more and more sus about the Show Text node

-> OUTPUT_IS_LIST = (True,) is in the code from this node

[hardy cloud]

Wait I found something too, very strange

[lone bolt]

Other question. You just want to extract {forename} {surname} from the filename?

[hardy cloud]

1. Load Image with File Name (Ruiner) -> TextSplit -> TextSplit -> ❌ CR Text Replace
   I tried to connect the test modifiers nodes TextSplit (Chibi Nodes) directly without Show Text and it didnt give a positive result. Error.

2. Load Image with File Name (Ruiner) -> ✅ CR Text Replace
   If I connect filename string with CR Text Replace directly, everything works. At this point you can decide that the problem is in Chibi Nodes, but wait....

3. LoadImageExtended (Chibi) -> TextSplit (Chibi) -> ✅ CR Text Replace
   If I take a different node LoadImageExtended (Chibi) and string filename from there, but since it contains path, I delete it with TextSplit (Chibi) and then give CR Text Replace, everything works with the new file without (1). Then it turns out that the problem is not in TextSplit (Chibi), but maybe in some of its states.

😖

[lone bolt]

Ok, but you just need forename and surname?

[hardy cloud]

Yeah, but I have different file formats, and since I'm not an engineer, I did this complicated shit. Formats:
name-01.jpg
name-01-ig.jpg
name-name-01.jpg
name-name-01-ig.jpg
name-name-name-01.jpg
name-name-name-01-ig.jpg
+
for each of them Load Image inserts (1) into the file name for some reason if it was not loaded for the first time

[hardy cloud]

But this is only when I need it for the style caption text in the second step. But in the first step I need to remove ".jpg" ".png" or " (1).jpg" " (1).png" where 1 can be any number -- this is the format I need for the .txt file name (CR Save Text To File).

[lone bolt]

Ok, got you

[lone bolt]

Sent you a DM

[wispy kelp]

Finally. With the swarm I might be able to digest this.

[olive yoke]

what is your goal?

[wispy kelp]

consistent characters from all direction and angles. the ability to generate different consistent characters, objects, backgrounds and then composite them in photoshop and then somehow image to image them together as if they were generated the same time and then paint in motion and lipsynching. this could cover 90% of all situations for a movie. for more action heavy scenes probably some controlnet action and somehow get motion from pre existign databanks or just rnadom videos online lol

[olive yoke]

oh yeah you wanted to make movies

[wispy kelp]

thast the goal i think for many of us

at leats music videos for now

something

I love how SD for 3D came out now - clearly AI is starting to udnerstand what 3D representation of things is. Very good sign

soon well be able to somehow lock a character or object down adn reuse it in infinite scenarios

[steep marlin]

I wrote a comfyui wrapper node for https://huggingface.co/roborovski/superprompt-v1 and his tiny model is surprisingly good. Way better than the gpt2 fine-tune that foooocus uses. I'll post it a few after I make it into its own node(it's in a custom node pack of mine right now with a bunch of other nodes of mine)

I should probably stop being lazy and make git pages and submit them to comfy manager so people can find and download them

[jovial field]

Okay, I'm a bit out of the loop!
I'm used to CLIP vision only for images, but now there's BLIP, SigLIP... 😅 What's the hot new thing, Any good benchmark results or practical use-cases to help me get up to speed?

[sly sundial]

I'm interested in trying this out, but I don't know how to set up superprompt to start with. Does it need to run as a separate entity and Comfy interfaces with it?

[steep marlin]

Here's the lone node. It's a prompt expander/enhancer. Put it in your custom nodes folder. First time it runs, it has to download the huggingface model (it's super small, like 300mb). Let me know if anyone has any problems with it

[steep marlin]

lol just in time

[sly sundial]

Thanks!!!! 😄

[steep marlin]

if you see replies getting cut off, raise the token number higher

you can use pysssss show text to see the output, i might work it into the node eventually, i'm just always pressed for time

[sly sundial]

It works, but there's something a bit buggy about the text input area. I'm trying to change the text and the cursor keeps jumping to the end of text.

[steep marlin]

Weird, I've noticed a few things like that lately in comfy. Well just right click it and convert the prompt to a pin and use a string node

[sly sundial]

I've only seen it in this node. The others still work fine.

How does this one differ from the "SuperPrompt" node?

[steep marlin]

I'll double check it. Not sure how it differs other than this is based off running a tiny llm in the background that takes basic keywords like "a dog in a park" and expands it to make a more meaningful prompt. And one thing I forgot to mention is that this is for SDXL based prompts and will work really well when SD3 comes out

Check the huggingface link to find out more about it

[sly sundial]

This is strange! I convert text to input and added a text box. The text box works fine, until I connect it to your node, and then it has the problem with cursor jumping to the end of the text, even if I delete your node it still does it. If I delete and re-add the text box it works fine until it's connected to your node again.

[steep marlin]

weird, yeah i'm not seeing a problem with it on my end and i've tested it on two machines now. try closing the comfyui browser tab and the command window, then relaunch it all. maybe press ctrl+f5 to force a hard refresh of the web page

[sly sundial]

I've actually got the SuperPrompt node to work with the same model, and it's not having any issues with the text box. 🙂

[steep marlin]

what is the node named? is it RX Superprompt-V1 Prompt Expander?

[sly sundial]

Isn't that yours?

The one that's working is just called "SuperPrompt"

[steep marlin]

Yeah thats mine, I don't know what the superprompt one is though and it's not on the comfy manager

The folder name is RX-Superprompt

that's mine

the node within comfyui is "RX superprompt v1 prompt expander"

[sly sundial]

It's from KJnodes

[steep marlin]

lemme check it out, maybe it's using the same variable names or something

[sly sundial]

It doesn't download the model automatically, and has a prompt input, but otherwise it seems the same as yours

[steep marlin]

yeah i see it now

return {
"required": {
"instruction_prompt": ("STRING", {"default": 'Expand the following prompt to add more detail', "multiline": True}),
"prompt": ("STRING", {"default": '', "multiline": True, "forceInput": True}),
"max_new_tokens": ("INT", {"default": 128, "min": 1, "max": 4096, "step": 1}),
}
}

it's likely a variable name overlap issue

like if you have his node and mine in the same graph, they will bug each other out

[sly sundial]

I've not loaded both together at any time

This one isn't glitching anyway

[steep marlin]

oh wait, they still are loaded together anwyays when you start up comfy and it parses through all the installed nodes

they dont have to be in the same graph together

[deep gull]

Does anyone know if https://github.com/castorini/daam has a script for this for comfy?

[steep marlin]

if they use huggingface based libraries, they are usually pretty easy to make wrappers for

[sly sundial]

That new one has the same problem

[deep gull]

I can do this, but it's in the form of showing an Attention in clipseg. I'd like to find a DAAM that forms in a latent state, like the a1111 DAAM, but unfortunately I don't have the ability to create one myself, so I was wondering if there is one.

[steep marlin]

yeah i missed some variables in there, i deleted the message, lemme double check again

[steep marlin]

Alright, one last time lol... I really do appreciate you helping me test this out

I kept missing variables.like.this that needed to have the rx tagged to them lol

[sly sundial]

Damn, still does it 😦

[steep marlin]

well just for final sanity's sake, temporarily move his node pack out of the custom nodes folder and try it again. this really sounds like some variable overlap issue and like i said, comfy parses through all the nodes at launch

[sly sundial]

When i first load it, everything works fine with the text already in the box. As soon as I cut and paste new text then the problem starts

[steep marlin]

are you making sure to delete the node after this tests?

[sly sundial]

Yes....aha!

[steep marlin]

and put a fresh one in each time

[sly sundial]

Try pasting this into your node a cinematic photo of a character intro of a {male|female} White Mage {|with white hair}

[steep marlin]

works fine

[sly sundial]

Now move the cursor to somewhere in the text and edit it

oh...you have 😄

[steep marlin]

yeah i did

[sly sundial]

2 secs...

[steep marlin]

thats why im finding this so weird lol

[sly sundial]

I disabled the other nodes, restarted, loaded your node...same problem 🤷🏻‍♂️

[steep marlin]

open comfy up in a private tab

where theres no browser cache

[sly sundial]

I did...

It just jumps to the end before I can edit anything in there

Could be it clashes with some other node, but the other superprompt works and I'll use that. Thanks anyway 🙂

[steep marlin]

yeah probably some other node then, i dont really know what to say. i have a lot of nodes on my main install and havent ever had an issue like that before

well his node does the same as mine, so if it works, it works

[rugged spear]

https://github.com/taabata/LCM_Inpaint_Outpaint_Comfy Does any Expert here have run this nodes successfully yet?

[olive yoke]

websocket_image_save.py in custom nodes...

[rugged spear]

HELLO

[pure salmon]

HEY THERE

[rugged spear]

This place is so deserted.Any friends who enjoy studying together

[wispy kelp]

I am trying to learn supir

and in permanennt search of 100% consistency.

[eager escarp]

there is no such thing

you either copy 1 pixel and turn it into a 2x2 pixel block, or you have to guess at some of the details

[rugged spear]

Cool

[wispy kelp]

well these new text to 3D generators prove that AI is starting to udnerstand 3d shape so if it can generate somethign from one angle eventually we will be able to generate the same stuff from different angles

hence consistency

[rugged spear]

eventually

it would learn the world model

[wispy kelp]

the idea is to be able to generate props characters and sets and then combine them in photoshop and then re AI it with lighting and paint in animation

i think this will be the pipeline somehow eventually

essentially this could let us direct a movie

with total control

not just loose promts

[rugged spear]

maybe the Photoshop is unnecesary

All in ComfyUI

[wispy kelp]

if you want total control over composition its a must

its the equivalent of dressing a set or tellign actors where to stand in real filmmaking

[rugged spear]

there are some nodes can do this

[wispy kelp]

well for someone very good at comfy

lol the rest of us would probably have to rely on PS

[rugged spear]

transfer from platform is time cost

[wispy kelp]

true however we can save time that's awesome

[rugged spear]

i am making all my process into ComfyUI now...

[wispy kelp]

wow

[rugged spear]

contains Programming, Text generations and ofcause Designing stuffs

[smoky ore]

so i would like to have comfyui installed on its own virtual enviroment so it doesnt interfere with other programs. Does anyone have any good turtorials for doing this?

[eager escarp]

do you mean a python venv?

i'm using the portable install - it doesn't affect anything else on my system

it's on its own little island

[olive yoke]

what is your goal?

[smoky ore]

i did, just spent an hour configuring it and it seems to work

but good to know that it doesnt affect the system if i just run a normal install going forward

[smoky ore]

i wanted to make sure it was in an isolated enviroment as to not effect other projects i have on my computer. But if it is the case that it is self contained and completely portable then i will probably just run a normal install later on

[sly sundial]

It is its own little box of contained magic 🙂

[smoky ore]

interesting, so if i install it on another computer, and transport it to another system that didnt have it before, would it still work?

[sly sundial]

In theory, yes. If you look inside you'll see it has its own python environment.

[olive yoke]

is that something you plan to do?

[smoky ore]

yea probably going to give it a try after i finish lunch

[smoky ore]

@sly sundial @olive yoke looks like it works. installed it on one machine, and ran it on another machine that didnt have any preinstalls on it without wifi

[spring jungle]

am i allow to post a meme song i made about comfyui

https://app.suno.ai/song/038726f6-f790-44e2-a025-ecf3f801ab2f

link wrong one xD

[surreal whale]

@steep marlinyou release that node somewhere?

[daring grail]

Don't you love it when everything works flawlessly?

[sinful cape]

Anyone know what the file model_metadata.ldb file is for?

[steep marlin]

Which one?

[olive yoke]

i don't think that belongs to base comfyui

[sinful cape]

I didn't think so either. But ir appeared in my ComfyUI/models/checkpoints folder for some reason. Since I didn't put it there manually, I assume that one of the custom nodes did . . . ?

[rugged spear]

https://github.com/taabata/LCM_Inpaint_Outpaint_Comfy Does any Expert here have run this nodes successfully yet?

Any one knows how to Skip the Nodes which input is None?

[smoky ore]

is comfyui supposed to be faster than a1111?

[halcyon thorn]

faster? sometimes
more mem efficient? yes

[halcyon thorn]

its swarm

swarm adds that

[vital summit]

Is there a way to replace someones head not with a picture, but with a LoRA?
Basically generating an image and after the fact apply a LoRA affecting only the head of the subject.

[rugged phoenix]

I have 4gb vram .is there a way to install comfyui and stable diffusion and run smoothly

[vital summit]

idk about smoothly, but it happened that I was reading a Reddit post with a guy telling how he run SD with only 4GB of VRAM:
https://www.reddit.com/r/comfyui/comments/18cf2sb/comment/kce8hm0/?utm_source=reddit&utm_medium=web2x&context=3

he may be able to help you with a workflow fine tuned for 4GB

[sick badger]

thanks for the new websocket script @storm folio
https://github.com/comfyanonymous/ComfyUI/blob/master/script_examples/websockets_api_example_ws_images.py

[olive yoke]

what are you using hte script for?

[olive yoke]

what is your goal?

you can use differential diffusion to make a "smooth" mask over the face and inpaint with various denoising settings with the lora enabled

[sick badger]

easy quick webapp, just took a few minutes 🙂

[olive yoke]

would this be helpful?

• https://github.com/hiddenswitch/ComfyUI/blob/master/script_examples/remote_api_example.py
• https://github.com/hiddenswitch/ComfyUI/blob/master/script_examples/basic_api_example.py

what is the webapp for?

@sick badger oh you are already aware of this 🙂

i think you took a look a month ago

i just did a lot of updates for better API support

https://github.com/hiddenswitch/ComfyUI/blob/master/comfy_extras/nodes/nodes_open_api.py

[vital summit]

Hii! My goal is to generate images with my likeness in SD1.5 since I think it's more "creative" than SDXL. I already learned how to make good LoRAs and have both 1.5 and XL LoRAS of my face.

Anyway, getting back on track: My goal is to generate images with my likeness in SD1.5, then >try< to fix the face and upscale, then, somehow, send that 1.5 creation to an SDXL Workflow where I'll do a second pass of "face fixing" with my SDXL LoRA.

I don't know if I made my point clear but if I had to make an TD;LR it would be:
SD1.5 > Harriet_LORA_1.5 > Fix Face and Upscale > Send to SDXL > Fix Face again with Harriet_LORA_XL > gg?

[olive yoke]

okay, and it's not essential that the generations look exactly like you, but as long as they're mostly like you and consistent?

it is kind of hard to do face transfer in a general way

or as you're calling it face fixing

i know you have this fine tuning that you've made

[vital summit]

I guess??? I mean, the more it resembles me, the better. But I also wanna have it to be creative with anything else besides my face haha

[olive yoke]

okay define creative

that's very general

[vital summit]

and idk if I'll achieve this through LoRAs or just by making an generic image that doesn't resembles my face on SD1.5 and THEN inpainting my face on it.

[olive yoke]

like what is something concrete you've tried, and it "hasn't worked" with "just" a lora or "just" ip adapter

the technology just isn't there yet for non-celebrities

[vital summit]

I tried make an image of me wearing Shreks clothes

[olive yoke]

LoRA fine tunings on celebrities work because the real stability datasets have a lot of images of the people depicted already

of course with enough data good data you can make something flexible but it's very hard, it's exceptionally hard for photoreal

[vital summit]

i never tried IPAdapter 👀

[olive yoke]

good data means you have hundreds or thousands of specially staged images, with contrastive captions

[vital summit]

oh, gotcha!

[vital summit]

Yesss, my SDXL LoRA is REALLY GOOD at generating my face in closeups and medium shots. But fails miserably when it reads the words "full body", ignoring the fact that I used 30 good quality full body pictures of myself in the dataset. 😑

@olive yoke anyway, do you happen to know any good workflows to make SD1.5 faces less shitty? Or even a good after-the-fact inpainting guide? 😬

[olive yoke]

have you tried manually creating a mask and seeing for yourself if inpainting will even work?

you would have to use differential diffusion to get a little bit of a chance of a good result. it's just 1 node, it does nothing more

[vital summit]

no, I never tried inpainting in ComfyUI before. Is there a way to draw a mask in comfy? 😮

[olive yoke]

i think so yes

have you tried achieving this in sdwebui?

or using one of the photoshop or krita plugins? if you want to manually edit stuff, it's great to use those interfaces

unless you need to generate a million images it's sort of better to just use a tool that is purpose built for inpainting

[vital summit]

gotcha! but unfortunately, no.

I use my partner's machine so I'm not allowed to install more programs 🤭

that's why I'm trying to make everything inside comfy

also, not tech savy lol

[olive yoke]

okay well i guess the punchline is that what you want to do is impossible in the general case. it will generally produce crummy results

there are a lot of face transfer tools online. you should try them first, just to see what is possible

if one of those work, then you know you can maybe achieve it in comfyui, because it's all the same BS

but if you can't find anything that works, it's impossible

[vital summit]

naah, it's possible! I just need to find the right guide for the steps I'm trying to make

SD1.5 > Harriet_LORA_1.5 > Fix Face and Upscale > Send to SDXL > Fix Face again with Harriet_LORA_XL > gg
OR
SD1.5 > Harriet_LORA_1.5 > Fix Face and Upscale > Send to an inpainting node > Inpaint my face on it

I believe! 🙌

but thanks Dr. Pangloss!

[sinful cape]

Ahh thanks.

[silver nebula]

How do you set this all up? Just add to Custom_Nodes directory?

[fallow jackal]

A workflow for hexagon comparison grid! https://civitai.com/models/368486/hexagonal-grid-or-comfyui-workflow

Took me a while figure out all the math.

[tranquil parcel]

I need an easy way of saving images in the dated folder in any directory I want, what are the best nodes for doing this?

[hollow bay]

Hey guys, is it possible to make a comfyui SVD workflow, and then use it as an api for inference by passing an image?

[smoky ore]

does anyone know how to adjust the size of this?

[sly sundial]

I use this from the WAS suite...

It creates the dated folder if it doesn't exist, and then dates the files inside it too.

[fallow jackal]

A bird told me it's already supported in vanilla ComfyUI, with a filenames like this: %date:yyyy-MM-dd%/%KSampler.seed%. I didn't check it tho.

[tranquil parcel]

Thanks 🙏🏾

[olive yoke]

is there a circumstance where someone would want to open an image or workflow, and not have the option to download the models it references?

according to the steam hardware survey, which is representative of a random sample of PC users with GPUs, 85% have more than 100GB of free storage space

for me it has exclusively been a meaningless hassle. i never got more joy out of visiting the websites to download the models, as opposed to having it download automatically

[sick badger]

it's just a python script you can run on a web server or command-line that connects to your comfy instance where you can put your workflow

[olive yoke]

"just"

[plain condor]

Hello 👋🏻
I'm new to Comfy (and SD, in general). Slowly working my way through online tuts.
My first objective is to take an existing photo of someone and change their facial expression / mood. Preferably without changing perspective or composition.
I'd appreciate any tips, guides or scripts you may have to offer. Thanks.

[fallow jackal]

Look for Impact Pack custom nodes, it basically designed around that task.

Well, technically, Impact Pack lets you inpaint a new face. Preserving it to look like an old face is another task, and one of the tool for that is Photomaker (included in vanilla ComfyUI).

[jovial field]

I use crl+ or - mouse

[plain condor]

Thanks, for the guidance @fallow jackal

[smoky ore]

anyone know where i can ge a decent upscaler model?

[untold pine]

Yeah zoom in on the webs browser window.

[untold pine]

SUPIR is my favorite right now.

[silver nebula]

Ctrl/Cmd +-

[surreal whale]

Certainly works, i use %date:yyyy-MM-dd%/%CheckpointLoaderSimple.ckpt_name%

[jolly warren]

i install stable swarm and get error how to fix i replace comfyui old to dlbackend

[olive yoke]

what is your goal? ordinary text to image? can you give an example prompt of something you want to make?

or something else?

[olive yoke]

it's telling you what the error is, but you're going to run into a lot of roadblocks like these

i don't know what the UI looks like in stableswarm-ui, but you are supposed to (probably) select a main.py file

[silver nebula]

What is NDI please?

[sly sundial]

Network Device Interface. I presume you're referring to this node?
https://github.com/ningxiaoxiao/comfyui-NDI

[granite shore]

I guys I've got a simple question: how I can decide the speed of a video generated with AnimdDiff stuff? Because i've generated 300 frames at 60fps and the result is super speedy. Is there a way to make a slow motion like video?

this is my workflow atm

[sly sundial]

Less fps, or create more frames

Half the fps will be half the speed

...or twice the frames

[granite shore]

if i lowered the fps the video is scattering and isn't fluid. And if i've generate more frames the video length change but not the speed of transition

but i'll give another try generating more frame, in any case many thanks for the answare!

[sly sundial]

Are you using interpolation to pad it?

[granite shore]

Sorry man i'm very unexperienced with comfy so probably mine are noob questions. The workflow that i'm using is attached above and i don't know where is and (if there is) the interpolation stage you're talking about

[sly sundial]

Do you have this node?

That can multiply the number of frames you have and fill in the gaps

[jovial field]

for comfy:
Hi all, I'm curious what CUDA version others are using with their RTX 3080 graphics cards. Are you on V11.8, V12.2, V12.1, or something else? I'm mainly interested in the best choice for overall performance and compatibility with popular extensions and libraries like those used in ComfyUI. Thanks for any insights!

[granite shore]

yes i got it, but don't know where to put it

mabye after FILM FVI?

[sly sundial]

Between the frames output and the video combiner.

[granite shore]

all right i'll give it a try

Many many thanks man!

really appreciated

[sly sundial]

Wait, I think Film VFI may be the same thing

Not familiar with it, but try swapping the two and see if it makes much difference. You could increase the multiplier a bit more to get more frames.

[granite shore]

ok!

[silver nebula]

RTX 2070 torch 2.2.1==CU121

[silver summit]

Hi everyone, i am trying to use the best resolution for controlnet, for my image2image. in A1111, the resolution is in multiples of 8, while in comfyui, it is in multiples of 64.

is there a node for me to use controlnet in multiples of 8? how does controlnet actually work?

I will prefer to use multiples of 8 as i can get a depthmap to match my original img. the shape of my original image is very important and i want to try not to have it off by even 1 pixel.

[red dagger]

https://twitter.com/beckylitv/status/1772407948481892396

This video is been making the rounds on socials. Looks pretty well done imo.

Wonder if something this good can be achieved manually using comfyUI. I've used reActor for face swap with good results, but what workflow would you recommend to achieve something as good as this (face swap, lip sync, etc)?

Beck 🎈 (@beckylitv) on X

It’s terrible but I still think it’s wild that this can all be done with AI. Imagine in 6 months…

Twitter · Mar 25, 2024, 11:39 PM

[vital summit]

Hii!
Is there a node that receives an image and, based on the size of that image, it outputs the closest width an height recommended for SD 1.5?

[olive yoke]

this is very dubious

they have example ads from more than 1 brand that i've worked with, and i cannot find the referenced ad on the claimed platform anywhere

i did find something that resembles what they are demonstrating, but it wouldn't make sense to use them for that purpose

[olive yoke]

@storm folio have you looked at https://discord.com/channels/1106248220687994911/1151019966120873995 ? this is ideogram's bug-report channel

[hollow frost]

Hey guys, I've been trying to get comfyui working and I'm attempting to follow this https://raw.githubusercontent.com/nerdyrodent/AVeryComfyNerd/main/workflows/SD15/BatchImageAnimate.png workflow.
I keep getting errors over and over, mostly referring to 'NoneType' object has no attribute 'lower'.
I seem to be struggling with the IPAdaptor models, specifically the clip vision. I've found some info online but Ive been struggling for hours trying to solve this and I Cant. Any one ouut there that oculd possibly assist me? ❤️

[red dagger]

Yeah someone shared this on another discord https://twitter.com/BenjaminDEKR/status/1772906015610024022
Also saw her IG on another post, its arielmarie.ugc

So she either sold her likeness to that company by recording a few videos which you can just lip sync on top by using their service, or she's doing 'em one by one personally (no AI involved) which would be quite the grift 😆

Benjamin De Kraker 🏴‍☠️ (@BenjaminDEKR) on X

So the "AI generated girl" in the car video?

Yeah, she appears to be a 100% real person named Arial who does gig videos on Fiverr

Some of the biggest AI influencers on X have been telling everyone this was AI generated, and two different AI video companies including…

Twitter · Mar 27, 2024, 8:38 AM

[olive yoke]

it's definitely using an image technology to produec the videos because it happens fast

[olive yoke]

the NoneType object has no attribute lower bug means you are missing the model referenced by the node that is throwing the exception. it means the file isn't there, or it's in the wrong place, or has a different name than the workflow you uploaded. @storm folio this is a persistent source of grief from end users. it took 1 line to fix in my fork. if you looked at ideogram's bug reports, you'll see that they're real and interesting, compared to the issues you receive, which is people hung up on absolute pointless BS that is easy to fix

[eager escarp]

Only the voice is changed

That's real video

[red dagger]

Thanks!

So, circling back to my question. Is there a good way to do this locally? Like, downloading a lip sync model and wiring it up on ComfyUI or something. Any idea is appreciated.

[eager escarp]

Swap your own voice and do it with yourself

Or swap the voice of a video that's about it

Not really comfy related but yeah

There's plenty of free voice cloning software that can be run locally with a good GPU

[eager escarp]

Soviet Union domain tells you all you need to know about that

[fallow jackal]

You gave me an idea to experiment with my old Photomaker workflow, and turns out it can be used to change face expressions! (workflow in metadata)

[hollow frost]

@uncut shoal I sent you a pm but I understand that might be BM. I saw with the search results that you're pretty good at figuring this out and I've been stuck for almost 2 days. https://www.youtube.com/watch?v=bH-56e3cR2g <- I'm trying to do this tutorial ( or at least get this effect) and I'm stuck on 0% after waiting over 30 mins, any idea what could be the issue?

[stiff forge]

For folks calling comfy from python API, do you generally prefer using Anaconda / conda or Python / pip?

[sick badger]

python/pip

[olive yoke]

lemme introduce you to pip install git+https://github.com/hiddenswitch/ComfyUI.git

🤏 you have to know this much about creating a venv

[olive yoke]

from comfy.client.embedded_comfy_client import EmbeddedComfyClient
async with EmbeddedComfyClient() as client:
  outputs = await client.queue_prompt(prompt)
  save_image_node_id = next(key for key in prompt if prompt[key].class_type == "SaveImage")

  print(outputs[save_image_node_id]["images"][0]["abs_path"])

^ if you want to "just" start comfy locally and run workflows, not as a server, but like, as a library, like any other piece of code

from comfy.client.aio_client import AsyncRemoteComfyClient
client = AsyncRemoteComfyClient(server_address="http://localhost:8188")
# Now let's get the bytes of the PNG image saved by the SaveImage node:
png_image_bytes = await client.queue_prompt(prompt)
# You can save these bytes wherever you need!
with open("image.png", "rb") as f:
    f.write(png_image_bytes)

^if you want to run it remotely

@stiff forge is this helpful?

[sinful skiff]

cat

[radiant wedge]

is there some way to save text to a file from withint comfyui and have the file be like ComfyUi_32059.txt for my ComfyUi_32059.png ?

[stiff forge]

Amazing, this is exactly what I was looking for, thanks!

https://github.com/hiddenswitch/ComfyUI/blob/master/script_examples/basic_api_example.py

[olive yoke]

you will have to use the fork, and you will have to do some light fixes for custom nodes, at least for the next few weeks until i have the time to test all of those

[stiff forge]

Wait, are you the author of the fork? If so, thank you so much for building this!

[olive yoke]

what is your use case?

[stiff forge]

Experimenting with automatically tweaking prompts with an LLM. Not too specific to comfy, but that's the UI I'm already familiar with.

[olive yoke]

are you using something like ollama?

like from the point of view of interacting with comfyui as a JSON schemas defined tool like OpenAI does

i am not super up to date on the open models

i am already doing this so it may be useful to exchange notes

[stiff forge]

That'd be great. I'm planning on starting with a better model like claude3 then scaling down one POC is working.

Going to start with Instructor and see how far I get with that

[olive yoke]

what exactly do you want to do? "just" rewrite prompts? invoke the prompts through a chat interface? or write new workflows in text?

[stiff forge]

TBH I don't have a super specific goal, just want to try out some ideas and see where that takes me

First experiment is generating variations of a prompt in a specific way.

[olive yoke]

okay it might be worth it to try to write it as a custom node first, using something like Phi2, because deterministic, small NLP tasks are well suited for the UX

if you are ambitious, use CogAgent / CogVLM

and don't do it using transformers / diffusers

or, write a wrapper for diffusers models to be compatible with comfyui's model loading and unloading, which would not be challenging

since people are constantly using them anyway

you can see in the chat here where i ask the comfyui author why he doesn't like huggingface's libraries and decide for yourself if it is a reason to avoid using them

[stiff forge]

I wasn't familiar with these, thanks for the pointers!

[stiff forge]

Why do you say this?

[olive yoke]

well you should

if you wanted to PR it it won't get merged

by upstream

are you on windows?

[stiff forge]

I use mac/linux for work, windows/wsl outside of work

[steep marlin]

Writing nodes that wrap HF models super quick and easy

[olive yoke]

@real valve re: comfyui repo work

[eager escarp]

There's a lot of nodes for adjusting images, blending, sharpening, color correcting etc.... but the most important node of all - curves - doesnt seem to be a thing. Have I just overlooked a version of this somewhere?

[olive yoke]

i was thinking of authoring a curve widget as part of changing values as sampling progresses

for loras, controlnets, etc.

to reduce the pain of having multiple samplers

but eventually it might make more sense to move the nodes into ane nvironment that laready has that stuff

rather than reinventing it all

for example blender's node editor

[eager escarp]

that would be a game changer imo

[olive yoke]

so should i just do that

it is a "low lift"

[eager escarp]

ive reallllly wished we had that so many times

[olive yoke]

rather than authoring more and more non-ML related nodes

curve or blender?

do you want all of The Foundry Nuke

[eager escarp]

the curve in comfy

[olive yoke]

architecturally what was missing was a contextlib Context that is available to all nodes

[eager escarp]

tbh, i havent used nuke

[olive yoke]

which would decouple a lot of this stuff in a good way

[eager escarp]

but ive done a lot of traditional image editing

gotcha

[olive yoke]

i'm not sure many people have discovered the superpower of changing parameters as a function of time (sample_t)

[eager escarp]

but yeah a few places where more manual control would be nice....

agreed

its extremely powerful and little explored

[olive yoke]

i wonder if this is already how dall-e3 works, and that they are cascading in the wrong direction

[eager escarp]

Imo mostly because implementing in most cases requires chainitg nodes

Yeah

Def

[olive yoke]

"DALL-E 3 has many improvements over DALL-E 2, many of which are not covered in this document and could not be ablated for time and compute reasons. "

that stability is cascading in the wrong direction*

[eager escarp]

Main limit for me with that stuff is too many stages= to many nodes= interface slows to a crawl

[olive yoke]

it should be 4 turbo models, each running 1 step, for that particular timestep

instead of thinking that the symmetry of the process is an advantage

it's a disadvantage

[eager escarp]

Yeah

Time stepping anything and everything has been interesting from my experiments

Even weird shit like skipping a denoising step at variable points

And now with the new ipadapter nodes... So many ideas

Being able to just draw curves would be so helpful and might get more ppl to experiment

[olive yoke]

hmm

well it's an ephiphany to me that Dall-E3 is probably "cascading through time"

instead of like stable cascade or IF/Imagen which is "cascading through space"

like i'm sure it does both, but the most important part is that it must be a latent space model "most of the time," and that there's a specifically trained model for each denoising step, instead of one model for all denoising steps

and i can't think of anyone who has published this

so maybe it's just crank nonsense

but considering how much work SDXL Turbo does in one step, and the idea behind that kind of distillation, i am pretty confident they distilled N distinct models each for t_i to t_i+1 for i <= N, and each model does the same amount of work as SDXL turbo

for some number of latent steps, then 1 pixel step like IF/Imagen

[eager escarp]

I think that makes a ton of sense

[olive yoke]

it also lines up with my experience for how much better control i get from fine choices throughout the duration of denoising

[eager escarp]

Absolutely

Best results for me have been when I do exactly that

[olive yoke]

this might also be why they don't have an image fine tuning product

and why midjourney doesn't either

[eager escarp]

And tweaking the scheduled
Scheduler alone at various points has been huge

[olive yoke]

this has also been the case for me

i think the biggest surprise is to always start with ddim uniform, because that is what is used for training

meaning to always start trying to build the pipeline around ddim uniform

especially if you use community assets

[eager escarp]

And we've all seen it
With previews... Something you like changes about halfway through

[olive yoke]

stable cascade is i guess a similar idea

[eager escarp]

So then you gotta play with the scheduler manually

[olive yoke]

it has two distinct denoising stages

[eager escarp]

Or chain kadv

[unreal reef]

I've just begun playing with the SDXL prompt styler and I really like the way it turns out various art styles.

I have a good workflow I already like that is producing images. I want to take the output of that workflow and feed it through the stylerto turn those images into the desired aesthetic. I'm not sure how to wire them together though. At the moment I clearly have something hooked up wrong. Any advice?

Is there a method of using the styler on an image input rather than a text prompt?

[olive yoke]

what is the "prompt styler"

you can put an image into a ksampler node using VAEEncode, aka img2img

[unreal reef]

https://github.com/twri/sdxl_prompt_styler

[unreal reef]

Thanks - I'm still learning Comfy and that's helpful!

[sudden gull]

x_x

Have no idea how to properly start up ComfyUI

[sly wolf]

does someone know how to upscale in comfy without having to do tiled upscale

[steep marlin]

Find a model that doesn't implode when going beyond its training sizes. Other than that, you need to go the normal tiled route. If you're going from a 1.5 image at 512x512 to 1024, you can always use a 2x upsampler and then denoise it between 20-30% with an SDXL model. But if you want to do something like double an SDXL image, you're going to have to tile it

[eager escarp]

Tiling generally gives better results anyway

[steep marlin]

Usually, especially if you use a tiling control net along with it

[lone cape]

If you use SD ultimate upscaler, and the tile_width and tile_heighth is the final upscaled size, + set model_tyle and seam_fix_mode to 'none' you will get upscaler what not modify your generated image just upscale. I using SwinIR model and working well.

[prisma fern]

i had a problem when just installing controlnet on comfyui with version "ComfyUI_windows_portable_nvidia_or_cpu_nightly_pytorch".

when i try install controlnet via Comfyui Manager Menu, having this kind of appearance, has a lot of conflict.

I repeatedly tried to install it, but it kept getting errors and failed.

next I tried with git clone.
still error.
then I tried using the manual method, by downloading and extracting it.
but it still has problems when running. can't read it

how should i fix it?

[robust dove]

These conflict messages are just cautionary, meaning that as long you don't mix the conflicted nodes on the same workflow you should be fine.

Here's a screenshot of my setup with controlnet installed. The warning is still there, even tho it works just fine

But that giant red and yellow warning scared me as well, and I had to look up online too.
Anyway, the Manager addon could do a better job of explaining that.

[prisma fern]

thanks for reply. you are quite helpful

and before installing custom nodes, I only need to update comfyui and pyhton dependencies. and wait for it to finish.

[hushed pike]

on automatic1111 when you put a png with the same name as a checkpoint model in the folder with the checkpoints, that image will show up as being the preview image, is there a preview feature for stuff in comfyui, and does it work the same

[stable bluff]

Does anyone know how to add an object's workflow

With the technology of fooocus, it is implemented in comfyui

[grand pike]

hello lovelies, is there a prompt or lora than can help randomize a completely new person with every generation? No matter what terminology i use, it will consistently put out the same person.

[olive yoke]

did you change your seed between clicks of Queue Prompt?

[grand pike]

I'm using comfy ui and it's set to randomize the seed every time.

Usually people have the opposite problem right? They want consistency lol

[toxic breach]

hey all, Is paid google colab good to run comfyui or do you recommend something else?

[frank path]

Guys

I get the error cuda operation not supported

What should i do?

[olive yoke]

you would have to share your workflow

it is more about understanding what you are asking the interface to do

you are still programming, in disguise, and it can be really tough if you don't have experience programming

[grand pike]

@olive yoke Just made it back home, here's the workflow. Extremely basic with no custom nodes, Like it just fuses all races and skin colors into one person, lol

yeah i feel like i'm just not using the right words to get the program to understand what I'm trying to do. I've tried other tokens as well, like different person, alternate person, etc.

it does make a male version of the lady some times, but that's as random as it gets.

[eager escarp]

"random" is not a real prompt

imo

use dpmpp ancestral, cut your cfg back to 6, steps up to 35

[grand pike]

will try that now and get back to you, thanks!

[eager escarp]

that won't necessarily solve the issue of people looking similar, but your quality will be a bit better and the odds will be higher

try using a single descriptor word for a person, and changing that when you really want someone new

that could be clothing, hair, body, facial feature, whatever

[grand pike]

that seemed to do the trick, 35 steps is pretty rough on my 1660 ti though

i'm basically trying to mass produce generic stock photos with varied people, but i don't want to have to change the prompt every time

hhmm so it changed from that other lady but now it's stuck on a different one

[olive yoke]

okay well

you are still learning how to do this stuff

you should start by reading prompts from CivitAI, for portraits that you like, which will help you understand how other people have authored prompts that make unique individuals

[grand pike]

@olive yoke yeah i've tried replicating prompt techniques from civitai, and they help to get the specific look of a lora or something. But I just want a different person with each new generation, without having to edit the prompt.

[olive yoke]

just
😎

sounds like you want to pay for a service that does this. many exist

[grand pike]

im even more confused now lol

[distant pendant]

try using wildcards in prompt, like just get a long list of random names and another one of surnames, and it will pick a different combination each gen, (probably) resulting in more varied people

wildcards for hair colors, clothes and such would obviously work too

(there should be a bunch of nodes for that in the manager)

[grand pike]

@distant pendant I did find this one here, but i'll have to learn how to use wildcards tho

https://civitai.com/models/232027/easy-random-wildcards-ethnicities-skin-tone-female-clothing-locations-poses-actions-hairstyles-and-more

is comfyui-stable-wildcards the best custom node for wildcards?

[distant pendant]

I haven't the foggiest which one is the best

[grand pike]

thats okay i'll just try them all.

thanks, i think this is the feature i've been looking for

[grand pike]

this isn't working ya'll, no matter what i do, no matter what wildcard i use, it refuses to make each person different. I'm going insane trying to figure this out. I never imagined something seemingly simple as a random person would be so difficult

[plain condor]

I'm running ComfyUI on a M3 MAX MBP with 128GB Memory running macos Sonoma 14.4. I seem to be running into one roadblock after another with it.

Roadblock #1 - latest versions of PyTorch and Sonoma 14.4 break GPU acceleration, so I need to run PyTorch 2.1.2 in a virtual environment running Python 3.11 in order to utilize my GPU. Otherwise, I'm stuck running Comfy in cpu-only mode.
Roadblock #2 - AnimateDiff Evolved crashes Python and is unusable. And the original Animate Diff Sampler only works for <10 frames. Otherwise, it just outputs black frames.
Roadblock #3 - Frame Interpolation (GMFSS Fortuna VFI) requires cupy... which no longer supports macos.

Are there any other mac users on here who have found solutions to any of these issues? It's feeling like for every roadblock I circumvent, a few more are just waiting for me around the bend. 😓

[frigid matrix]

anyone know how to insert a civitai prompt directly into comfyui like you can with stable diffusion

[flat talon]

I think you can try the Portrait Master, which has many settings for facial details and random modes for each part (sorry, English is not my native language).

like this.

[grand pike]

@flat talon oh wow. that looks amazing. Already downloading lol

is portrait maker related to it? saw that on the list as well

[flat talon]

Sorry, I don't know, I rarely pay attention to the author of the node

[grand pike]

That's okay, I appreciate your help ❤️

[flat talon]

😆

[grand pike]

do i gotta connect it to anything? positive and negative nodes are greyed out.

[flat talon]

Right click on ClipTextEncode

[grand pike]

I found a video tutorial homie, thanks. Actually the video does it differently, not even sure if it's method will work or not

[flat talon]

Ok, bro, trying more will help you better understand comfyui.

[grand pike]

ugh you're a genius

its so random now

like this, right?

[flat talon]

yes, that's right!

[forest birch]

Hoi, do you guys know of a workflow/models that can achieve really good video interpolation? Where i can choose how accurate it needs to be to achieve the least amount of artifacts?

[olive yoke]

what is your goal? what kind of videos are you trying to make?

[hazy cliff]

https://www.youtube.com/watch?v=b6TbdBJBI4Q

[forest birch]

Any really.

Content depends on what funny/interesting video i want to interpolate. Could be a small clip, or an entire movie. live action, pixar style or anime/cartoons

[olive yoke]

okay that's pretty inspecific

you are going to have a bad time

it sounds like you want to use a ready-made plugin in an adobe suite program

[forest birch]

Because it will be used for any kind of video. Like flowframes, but more "in depth configs" and accuracy/targeted content, If there's different models per kind of media, do list them, as just with the different upscalers, some are best for anime, some for real, other for art, some for text etc

[olive yoke]

is this for you or for someone else? like a service?

what is the idea?

[forest birch]

For memes/jokes, as well as "seriousness", but personal use to interpolate movies that i'd want to watch with 2x and higher framerate for smoother motion

Avatar way of water being one for instance. As i hated bluray release never got a 48 fps release

[olive yoke]

okay, have you tried using a non-neural frame interpolator, like dlss3? or RIFE?

or FILM?

you can download programs that do this for free, in one click

adding a "little bit of motion" to "memes" is only really achievable by SVD style stuff, but that is also available "one click"

[forest birch]

Can't use DLSS3 due to 30 series card. And yes, flowframes, as well as SVP. SVP is too artifactry, and flowframes is way better, but nto as accurate. So i'd want for instance a workflow for comfy that would take a video apart frame by frame, analyze and interpolate each extracted PNG, then throw it all back into a mkv/mp4

[olive yoke]

it sounds like you are asking, is there a thing you can download off civitai, which is a workflow, and also works with all this comfyui manager stuff you are familiar with, which gives you "one click" solution to your inspecific task

that does everything for you

[forest birch]

Literally what i asked lol

[olive yoke]

well did you search civitai?

comfyui is the opposite of what you are looking for. you are looking for a ready made application

a ready made application disguised as a bunch of comfyui artifacts, that's fine, but it is going to be inferior to anything you find on google

[olive yoke]

anyway i think there are a lot of good results for "comfyui animation interpolate frames"

[small wagon]

Made this workflow to generate the soft opposite of an image

[weak urchin]

I have an Interpolator at the end of this workflow, but I have one just for upscaling and Interpolating as well. https://civitai.com/models/367412/geeky-ghost-vid2vid-organized-v1

Let's you manipulate video in a variety of ways with 2 images.

I'm going to do another update to it soon now that I got loras working well with animatediff to change the style.

[clever thunder]

hello everyone. is there a node that can talk to an API? i would like to send an image to Replicate and getting it back all in comfy. any tips would be helpful. thx

[olive yoke]

can you be more specific

[clever thunder]

like this ..i would like to use replicate for some stuff, i use some of those in my own workspace in Replicate. but would like to integrate them into comfy https://replicate.com/tencentarc/gfpgan?input=python

so for example send a picture i generate locally to an upscaler or something like that, but using API

[silver nebula]

I updated my InstantID, and produced this "Splendour" SDXL Turbo/Pearl Mountain Collage Maker

[grand pike]

Hello lovely people. If I wanted to insert my own custom background into a character image, how would I do that? Like using a picture of a real location and the AI will sort out all the messy pixel edges and whatnot.

[eager escarp]

you could try outpainting

you could try ipadapter plus

[lapis garnet]

Hi guys is there any workflow available for upscaling images similar to Leonardo AI within Comfy UI

[unreal reef]

Anyone ever run into an issue with not being able to type into CLIP Text Encode? My cursor keeps jumping to the end of the box.

[sly sundial]

Yes! For me, it was caused by a custom SuperPrompter node.

[unreal reef]

Yeah, it only seems to do it in this one workflow - I'll have to see what else I've got loaded in here and rule stuff out.

[sly sundial]

I removed it and installed the one from the Manager...then it worked perfectly.

[unreal reef]

I'm still trying to figure out how to use the SDXL Prompt Styler. I wish I really understood what the hell I was doing rather than just copying examples and pushing buttons at random lol

[sly sundial]

This is when I hit the problem #🧣｜comfy-ui message

[sly sundial]

I use that all the time. What's troubling you?

[unreal reef]

Well, I have a workflow that I really dig - it's basically a flow of pony 2.5 going into a more realistic model so that I get the filth of pony with the photorealism of something better, but I was trying to then hook the output of that in as the source for a styler... i.e. take a completed image and apply a style to it. As far as I can tell, the styler is really designed to sit at the beginning of the process, not the end... it styles the prompt, but can't be used directly on img2img... but I might be doing it wrong.

[sly sundial]

It is used just to add styling text to the prompt, so needs to be at the creation of the initial image.

The styles are just pre-written blocks of prompt additions.

[unreal reef]

Yeah, I think I may need to figure out how to hook it into that second stage of the process where I convert the pony output to photoreal.

[sly sundial]

Possibly something like this, but just change the prompt input for the next model sampler https://civitai.com/images/6746526

That's Cascade with SDXL refiner, but the principal is the same.

https://civitai.com/models/119257?modelVersionId=347573

[unreal reef]

Nice - I shall give that a spin!

[crisp ember]

is there a way to load checkpoints from a string?

[lusty pewter]

Grab this set of nodes:
https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes

Then use CR String To Combo:

[maiden laurel]

are there any nodes that remove letterboxing in outputs?

[smoky ore]

Is anyone familiar with this error? finally got UnifiedLoaderFaceID to work yesterday but now im getting this

[lusty pewter]

Update ComfyUI

[crisp ember]

Thanks for the tip about CR String To Combo, that was exactly what I was looking for!

But now I have a different problem. I'm filling up my queue A LOT. Like, I'm queuing up about 1k images in my workflow. The thing is, the queue gets really slow after the first 20 or so items are added... it starts to execute the flow before all items are queued up, which results in a traffic jam of sorts. When the queue finally fills up and new items stop being added, it becomes fast again...

As for my hardware stats, everything is basically idle.

• GPU is practically idle with plenty of VRAM to spare (as we speak, it's at 7Gb used out of 24Gb available and sitting at 9% utilization - I'm using --highvram already).
• CPU is at 9% usage with roughly 2 cores (4 threads) spiking, but not much.
• Ram is at 65% utilization, stable
• No disk activity

And it's just choking and choking to fill up the queue

[serene trench]

古代皇帝，书房，写字

[surreal glade]

hello does anyone know if it's possible to show image on google colab and if it's possible how should i done it?
PS: here what i've tried

[smoky ore]

it was updated. I found that the problem was i have the IPAdapter Model Loader connected to the unified loader. which for some reason was causing a conflict. once i disconnected it, it started working

[olive yoke]

is it expected that clip set layer does absolutely nothing?

is it just broken?

[lusty pewter]

If you mean CLIP Set Last Layer, that's also known as "CLIP Skip" in A1111, and it works just fine. Take a model that requires a value that's not 1 and watch your output not work.

[olive yoke]

i see

does it have no affect on models that have no clip skip specification?

[lusty pewter]

It has an effect, but that impact may vary depending upon the model.

[smoky bear]

hey everyone, can someone maybe help me with paths? WARNING: The script transformers-cli.exe is installed in 'C:\ComfyUI\ComfyUI_windows_portable\python_embeded\Scripts' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.

i get this

WARNING: The scripts convert-caffe2-to-onnx.exe, convert-onnx-to-caffe2.exe and torchrun.exe are installed in 'C:\ComfyUI\ComfyUI_windows_portable\python_embeded\Scripts' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.

and this

looked it up and stumbled upon environment variables

but i dont want to mess up even more

[fierce hornet]

Those shouldn’t be an issue

Everything still works for you right? @smoky bear

It just means if you open a cmd prompt & try to run those commands from someplace other than being in the same directory the system won’t recognize them

[placid axle]

ugly, deformed, noisy, blurry, distorted, out of focus, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, nudity, nude

[fierce hornet]

Who are you calling “ugly, deformed, noisy, blurry, distorted, out of focus, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, nudity, nude”?!

[wanton hull]

anyone have a good google colab pro repo for running comfyui?

my old one is no longer working

[silver nebula]

Try this ...

[olive yoke]

you can ignore this

[viral torrent]

Heya, has anyone here figured out how to add IP Adapter to Searge's latest workflow? I think I'm feeling my brain simmering as I stare at this spaghetti

[weak urchin]

got a link to it?

[viral torrent]

Yep! Here's Searge-SDXL https://civitai.com/models/111463/searge-sdxl-evolved-v432-optimized-workflow-for-comfyui-2023-11-13-txt2img-img2img-inpaint-revision-controlnet-loras-freeu-v1-and-v2

And here's IP Adapter Plus
https://github.com/cubiq/ComfyUI_IPAdapter_plus

[weak urchin]

I'll take a look at Searges. You just need the IP adapters integrated, and what were you looking to do with them? I'm familiar with them and use them for crazy things lol. https://civitai.com/models/385959/geekyghost-3-subject-cascade-and-svd-everything-but-the-kitchen-sink-workflow

[weak urchin]

And I see the issue lol. It would be extremely difficult to incorporate it into his workflow lol. There's a lot of merged nodes and custom nodes, the settings and flow seem delicate and already using a ton of resources. I would not be able to add it lol, well, I coulld probably, but the level of effort is beyond what I'm willing to expend.

[viral torrent]

I’m glad it’s not just me 🤣 I wasn’t looking to do anything crazy, I’ve actually yet to try IP Adapter in any form, so I was just looking to get its basic functionality going. Sounds like if I want to do that though, it’s going to involve a new workflow - which I can’t say I’m thrilled about, but such is life. Do you have any workflow you’d recommend that wouldn’t be an unholy nightmare to stick IP adapter into?

[weak urchin]

Ah, once you dive into workflow creation though you'll realize it's much easier than it looks. I had to turn off the spaghetti before I could even look at a workflow lol. But now I make them for fun to do random weird things lol. And I get weird with their layouts lol.

Supposed to be homer and a TV lol

[viral torrent]

Probably is easier than it looks, I just unfortunately don't have the free time to devote to getting good at it at the moment. That's where my attachment to Searge comes from, I learned that workflow when I had a lot more time on my hands, alas

[weak urchin]

Fair

If it helps, the IP adapter node comes with example workflows to use to check it out.

Works great for image to image, and I use it in video2video with animateDiff with good sucess.

[viral torrent]

Glad to hear it lives up to the hype, I'm looking forward to taking a look at it at some point

[weak urchin]

I've been using it to turn my Barbie Captain Marvel Image Gen into different styles lol

[viral torrent]

How is it for things that aren't people? Landscapes, architecture, that sort of thing?

[weak urchin]

For non people, I'd recommend using controlnets alongside it. It's great for subjects, but when you combine subjects with background images, things are going to bleed over.

Controlnets help force shapes, IP adapter helps force Style.

[viral torrent]

Hmm - last I heard, SDXL controlnets were, uh, not great. Have they improved much in the last several months?

[weak urchin]

Ah, that I don't know lol. I use 1.5 mostly. I have a custom Geeky Ghost model I started training when SD first dropped. I refuse to give it up, I test with LCM's for speed, and use mine when I need quality lol.

Plus all my good loras are for 1.5 and I can train my own controlnets, motion models, loras, embeddings, from scratch models, and etc. with my 3D renders and etc.

My render library ain't no joke lol. Here's 3D models I made of me and my wife lol.

[steep marlin]

They work fine, people just use them incorrectly coupled with other vestigial baggage like having a novel's worth of negative tags and incorrect prompting

But hey, if the lab rat food dispenser ejected food when the rat spun, it is always going to think it has to spin to make the food come out, right?

Even if it was just set on a random time interval lol

[woeful grove]

Guys, I'm excited to share that I just integrated Groq and OpenAI APIs inside comfyui, would appreciate it if you guys could check it out. https://github.com/ronniebasak/ComfyUI-Tara-LLM-Integration/blob/main/README.md

https://www.reddit.com/r/StableDiffusion/s/HbOA5xdG8J

[weak urchin]

Adding a bit more info about the nodes for this workflow since I combined some and use some lesser used nodes. https://civitai.com/models/388864/geekyghost-aio-workflow

[eager escarp]

some are indeed not nearly at the same level as SD15 versions

qrcodemonsterXL is a great example of that

[silver nebula]

I'm currently successfully using IPAdapter (original version). Is it yet possible to run IPAdaptr v2 alongside - and separate from - the original version without breaking it?

[lusty pewter]

You can have both installed and even in the same workflow.

[eager escarp]

anyone know if there's any nodes that can be used like chained ksampleradvanced, except which actually allow transferring momentum?

i'd like to be able to switch models (ipadapter pipelines, specifically, to save vram, as i'm hitting the 23.3gb limit where my system basically freezes up) at various points during sampling but that doesn't work too well with ksampleradvanced as momentum is lost

that, or something that allows timestepping models

[lusty pewter]

Something I've been looking for was under my nose this whole time...I've wanted a preview node that didn't write to the temp folder.
Turns out CR Image Output has an option for UI only. No temp file, no output file, nothing written to the drive that I can find.

[steep marlin]

Why? The temp cache clears out every time you relaunch comfyui

[lusty pewter]

And writing tons of images to SSD adds unecessary wear.

[eager escarp]

good point...

[smoky bear]

hey everyone, i cant really figure out this path thing or i dont even really know what it is. installed python 3.10.10 and comfy worked wellfor a while. but for some reason when i check in python_embeded it says python version is 3.11.8. is this because i ran this "update_comfyui_and_python_dependencies"? faceinsight doesnt work and i get all kinds of error messages, like this: Error occurred when executing IPAdapterUnifiedLoaderFaceID:

IPAdapter model not found.

when i installed python i didnt add it to path

i only checked the vm box

im down to do everything from scratch again, i factory reset my pc a couple times already on my journey to figure stuff out. im kinda learning by doing so i broke a lot

can someone maybe guide me into the right direction on how to set it up properly for the future?

[distant pendant]

guessing it only updated dependencies for the main repo but not all your custom nodes, if you're using the manager you could probably reinstall all the dependencies for those

[smoky bear]

i am

how do i do that?

[distant pendant]

there should be an 'update all' button in settings I believe

I don't really use it myself so not positive on the specifics

[smoky bear]

did it close it opened again

Install: pip packages for 'C:\Comfy V2\ComfyUI_windows_portable_nvidia_cu121_or_cpu\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Manager'
Install: pip packages for 'C:\Comfy V2\ComfyUI_windows_portable_nvidia_cu121_or_cpu\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_essentials'
Collecting pixeloe
Using cached pixeloe-0.0.7.tar.gz (13 kB)
Preparing metadata (setup.py): started
Preparing metadata (setup.py): finished with status 'error'
[!] error: subprocess-exited-with-error
[!]
[!] python setup.py egg_info did not run successfully.
[!] exit code: 1
[!]
[!] [6 lines of output]
[!] Traceback (most recent call last):
[!] File "<string>", line 2, in <module>
[!] File "<pip-setuptools-caller>", line 34, in <module>
[!] File "C:\Users\lukab\AppData\Local\Temp\pip-install-nsei9_c8\pixeloe_8b7be19700e240bdbf9ba7e04baf5147\setup.py", line 2, in <module>
[!] from pixeloe.cli import command_map
[!] ModuleNotFoundError: No module named 'pixeloe'
[!] [end of output]
[!]
[!] note: This error originates from a subprocess, and is likely not a problem with pip.
[!] error: metadata-generation-failed
[!]
[!] Encountered error while generating package metadata.
[!]
[!] See above for output.
[!]
[!] note: This is an issue with the package mentioned above, not pip.
[!] hint: See above for details.

got this

hmm can you maybe teach me how you would do it? instead of trying to fix the chaos i created

[distant pendant]

I'm not sure what pixeloe is tbh, personally what I do is use micromamba to create an env, then install torch and comfy requirements in it, then clone any nodes I want and just install their requirements manually

[smoky bear]

honestly no clue what you just said but ill look into it and try to replicate

maybe it gets me somewhere

[distant pendant]

conda/miniconda/mamba/micromamba - package managers, pick any flavor you like, micromamba is literally just a single executable, you can do python venvs too but that's under assumption you have python installed

package managers don't need anything at all

[smoky bear]

lets say you only want to run comfy on your pc, would you use mamba or python venv?

[distant pendant]

I use micromamba for everything personally since it's just so minimalist

but like assuming you had one of the above:

mamba create -n comfy python=3.11 git
mamba activate comfy
mamba install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
git clone https://github.com/comfyanonymous/ComfyUI Comfy -b master --depth 1
pip install -r Comfy/requirements.txt

now you have a clean comfy install

oh forgot torch, one sec

there, done

[steep marlin]

SSDs have a higher average lifespan than HDDs do these days and most can handle many terabytes written. Talking like 500TB for your average 1TB drive. The capacitors will go bad before the drive's life dips below 90%... Now watch my SSDs die on me soon lol. Oh and to put that into perspective, a terabyte would be like 300-500k 1024x1024 pngs. To do 500TB worth of wear on your SSD, you'd need to generate like 150-250 million pngs

[smoky bear]

how do i interact with mamba?

cmd?

or is it a programm?

basicaly where would it put that text

[distant pendant]

terminal/cmd
https://mamba.readthedocs.io/en/latest/user_guide/micromamba.html

assuming you have the micromamba executable, I mean

if you're on windows you can get it with winget

[smoky bear]

im on windows

[distant pendant]

winget install Mamba.Micromamba in terminal, default keyword might be micromamba not mamba, I forgot

[smoky bear]

should i switch to linux? was unsure about that too but then i thought if i also switch operating system it will become even more confusing

[distant pendant]

if you don't know your way around linux already it won't help you with anything

[smoky bear]

okay so i stay on windows:)

linux maybe for later then

how can i reset everything i did so far?

both python versions the paths etc

factory reset again? haha

[distant pendant]

I'm not sure how you installed everything but factory reset is way too drastic to get rid of a single repo and some python exe

[lusty pewter]

While that may be true, there are other reasons why I don't want to move that data onto the drive unless I'm committing the file.

Scale is a thing.

[smoky bear]

okay i see

[steep marlin]

Oh paranoia and tinfoil hat stuff, worried people are going to see your waifus

[smoky bear]

is there some way to have a look at all my repos?

[distant pendant]

if you just have python installed system-wide you can leave it alone, package managers won't interfere with it

[lusty pewter]

Your mind went there, not mine.

[smoky bear]

i do but not added to path

whatever that means..

im sorry man

im really noob

but thanks for your help

[steep marlin]

lol, but the shoe fits for a good 80% of SD users. stereotypes are a thing for a reason

[smoky bear]

more like 90 if you look at civit

[lusty pewter]

Okay. 🤷‍♂️
I guess you do your thing and I'll do mine? ☮️

[distant pendant]

'in path' just means it's visible to all the scripts and you can call it from anywhere, if you search for 'edit environment variables' in windows you'll get the popup for it, just have to modify the 'Path' variable there, but like this is not strictly needed for what you're doing

[steep marlin]

yeah, i'll stick to making my desolate and abandoned desert landscapes and cities, you keep makin your anime t----s (rhymes with cities)

[smoky bear]

i see how come comfy has a different python version then there is installed on my pc?

[lusty pewter]

Well...though I enjoy a nice pair as much as the next person, like I said earlier, that's not where I went with it; nor am I planning to. I'm working on something at a scale that requires every ounce of performance. Writing to the drive adds to that at scale.

[smoky bear]

my guess would be comfy has some kind of venv? running pytrhon?

[distant pendant]

if you install comfy by means of downloading the portable version from github it comes with its own python executable and venv

[smoky bear]

i did that

[distant pendant]

it's in the python_embeded folder or something like that

[smoky bear]

so i could uninstall python from pc and it would still work?

or?

[distant pendant]

yeah

[smoky bear]

okay this makes me think my issue is something else then...

[distant pendant]

you don't need anything other than the archive from github to run comfy

[smoky bear]

i just create new folder and git pull this https://github.com/comfyanonymous/ComfyUI.git into it?

and dont mess with anything else?

[distant pendant]

your initial issue was that your comfy update script modified python version and all the custom node dependencies went to hell because of that, I think
not sure about current issue

[steep marlin]

Not really if it's done with a separate thread. Also, I think comfyui now natively has support to not save previews/temp stuff in API mode. I'll dig up the pull off git

[smoky bear]

that sound like it makes sense

[distant pendant]

you git clone (not pull) it if you want to install everything manually, otherwise you should use the archives in the releases tab
like if you just want a refresh download 'Download Link with stable pytorch 2.2 cu121'

[lusty pewter]

Not using the API.

[steep marlin]

Well either way, the <1ms time that it might stall your pipeline to save a file is likely going to be the least of your worries. Even if you're doing some realtime video stuff, the highest you'll need to worry about might be 60fps(16.7ms per frame)

[smoky bear]

i dont find the archives in the releases tab do you maybe mean this? https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_cu121_or_cpu.7z

thats the direct download link

mb

[distant pendant]

yeah, that one

[smoky bear]

okay

so uninstall comfy uninstall python cause why not, then download this

how do i prevent it from upgrading python version in the future?

so this doesnt happen again

[distant pendant]

you don't technically need to uninstall anything since it comes with its own python and everything like I said before, and I mean the easiest way would be to simply not click on the updater...

[smoky bear]

okay:)

i installed faceinsight and something called onnx or something, will that cause any issues moving forward?

had to install visual studio for it to run

[distant pendant]

I mean, depends on where you installed it, but the comfy venv is just its own separate thing

if you compiled(?) those outside of the venv you won't be able to access them inside it (most likely)

[smoky bear]

probably did

how do i get rid of them?

look up how to get rid of certain dependencies?

is that the keyword?

[distant pendant]

pip list
then pip uninstall whatever-package

[smoky bear]

C:\Users\lukab>pip list
Package Version

---

pip 22.3.1
setuptools 65.5.0

[notice] A new release of pip available: 22.3.1 -> 24.0
[notice] To update, run: python.exe -m pip install --upgrade pip

feel like i update that pip about 100x btw

[distant pendant]

that means you basically don't have anything whatsoever in your system python install, might want to go with the fine-grained issues to tech support at this point tbh

[smoky bear]

okay thank you for your help

appreciated

[distant pendant]

no problem

[smoky bear]

one last question if you dont mind

any good reasons not to use the direct download link?

[distant pendant]

I mean, it's just a matter of preference, cloning the repo, installing torch and all the dependencies will likely take more time than downloading and extracting an archive but you have less control over it

it's just there if you're not very technologically inclined or just want an 'exe' to click on and gen images

[smoky bear]

okay, but i can still install dependecies individually within it right?

[distant pendant]

yeah but if you're doing that without the manager you have to use the python from embeded folder

like just calling pip in the main dir won't work, you should do python_embeded\python -m pip install whatever

[smoky bear]

okay i see

do i not have to enter full path?

i guess since there is only one python_embeded on my pc it finds it like that? is that how ti works

[distant pendant]

it's relative to current dir, I don't really have the full structure before my eyes but you can probably just go to the embeded dir and call python from there, it should take precedence over path

[smoky bear]

C:\ComfyV1\ComfyUI_windows_portable\python_embeded>

[smoky bear]

i dont understand this but ill look it up

thank you very much really

[distant pendant]

you're welcome, and sorry, I'm not very good at explaining things 😅

[spring jungle]

so question i tried to load a workflow it says i have missing nodes but i cant install any of these node from manager so what do i do xD

i fix that problem i deleted the folder reinstall add on

[wild thicket]

cfsccv

[hazy ivy]

Folks, what is the best way to automatically crop face while preserving the latent size to 512x512 in ComfyUI?

[storm folio]

https://comfyanonymous.github.io/ComfyUI_examples/edit_models/

[sly sundial]

How long does it normally take to extract the windows portable version of comfyui? Mine is estimating 8.5 hours...?!?!

[lone bolt]

What zip program are you using?

Windows was painfully slow. 7zip solved that.

[sly sundial]

I am using Windows. The time has now updated to "More than 1 day" 😢

[sly sundial]

Thanks! 7zip took 1.5 mins

I had it installed before, but forgot to re-install after replacing my C drive

[smoky bear]

Hello everyone, I started using comfy about a month ago, slowly getting used to it. So far I really enjoyed the journey and my jaw dropped more then once by the capabilities there are. Do you guys think comfy/stable swarm/auto1111 is here to stay and maybe become industry leader over the next couple years? I guess diffusion models in general are here to stay, so learning something like comfy will always be advantageous? The thing is, I'm down to put all my energy into it over the next couple of months I'm just a little scared I might waste my time if something else comes along. Does anyone maybe have some insight on what's on the horizon and how the future of diffusion models might look like?

[robust bloom]

hello, i use mac, why my image is pure blue color image. vae:https://huggingface.co/stabilityai/sdxl-vae/tree/main, base:https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main,refiner:https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/tree/main, my comfyui log is : got prompt
model_type EPS
Using split attention in VAE
Using split attention in VAE
clip missing: ['clip_l.logit_scale', 'clip_l.transformer.text_projection.weight']
Requested to load SDXLClipModel
Loading 1 new model
Requested to load SDXL
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:08<00:00, 2.34it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 18.44 seconds

thanks to everyone.

[eager escarp]

i don't think it'll ever be a poor use of your time

here's the huge thing with comfyui: even if a new version of it and/or stableswarm comes along that overhauls the interface completely, you'll still have learned a ton about how the process actually works under the hood so to speak

i think everyone should learn how to use it and use it a lot for that reason

[robust bloom]

thank you. do you have a reminder of this issue? so i can learn from this issue. this workflow is a example of what comfyui offers.

i setup a svd workfow before,it successed. then i changed to text_image,it generates no image.

this is log of svd which can generate a two seconds video:To see the GUI go to: http://127.0.0.1:8188
got prompt
model_type V_PREDICTION_EDM
Using split attention in VAE
Using split attention in VAE
Requested to load CLIPVisionModelProjection
Loading 1 new model
Requested to load AutoencodingEngine
Loading 1 new model
Requested to load SVD_img2vid
Loading 1 new model, and, this is log of text_image : it seems that i missed something:
model_type EPS
Using split attention in VAE
Using split attention in VAE
clip missing: ['clip_l.logit_scale', 'clip_l.transformer.text_projection.weight']
Requested to load SDXLClipModel
Loading 1 new model
Requested to load SDXL

so the key point is :Using split attention in VAE
clip missing: ['clip_l.logit_scale', 'clip_l.transformer.text_projection.weight']. ?????

[robust bloom]

the same issue, anybody had this issue? https://github.com/comfyanonymous/ComfyUI/issues/3161

[olive yoke]

what are you trying to do?

[eager escarp]

author of ImpactPack and InspirePack has patched the unsampler and tiled_ksampler nodes so they work again

https://github.com/BlenderNeko/ComfyUI_TiledKSampler/commit/8e4c354db1c94f915187170382b5cbdc4e7d81ef
https://github.com/BlenderNeko/ComfyUI_Noise/pull/28/files and another patch for the unsampler etc

[robust bloom]

anybody here can remind me? it is pure color image, i think there may be a low level mistake in my workflow setting.

[olive yoke]

this is a little vague

[olive yoke]

do you mean this issue?

[robust bloom]

yes.

[olive yoke]

are you on an M1 mac?

[robust bloom]

yes. it is true

m2 max, 32 g ram

14.4.1

[olive yoke]

you should share all the logs, including the startup logs

[robust bloom]

use startup.log.1 , i don't know what is spoiler which is automatically generated by discord

[olive yoke]

add --fp16-vae to your start arguments. it is unusual that you are invoking python this way

you should be using a virtual env

python3.11 main.py --force-fp16 --fp16-vae so try this.

[robust bloom]

i installed python in homebrew already, so i just took on this environment.

[olive yoke]

you can ask a chatbot what i mean by virtual environment

and walk through the process on macos

[robust bloom]

thank you.

[olive yoke]

you will do something like

cd ~/Documents
python3.11 -m venv venv
source ~/Documents/venv/bin/activate
...

[robust bloom]

if i remove --force-fp16 . do i need to use --fp16-vae ?

[olive yoke]

use the command i sent you

you should also do

python3.11 -c "import torch; print(torch.__version__)" to show your torch version

just in case

and share it here

[robust bloom]

python3.11 -c "import torch; print(torch.version)" to show your torch version
2.4.0.dev20240401

[olive yoke]

i think it's a little ambitious to use a prerelease torch version

[robust bloom]

https://developer.apple.com/metal/pytorch/

[olive yoke]

pip uninstall -y torch torchvision torchaudio && pip install torch torchvision

use the command i shared

[robust bloom]

i just use this to install pytorch

[olive yoke]

hmm