🧣｜comfy-ui

[merry ermine]

wow I gave it slot machine and craps table as extra prompts and it got them!

this is really cool

[hollow bay]

This is acutally not to bad, but its not that my image contents are random. It must be possible to set the mask without yolo or not?

I always want the hexagon in the middle

to be masked

[merry ermine]

is the hexagon always exactly the same pixels

in coordinats

[hollow bay]

yes!

[merry ermine]

ah yeah this will be doable without a model

[hollow bay]

Thought so, but i couldnt figure out how

[merry ermine]

can't you go into photo shop
add a second layer, and draw mask?

[hollow bay]

I want to automate the process

Or what do you exactly mean?

[merry ermine]

not sure how to automate it

well I would do it will yolo but apart from that

[hollow bay]

I cant believe that its that hard to pre set a mask given on a shape 😄

[hollow bay]

I mean thats a nice way, but i gotta keep the computing low

and the ifnerence as fast as possible

[merry ermine]

I don't think that will be an issue
simply because running yolo is so much faster than running stable diffusion

if you can generate the image then you can definitely run yolo in a fraction of the time

[hollow bay]

thats good to know, ill keep it in mind if i cant find another (simpler) option!

thank you so much man!

[merry ermine]

no problem

there will definitely be a classical method but it will be more complex

[hollow bay]

Yea i was thinking about sth like exporting the hexagons one by one, (i guess then it should be easier to mask them ) and simply but them back together when inpainting

But maybe its even easier. Ill let you know

[hollow bay]

@merry ermine by any chance do you know ho i can prevent my generation from cutting of / being out of frame?

[merry ermine]

what does your latent look like before VAE encode?

[hollow bay]

Initial image + mask input, get mask in 1024x1024, then forwarding it to the inference

Sorry for the mess, im pretty new with comfyui @merry ermine

And i dont think i need the Mask to region part however i was happy it was running like this and didnt want to make too much changes to it

One hexagon should be 1024x1024 whereas the picture with all 7 hexagons has 3072x3072

[merry ermine]

not sure this inpainting method works

you could try a different method like powerpaint2 node

[hollow bay]

awesome ill give that a try, as well as foocus. Thank you !!!

[proud vale]

hello, i am new to Comfy-ui, i want to download some model from hunging face , but i find it always great amounts of files.all of them need to be saved or just part of ?

[hardy tree]

if you just want the "models", and dont want to worry about those files and folders, just use CivitAi instead

unless what you want is exclusive on huggingface, in which case you only want the biggest file, models are usually anywhere from 2GB to 6.5GB, depends if SD 1.5 or SDXL, etc

[proud vale]

like this , there are different branch folder. Can i ignore them and just download the biggest .safetensors file in the root directory?

[hardy tree]

yes just download the biggest file in safetensors format

but you can find all that on civitai and it has previews, etc

also, i recommend you download finetuned models instead of base models

but then again, experiment with everything 🙂

[proud vale]

yes i tried civitai and it actually easier to operate. But sometimes i prefer base model than finetuned models.Anyway thanks for you help!

[hardy tree]

np

[upbeat charm]

Hey, I'm a complete noob, been trying my way with AI for a couple of weeks now and I wanna change my work flow excessively. I want to use multiple prompts to get more control over the end product, and I found a node to do that. But somehow I can't connect anything to the node, nor can I find a one-stringed text box. The node I want to use is NoxinNodes Split Prompt Organizer. I feel like I might be missing something. Could someone help and and explain to me like I'm stupid? Thanks a lot.

This is the node.

I saw someone use it as so:

And I can convert the widgets to inputs, but then I still can't connect a thing to them. The person in the teal-colored picture has one-string text boxes connected, but I do not know how to do that. I also don't know if this workflow really works cause don't things have to flow in, and then also out again? How can it do that if it's one-stringed?

So lots of questions lol. I hope someone can help me 🙂

[sly sundial]

The input (left side) are just entries that go into the fields, so you can type them directly into the widgets, or use a text node to add them. The output (right) needs to be connected to whatever you want to use those inputs on.

This text box should work

[upbeat charm]

I can't find this one in my nodes. I even searched with the searchbox but this one does not come up.

These are the only ones that come up

[sly sundial]

It's a custom node, so you need to install it from the manager

[upbeat charm]

Ah

That makes sense

[sly sundial]

The name of the nodes is top right

"WAS Node Suite"

Install from the manager

[upbeat charm]

Installing now

I only installed the manager yesterday so I'm not fully used to that either yet

Thanks a lot 🙂

I know the Combo+ and Combo- should be connected to something to flow through the workflow, but I got no idea what it should be connected to, lol. KSampler?

Cause it won't connect to that

[sly sundial]

I'm not familiar with the node, so maybe look at the repo for the instructions on using it. Probably need to got to a node to be combined and then fed into a text conditioning node.

[upbeat charm]

Got these two connected now, thanks 🙂

[upbeat charm]

I checked the read me, but it doesn't really tell you anything on how to use it

So it looks like I'm on my own

[sly sundial]

Doesn't it have example workflows?

No directory for them?

[upbeat charm]

Let me get the link, cause I coudln't find the examples but maybe I'm looking wrong?

https://github.com/noxinias/ComfyUI_NoxinNodes

[sly sundial]

The file that has "workflow example" in the name 😄

Drag it into your ComfyUI

[upbeat charm]

This is awkward, I missed that completely

Apparently this is the example then, but I don't see the prompt organizer in this, lol.

This is for the library it seems

The workflow did give me the idea to do this... and I'm gonna see if it works

It worked.

https://tenor.com/view/wolf-wallstreet-gif-19616559

The flow works, now tweaking 🙂

Couldn't have done it without you, thanks @sly sundial

[upbeat charm]

Next "problem" I run into is that it is taking colors from different prompts. I've had this problem since I started with AI. I want the girl to have grey blue eyes, but AI keeps grabbing her eye color from her dress color. No matter what color dress I choose, her eyes will match. Is there anything I can do to fix this?

[flat dirge]

What you're probably looking for is conditioning concat

[upbeat charm]

I honestly don't know what that is but I'll google. Right now I fixed the order of loading first.

[queen temple]

that happens because the tokens influence each other in cascade, influencing less and less the tokens coming after

pretty sure that node doesn't account for that

[upbeat charm]

I'm looking at the node, but I also can't figure out how that's supposed to stop the colorbleeding.

[queen temple]

try to add BREAK at the end of prompt

[upbeat charm]

Isn't that A1111?

[queen temple]

some nodes use it too

[upbeat charm]

When I search online they refer back to concat again

What is BREAK supposed to do?

This is my workflow right now. Just for reference

[queen temple]

id it doesn't, you need to add a concat properly done to split the colored element from the rest of the prompt

[upbeat charm]

Face is already better thanks to the image reference, lol.

[queen temple]

concat is wot's is used to imitat BREAK of comfy

[upbeat charm]

Yea, but I got no idea how to use it, lol.

[queen temple]

it's easy, just connect 2 prompting node to a concat node

you can even add a concat node to a concat note

to BREAK more prompt nodes

there are different concat types

[upbeat charm]

Yea, but what will that do? It says "conditioning from/to", but what will the outcome be?

I might try it in a clean default thing to see if I can figure out how it completely works, lol

[queen temple]

you should test what answer better to your necessities

for now just add a concat to face and add 2 prompt nodes

[upbeat charm]

I don't think that's going to work for my split prompt organiser though

[queen temple]

somewhere i have a concat for strings

too bad i deleted comfy and just using krita lol

testing it

[upbeat charm]

I can't connect CONDITIONING to face. And the separate nodes for my prompt shouldn't be mixed. I am going to test it, but I can not connect it to my prompt cause the nodes don't connect.

[queen temple]

there are concat that use strings instead of conditioning

[upbeat charm]

I could only place it between CLIP Text Encode (Prompt) and KSampler

Okay, I don't think I got those types of nodes yet

I put "green eyes" in my specific negative prompt now and I got brown eyes...

Not exactly grey/blue yet, but getting somewhere I suppose

Negative prompt does seem to be working for me.

So maybe the concat stuff isn't necessary at all 🙂

[queen temple]

if you can gen fast enough you can try to gen face separately nd than use it to influence the final gen

[upbeat charm]

I tried that, and I can generate fast, but I didn't like how it came out

[queen temple]

sometimes it works

oh ok

did you use inface? or adetailer?

[upbeat charm]

I tried that like two weeks ago or something, a friend send me his workflow for it

No idea what he used, lol

[queen temple]

or ipadapter

uahahah

[upbeat charm]

This is about the color I was aiming for

For the eyes

Very happy with my flow so far 🙂

Glad I decided to change it

[vocal vortex]

how do I remove, or disable, wildcards in the impact pack?

[hardy tree]

@storm folio this made me laugh https://new.reddit.com/r/StableDiffusion/comments/1e779np/why_my_comfyui_is_showing_this_is_there_anyway_to/ 😆

[chrome ether]

Does installing ComfyUI on windows actually installs anything (in the registries, environment paths, etc.) or is it akin to a portable application? (Beside the Microsoft runtime you have to install that is.)

[vocal vortex]

I installed the real way to install python based programs not the portable thing. venv which means it doesn't touch your system it is is a virtual environment.

[wispy kelp]

Hello. Currently as of today what is the best upscaler for comfyui other than supir.

like there is a lot of upscaler models for the sd upscaler... which one is best

for realism even

[livid gust]

They all seem the same to me. I typically stay with 2x.

[chrome ether]

Alright, thanks!

[vocal vortex]

you're welcome.

[chrome ether]

Is there a way to get which seeds were used for a given generation?

[merry ermine]

I use the portable version on windows with no issues. YMMV.

[chrome ether]

Yup, it's been working well for me too, at least for now.

[wispy kelp]

yeah all the youtubers make somethign seem aamazing but they all seem more or less the same, they all do an ok job..

[tardy hornet]

Hi everyone 👋🏼, looking for some AI image experts to join a soon to be launched virtual companion app. We're building a virtual companion app to cure loneliness and help people grow and feel heard. Our app allows users to customize their AI friends and we're very close to launching. Our goal is to have hyper realistic images of AIs including our items. We're using SD 1.5 as a base model. Looking forward to meeting people who want to join part-time, as freelancers or core team members. DMs are open! 🚀

[wispy kelp]

🤔

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

hehe

[brazen moat]

What is the best way to enhance an image?
Highres fix doesnt give good results

[steep marlin]

What do you mean by enhance? Upscale? Refine? Inpaint?

[brazen moat]

I mean to refine the result, faces don't look human, etc

[vocal vortex]

btw, wth is even the point of the portable since venv is 100% portable? I never understood why it exists.

has to be a reason

[chrome ether]

The portable version has an embedded python version.

[vocal vortex]

ahhh

[vocal vortex]

Since my system has python it made no sense to me. I had more issue with it, and if I wanted to do my own things it was so jacked up I finally deleted and went to the github venv version. No issues with it since.

[chrome ether]

Ahhh

I'm not a python guy ( 😲 ), so I'm fine with that version 😄

[upbeat charm]

Which tags do y'all use to create higher quality images? Right now I use 4k, HD, RAW, Photorealistic, photo, real (positive prompt)

bad quality, low quality, worst quality, blurry, incorrect hands, bad anatomy, extra digits, extra hands, extra legs, bad face, ugly, ugly face, extra limbs, bad anatomy, bad eyes, cgi, airbrushed, plastic, deformed, watermark, bad teeth, mole, moles, glitch, (negative prompt)

Is there anything I could do differently?

[forest birch]

Hoi, anyone knows how to fix this?

[steep marlin]

Try reinstalling it with comfy manager

[forest birch]

that's with comfyai manager

Unless you meant remove it, then reinstalling. ended up doing that manually, then i could continue with what i was doing. But comfy errors like that with everything

[merry ermine]

yeah I don't use any of those

just the word "photo"

and then the subject

and no negative

if you want higher image quality its better to pick a model that is focused on that and then just prompt "photo"

e.g. SD3, RealVisXL V4.0 etc

[steep marlin]

Yeah full uninstall, delete the residual folder in the custom nodes folder if it's still there, then reinstall it

[fading mural]

guys what sampler and scheduler do you use?

[cursive ridge]

What is this crap? 😦
"Error. No naistyles.csv found"
everything was working great until I started running SD3 stuff

[half hedge]

my whole pc freezes when stable diffusion gets too 99 percent and im not sure whats wrong , how do i fix?

[flat dirge]

You could look at the logs to find out more info, also monitor your system resources as it's happening.
My guess is it's happening when you get to the VAE decode. Maybe you're running out of memory there. You could try swapping to tiled VAE decode and see if that helps.

[mystic plaza]

i am guessing a model is missing here ?

[mystic plaza]

@steep marlin @dry rock any inputs on how i can fix this ?

[dry rock]

apart from telling you it s due to one of those nodes not finding a model. cant tell much more.

[mystic plaza]

so do i need to download that model or write a code for it cause these are all the models present in that directory ?

[dry rock]

dont think you have to write any code

more likely you re using the nodes in a wrong way.

or they re outdated

[mystic plaza]

https://www.youtube.com/watch?v=pFYqPf1Z7x8&t=107s

[mystic plaza]

trying to do something like this mainly @dry rock

[dry rock]

maybe it failed to download its models / didn t have enough room to download them. can t tell you for sure

[mystic plaza]

so should i try a uninstall and reinstall through the comfyui manager ?

[dry rock]

you can try, make sure to check the log for any errors

and make sure to have free space (25gb according to the video) on both your f: and c:

[mystic plaza]

it has plenty of space already tho in both drives

[livid gust]

@upbeat charm One thing I always add to the positive prompt when making characters is "realistic skin, natural pose." Realistic skin will trigger a close up, so be warned. Don't use it for landscapes or you'll get people.

[upbeat charm]

In my default positive I got: detailed eyes, detailed lips, detailed face, detailed nose, detailed body, beautiful, detailed iris, symmetrical face, symmetrical iris, realistic eyes, detailed skin

I got really weird faces without this all, but I can add realistic skin & natural pose as well

Also got 'cleft chin' in my specific negative cause I got super weird cleft chins on all my models lol

[livid gust]

Are you overdriving the model's resolution? Try backing off on the size or trying an alternate orientation (i.e. portrait vs landscape) I also often throw in "camera rule of thirds" in my positive, if subject framing is problematic.

[graceful pewter]

is there a node floating around that takes an image and an input mask, and returns the image with the masked area as transparent?
or do I need anynode to make a node to do that?

[steep marlin]

Stable diffusion doesn't use an alpha channel for actual diffusion unless you use that specific custom transparency model workflow, but it's a custom addon. An alpha mask is usually all you need to direct things like cnets, ipa and inpainting

What are you trying to do? Like why does it have to be transparent?

[graceful pewter]

I have the image generated, I just want to remove the background on it for placing on a website so that I can use css to control the background. Its a final process - I could use the transparent-background node from the LayerStyle pack, but it seems stupid to have it figure out where the background is when I already have a mask from earlier in the workflow

[tall prism]

hi

how to change stylize image A to image B Stable diffusion comfyui

[steep marlin]

I'll check in a minute, but I think there's a join alpha node or something, but you might have to invert the mask before it

If not, just save the mask separately and use krita or photoshop to apply it

[livid gust]

@tall prism It's called the IPAdapter. Watch Mateo's video posted at his GitHub.

[trail trench]

Hi i have a question. Is it possible to inpaint an image from an ipadapter? Like for example i have a base image of a empty interior and I want to inpaint a chair from ikea, is it possible?

And do we have a tutorial for that? I'm not really familiar with inpainting, just with ipadapter and controlnets.

[merry ermine]

powerpaint and brushnet 2 nodes

[rapid elbow]

auto1111 has prompts from file or textbox how would i get that on comfyui?

[merry ermine]

there's probably a node

WAS Suite "Text Load Line From File"

[covert bough]

https://x.com/fofrAI/status/1815786739362918751 FoFr has added Llama 3.1 405B into the Replicate ComfyUI custom nodes repo

fofr (@fofrAI) on X

I just dropped Llama 3.1 405B into the Replicate ComfyUI custom nodes repo.

So now you can run 405B straight from your ComfyUI:
https://t.co/wSAnwESe9U

Example workflow included in repo.

Twitter · Jul 23, 2024, 4:31 PM

[merry ermine]

awesome this might be the best prompt writer around potentially

[hardy tree]

@storm folio not sure if this was asked, but will Photomaker 2 be natively supported, it just released 😮

[storm folio]

probably not in the short term since it would require me to add insightface to base comfyui

[merry ermine]

that insightface license causes trouble for so many projects lol

[graceful pewter]

The github repo is MIT?

Ah its the model weights that are non-commercial

[livid gust]

Is there any way to color fill a masked area? I'd like to tint my overlay.

[distant pendant]

just load a second image with color maybe

there's probably some node that can generate image from color too

[livid gust]

It looks like ComfyUI_Fill-Nodes has a way to fill a solid color using a masked image. This lets me composite an overlay with a drop shadow.

[covert bough]

https://github.com/Acly/comfyui-inpaint-nodes

https://github.com/nullquant/ComfyUI-BrushNet

[steep trout]

Images d une terre aride dessèche et poussiéreuse. Avec un homme a la peau blanc âgé de 50ans avec des lunette une blouse médicale . Un policier noir une deux femmes . Nous sommes en Afrique dans un monde post apocalyptique

Images d une terre aride dessèche et poussiéreuse. Avec un homme a la peau blanc âgé de 50ans avec des lunette une blouse médicale . Un policier noir une deux femmes . Nous sommes en Afrique dans un monde post apocalyptique

[covert bough]

vous ne pouvez générer que dans le canal artisanal. cliquez sur ce lien et lisez les informations. #artisan-faq

[vital root]

please help me

[covert bough]

you'll need to either use the manager to install missing nodes, or install it by hand. let's do the manager first - go into it, and update all, then restart comfy. if the node is still not there, go back into the manager and choose install missing nodes. if it can't find any, then you'll need to go to the repository for the node, read the by hand install instructions, and follow them. you'll need to restart comfy after either method

[obtuse latch]

anyone have any script that pulls embeddings from images?

[steep marlin]

What do you mean, something like an image captioner?

If so, Florence-2 is really good and there's a wrapper node on comfy manager

[obtuse latch]

No I want to be able to see / view my prompts in my image workflows (embedded workflows in images)

[obtuse latch]

CLI / Script

[steep marlin]

Oh that's a pain in the ass for comfyui since it's node based. You'd need to do something like give your pos/neg prompts some unique names and then parse the metadata for them

You could have 20 different positive prompts for various stages of a workflow

See where I'm going with this?

[obtuse latch]

For sure, I just need to be able to access the actual json workflow from an image, I can parse the prompts.

[steep marlin]

Should be doable with PIL I think, but there might be another library needed to read the metadata and probably a json library as well

Just Google for python image read metadata json or something like that

[obtuse latch]

Got it, metadata = img.info thanks

[steep marlin]

sweet, yeah, you'll still likely need some kind of json library to make things easier

since it's stored in that format afaik

[obtuse latch]

Yeah, all good on my end, thanks again!

[true onyx]

I'm getting this error using this workflow, using the same image in a previous workflow with VAE encoder for inpainting doesn't give the error. How can I fix this?

[forest birch]

tested out tensorRT in comfy, and it was interestingly slower than automatic1111's tensorRT from almost a year back. Only gained 3-4 steps in comfy, while automatic's boosted it 3x faster than regular models

@storm folio Hoi, do you know why it's only giving 3 more iterations per sec with 512x512 1.5 base model? As automatic's addon gave 3x higher it's than stock, and it also needed the tensorRT library. Doesn't comfy use those?

[storm folio]

GPU?

[bleak stump]

New audio-reactive geometry system!

[Architecture - Archaic / Ancient - AI]

https://www.youtube.com/watch?v=IOD-eTIm9g0

[covert bough]

https://x.com/cubiq/status/1816830400339030235

Matteo Spinelli (@cubiq) on X

clipvision enhancer is now compatible with image batches. Take this weird animation made with it. We lose some of its power because I have to average composition/details embeds, but maybe you can do something cool with it. Also thanks to @comfyicu for becoming a silver sponsor!

Twitter · Jul 26, 2024, 1:38 PM

[obtuse latch]

Trying to get a tiled teture working, anyone know the best way?

[radiant sundial]

Hi guys, currently trying to install ComfyUI on Windows with RTX 3090.
Git clone, install pytorch stable, pip install requirements then when I launch python main.py I got this error :

PS C:\Users\Raphael\ComfyUI> python main.py
Traceback (most recent call last):
  File "C:\Users\Raphael\ComfyUI\main.py", line 83, in <module>
    import comfy.utils
  File "C:\Users\Raphael\ComfyUI\comfy\utils.py", line 1, in <module>
    import torch
  File "C:\Users\Raphael\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\__init__.py", line 148, in <module>
    raise err
OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\Raphael\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\lib\fbgemm.dll" or one of its dependencies.

Any idea how to fix this?

[obtuse latch]

You don't need to install anything the .bat installs eerything for you.

[radiant sundial]

with the direct download link? it's the same than doing the git clone?

Oh this: https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.0.2 ?

[obtuse latch]

Unless you need to rebuild / modify, I wouldn't use the repo.

[radiant sundial]

Okidoki

To be honest I'm not sure if I need to rebuild / modify, so probably I don't 😄

Everything will be optimized for Nvidia GPU already with this install?

[obtuse latch]

You may need to install uda

[covert bough]

you should probably just watch Scott's tutorial on how to install comfy

[eager escarp]

posting this here for anyone that's using UltraPixel for cascade https://github.com/2kpr/ComfyUI-UltraPixel/pull/27

[covert bough]

did you notify the ultrapixel author?

[eager escarp]

gotta do that too, yeah

[radiant sundial]

Yes I did and it works like a charm now! 🙂

[radiant sundial]

I see this when I launch comfy:

Device: cuda:0 NVIDIA GeForce RTX 3090 : cudaMallocAsync
cuda:0 means it's not installed I guess? what's the best way to install it?

[eager escarp]

https://github.com/catcathh/UltraPixel/issues/12 done, or close enough to it

they're the ones that originally linked to the non-safetensors safetensors

[eager escarp]

something that really should be a thing:

we've got checkpointsave, clipsave, vaesave....

but we don't have modelsave/unetsave

[pliant sluice]

I'm having trouble finding the "Manager" button in ComfyUI. Has anyone else encountered this issue or has any suggestions on how to resolve it? Any help would be greatly appreciated!

[covert bough]

i had that. clear your browser and windows (if using windows) cache - and i don't mean hit F5, i mean go into the browser's settings and clear cache files

[pliant sluice]

got it thank you very much

[covert bough]

you're more than welcome, but it's not my solution. it's what @uncut shoal had me do

[trail trench]

Anyway to improve the workflow? I want to put my chair inside a base empty interior room (not a generated one) :

i cant seem to make it work

[long flame]

@covert bough @dry rock what i did wrong? the image is so ugly

[covert bough]

why is your CFG set to 0?

[wispy kelp]

so initial render with pony, refine with sdxl- replace background with sd3, - handetelair - facedetailer - upscale with Sd upscaler - final sd15 pass

[long flame]

its 9

[wispy kelp]

early attempts - also there is a face ip adapter in the facedetailer part and two initial optional openpose and whatever else controlnets

so well render these movie scenes and eventally when video gets good enough we make movies

[covert bough]

can you upload your workflow please

[long flame]

workflow?

who can i see that? do you want me take a screeshot of my comfy?

[covert bough]

the first thing you need to do is install the comfy Manager. then you'll have access to several extremely important tools, including the ability to save your workflow

some good comfy tools https://github.com/lucianosb/awesome-comfyui

[surreal whale]

you can save json with the save button and save image of workflow right clicking the screen, not part of manager, but ya manager is super nice to have

should be default

[covert bough]

sort of the difference between trying to fix a car with just a wrench and screwdriver - and trying to fix a car with acess to a mechanic's shop

[surreal whale]

pretty accurate

[long flame]

who can i find this workflow bro

[surreal whale]

right ckick save png

[long flame]

i dont have?

[covert bough]

did you watch the full tutorial i gave you the link to yesterday? the first thing scott has you do is to install comfy manager

[surreal whale]

i guess i was wrong i havent had vanilla comfy for a long time

[wispy kelp]

just render the image and save it with comfy

itll have the workflow in it

[surreal whale]

the first pic you posted looked somewhat reasonable besides the upscale part, you should put the denoise at 1 and try again

[long flame]

gonna do everything he said and generate a image, after i go back here

[merry ermine]

The following 5 things in your workflow will lower image quality:

1. sub optimal prompting
2. CFG too high
3. DPM++ SDE does not converge well at 30 steps. Switch to DPM++ 2M if you just want a single pass with 30 steps.
4. You are doing 0.46 denoise on an empty latent
5. You do an 8x GAN upscale

[trail trench]

Hi people! How do we apply a specific part (example a chair in a room) and apply an ip adapter to that changing its style?

[covert bough]

inpaint

[trail trench]

will try that

do you know how to copy nodes while still being connected to other nodes?

tried copying but they are untethered to their original connected nodes

[covert bough]

not a clue. i see people do that in videos but i can't get it to work

[trail trench]

alright thanks!

[trail trench]

@covert bough tried inpainting and it only resulted in this image

im assuming inpaint by model also requires a specific model right

[covert bough]

you have to mask out what ou want to change, and then change your prompt to specify what you want in the masked area

[long flame]

i did everything in the video and continue ugly @covert bough

[covert bough]

i'm not sure what you have in that negative prompt, but remove all of that and leave that blank. only use negative prompts if you absolutely need to, and then be super specific about the terms you put in, and only put in one or two, to start with. upload your workflow please

[wispy kelp]

why 512 tho

[covert bough]

probably how it loaded. he's very new to comfy

[long flame]

i dont have workflow

[merry ermine]

there are still 4 big problems:
prompt issues, steps too low, CFG too high, euler is not a good sampler

[long flame]

what i put in cfg?

5?

better?

[merry ermine]

this is getting better yes

you don't want to try to predict in advance which values would be good for setting like CFG
instead hold everything else the same and do generations with different CFG values and see the result

with same seed

[long flame]

so i put different values in cfg and the same seed?

[merry ermine]

yeah and see what happens

[long flame]

what is you suggestion in cfg?

you said its too high

[covert bough]

don't put cfg higher than 4 to start

[covert bough]

no. that's a screen shot. i can't load that into comfy and see your values. save the workflow with the SAVE button in the manager OR save an image - images saved from comfy have the workflow in them - and upload either or both of those

then the rest of us can load your workflow, look at the settings, see what might need to be adjusted, and actually help you with this

[long flame]

finally i did

[wispy kelp]

haha poor euler always gets trashed

[covert bough]

it's usually a good idea to give your workflow a unique name when you save it

[covert bough]

you still haven't set cfg down to 4

[long flame]

now i did

[long flame]

i have to rename it?

[wispy kelp]

send me the workflow

ill tinker with it

[covert bough]

is this the workflow that scott had you build or did you get this from somewhere else?

[long flame]

scott

[covert bough]

okay. i don't have that specific dreamshaper's model. i'm using dreamshaperXL_v21TurboDPMSDE. by switching to that model, slowing CFG to 4 on BOTH ksmplers, changing the sampler on the first one to euler_ancestral (leaving the second on alone) this is the image i get

by changing the sampler to euler_ancestral on both of the ksamplers (changing nothing else), this is the image i get

so at this point, you have comfy running, now you need to learn what all the values on all the nodes are for and what they do. start by just playing around with different samplers

and different models

[long flame]

gonna try this

the steps still the same ?

scheduler is normal im ksamplers?

[covert bough]

if you 1. click on one of the images i just uploaded 2. once the larger image opens, click on the words open in browser 3. once the image opens in a new tab or window, right click on the image and save it - then you can just drag and drop that image into comfy, and you'll get the workflow - but you will need to go find and download the model i'm using or that node will get a red outline and you'll get an error that comfy can't find that model when you try to run it

[covert bough]

karras

eular_ancestral sampler. karras for the scheduler

[long flame]

okk

[covert bough]

here are the settings for both of the ksamplers

[long flame]

now i just need to wait, its take too long to generate a image

[merry ermine]

there are ways to get the step count down but try to get your image working first

[wispy kelp]

turbo

not turbodong

[merry ermine]

I meant stuff like Align Your Steps or GITS with UniPC

[covert bough]

that's a function of both how much you have it doing, and what hardware you're running on

[wispy kelp]

so whats the deal with unipc

[covert bough]

no, that's the n-sfw version

[wispy kelp]

i kinda used it by accident and it seems like a hidden gem

[merry ermine]

unipc is one of the strongest samplers

it can do ok at rly low steps

with the right schedule

[wispy kelp]

👍

[merry ermine]

oh its turbo

that changes quite a lot LOL

[covert bough]

it's also a great model

[long flame]

the ksampler normal

[long flame]

you have the link of the model?

[covert bough]

your prompt is confusing, btw - i'm not sure what you mean, and the AI is probably even less certain

[covert bough]

search for it with google. you should find it on civitAI and probably a few other places like huggingface or github

[long flame]

i just copy the prompt from civitai, have i put other prompt?

[merry ermine]

I don't know turbo very well but 20 might be too much steps for turbo

if I remember rightly 4 CFG is ok for turbo but not for lightning or hyper

[covert bough]

look at his image again. i'ts not overcooked, she's in a wierd pose. that's the prompt

[wispy kelp]

the sad truth is that even with the best settings u can get duds like that floating shoe

[merry ermine]

yeah that's true it doesn't look overcooked so I guess 20 steps turbo is fine

it surprised me cos I thought turbo was closer to lightning in terms of steps

[covert bough]

20 steps is probably not necessary. but the prompt is confusing to the AI

however, i was demontrating the difference in cfg and sampler, so left everything else he had alone

[merry ermine]

yeah thats fine
its good to change only one thing

I'm essentially permanently stuck on juggernaut
cos they have a lot of sci fi in there

compared to the other big fine tunes

[long flame]

@covert bough well i downloaded the model

[covert bough]

did you follow the steps i said to download the correct image from discord? or did you just right click and save the image you saw in this chat?

[long flame]

yes

i did

[covert bough]

did which?

[long flame]

the steps i did

[covert bough]

this is a good, basic, SDXL workflow. start with this

or this

if comfy can't find the model, just pick a model you do have

[long flame]

works

i just have to wait now

[covert bough]

when you get this, you go into the comfy manager, which i see you still have not installed, and you run the install missing custom nodes function

go install comfy manager

[long flame]

@covert bough bro finally its generate

its soooooo long lol

[covert bough]

nice image and how much vram did you say you had?

[long flame]

16 ram

3200mhz

[long flame]

@covert bough i downloaded comfy manager

[covert bough]

restart comfy, and manager will show lup

[long flame]

its already show

[covert bough]

okay, so now you need to go back to scott's youtube channel, rewatch his first video to make sure you didn't miss anything, then watch all of the rest of his comfy tutorial videos and work along with him, doing the steps as he does them

and don't skip anything

watch from the very first comfy video i gave you the link to, in order, till you get to the last one

[long flame]

okay, the one problem i have now is it take toooooo long to generate

[covert bough]

that's because your trying to run the program on hardware that isn't really powerful enough

[long flame]

who can i change it

[covert bough]

you buy a better computer

[long flame]

but my pc in games is good

[covert bough]

so what? those are games. totally different subject. what you need for games is NOT what you need for graphics. those games use a combination of pre-rendered graphics, and low polygon textured objects. they are not processing graphics for the most part. not sure why everyone thinks that if their computer will run games, it will also be a graphics powerhouse rendering machine, but that's not the case.

[long flame]

ooohh 😭

your gpu is nvidia?

[frosty dock]

Hello, I'm not sure if this is the right chat to ask, but has anyone encountered a white blank screen on ComfyUI? Everything was working fine, but it suddenly stopped. I checked the console with F12, but there are no errors or anything interfering.

[covert bough]

if you want to run CUDA, you need nvidia

[covert bough]

reboot first to reload everything fresh (use restart from the manager, or just kill the command line window and close the browser tab) and then, if it happens again, try clearing your browser cache

[frosty dock]

I uploaded to my other drive from scratch it still did not work, looks same situation. I deleted all my history from browser and updated everything still same.

[covert bough]

turn off your browser's ad and popup blockers, and your anti-virus if you're running one, and see if that helps. what browser are you using

[frosty dock]

I ve just turned off all no luck. I use chrome and tried Firefox did not work.

[covert bough]

can you do this: set your anti-virus not to run at startup. make sure you don't have anything on your system that will block things that try to open in your browser, and then shut your machine down, leave it off for 10 seconds to totally discharge, boot it up, make sure you don't have any thing running in the background that could interfear. And then try runnign comfy again

[obtuse latch]

Anyone runnign cui on linu?

[obtuse latch]

Don't turn your Antirus off @frosty dock yo should neer need to, you most likely installed a node that's breaking, just roll bak and test a new install.

[covert bough]

sigh.

[obtuse latch]

He said it was working before, turning off Anti wouldn't do anything, the ports are not bloked by anti anyways.

[covert bough]

his tech support is all yours at this point.

[obtuse latch]

Turning off av isn't support 😄 but sure.

Just dl a new instane of ui and see if that will run.

[frosty dock]

I don’t think antivirus cause this I ve checked/tested multiple times nothing is wrong. I installed new comfiyu which is no custom node , still not working. Comfyui icon is working I can see but not responding . I have encountered so many problem but this one I have never seen

[obtuse latch]

it's not anti, it's neer anti, unless it's bloking a port (404) and that's fire wall.

Never turn anti off.

[obtuse latch]

Your inspector isn't showing any network, so you're not actually hitting the host.

No data is being GET

try localhost:...

[frosty dock]

Is that bad 😅 I’m not really tech person but trying

[obtuse latch]

There's no data being sent from the instane.

So sure.

[frosty dock]

Remember last week there was windows problem all around the world. I think I also affected . After windows update everything is working. Thanks for your information guys.

[wispy kelp]

block updates.

[wispy kelp]

Bill Gates moment.

So this is the changes we will enforce. And these are the ways you can circumvent said changes. will you please love us? Apple will not allow such transgressions. The multitudes waving those phones will not understand.

[covert bough]

it was not windows. it was crowdstrike for windows. and since crowdstrike runs where the windows OS files run, it crashed windows when it crashed.

[wispy kelp]

We are at the end of time.

It is one of the reasons AI is allowed to proliferate. This is to show how easily our senses are fooled and how superficial what we think of as reality is and how much deeper truth is.

Eventually everything out there will deceive us and we can only find truth within.

This is it folks.

It will only get worse.

The wars, the climate, the nonsense. The pain.

The bullies,

The idiots.

When you can make your deceased ones talk as if they were still here. Thanks to AI.

You will know how fake this is.

All of it.

And you will be angry you ever fell for it and angry you did not punch the bullies and liars in the face way before.

And moved on to better and greener pastures.

Every batch of apples needs a rotten apple.

But not forever.

Damn am I glad I live now at the end of time to see all of the evil shitheads go down.

We will all celebrate yes?

Puke drunk.

Remember.

Iran, North Korea, Hamas, Hezbollah, all the bullies, in every classroom, in every Discord server, in every family. All evil. All evil. all of it.

Gone forever.

We won't shed a bitter tear.

[hollow bay]

Hey guys, I was wondering if you can use SD 1.5 Loras with SDXL models or even sd3?

[sly sundial]

No

[livid gust]

@frosty dock From any open browser, you can always just navigate directly to the port.

http://127.0.0.1:8188/

[radiant sundial]

Hello guys, I'm tryng to use FaceDetailer in a txt2img workflow, I do VAE Decode --> Image batch to Image List --> FaceDetailer (pipe) but I have this error :

Error occurred when executing FaceDetailerPipe:

mat1 and mat2 shapes cannot be multiplied (154x2048 and 1280x768)

File "C:\Users\Raphael\ComfyUI_Portable\ComfyUI\execution.py", line 151, in recursive_execute

Probably missing something but not sure what

Also tried using the Rebatch Latents between my kSampler and VAE Decode but having same kind of error

[vestal thicket]

https://tenor.com/view/chaos-chaotic-chaos-reigns-anti-christ-fox-gif-5728648

Anywho, new to comfy, old hat with modular programming, though. Trying to figure out how to implement refiner. Anyone got any advice?

[wispy kelp]

what does this do?

why rescale so much

[vestal thicket]

Creates a random choice of image size from a set of predefined ARs. There are easier ways to do it. I'm just being goofy. It's sort of a joke.

[deft saffron]

I am looking for someone who can generate pics of 2 different person in same picture using multiple ip adapter

For example me and Zuckerberg shaking hands

[vestal thicket]

Your example makes me immediately suspicious.

[deft saffron]

Haha

Bdw is it possible?

@vestal thicket is it possible using two different ip adapter of 2 different faces?

[deft saffron]

How 😮 @vestal thicket

[terse talon]

Hey everyone I am looking for a ComfyUI expert: https://www.upwork.com/ab/applicants/1815333397113988233/job-details

[deft saffron]

Can you please help @vestal thicket

There's no where anything on internet about two PPL in same photo

Everywhere there is lora

[vestal thicket]

Yeah, I'm not comfortable helping you with this. But we'll always have Zuckerberg-Meringue pie.

[covert bough]

SD3 can successfully do more than one person in the same image. the other models have a problem with that concept

[vestal thicket]

Moving on... Is there a custom save image node that'll store and output said image when called?

[covert bough]

yeah. i'ts called save image

it'll display the image like preview image does, but also write it out to the outputs directory, and the workflow is included in the image

[vestal thicket]

Poor wording. Trying to cut out a step. I'm looking for a save image with a switchable outpat. So I can say "Hey, I like that output, let's img2img it." without going to explorer.

[covert bough]

no, you just run the image connection from VAE decode into another processing set of nodes

[vestal thicket]

I hate that. Time to learn how to make nodes.

[covert bough]

here - this is my SD3 to SDXL character generator workflow. create the character with SD3, process/refine/etc in sdxl

use that to start with

[vestal thicket]

Thank you, o wise wizard.

[covert bough]

welcome.

[surreal whale]

https://github.com/chrisgoringe/cg-image-picker

[vestal thicket]

https://tenor.com/view/lucy-vanpelt-charlie-brown-thats-it-piano-peanuts-gif-16784423

Holy carp this is even better than what I had in mind. This will save so much time. Thanks much.

[surreal whale]

NP, this set also has a bunch of different types of switches, might be useful if your gonna build out a big workflow https://github.com/rgthree/rgthree-comfy

[vestal thicket]

Lol. I already have it. I need to learn what f* nodes I actually have. Between Comfy an VCV I have no idea what kind of power I'm sitting on.

This dude's "Sillynodes" pack looks wild, too.

[vestal thicket]

So, noob issue. I git-cloned CG-Image-Picker, but it's not showing in my custom nodes folder. Thanks again fro recommending RG. Already very useful.

[vestal thicket]

Nother issue: I'm trying to select random images from a source image directory. I'm going random prompts(1-500 as string)->Anyconvert to int-> batch index input. Keeps erroring out saying int is invalid input for index. Tried inputting the raw string, and it hates that more. Is there a way for me to find out what kind of input batch index is looking for?

[surreal whale]

dont really know what or why to either question, one more helper package makes installing things easier. might help in general if you don't have it https://github.com/ltdrdata/ComfyUI-Manager

[vestal thicket]

My dumb ass. I didn't git it with manager.

[vestal thicket]

"This action is not allowed with this security level configuration". Run as admin, maybe?

[surreal whale]

mmm ya i dont know, windows thing i guess, never got that

[wanton rain]

Hello everyone, can I generate images by making requests to comfyui via API?

[maiden wigeon]

Hello everyone, does anyone know a way to change the folder where ComfyUI looks for the checkpoints I downloaded? I am running out of space on my current drive. Thanks

[steep marlin]

You can use either symlinks or I think you can add a launch flag or edit a config file somewhere to set the directory

A symlink is like a virtual shortcut from one folder/file to another. So if you symlink comfyuis checkpoints folder to some other folder with your models, when comfyui goes to access the folder, it gets virtually redirected to the other folder

oh and if you don't feel comfortable with that:
https://github.com/comfyanonymous/ComfyUI/blob/master/extra_model_paths.yaml.example

[radiant sundial]

Hello, I am working on creating a consistent character and was wondering if I should use IPAdapter FaceID or InstantID?

What I want is to be able to pass multiple images of the face of my character

I am not sure to really grasp the difference between the 2 of them so if anyone could enlighten me it would be amazing!

From what I understand InstantID do face + style consistency while IPAdapter Face is solely focused on face consistency?

[merry ermine]

both

[chilly kayak]

is this the right place for me to ask for help when encountering an error when queuing a prompt? I was hoping someone could break down what my error means so I can better understand it and fix it myself in the future.

[covert bough]

sure. what's the error?

[chilly kayak]

I actually googled it and figured it out on my own 😅

But as long as I know that this place is one where I can ask questions regarding errors then that'll help in the future

[tall prism]

How to fix it

[vital root]

Please help me to improve my workflow. This is to slow.
Best regards

[livid gust]

@tall prism I typically search for the model name in google, but you can try the Manager's Install Models button. After you install the model, click the Refresh button so Comfy repopulates all node dropdowns with new values.

@vital root With all those detailers, you have to wait it out. One thing you can do is to reduce that sample count. 60 on every detailer seems ridiculous. Try running 9. That can speed things up.

[vital root]

Thanks 😀

[shrewd bobcat]

Does anyone know why ComfyUI_StableAudio_Open would completely wreck my install? ive tried to install it twice and its pretty bad.

its either that one or comfyui-sound-lab idk tbh, should be reported or something

[covert bough]

https://x.com/fabianstelzer/status/1818305254909149621

fabian (@fabianstelzer) on X

introducing: ComfyAGI 🧙‍♂️😉

we've taught Claude to generate ComfyUI workflows, so you can now build comfy workflows just with prompts...

We're open sourcing the entire prompt chain for this (which is a glif too), so u can remix & adapt it

join our Discord & come hack with us:

Twitter · Jul 30, 2024, 3:18 PM

[surreal whale]

neat its gonna suffer from knowledge cutoff at times till retrained

hum or is it smarter than i think

[surreal whale]

dont know what thoes nodes are comfy supports stable audio open by default, dont need extras

[merry ermine]

RAG might help

with newer node docs

feed it the python files as well

this is speculative though I am not sure

[chrome flint]

anyone doing architectural rendering enhancements with artificial intelligence?

I don't know where to download comfyui prepare image and mask for inpaint, because by double clicking with the right mouse and typing it it doesn't appear, does anyone know?

[merry ermine]

brushnet with power paint nodes

[merry ermine]

anyone has idea about why comfyui update is stuck at Downloading opencv_python_headless-4.7.0.72-cp37-abi3-win_amd64.whl (38.1 MB) ?

hmm its progressing now, weird that it took so long for 38mb

i hate linux oriented frigging update process

[shrewd bobcat]

Oh thanks. I ended up moving from portable windows to install on WSL ubuntu. I think this was a little smoother too.

[long flame]

if stable difussion in my pc dont work, civitai helps

[wispy kelp]

We really should have a contets with the most complicated workflows, has anyone ever run out of space in comfyui, like no more room to add nodes?

[covert bough]

in case you haven't seen this site https://comfyworkflows.com/

[wispy kelp]

https://www.youtube.com/watch?v=rtSMGm9nSJk

this workflow would probably make top 10

[primal otter]

That looks REDONCULOUSLY complicated

Iiii just got comfyUI Up and running for the first tiiimee YAAAY! :D

[wispy kelp]

It could be simplifed probbaly a lot for same results...

[merry ermine]

yeah a lot of the youtuber workflows, or ones that go viral on civit or reddit, are like that

[wispy kelp]

ikr!

maybe they want to seem impressive

its like comfui bling

comfy flexing

Flex those noodles!

[merry ermine]

yeah massive workflows like this in comfy isn't a good idea

if a workflow gets this complex it should be done with python scripts really

[queen geode]

Hey guys, do you know which workflow I could use for Personal Image generator? Thanks

[wispy kelp]

what do you mean?

[queen geode]

So I'm new to Comfy ui and Stable Difussion, my goal is to play around with personalized photo profile (like Image to Image generation), based on what I learn we could use some workflow for these, do you know any?

[wispy kelp]

https://comfyanonymous.github.io/ComfyUI_examples/img2img/

🤷‍♀️

[queen geode]

Great, but what I mean is generating photo realistic from several photos

[livid gust]

Check out some tutorials on the IPAdapter node (Mateo's work on Github is a good place to start). It's good for photo reprocessing and nabbing the style of one photo to apply it to another. There are nodes that can read images from a folder as well.

[distant pendant]

anybody got flux working in comfy? according to the hf page

FLUX.1 [schnell] is also available in Comfy UI for local inference with a node-based workflow.

but I'm just getting key errors trying to load the checkpoint

[woven rivet]

Same

[distant pendant]

yuup

[woven rivet]

I'm surprised there's no example comfy workflow somewhere

[distant pendant]

oh, I got further by changing the extension from sft to safetensors

but now it's saying it can't detect the model type

[woven rivet]

I tried that at first

But if you look in the comfyui commit, they updated the file extensions to include .sft so I think that's on purpose

https://github.com/comfyanonymous/ComfyUI/commit/1589b58d3e29e44623a1f3f595917b98f2301c3e

[distant pendant]

oh huh, there is a commit for that, how did I miss it

[woven rivet]

But no workflow

Hmmm

Not the right clips but flux is a type in theer

*there

I wish I was smart with this stuff but I only know how to click and drag 😭

[distant pendant]

dual clip is nice but I can't really find a node to load the unet/transformer or whatever they use

maybe the checkpoint should come in separate pieces like sd3 did

[woven rivet]

Check in #🆕｜sd3 , someone figured it out 👍

[distant pendant]

ah I see now, thanks

time to download those encoders again

[wispy kelp]

flux!? another model??

[primal otter]

Whatching a video, I found that there is a differentiation between a CLIP text Encoder, and one SPECIALLY made for SDXL... whyyyy is this a thing? Comparing images there doesn't seem like much of a difference. (Though, I haven't done a billion experiments so there could be something i'm not seeing)

[wispy kelp]

hmm after my latets update the clear resources arrow is gone 😦

[covert bough]

wait till you find out there are 3 different encoders for sd3

[primal otter]

Oop-

Do all of them... serve a purpose? Like. Do they make a big difference?

[covert bough]

yes. that's why they are there. i can go into detail on what all of them do if you like. Dm me, it's a long discussion

[hardy surge]

Hi, I am running Flux but it unloads from VRAM after generation, filling all of 32gb RAM and making PC barely usable, is there a way to keep model partially loaded forever? I don't have --lowvram argument in .bat file but it is still runs in lowvram mode

[covert bough]

https://x.com/_akhaliq/status/1819385104394809502

AK (@_akhaliq) on X

Live Portrait Updates on Animals mode

https://t.co/QXwhypZrob

Twitter · Aug 2, 2024, 2:49 PM

[cursive ridge]

anyone have a flux workflow I can use? Apparently just plunking the model into an SD3 workflow doesn't work

[minor plaza]

Anyone may know what is happening here? The vae isn't selectable/showing up within ComfyUI.

Prompt outputs failed validation
VAELoader:
    - Value not in list: vae_name: 'ae.sft' not in []
DualCLIPLoader:
    - Value not in list: type: 'flux' not in ['sdxl', 'sd3']

It still can't recognize it even after installing the VAE directly within ComfyUI. 😦

[covert bough]

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

[covert bough]

did you follow the install instructions on that page for the models?

[minor plaza]

Yes? I've triple checked it too.

[covert bough]

you're missing the encoder - you can get that from the SD3 hugging face page and you need a VAE - you can probably use the same one you use for SD3

[minor plaza]

I'm not sure what to download for the encoder or what folder it would be placed in then.

And I've downloaded the VAE twice. It's not letting me select it in ComfyUI

[covert bough]

you need the t5xxl encoder. zoom in on the node on the left side of the workflow and see what it is that's listed there. yo uprobably ahve it if you have SD3 but it might be in a different directory. left click on it and see if you can find t5xxl in the list

[covert bough]

did you put it in the right folder?

[covert bough]

okay, put the ae.sft file into ComfyUI\models\vae and rename it to ae.safetensors

you have to rename sft files to safetensors

[minor plaza]

Ohhh duddeee. I did not know that! 😅

[covert bough]

make sure you rename the other flux files from sft to safetensors too

[minor plaza]

Yep. Now the last error:

Prompt outputs failed validation
DualCLIPLoader:
    - Value not in list: type: 'flux' not in ['sdxl', 'sd3']

I am still missing a file, yeah?

[covert bough]

restart comfy

[minor plaza]

The error is still present when I try to generate an image. It's not in this list, sorry:

[covert bough]

left click on each of those, and see if they are in the list of stuff you can pick from

those are your encoders

[minor plaza]

The only issue is coming from the type in the DualCLIPLoader, which is missing the option flux.

[covert bough]

you can download them from here https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders if you need to

[minor plaza]

I have them.

The latest ComfyUI now supports FLUX and allows the .sft extension to be used interchangeably with .safetensors. If your ComfyUI doesn't recognize the .sft extension, it means your version is outdated and needs to be updated.

But mine is updated lol

[covert bough]

mine is too. easier to not fight with it and just rename them

[spring jungle]

question is their a way to get avd to control start and stop of steps ?

[minor plaza]

Yeah, that's fine, but I dunno why I cant generate an image still.

[covert bough]

open an issue on comfy's github page

[dreamy smelt]

Just update your comfyui and change the CLIPLoader type to Flux

[minor plaza]

What is the latest version? This is as far as I could update to.

@covert bough Thank you for opening this bug report - https://github.com/comfyanonymous/ComfyUI/issues/4179

[dreamy smelt]

2024-08-01

[minor plaza]

My ComfyUI refuses to update then? :3

[flat dirge]

I updated with "D:\SD-ComfyUI\update\update_comfyui_and_python_dependencies.bat"

[minor plaza]

I do not have this. I am using a forked version of ComfyUI called ComfyUI-Zluda, since I have an AMD GPU. :c

[minor plaza]

Updated

[covert bough]

welcome. it's apparently hitting a lot of users

[merry ermine]

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

[small gate]

i’m getting this same error, my previous stable diffusion ui was just automatic1111. can you please share how you fixed it?

[fair tiger]

how do u get ur comfy to look like that

oh nvm lololo

[mystic plaza]

https://www.instagram.com/reel/C-JBb1ZAnha/?utm_source=ig_web_copy_link&igsh=MzRlODBiNWFlZA==

enigmatic_e

As promised, here is a breakdown of how I did the Deadpool animation I recently posted. #deadpool #wolverine #marvel

Likes

796

Instagram

[mystic plaza]

can i get some help on creating this workflow bros ?

[merry ermine]

it uses viggle so we cant

viggle is closed source

[mystic plaza]

i mean the next part of it where he upsclaes and fixes the animation using animatediff

[mystic plaza]

already did the animation on viggle

[merry ermine]

oh he uses runway and adobe after effects as well

[mystic plaza]

also comfyui

[broken spoke]

Guys

why is it like this

I used the 12gb vram workflow for flux

Do I need to use the schnell model?

[sly sundial]

You can use either https://huggingface.co/Kijai/flux-fp8/tree/main

[broken spoke]

Im using it inside Pinokio

but the dev version and the workflow to it was so slow as I need to switch to something else

[sly sundial]

The models are the same size if you use the fp8 versions.

workflow is the same for either

[broken spoke]

let me try

I have this models

I think the workflow is for the schnell model which I use because it only use 4 steps

[mystic plaza]

@sly sundial can you pls guide me on how to create the upscale work flow he is using in comfyui ?

[sly sundial]

No

[mystic plaza]

thanks it looks complex as hell tho but any alt to achieve the upscale or polish the animation a bit ?

[broken spoke]

why it can't load the schnell model

[wispy kelp]

There goes 50 GB of hard drive space.

[wispy kelp]

Update and RESTART comfyui - as in close everything and start it fresh.

and someone showed that upscalers work with no problem with flux..

[merry ermine]

he's got a patreon https://linktr.ee/enigmatic_e

if you sign up to their patreon they probably have the workflow there

[mystic plaza]

the workflow is paid i seen it

[broken spoke]

I restarted it already

[minor plaza]

I've done made the 50900 Ti with Flux Schnell

[vital root]

https://civitai.com/models/620166/elevate-your-creativity-with-my-flux-professional-workflow
This ist my workflow for flux. This model is amazing. Inclusive refine and upscale in krea.ai style. Enjoy. 🥰

[desert rapids]

Anyone know how to enable swarm option for a listen flag so I can connect to my swarm ui from another computer on the home network. Running portable

[desert rapids]

Can't get it to work, missing pretty much every node it seems like:

[fathom lake]

Did you try installing them using ComfyUI manager?

https://github.com/ltdrdata/ComfyUI-Manager

[covert bough]

comfy posted a response on that issue - you have to use the update script in the comfyUI/update folder, not the manager update function. you also have to update pytorch - and that forced me to update xformers and a lot of other files as well

[minor plaza]

It turns out I had some weird outdated forked version of ComfyUI-Zluda that someone in the #🤝｜tech-support channel put me onto. 😅

[covert bough]

that'll make it mad, too

[minor plaza]

So I git cloned the actual ComfyUI Zluda version and things work fine. 😛

[covert bough]

is it working now then?

[minor plaza]

Yeah, I had it updated and working fine here last night. My apologies. 😅

[covert bough]

🙂 i'ts fine. i'm still updating stuff. pip is throwing a fit

[fathom lake]

Anyways, I got this error when trying to run NerdyRodent's Reposer workflow:

!!! Exception during processing!!! 'NoneType' object has no attribute 'lower'
Traceback (most recent call last):
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 152, in recursive_execute
    output_data, output_ui = get_output_data(obj, input_data_all)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 82, in get_output_data
    return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 75, in map_node_over_list
    results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 657, in load_ipadapter_model
    return (ipadapter_model_loader(ipadapter_file),)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\utils.py", line 130, in ipadapter_model_loader
    model = comfy.utils.load_torch_file(file, safe_load=True)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\lenovo\Downloads\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 14, in load_torch_file
    if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
       ^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'lower'

_ _

I feel like it's probably something obvious, but I'm not sure what

[steep marlin]

press f5 and in any slots where models are loaded, manually update the selected model in the options

his model checkpoint name might have been different than yours

[fathom lake]

Yeah I'll try that real quick

I think I just have the ipadapter model in the wrong directory. The video showcasing the workflow said to put it in here, but it looks like ipadapter's changed (?) since the video came out. Where do I put it now?

[steep marlin]

you'll have to check his repo

the ipa author cubix or something like that

https://github.com/cubiq/ComfyUI_IPAdapter_plus

[merry ermine]

how much it tool to render ?

[broken spoke]

few second with the schnell model and 12gb vram but the output was bad anyway

[merry ermine]

i just saw comfyui for blender and looks intersting.

[fathom lake]

Okay I think IPAdapter's working just fine now, but now it's getting stuck on this

[merry ermine]

i hope it support direct animation rendering. i will try it tonight.

[fathom lake]

Nevermind, it's loading, but this is the current progress

Is this normal for running on CPU?

For reference, I have an Intel Core i7-6500U CPU @ 2.50GHZ

Not much but

[vocal grail]

I just tried your workflow and it worked. But after restarting comfyUI it doens't work anymore.

ComfyUI is not even starting

Do you know what could it be?

[surreal whale]

reinstall pytorch with cuda support, i dont know how with embedded https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#nvidia

[vocal grail]

Thanks it worked by updating COMFYUI

this one specifically

[fresh vault]

can the unet folder location be changed? (something in extra_model_paths.yaml perhaps?)

[vital root]

This is my first workflow I've created, and even though it works, it's all messy. I'd like someone more experienced than me to tidy it all up for me to have it nicely organized. I've made so many connections that I don't know how to arrange the nodes anymore. And if you could please add OpenPose to my workflow as well, that would be really fantastic. If someone helps me, I'll give a tip on Civitai. ❤️

[storm folio]

https://comfyanonymous.github.io/ComfyUI_examples/flux/#simple-to-use-fp8-checkpoint-version

[merry ermine]

its 17.2 GB do we have less then that ?

[sly sundial]

https://huggingface.co/Kijai/flux-fp8/tree/main

11.9GB

[merry ermine]

not just updating torch, updating nodes and extension akways lead to an error.

is there a solution for this ? is there a torrent link of comfyui full updated with all nodes?

[small gate]

dealing with the same issues, is beyond frustrating i just want to get flux running and everything is so bad and dont know where to turn to for directions

[distant pendant]

here's a node that just calls the flux hf space and returns an image for anybody with a potato

(might need to pip install gradio_client)

[wispy kelp]

So I am not exactly 100% sure how to run the magic clothing thing.

there is this lebif thing missing or something?

Birefnet?

[wispy kelp]

I guess not just windows updates are hell but also Comfyui updates too.

[covert bough]

if you update torch, you also have to update xformers AND all the files that numpy will then be incompatible with.

(and then cocktail peanut tells me yesterday flux doesn't really NEED torch updated after comfy said it did)

[merry ermine]

the issue is comfy extensions update torch then mess with the whole app.

i never updated torch.

[vital root]

Better than midjourney with y workflow? 😀

Can you make it more user-friendly for me? I'm afraid to touch any links. I'll leave it to you experts. Can you organize everything? I think the quality I've achieved even surpasses Midjourney.

[merry ermine]

do you guys have full comfyui as torrent or direct link?

or only via github

[vital root]

This is my workflow

If you can reorganize everything neatly, you'll make me very happy. There are too many spaghetti around.🤣

I created so many connections that luckily I managed to get what I wanted: an upscale that usually requires paid services like krea.ai or magnific.ai. But I don't even know how I did it. Surely someone more skilled than me will be able to organize my entire workflow.

sorry for my english I use google translate🤣

[vocal grail]

Yeah I just grabbed your workflow and bypassed most of it though. I am just trying FLUX without any img2img or upscaling, just to try it out. I didn't find the simple workflow for that so yours is what I found.

Anyway, is there a FLUX discord? I don't want to spam FLUX content into stable diffusion discord as I don't know it is the right place

[vital root]

you're right, but since I'm using comfyui I think this is the right place to talk about it. Then I don't know for sure

[lavish stone]

anyone knows what this is doing, i feel like it changes the image in each step

[merry ermine]

TBH constantly having to mess with python, torch, xformers and numpy versions is a thing in every area of maching learning

cos I use a wide range of models I have dozens of versions of this stuff installed in different areas

very confusing

[fathom lake]

What can I do about this?

[merry ermine]

that looks like it is scheduling the shift to linearly (or log-ly) go from one shift to the other

over the course of the steps

just guessing from the node wording though pls verify

[steep marlin]

It does that under the hood with or without the node. Those are the default values except for the resolution values you changed. The node just lets you tweak them if you want

[merry ermine]

what I am not understanding yet is what makes this different to SD3 shift

and does using adaptive solver nullify this anyway?

[steep marlin]

they both use the same sampling method: FlowMatchEulerDiscreteScheduler

but it's likely due to how flux does a few things differently in the model

like with embed pooling and whatnot

[merry ermine]

ok thanks

[steep marlin]

i'd have to compare the two codes to really tell you for sure

[merry ermine]

don't worry I will start looking into it later next week

I'm just doing some random SDXL experiment for a week first

I'm trying to make a custom node for SDXL that tweaks rescaleCFG/tonemap/thresholding to my liking

[steep marlin]

check into some of Extraltodeus's nodes. he makes the automaticcfg node, but he's got some other new ones like https://github.com/Extraltodeus/Skimmed_CFG and https://github.com/Extraltodeus/pre_cfg_comfy_nodes_for_ComfyUI

the skimmed one looks pretty neat, but i havent tested it yet

[merry ermine]

oh he finished the skimmed one wow
I was reading his comments on reddit about it while he was making it

yeah I love Extraltodeus's stuff
I just wish he was easier to understand

[steep marlin]

yeah, he words things pretty vaguely, but just trial and error the hell out of things and see what sticks

[merry ermine]

his node vector sculptor is really really fun on certain images

you can sort of use it to exaggerate concepts

so like you have a cocktail bar with colourful bottles
and then you boost the vector sculptor and then suddenly the bottles are all lined up in the foreground

or you have a car in a jungle, in a medium shot
and then you boost the vector sculptor and suddenly the car is filling the frame

[steep marlin]

I'll have to check it out. I hear it's good for breaking up the "sameness" problems that can happen. Like people having the same faces and shit

[primal otter]

Edit: Nvm I could be dumb. Experimetning is good

[merry ermine]

CADS is the best for lowering sameness

but CADS can get a little too crazy for some uses LOL

[wispy kelp]

so whats this birefnet

[merry ermine]

its a seg model

[wispy kelp]

so that new magicclothing all in one node is buggy as hell

[merry ermine]

ye I tend to avoid all in one nodes

[wispy kelp]

sadly i cant cobble together the face ip adapter into a workflow with magic clothing and also idk is it even better than just a plain adapter for clothing...

i guess your mileage may vary.. also onyl for sd15? tsk tsk

the quest for consistency continues...

[tall prism]

Hi where of file location to change IP network of comfyui?

[crimson root]

oh here it is hahaha!!! hey guys !!!!!

i have some questions and some issues im trying to solve is anybody interested in being like my teacher hahahhahaa

[flat dirge]

Try these args.

• --listen [IP]

  • Type: str
  • Default: "127.0.0.1"
  • Description: Specify the IP address to listen on. If provided without an argument, it defaults to "0.0.0.0" (listens on all interfaces)

• --port

  • Type: int
  • Default: 8188
  • Description: Set the port to listen on.

[covert bough]

@thorny lily how about adding a way in the comfy config files to point to a different directory other than comfyUI/output for output files?

[flat dirge]

I think that's also something you can change with an arg

[covert bough]

and a symlink, but it would be real nice to just have it in the extra paths file

[surreal whale]

Its in here https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cli_args.py#L36-L117

[covert bough]

that's not the extra path's file, so not real convient for most people

[sly sundial]

It's just as easy to do this with the save image node, probably easier for some users.

[lone cape]

Hi, how to solve this error:

ERROR: Could not detect model type of: H:\ComfyUI\models\checkpoints\Flux\flux1-dev-fp8.safetensors

when I trying to use safetensor as checkpoint instead of unet?

[merry ermine]

that file must be in clip folder i geuss

[merry ermine]

when i load a workflow it says to me that some node types were not found

while i actually i installed every updates of it

[flat dirge]

It could be a situation of the custom node updating and either removing or completely changing a node. If that happens older workflows will show a missing node. In that case you have to replace the missing node in the graph with the updated version, or find some other replacement.

[merry ermine]

I actually instaled it from comfyui manager

but strangely when i restart it, even though it's updated, it still has the problem

"Cannot execute because a node is missing the class_type property.: Node ID '#107'"
That's what they tell me when i can't launch it

[livid gust]

That just means you missed a node, or the missing node is not part of the node registry (like recent custom nodes). If you can't find it in your workflow, try moving nodes a round. I've had missing nodes hidden under other working nodes, leaving me perplexed why I can't see them.

[long lodge]

hello all. Relative new to comfy. Can you set somehow in the node (example load model) were node has to load from (directory etc)?

[merry ermine]

is there a civitai browser inside comfyui?

to import generated data directly

[frigid shale]

Can some one help me with Lora training in ComfyUI? I'm getting this error while executing lora training. I'm using SDXL checkpoint

[tranquil scroll]

Trying to bite the bullet and learn Comfy UI, but when I unpack the file it's missing the python_embeded folder, so the installation fails.

Why is that folder missing?

[merry ermine]

there is no python installed , not sur why its not using the local venv

[tranquil scroll]

except that it's definitely installed 😄

[distant pendant]

you probably downloaded source code instead of the portable build

[tranquil scroll]

that sounds right

[distant pendant]

just scroll down in releases and you'll see 'ComfyUI Standalone Portable Windows Build (For NVIDIA or CPU only)'

that's the one you want (if you don't want to use your own python)

[tranquil scroll]

yep that's what happened

thank you 🙏

[deep rapids]

stupid question: I need this to try Flux,right?

I found some instructions: it included Comfyui install.... I also am wondering if my 1660 ti can run it.... I think I have to account for the low vram...

[thorny lily]

i've been suggesting to the rest of the comfy org team that comfy's core settings should have an actual coherent interface like Swarm's settings, vs. just random cli args and files wherever

[covert bough]

i will back you on this 100% - random files are fine if you're actually a programmer - but for the less technical people, a configurator is much preferred - and comfy is now mainstream

[merry ermine]

my biggest issue with comfy ecosystem as a whole is that there is not always a lot of comments in the code, or documentation outside of the code
so for example I will often find a custom node with a new sampler with some really long name like DPM-restart-improved-4m-SDE-CFG++
and it can be tough to work out exactly what it is doing

will the coming changes in comfyui make it a standalone app ? no webui

they are changing the frontend.

native ui would be amazing yes

that would be a large undertaking though for something the size of comfy

[primal otter]

Any suggestions on a good Node pack with a Multiple Lora Loader? Or maybe a LoRA string converter? A Lora_Stack to clip and model node?

[primal otter]

I have not. Let me try that one.. I have seen the Impact and Inspire pack, but never tried them yet. They look good and if they were from the same guy who make the manager it's PROBABLY good... Imma go try them both!

[obtuse latch]

Anyone have a simple workflow that will output 1024x2048 properly?

[merry ermine]

Recurrent Attention U-Net and Multi-Scale Windowed Multi-Head Self-Attention

[primal otter]

Having installed the module, I do see a couple that I could use... is there a specific node you'd have in mind?

[merry ermine]

you can get them here ```https://github.com/blepping/comfyui_jankhidiffusion````

[primal otter]

Sweet! That helps! I shall check him out

[merry ermine]

PatchModelAddDownscale with res-adapter also works
https://github.com/jiaxiangc/ComfyUI-ResAdapter

[primal otter]

TYVM!!

[obtuse latch]

That doesn't give me control of the uscale dimesnions

currently what I'm dealing with:

[merry ermine]

it does
just pick the latent size that you want

[obtuse latch]

The model was trained on 1024.

[merry ermine]

yes that is true

but that doesn't mean that you, the end user, have to use 1024

[obtuse latch]

I need to create a image at 1024, then upscale.

[merry ermine]

why do you have to create the image at 1024 and upscale instead of just creating the image at 2048?

[obtuse latch]

Becase, the image was trained on 1024, not 2048.

[merry ermine]

but you don't have to use the model at the resolution it was trained at

[obtuse latch]

If I don't want duplicate objects in the image, yes I do. Re:

[merry ermine]

the idea behind the nodes I suggested is that they stop you from having duplicate objects

[obtuse latch]

I don't understand how that works, seeing how you should prompt at the dimensions a model was trained on.

[flat dirge]

There are a variety of methods that have been created to avoid the issues that appear when generating images outside the range of the model's trained resolutions. Kohya Deep Shrink, HiDifusion, etc.

[merry ermine]

you can read the original papers if you want

[obtuse latch]

Since there's no input into the initial dimenision size, I don't see how they would auto know.

[merry ermine]

there is an input for the initial dimension size
that's what the latent is

[obtuse latch]

Just tried, it doesn't work, even their example is 💩, but thanks.

If anyone can reccomend a simple fast upscaler, just need 1024x2048. .

currently, (unless someone can suggest better)

[merry ermine]

it does work

there must have been some issue

these techniques are over a year old at this point

[obtuse latch]

Their example doesn't work, so... 😄

[merry ermine]

I'm just going to assume you are trolling at this point

[obtuse latch]

Yes I'm the one trolling: https://github.com/blepping/comfyui_jankhidiffusion/issues

[compact sluice]

im confusion. i want use flux schnell/dev model but with this is really new to me. can anyone help me?

[obtuse latch]

You need a checkpoint.

[compact sluice]

how can i get it?

[obtuse latch]

Google

[compact sluice]

some youtuber said patreon patreon

[obtuse latch]

Watch a different YT

[flat dirge]

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

[compact sluice]

i putted in comfyui

50-60gb was it

[flat dirge]

your using a repo with the word jank in the name. just use the default comfy node for this. PatchModelAddDownscale aka Kohya Deep Shrink

[compact sluice]

how can i put. i dont know how to do this?

[flat dirge]

Go to this website here and drag the image into ComfyUI. The image contains the workflow. The page also explains what to download and where to put the file.

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

[compact sluice]

oh

[obtuse latch]

Yeah the lad I was talking to doesn't seem to understand what upscale artifacts are and assums the "gems" are part of a good prompt, and not noise. I'll check it out, thanks.

[merry ermine]

I literally suggested PatchModelAddDownscale to you lol

[compact sluice]

i putted clips in the folder

[merry ermine]

that's how I know you are trolling
you take the suggestion for the same node when he gives it but not when I do

[compact sluice]

which then

ahh i see

[obtuse latch]

Mate you literally suggested jank, I checked it out then when I pointed out that the noise in their own example workflow (resolves to gems) was broken you didn't understand.

This is upscale noise, it's broken.

[merry ermine]

I suggested the same node he did PatchModelAddDownscale
and you accepted it when he said it but not when I said it

[obtuse latch]

Because your track record (even though I'm thankful for you trying) hasn't helped, I tried your suggestion, I'll try the other.

[primal otter]

Ngl it looks kinda cool though with the infocus image

[compact sluice]

???

[merry ermine]

hi-diffusion actually does work

I've used it on SD1.5

if you search on reddit or youtube you can find dozens of people using it

[compact sluice]

this is annoying....

for the checkpoint

[merry ermine]

flux dev is huge yeah

[compact sluice]

well i tried this:

[merry ermine]

not sure what those are

[compact sluice]

HUH???

[flat dirge]

14 min

[compact sluice]

with rtx 3070 is really slow?

[merry ermine]

yeah rtx 3070 will be slow

[compact sluice]

no option for faster mode?

[flat dirge]

Flux schnell is faster

[merry ermine]

generally flux dev is for 24gb vram
but with optimisations like quantisation and some offloading it can work on lower vram hardware

[compact sluice]

i just want make only anime pictures

[merry ermine]

it depends because some people are okay with waiting

[compact sluice]

are both the same or different?

[merry ermine]

dev and schnell are very different

schnell is a lot faster

[compact sluice]

no this is not what i mean

the name of data are same

[merry ermine]

looks like the name of the dev safetensors files are the same yeah

unless I misread it

[compact sluice]

im downloading schnell with 17gb but i have another in unet folder. why are both different in gb?

[merry ermine]

is one a quant and the other not a quant?

[compact sluice]

i should download this for checkpoints

[flat dirge]

The 17 GB one from Comfy-Org huggingface is an all in one, checkpoint file version,

[merry ermine]

oh its got T5 included I guess?

[flat dirge]

it contains model, vae, clip_l, and t5xxl all in one file

[merry ermine]

ah okay yeah

[compact sluice]

so i wasted my time to download for other folder?

[compact sluice]

wtf is this with flux schnell

any idea why this?

[obtuse latch]

I'm getting very simmilar faces, anyone know how to get more different ones with the prompt?

[obtuse latch]

any clue why I'm getting these facial artifacts?

[merry ermine]

use auto re seed, if u detailed face in prompt it will not change alot

[boreal stirrup]

I gave flux a go on my 6600 xt with zluda instant blue screen when I hit Queue

[steep marlin]

You probably only have 16gb system ram and a page file that isn't big enough

The model will need around 30gb of system ram or system ram+page file

[boreal stirrup]

Man I do have 16gb

[steep marlin]

Jack your page file size up in the windows options, set it to like 16gb. But if it's on a mechanical HDD, forget about it

[boreal stirrup]

I had nvme

*have

[steep marlin]

Yeah you can use that, just be aware that the constant writing will wear the drive down faster

[boreal stirrup]

Sadge

[steep marlin]

Pick up a matching pair of ram or upgrade. Rams cheap if you aren't getting bleeding edge versions

[boreal stirrup]

Give me a good ddr4 kit real quick

If I can get 64gb of ram for relatively cheap send me a link if possible please

[steep marlin]

what motherboard is in the computer if you know off the top of your head? it will be something like b660 or something

[boreal stirrup]

I have 5600 it's a asus tuf b450m gaming plus (I got it back when ryzen 3000 launched)

[steep marlin]

so probably some ddr4 3200 would do

on amazon, just search for ddr4 3200 32gb or something. should run about 50 USD for 32. you might find cheaper deals on slightly higher versions

[merry ermine]

there is a weird thing where

ram speed matters much more for AMD than for Intel

[boreal stirrup]

Their CPUs loves speed

[steep marlin]

well i can almost promise that the difference between something like 2666 vs 3200 isn't even going to make a 10% difference in it/s speeds (talking about ddr4 ram speeds)

[boreal stirrup]

Can I assume for ssd having one that has got Dram matters also?

I may get refurb 3090 for £550 GBP or so with 2 year warranty

[steep marlin]

that would help, but realistically, you still have to have more system ram. 32gb really is a minimum for these newer LLM based diffusers

they need to be able to offload the models like t5 to sysram

[boreal stirrup]

Unfortunately I wish I could get 4090 for £1.8k but I don't have 1.8k chilling for a GPU 😥

So ig refurb 3090 will do

[steep marlin]

yeah, the 24gb vram is really good

[merry ermine]

refurb 3090 is easily the best option

[steep marlin]

just make sure the warranty is solid

if its a refurb

[merry ermine]

I don't want to sound like I am always pushing cloud but on cloud you can get 24-40GB vram for about £0.20 per hour
just to let you know

[boreal stirrup]

I do have vast but importing models take like 3 light years

[merry ermine]

are you in US?

[boreal stirrup]

UK

[merry ermine]

sometimes Sweden data center servers have faster internet

[steep marlin]

yeah it really depends on how much you plan on using it. if you do diffusion for maybe an hour or two a week, cloud options are likely better in the long run. but if you also game and stuff like that, the card is likely a better value

[merry ermine]

ah yeah that's the thing I don't game
if you game then its different

[steep marlin]

you'd need 2750 hours of cloud diffusion @£0.2 per hour for them to break even

vs the £550 cost for the 3090

so that would be ~4h of diffusion per day for two years straight, if you would normally have gotten two years out of the card before upgrading again

[merry ermine]

hmm yeah its tricky

[merry ermine]

bare in mind if you are willing to go for interruptible, non-data-center, non-verified GPUs
then the price is as low as $0.035/hr

but if you go that route its important to use anonymous email and payment, and never show your info on screen

sadly this won't work forever, as models get larger the download times will become too big for this method

[boreal stirrup]

I think I'll get 3090 because then I can also game on it and not worry and not keep paying monthly

[merry ermine]

okay nice