🧣｜comfy-ui

[hard marsh]

Now if I wanted to dabble in linux, it'd probably be better for resource management.

But linux seems like a hot mess so I'm leaving that can of worms alone.

[cursive ridge]

have you tried mage.space and civitae for fun? 🙂

I hope SD3 works well with 8gb vram/16gb ram!

[hard marsh]

I don't have money, so uh...no. 😆

I dunno if either of those are free, but I feel like they aren't, so.

[cursive ridge]

there's a free tier on both

[hard marsh]

Ahhh.

Well, I like system run better.... It feels better than being limited by 10 free gens a day (if that's the case, not sure though)

[cursive ridge]

mage free is unlimited, but you only get the simple mode, not the add any model or lora you want mode.
Civitae is limited by how many points you have accrued I think.
Local is WAY better set a comfy randomizer workflow with 200 different variations, go do errands 😄

[silver nebula]

How do you run Mobius in ComfyUI? 🙂

[hard marsh]

Why is my inpainting removing the object?

I can't share the image here as it's mild nsfw, but it's...removing the entire hand i'm trying to fix.

Like sometimes it adds in an attempt of what i'm looking for, and other times it just completely erases what i'm trying to fix.

Even a1111 is doing it? What the fuck?

I might've been using the wrong nodes...agh.

[lone bolt]

Quick PSA if nobody brought it up already. If you are using those nodes here, there is a high chance that you might have been hacked. https://github.com/AppleBotzz/ComfyUI_LLMVISION

[distant pendant]

I dread to think how many people randomly installed this and possibly other nodes (according to the readme) without even looking at the sketchy wheels in the requirements

[verbal sky]

Brand new Comfy training video is available to all! (Layer Composition, Photopea, Black Cut-out Effect, Sliders and Re-route). Enjoy! https://youtu.be/2I1O-NWbk6A

[unborn dune]

hey guys, how do i set my workflow in comfy to run multiple times like in A1111 (i think it was called Batch Count)

in a1111 it could even be set to run forever. anything like that in comfy?

[sly sundial]

There's a little square tick box for extra settings in the toolbar. There are options in there for batches.

[unborn dune]

yay ty!

[cursive ridge]

You can run comfy on your phone now!
It works 😄 Though the models that come with it kinda suck, but it still works. Perhaps I'll see how to add all my fave models to it when I have some free time.
https://auth.fal.ai/u/login?state=hKFo2SBoUi1qNnRHRW9XdmhqSnJGVHpuQnZWOXRlazdhS0o1TqFur3VuaXZlcnNhbC1sb2dpbqN0aWTZIFQxSElGdURVWE4yOTg3YUZocHY3dFZZTlRnNV96WHpDo2NpZNkgdmpmRndtS1llV1FuT0RIa2puQnpIeUZNcDJpOGxzazU

[hard marsh]

I'm kinda lost as to why my workflows are suddenly super slow...

i'm hoping it's not fooocus.

like it's... a whole 2 seconds slower.

instead of 1-2s/it it's now 3-4s/it

...could it be the model?

it's not the model. i dont think it's the model.

what the hell is going on...

...maybe i'm just trippin out.

[sly sundial]

Well...you are talking to yourself as well 🤣

[hard marsh]

i like to update people when im having a problem

[sly sundial]

Very thoughtful, thank you.

[hard marsh]

You never know who may waste their time just in case i fix my problem 😆

[cursive ridge]

Some models definitely take longer than others. A lot longer.

[livid gust]

Comfy runs slow when you exceed memory.

[tame valley]

https://www.youtube.com/shorts/KgLVO0Ms5lc

[obtuse raptor]

https://www.reddit.com/r/comfyui/comments/1dbls5n/psa_if_youve_used_the_comfyui_llmvision_node_from/

[mystic plaza]

@sly sundial @hard marsh is there a wayt to add a bg image if there is a croma screen behind the image in ai ?

[sly sundial]

Yes. You can do it with or without a chroma screen, just need to be able to mask the subject.

[mystic plaza]

can you pls guide me on this bro or send a yt video will checl it out ?

[sly sundial]

Use this node to remove the background and then blend that with your background image.

[mystic plaza]

cause this the image i have and there intricate details as well in the image around the leaves portion and i mainly need to generate an outdoor image after that

[sly sundial]

Something like this

[sly sundial]

With this, you could colorize the image with a depth map and change it that way.

For example

[mystic plaza]

but the leaves are going to be tricky i feel to get done

[mystic plaza]

thanks for the idea will try it out

[sly sundial]

Does all of the outside view need to be there?

Could replace it all

[mystic plaza]

yes

[sly sundial]

It'll be harder to mask them, but possible

[mystic plaza]

will see what i can do asking one of my friends to do a keylight mask and then use that png and generate a outdoor image

[sly sundial]

Do you want the workflow for that anyway?

[mystic plaza]

yes pls

[sly sundial]

Embedded in this

[mystic plaza]

i am not getting your latent empty image bro @sly sundial

[sly sundial]

You can drag and drop that image into ComfyUI to get the workflow.

[mystic plaza]

so will this mask work ?

[sly sundial]

Yes, but it will need to be reversed because it changes the white areas

There is a disabled node in the workflow to do that

The mask still causes some slight blue areas though

[mystic plaza]

for some reason it aint working for me

[sly sundial]

Just need to load the mask into that invert image node

[silver nebula]

You want intricate details? Use this in your prompt ... microscopic ::1 closeup ::1 backlit jewel tiffany ::3 intricate needlepoint closeup murmuration ::3 photographic underwater lighting,

[mystic plaza]

this is what i am getting tho ?

[sly sundial]

You're not loading the original image

Where the mask is loading should be the original image

[mystic plaza]

but how to add this mask tho ?

[mystic plaza]

can you pls send the workflow once i am getting confused here

[sly sundial]

I already have

[sly sundial]

If you're using your own mask, this is all you need

Workflow is embedded in the image

[mystic plaza]

thanks finally understood how this works

[mystic plaza]

by increasing the steps will it give different results ?

[sly sundial]

That's only creating the background image, so it will change that image only.

[mystic plaza]

sometimes i feel theres a lot of things to learn of comfyui but its way too complex i feel

i guess the more i use a work the more i learn and understand how it works

[sly sundial]

It's fairly simple really. Follow the lines from the ksampler (where steps are set) and you'll see it leads to a preview image of the background, which is all that it is doing.

[mystic plaza]

i meant the 3d passes and all that workflows which people use or the stable diffusion video

[mystic plaza]

so just wondering but where do the pony models go into in comfyui our check points or ?

[sly sundial]

Yes, same as any other checkpoint

[mystic plaza]

i see and do you have some exprience with comfyui for 3d stuff or renders ?

[sly sundial]

No

[cobalt halo]

Is it planned to add stabel audio open Support?

[bleak crypt]

There already is a node for stable audio open that you can use.

I've used it, at least. It seemed to work completely fine.

https://github.com/lks-ai/ComfyUI-StableAudioSampler

[cobalt halo]

nice but i wonder if they plan to make a offical implementain

[hardy tree]

oh this would be cool 😮

[fluid pumice]

To all Comfy users from a Forge user with luv' https://youtu.be/O3NzGSHjj4s

[wispy kelp]

Rears lubricated?

I have 0 expecations. SO I am OK.

well i hope it's about as good as sdxl with sigma T5 or somthing

actually i expect it to be somwhat better than Cascade.

[stoic gulch]

what plugins do you guys suggest, particularly that might be compatible with the upcomming release

[wispy kelp]

probably nothign works with it

you can just load it and sample it and just about it

[stoic gulch]

any plugins/nodes work with the audio model?

[sharp raptor]

Hey I installed this from the comfy UI manager and now I cant open Comfy anymore. The cmd window shows for a second and then disapears

What is the issue?

[stoic gulch]

lol u answered my question ty

[sharp raptor]

Now tested to delete it from the custom nodes folder and comfy still wont load. Was using it without issues prior to installing the StableAudioSampler addon

[cursive ridge]

help? I have no idea how to connect these 😦

[hardy tree]

according to comfy, you dont even need all that, and besides, just wait for some workflow :3

well... workflow/instructions 🙂

[cursive ridge]

Thank you 🙂

[sharp raptor]

Do you know what I can do to be able to run comfyUI again?

[hardy tree]

again?

you mean it stopped working?

[sharp raptor]

My comfyUI was running properly before installing a addon from the manager

[hardy tree]

if you know the addon, just disable or uninstall it

[sharp raptor]

I did remove it from the folder

issue remained

could it be some of the requirements that it installed?

The addon is StableAudioSampler

[hardy tree]

yea python dependency hell is sadly a thing that exists

one way to fix it in your case is to just get a fresh copy of comfy and then just transfer over the models, loras etc to the working copy

[sly sundial]

I'm not waiting...at all

[sharp raptor]

ok will do this

[hardy tree]

or maybe just update within comfy, the update button

but not sure if it will fix it

[cursive ridge]

Thank you very much 🙂

[sharp raptor]

I cant open comfy

can test the new_updater.py file within the comfy folder

[sly sundial]

Where did you get that top right node?

[hardy tree]

@sharp raptor welp... then do what i said earlier... cause you seem to be fucked... or maybe if you have git then git pull to get latest files, but idk that wont work i think in your case

[sly sundial]

"SD3 select latent resolution"

[sharp raptor]

yeah that is the angle

[cursive ridge]

What is that awesome looking (orange) exteded image save called? 🙂

[cursive ridge]

I just searched for SD3 in comfy manager install nodes lol

[sly sundial]

It literally tells you all the details on it 😄
I chose my own colour

[cursive ridge]

I think I"m blinded for the next hour lol

[sly sundial]

Workflow is embedded, if anyone wants it.

[sharp raptor]

@hardy tree Dang I cant even start comfy in the fresh git clone in other folder now

[hardy tree]

blinded by love? :3

[sharp raptor]

here

[hardy tree]

huh

[sly sundial]

I get that a lot 🤣

[hardy tree]

baby dont hurt me

:3

[hardy tree]

dont be scared, but i think you entered the twilight zone 🙂

i actually need some sleep, hopefully you figure it out ❤️

[sharp raptor]

thank you, goodnight 😄

[sly sundial]

What does this command show? nvcc --version

[sly sundial]

Are you using windows portable version?

[sharp raptor]

no cloned the standard one

[cursive ridge]

How can anyone not like comfy, when you can do this?! 🙂

[sly sundial]

Run pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu123

Make sure you have latest Nvidia driver

[sharp raptor]

ok, ill check

[sharp raptor]

Have latest drivers and did that pip install. Still get "AssertionError: Torch not compiled with CUDA enabled"

[sly sundial]

I'd highly recommend using the portable version 😄
I used to use the cloned repo, but have had much less problems since changing.

[sharp raptor]

Okay try portable

Seems problem is with comfy since I tested to start Fooocus UI and that worked

[sly sundial]

You could run this in python to see what you get back

import torch
print("CUDA available:", torch.cuda.is_available())
print("CUDA version:", torch.version.cuda)
print("PyTorch version:", torch.__version__)

[upbeat needle]

Hi I am trying to run PixArt-Σ I am getting error messages has anyone able to help. Thank you.

[sly sundial]

Isn't there a sigma model choice in the loader?

I think your choice there is the problem.

Same in the resolution select

[upbeat needle]

I have
PixArt-Sigma-XL-2-1024-MS.pth -in >models>Stable-diffusion
t5-v1_1-xxl.safetensors -in >models>T5
diffusion_pytorch_model.bin -inmodels>VAE

[sly sundial]

Save it in a file (eg. cuda.py), run python cuda.py

[sly sundial]

In the dropdown menus

[upbeat needle]

Thank you I will try

[upbeat needle]

Yes model is there I can select

[sly sundial]

Change these to sigma models

[sharp raptor]

Says Cuda False but I can run fooocus and also use cycles render engine with CUDA set in blender

[sly sundial]

That'll be your problem then

Your system doesn't recognise it being installed

ComfyUI uses your system variables, when you use the cloned git version.

Portable has it's own setup, so is easier and you don't need to worry about stuff like this.

[sharp raptor]

Should I download this?

[sly sundial]

I don't know if there are any version limitations, but that's not what portable uses. If you're not sure what you're doing, then I'd still say use the portable version.

You can drag your models directory into that and use it straight away.

[sharp raptor]

Thank you sir

[cursive ridge]

What did I do wrong/not do yet? Or is it just my impatience and SD3 hasn't actually dropped yet? 😄

[sly sundial]

Looks like a problem with that node's code

[cursive ridge]

Ty, will reinstall

[sly sundial]

No, it's the code

https://github.com/GavChap/ComfyUI-SD3LatentSelectRes/issues/1

I just opened that 😄

[cursive ridge]

Thank you for that

but the workflow worked just fine for you right?

[sly sundial]

No, but I didn't test with no SD3 yet. I just have and get same error.

It can be fixed by adding a line to import comfy, but then it errors with:

!!! Exception during processing!!! SD3LatentSelectRes.return_res() got an unexpected keyword argument 'batch_size'
Traceback (most recent call last):
  File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 151, in recursive_execute
    output_data, output_ui = get_output_data(obj, input_data_all)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 81, in get_output_data
    return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in map_node_over_list
    results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: SD3LatentSelectRes.return_res() got an unexpected keyword argument 'batch_size'

[cursive ridge]

Hopefully it will be updated by the time SD3 comes out

I'm still amazed at that fancy image save node! 300dpi option ❤️

[sly sundial]

I've fixed it and created a Pull Request in the repo.

[upbeat needle]

I have made some progress no error messages but I got only some noise

[sly sundial]

Wrong VAE? https://huggingface.co/PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers/tree/main/vae

You're better off running the T5 model in CPU, not GPU

[silver nebula]

Go to Pytorch.org and d/load and install wheel from there

[upbeat needle]

Thank you I will try👍

[wispy kelp]

I make my nodes red just to freak myself out

custom red that is like a misisng node color

[upbeat needle]

I downloaded the vae this time I got black noise I think I my models are wrong. my check point is "PixArt-Sigma-XL-2-1024-MS.pth " your one is safe tensors" my T5 model name also different "t5-v1_1-xxl.safetensors" Could you sent me correct models links. Thank you

[dense mica]

Guys is there a defacto best way to do seperate character composition?

such as if im trying to make an image with two characters each wearing a seperate outfit or having distinct different features how am i to try to keep this consistant across different seeds without pot luck

ive seen things such as multi area conditioning or regional prompt but what do you guys do

[sly sundial]

Have you tried searching Huggingface?

https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/transformer

[cursive ridge]

Wow, ty so much 🙂

[cursive ridge]

how did you guys get it to work? 😦
I got this

[sly sundial]

Some context?

[cursive ridge]

That same workflow (that you posted above). I updated everything via manager....

[sly sundial]

The Pixart-Sigma one, or the SD3 one?

Can you post your workflow?

[cursive ridge]

Though I didn't use the proper vae yet (can't find it on gimp or via comfy manager), but the other person at least got at an output with a random vae lol

[sly sundial]

This won't work yet

[cursive ridge]

Needs the proper vae to do anything?

[sly sundial]

1. The latent select node is broken.
2. The text encoders will be looking for the SD3 CLIP.

[spring jungle]

is sd3 github out yet xD

🤔

[sly sundial]

wtf is that? 🤣

[spring jungle]

o im in wrong tab

xD

sd3 come out today xD

[sly sundial]

It doesn't require any git repo

It'll be on Huggingface as a model.

[spring jungle]

oh must of mixed them up lol

[upbeat needle]

Thank you for your help, Finally worked I replaced all models from HuggingFace. Here it is my first image.

[sly sundial]

https://tenor.com/view/despicable-me-minion-minions-clap-clapping-gif-561643868428822739

[upbeat needle]

🙂

[unborn dune]

hey folks. i tried loading the example comfyui workflow from the sd3 hugging face folder and see im missing nodes

but i ddnt see them in the manager

(ive been using comfy for like a week and love it but still rly fresh w/ it)

[merry ermine]

does anyone have some good workflows + models to use for generating unique images with complex contexts?

[spice wharf]

same here!

[merry ermine]

is SD3 a new model?

[sly sundial]

That latent select node doesn't work

[spice wharf]

ah... i was stupid...

[merry ermine]

where do I find a guide/tutorial on setting up SD3

[spice wharf]

You need to update your comfyui. Dont forget to refresh the browser (that was my problem)

[unborn dune]

oh lol ty

[hexed pumice]

Trying out the workflow for SD3, I replaced the CLIP loader and trying to use the model with clip included, but I got this error, any idea?

[craggy basin]

I have the same issue as above. Missing Nodes. What am I supposed to do to resolve this?

[hexed pumice]

Update your comfyui

[merry ermine]

anyone have a guide page to set up SD3?

[craggy basin]

This was the latest build

Ok, will try the nightly build

[hexed pumice]

This is the one I have

ComfyUI Revision: 2244 [c8b5e08d] | Released on '2024-06-12'

[cold mist]

have an error with SD3, default workflow

[storm folio]

https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[craggy basin]

C:\Users\silve\Applications\new_ComfyUI\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Traceback (most recent call last):
File "C:\Users\silve\Applications\new_ComfyUI\ComfyUI_windows_portable\ComfyUI\main.py", line 73, in <module>
import comfy.utils
File "C:\Users\silve\Applications\new_ComfyUI\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 1, in <module>
import torch
File "C:\Users\silve\Applications\new_ComfyUI\ComfyUI_windows_portable\python_embeded\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\silve\Applications\new_ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\lib\fbgemm.dll" or one of its dependencies.

This is with the nightly build

I don't have Python installed separately for the record.

[storm folio]

that's why it's the "nightly" try the stable instead

[craggy basin]

I did and got the missing modules error mentioned above

[storm folio]

that means you have to update with the update/update_comfyui.bat

[craggy basin]

ok let me try that. Someone said update, but I had no idea what that meant, so updated by getting the nightly builds

(there was no UI update button or the like)

Ok, it worked. And is now telling me that my sweet innocent self was deluded in thinking that 10gb file was enough. lol. Let me get the other missing ones.

Small question though:

[storm folio]

the 10GB one should be enough

https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[craggy basin]

What folder am I supposed to place them in?

[storm folio]

read that examples page^

everything is explained there

[craggy basin]

ok, I simply used the Basic workflow in Hugging

In the Comfy workflow folder

Do the above files add anything? Or are they only for specific edge uses?

[craggy basin]

Thanks for the help and the sample. I got it to run as well as the prompt in the Hugging even

And for any curious this was on a modest gaming laptop with a 4060 and 16GB system Ram

@storm folio It still complains, with your sample, about the missing clips, even if I use the embedded larger model. To fix this, I attached the Clip point of the checkpoint directly to the Prompt and removed the connection from the TripleCLIPLoader

[craggy basin]

I was able to test one of the prompts from the announcement page. Dang. Text can be hit and miss, but this is still seriously impressive

24 seconds on this very modest rig

[broken spoke]

Is there a prompt generator

[spare monolith]

hi, i have a question. is there something wrong with my workflow?
its keep getting RuntimeError: mat1 and mat2 shapes cannot be multiplied (2x768 and 2048x1536)
i tried updating comfyui and python dependencies and it didn't works

[storm folio]

look at the second example workflow on: https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[craggy basin]

The second image was what I had loaded

[spare monolith]

i tried the example, load single clip_l and still didn't work, so i tried downloading the model including clips and it worked! thanks!

[craggy basin]

Tell me, why am I getting different results when I use the separate CLIP files as opposed to the embedded ones? Is that normal? Just some weird RNG effect (yes, the seed is fixed and same)?

This was with embedded CLIPs

This is same exact everything but using the separate CLIP files:

If I use the CLIPless checkpoint it is the same as the last

[polar dove]

I updated to use sd3 and broke my comfyui impact nodes.

Reinstalling the node is not working, and everything is fully updated with the manager.

I fixed the reactor nodes, but the impact ones are simply not cooperating.

Any suggestions?

[craggy basin]

So the biggest difference is that one uses the fp8 text encoder and the other uses the fp16. If you isolate one compared to the other there are often night and day differences, at least with text reproduction

I readily admit this could just be noise, but multiple attempts keep showing this

Here is an example. I am using only the fp8 and fp16, with no other clips. Here is fp8:

Prompt: photo of a baking tray of cookies shaped as letters to make the words "Fresh from the oven". The photo is taken in a bakery with artificial lighting.

This is fp16:

[storm folio]

yes that's normal

fp16 is better but a lot bigger

fp8 is not much different from fp16 if you use the other CLIP models

if you use only T5 you should use the fp16

[craggy basin]

Can I make a suggestion? Offer a larger all-in-one file with the fp16 as well

[storm folio]

you can make one with the CheckpointSave node in ComfyUI

[craggy basin]

Oh I have no issue just loading the files directly in the ComfyUI, don't get me wrong. The issue is that many people simply won't even know there is a difference, and the sample page you linked to in no way suggests there is a possible difference either

I certainly did not and only discovered it via my quite anal testing (same seed, this on, that off, etc) as I always do with such things.

[storm folio]

oh I know about it, we just thought a full fp16 t5 was going to be too big for people

[craggy basin]

I don't doubt you know about it, as you confirmed it, meaning you did. But as far as I can tell, the difference is 5gb in size. Let people at least know what is what and they can decide if it is too big or not. I would have liked to at least know the all-in-one was not the ultimate experience, just the most practical one.

[severe pine]

I getting this error I think I doing something wrong

this is my first try comfyui

[storm folio]

AMD?

[stoic gulch]

no to karras

[severe pine]

yes AMD also even I installed amdsmi still

If I don't install amdsmi I got error in autmatic1111 too but I also installed it to comfyui still same
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/15874

[steep marlin]

@storm folio what's the point of the conditioning stuff on the negative prompt for SD3? like i see that it seems to zero it out after 10% of the way, but why? is it actually necessary?

[severe pine]

also my VRAM loaded %80

what can I do fix this

[storm folio]

better images in some cases

[lavish stone]

would be nice the T5 loader to work with cpu

is there any way to load the T5 with the old T5 loader of pix art sigma ?

it does not

it does not work for the moment

[severe pine]

opened issue at github
https://github.com/comfyanonymous/ComfyUI/issues/3698

[steep marlin]

ahh, alright. yeah i've noticed that it seems to absolutely butcher some images with it. even with an empty negative prompt

[steep marlin]

with and without the conditioning on the negative prompt:
prompt: "photo of an apartment living room at night, dim lighting" and with an empty negative prompt

like it gave everything some kind of cartoon outline or something

[merry ermine]

@storm folio Thank you so much for the implementation you did of SD3 in ComfyUI

I can run SD3 in about same speed as SDXL, and even do batches of 4!!

With --lowvram, the encoders are all running on CPU, and the sampling takes about as long as SDXL's sampling would, It was a fantastic surprise to see that I wouldn't need to get a new gpu to play with SD3, again, I don't have enough words to thank you enough!!

[storm folio]

do you get errors without --lowvram?

or is it slower?

[merry ermine]

I haven't tried --lowvram yet,
I always run it in lowvram because I always prioritize stability over speed

My hardware is a RTX 2060 with only 8gb of vram

[severe pine]

I got radeon vii and it won't run it load ram but it fail

[steep marlin]

im running on a 2080 with 8gb vram and have no issues using normal mode. every once in a while, it will say in the console that it's "loading in lowvram mode 4582.307691574097" but other than that, it works great and is fast. thats with all three encoders and 32gb sys ram

[merry ermine]

Ah, I will also experiment with SD3's tensorrt version later, for rtx cards it could run much better than the usual workflow

[merry ermine]

Oh damn thats rly nice, I ougtha try out normal mode more times

[lone jay]

Is there a way I could generate the image using multiple values for shift in ModelSamplingSD3? I want to use the same seed and stuff, just changing that setting without having to do it manually and generate an image for different values of shift.

[steep marlin]

convert the value to a pin and then use some other nodes to handle that part, there are a lot of ways to do it. like you could use an incrementer that starts at like 1.0 and then increments 0.5 each time you queue. so if you queue up up 9, it will go from 1 to 5 in 0.5 increments

[torpid hawk]

Hello dear comfy-ui users, I was wondering if anyone has something that replaces ComfyUI_Dave_CustomNode as the project appears to have been archived? Specifically, I'm looking for a node that does this:

[stone bolt]

hi can someone help me get started on SD3 with comfyui?

[stoic gulch]

yes

i am live now i can take your questions if u have, i jsut put out a (shitty) tutorial on sd3 with windows on amd cards

[stone bolt]

im using ubuntu with nvidia unfortunately, can i just use the checkpoints from the official hugging face page out of the box?

also what about other modules like ipadapters and controlnets can they be used currently with the new model?

[stoic gulch]

try it vanilla first, 22.04 is the most capable

[wispy kelp]

do i need a special model loader node?

what do

do i have to download the clip and text separately? cant comfy handle the all in one package?

looks like i just had to update

i hope the update didnt break anythign else. im always scared of that

so if the t5 text encoder is included does this make sigma obsolete?

[craggy basin]

You may want to use the fp16 encoder instead of the embedded fp8

[wispy kelp]

how do i do that

[steep marlin]

In my testing, no. Sigma is still reliably beating SD3 at prompt comprehension in a lot of the more technical prompts I've tested. But obviously, SD3's quality is light-years ahead.

Oh and I've tested both with both t5 weights. Even tried sigma all the way down to bnb4bit(computes in fp16) and it was still reliably beating sd3. I even tried every combination of the three encoders with SD3 as well

[craggy basin]

in the Encoder folder download the fp16 file, as well as the others and place them in the model/clip folder of Comfy. Then connect the clip nodes in the TripleCLIPLoader

you can use this image to load the basic setup (it is modified from Comfy's page)

[steep marlin]

I'm about to make a sigma to sd3 workflow lol... Need to slap another 32gb of ram into my PC probably

[craggy basin]

if you are completely new to Comfy, the idea is simply: save this image to your PC, then drag and drop it into the Comfy screen in your browser. It will load all the parameters automatically, including the prompt and seed even

https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders Get the first three and place in model/clip

[wispy kelp]

wow

hehe its another 10 gigs

-_-

so all in all 25 gigs for the sd3 fun train

I can't imagine what the 8B model requires

50-100 gigs?

:))

[craggy basin]

You don't actually need the 10gb model if you are using the separate text encoders

The sd3_medium.safetensors at 4GB is fine as you are using the text encoders separately. The thing is that the fp16 text encoder is better than the fp8 one, so it makes since to use it instead.

In pure images it may not be obvious, but as soon as you render text it becomes glaring

[stoic gulch]

how do i make clip input text larger

[merry ermine]

does anyone have workflow to share for comfyui that works with SD3 and Face Detailer

[wispy kelp]

i assume if you refine it with sdxl u cna add facedetailer

i doubt anythign works with sd besides just loading and sampling lol

right now

[merry ermine]

oh snap, guess i should wait then, cause comfyui is kinda complex .. i usually work with auto

[wispy kelp]

just decode the sd3 output then encode it for sdxl

do 20 steps in sd3 then another 20 on sdxl

i mena u have to experiment

we all have to experiment now anyway

find out sd3's quirks

[marble venture]

I made a small guide https://www.youtube.com/watch?v=W14LRlDc_F0&t=51s

[limpid nexus]

Hi, can someome here tell me what the sd3 model sampling node is for? What does the shift do?

[storm folio]

it shifts the timesteps up

[limpid nexus]

timesteps? What does that mean?

[storm folio]

so more steps will be spent sampling structure instead of detail

[limpid nexus]

oh okay. interesting. thanks

[marble venture]

@storm folio so how does that work? If we say start 0.1 and end 1 meaning it spends 10-100% of steps on structure?

[storm folio]

for 10 steps it might sample: 1.0, 0.9, 0.8, 0.7, etc...

[dawn zenith]

Noticed a slight difference when importing a image into comfy and the image produced

this was using a image from a slightly older version of comfy wonder if anyone else has had this issue?

[storm folio]

with higher shift it might sample 1.0, 0.95, etc...

[marble venture]

and in the example workflow, why do you split the negative into two? in one you zero it out,a nd in hte other it goes straight. What difference does it make?

and what does modelsamplingsd3 do? the default shift is 3.0

[storm folio]

in that example it's because some images look better when the lower timesteps are done with a zero negative

[marble venture]

I see! so it's not required, thanks

and the model shift?

and the purpose of splitting them?

[fast summit]

From my albeit limited testing yesterday, I honestly preferred 90% of the images I did when removing modelshift and the splitting of the negative. Aso setting the steps MUCH higher helped a ton (like 60+ high)

[maiden wigeon]

Hello, does anyone know a way to save image names with Comfy using the same name as the prompt? or beginning of the prompt

[lusty pewter]

It's a node system, so you can pretty much do whatever you would like with it. Connect the text output from your prompt box to another node that truncates the text to what you want it to be length-wise, then change a save image node's widget to an input and connect it to the node that truncated the text.

If you don't find nodes in the default stack that do what you want, search ComfyUI Manager for a node pack that has nodes you want that will accomplish it.

[maiden wigeon]

Thank you very much for your advice, do you know the name of a node that does this?

[lusty pewter]

No, but that's neither here nor there...just look through the nodes and find something that does it.

[wanton drift]

for some reason TRT with SD3 makes black boxes even though it should support it (I compiled the engine for the precise settings I use)

[warm vapor]

Both clip_g and clip_l are able to load on my ComfyUI instance but not T5XXL. Does anyone know why?

clip missing: ['text_projection.weight']
Requested to load SD3ClipModel

[heavy isle]

no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
actually means its working correctly 😂

[warm vapor]

photo of three antique dragon glass magic potions in an old abandoned apothecary shop: the first one is blue with the label "1.5", the second one is red with the label "SDXL", the third one is green with the label "SD3".

Nevermind, I got confused by the logs. It takes a few tries to get it right. Sample images are very cherrypicked...

[heavy isle]

yeah, it's just a warning that you downloaded the 4gb smallest checkpoint so you need clip/t5 in seperate files.

[warm vapor]

yeah, i'm loading all three through the tripleclip node

[wanton drift]

first try on my workflow, non-cherrypicked

[warm vapor]

Then something's wrong with my sampler settings...

[warm vapor]

Are you using both CLIPs + T5?

[wanton drift]

I use the one with the 8 bit T5 included

okay, second try it forgot the dot

but it's really easy to cherry-pick one with nothing wrong

third try succeeded

[warm vapor]

Hm...

What's wrong with mine? I'm literally using the workflow example from SAI's HF.

[wanton drift]

I'm using my own workflow, SAI's workflow might just suck

[heavy isle]

🐲

[wanton drift]

that's similar to my workflow with the ksampler settings, but I don't do that weird thing with the negative cond and I use 2 ksamplers

[warm vapor]

I don't think my Comfy is really running T5... something's failing on my end.

[wanton drift]

that's so much closer than before though

just a missing L is seed difference

[warm vapor]

hm

okay, nevermind. it's being loaded on my vram

[craggy basin]

Not cherrypicked. Taken from an entirely different model. (not being funny)

[mint igloo]

has anyone managed to run any SD3 lora with comfyUI?

ah, that's fine
I just saw it here https://github.com/comfyanonymous/ComfyUI/issues/3701

[echo jay]

Anyone here using a custom workflow for SD3??

[unborn dune]

hey guys, anyone have a suggestion if I have a finished image and I wanna increase the saturation within the same workflow?

[topaz pollen]

This exists in the hf repo now btw

It just lagged slightly behind

[wispy kelp]

There are amazing workflows for sd3.

but no one wnats to share

[flat bobcat]

can we use this with comfyui or we have to build engine from normal model sd3 ?

[wispy kelp]

since pony cant save SD3 maybe comfyui workflows can

[pulsar ivy]

Curious if anyone has any thoughts on how https://botika.io/ was built, I assume it runs on the Stable Diffusion architecture, as I'm not aware of anything else with the necessary levels of customization.

I could arguably build a workflow that achieves the same result, but not without getting deformities occasionally. This is leading me to wonder whether they have someone manually improving each image, or if there's some secret solution that I'm not aware of to ensure this level of high-quality consistency.

[sick badger]

realize comfy "is" a webui in itself, but is there a commandline switch to start it but disable the webUI?

[stoic gulch]

https://i.imgur.com/BuDDE3T.png i dont know how to use any of these but now they load

[hardy tree]

wait what you mean... you installed all those without knowing what they do?

[sage timber]

Hi! Pretty new to comfyUI and IP adapters. I want to do outfit swap with this anime guy (left) but after playing around the IPAdapter advanced I always got this. Any tips on this? It seems even controlnet doesn't fix it.

[stoic gulch]

i am getting: Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.... i feel like ive solved this before but i cant remember how to fix it

[craggy basin]

Can't you just close the page?

[wispy kelp]

whats that and where do i get it?

[sly sundial]

Could be custom

Possibly a clip merge...?

There are some that are part of ComfyUI

[flat bobcat]

where did you get image

[wispy kelp]

some latent man gave me a workkflow he keeps showing off on sd 3 with his workflows

so this is a par tof his secret workflow he shared

in the sd3 channel that is

🤷‍♀️

[wispy kelp]

its my image im trying to make latent man's workflow work

[sick badger]

Lol yes. I want to continue to run my workflows via web socket API but not have a UI exposed. Don't think it can due to the nature of what it is but had to ask.

[silent knoll]

Hi, looking for someone to help me deploy a comfy WF to replicate. Ping me if you're interested and I can share more details. Thanks

[wispy kelp]

whats the fee dear?

[robust lily]

fee for commercial, free for me

[stoic gulch]

anyway to increase the text size of the prompt area?

[covert bough]

@storm folio do you have technical documentation somewhere how exactly how the ModelSamplingSD3 node works? other than shifting the timestep?

[storm folio]

what do you mean by that?

https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_sampling.py#L177

this is the code

for the sampling part

[covert bough]

this node. i'm looking for a deep dive technical document that tells me eactly what it's doing under the covers. do you have one?

[storm folio]

it sets the sampling of the model to that code I linked above

[covert bough]

okay, thank you. i'll ask if i've still got questions after i go through that. kudos, btw - the updates for sd3 are working exceptionally well

[steep marlin]

he probably means more on what it actually does. like lower values focus more on shape and primary forms, high values focus more on the fine details (i think i have that order right? might be the other way around). i do know that you don't want to go below 1 though or weird stuff starts happening

[storm folio]

higher shift means more focus on structure

[covert bough]

question on this part:
def percent_to_sigma(self, percent):
if percent <= 0.0:
return 999999999.9
if percent >= 1.0:
return 0.0

that seems backwards to me

why those values?

[steep marlin]

oh okay, i had it backwards then lol. knew it was one way or the other. 50% of the time, i'm right 100% of the time

[covert bough]

i just barely had a handle on u-net - i'm way out of my depth on this new network

[storm folio]

that's not for SD3

[covert bough]

ah. okay

[tranquil parcel]

Any plans to make SD3 medium compatible with more Samplers and Schedulers?

[dusk grove]

Is it possible to use a mask for any tiled sampler node?

[craggy basin]

so before I waste time on this: will Hunyuan run on my 4060? Or do I need a GPU with more VRAM?

from within Comfy

the page about it left me unsure

[eager escarp]

it probably will

you can offload the T5 to RAM

[steep marlin]

Guys, differential diffusion works with SD3... Why didn't I catch that memo...

i'm going to play around with it more, but shit, i wasn't expecting it to just work lol. wanted to try it out to see if it would work or not

[craggy basin]

ok thx, will give it a go. Can't not be curious

[eager escarp]

hunyuan 1.1 is pretty good

they just released full training code yesterday too

[craggy basin]

and Lora support etc. Can I just stick them into the folders or do I need to run all the install scripts on the Git page?

[eager escarp]

i haven't tried any loras

[craggy basin]

I meant Hunyuan files and models in general

[steep marlin]

Wait, they updated it? I'll have to update my little standalone py app for it. I like hunyuan for brainstorming(scratch what I said before, was confusing with lumina's data set)

But it listens extremely well to prompting and is neck and neck with sigma for me, in that regard

[eager escarp]

yep, agreed, and yeah new version as of yesterday

they also released a "distilled" version that takes fewer steps, and a tensorrt version

[steep marlin]

yeah i've been using the diffusers-distilled version. i see they also have the 1.1 version of that as well. sweeet

[craggy basin]

distilled meaning same quality and faster? Or a slight tradeoff in quality?

[steep marlin]

usually just the former, especially if you're still running the same step counts

sweet, it uses the same T5, I'll just let python start to pull the repo to get the folder structure and move the T5 over from the older version. saves redownloading it lol

just checked the SHA hashes to make sure

[craggy basin]

and I thought SD3 was big

(in terms of space)

[eager escarp]

it/s are the same with distilled

but it gets a usable image in fewer steps

think they named it funny

[steep marlin]

yeah i should have specified that. it can be faster in the sense you can use 20 steps instead of 30to get roughly the same quality image, for example

[eager escarp]

yeah i haven't used distilled anything much at all tbh so i'm not too familiar with it myself

[steep marlin]

i only have an 8gb gpu, so i run it. my little app even tracks peak vram usage and it peaks to ~7.5gb with the t5 in q8 mode

[craggy basin]

ok, that's cool

[steep marlin]

the app is slower because it does a bunch of unloading/loading stuff to try to only keep the current step in vram, but i dont mind. i just use it to screw around with while doing work

[craggy basin]

Any notes on what the update involves? I mean, you can assume 'better', but in what I wownder?

Well, I plan to run from Comfy

[steep marlin]

apparently contrast issues and watermarks

[craggy basin]

contrast issues. Interesting. Like the early models of MJ

[steep marlin]

"HYDiT-v1.1 version is released, which mitigates the issue of image oversaturation and alleviates the watermark issue"

well saturation i should say

but i definitely noticed both these things a lot

[craggy basin]

oversaturation, ok that is different

T5 means it takes the same natural language prompts?

[steep marlin]

yep

their model is a bit slower than say pixart's though. with pixart sigma, i usually see like 1.6 it/s, whereas with hunyuan, it's more like 1.6 s/it. Oh and I got my models mixed up again, hunyuan peaks to 3.75gb for me using all these optimizations. my pixart app does similar stuff, but peaks to 7.5gb with t5 in q8 mode

[craggy basin]

I have not followed those last ones: Pixart for example

I last dabbled in homebrew when XL came out

[steep marlin]

pixart sigma is a beast at following prompts, it just lacks quality

[craggy basin]

Usually use Dall-E3 (MS's less censored one) and Ideogram

Dall-E3 on Copilot is way better than OpenAI's, which is a colossal pain with all its restrictions

but that is another topic. Anyhow, will test out this Chinese flavor for sure

The most exciting thing about SD3 (even this hobbled version) IMHO is the text rendering ability

[steep marlin]

I don't remember where i got this prompt from, but it's a really good one for testing prompt cohesion. SD3 fails this almost every single time. I think I ran a batch of like 50 and it got it right maybe three times. It would always show ears and wings

here's hunyuan 1.1 nailing it on the first try lol

[craggy basin]

Can you paste it here?

[steep marlin]

paste what?

oh the prompt

In a soft, creamy haze, a mystical creature stands, its velvety blue fur glistening in the gentle light. Oversized butterfly wings, delicate and intricately veined, sprout from its head, functioning as ears. Large, bright blue eyes shine like sapphires, filling the space with wonder and enchantment.

but both pixart sigma and hunyuan nail this prompt 95% of the time from what i've seen (it's the default prompt for the two little apps of mine lol)

[craggy basin]

Ideogram does a solid job

Dall-E3 was solid but not quite as good. IMHO

[steep marlin]

yeah but see how that one still has ears? that's what sd3 does

[craggy basin]

Ahh... functioning as ears. gotcha

so D3 scored

[steep marlin]

yep for sure, subtle little nuance to prompt comprehension though

[craggy basin]

But to Ideogram's credit, second try got it

or not

damn those ityy bitty ears

lol

D3 got it twice even, so no fails

[steep marlin]

i've tried other more literal variations of the wording like "has wings for ears," and sd3 fails at the same rate as with the regular wording

but i actually kind of love the prompt, i really should figure out who posted it. the critters are cute and the kids love em

[craggy basin]

yes, and it is challenging for the AIs so a good lithmus test

nothing too obvious or easy such as green ball under left foot etc

[steep marlin]

yeah, i have a bank of other litmus test prompts, but this one has been a go-to-first for me lately

[craggy basin]

damn. JSON for Comfy is not working

Anyone have an image with workflow for hunyuan in Comfy?

[stoic gulch]

looks like dinner

[steep marlin]

I don't use it in comfy because I accidentally nuked my comfy install with dependency conflicts that I was too lazy to undo, so I had to reinstall it lol

[covert bough]

SD 3 medium

[craggy basin]

so I got it all sort of running but ran inbto snag

Error occurred when executing HYDiTTextEncoderLoader:

T5Converter requires the protobuf library but it was not found in your environment. Checkout the instructions on the
installation page of its repo: https://github.com/protocolbuffers/protobuf/tree/master/python#installation and follow the ones
that match your environment. Please note that you may need to restart your runtime after installation.

I have Protobuf in my normal install

but it is using the embedded Python I think

[steep marlin]

Still has ears. The ears are supposed to be the butterfly wings. Maybe sd3 ultra will be able to do it

[steep marlin]

Yeah I remember something like that happening to me and I ran a force reinstall of the requirements and it bricked my comfyui venv

I hate python so much. It's so handy, but so f'in annoying sometimes

[covert bough]

🙂 and yet, when i ask for butterfly wings i usually get them on someone's head.

[craggy basin]

I managed toi bypass it but then got the following bug:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

[steep marlin]

That's a vram issue

[craggy basin]

and it crashed when it got to the Ksampler issue

[covert bough]

stop using your cpu to render with

[craggy basin]

I wasn't trying to do anything

[steep marlin]

You'll need to run it in a lower quantization like bnb8bit

[craggy basin]

I just ran Main.py

[covert bough]

what machine do you have?

[craggy basin]

plain Intel with 4060

[covert bough]

do you have a GPU?

[steep marlin]

If CPU offload is enabled in the code, it will cause errors like this when you run out of vram

[craggy basin]

how do I enable lowvram?

[steep marlin]

In comfy?

Or the standalone

[craggy basin]

well I am using standalone but not its python

[steep marlin]

You might have to edit the code, I'll take a look. It's the hydit folder app right?

[craggy basin]

No, this is with custom-nodes

but yes, there is a hydit folder in it

[steep marlin]

alright, lemme check real quick

oh i see why, shits not using the distilled version. the version you're using probably has the 20gb T5

[craggy basin]

that is correct. I can change

but when I ran the lowvram flag I got this BTW

Error occurred when executing HYDiTTextEncodeSimple:

"addmm_impl_cpu_" not implemented for 'Half'

[steep marlin]

lemme look through the code some more to see if it does any torch casting to 16 or bitsandbytes

yeah code looks like it wants to do t5 on the cpu:
class mT5Model(torch.nn.Module): def __init__(self, textmodel_json_config=None, device="cpu", max_length=256, freeze=True, dtype=None): super().__init__()

lemme see if it does the slightly hacky thing where it precalcs the tokens before sending them to the cuda pipe

i would say try loading the model on the gpu with bnb8 or 4, but idk if you'd still have enough room for the model and vae afterwards. bnb is awesome, but once the model is in vram, it can't be offloaded back to ram. it stays locked like an anchor

maybe try launching comfui with the --lowvram flag?

maybe it will shuffle things around more, idk

[craggy basin]

when I ran the lowvram flag I got this BTW
Error occurred when executing HYDiTTextEncodeSimple:

"addmmimpl_cpu" not implemented for 'Half'

I can DL and use the distilled versions

it is no sweat

[steep marlin]

you can't do BnB on the cpu, it's gpu only

[craggy basin]

as I am not a coder, I am guessing you don't mean Bed'N'Breakfast

[steep marlin]

oh bitsandbytes

it's a way of quantizing a model on the fly essentially

[craggy basin]

Ok, let me install all the distilled versions and see what happens

but the error I got with the ---lowvram does not make me hopeful

[steep marlin]

so if a model is like 10gb at float16, you can compress it down to ~5gb at 8bit and ~2.5gb at 4bit. LLMs usually run perfectly fine at 8bit(virtually no difference between 8bit and 16bit)

lemme make sure that flag is correct

yeah it's --lowvram, you add it to run_nvidia_gpu.bat (edit the file)

when editing it, the first line will look something like .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --lowvram

[craggy basin]

I cant use the embedded. It complains I don't have protobuf installed

[steep marlin]

oh and that addon isn't compatible with the distilled diffuser versions i dont think, kind of remember trying it and it not working

[craggy basin]

ok, then I will write this off as a no go for now. I wonb't lose any sleep

[steep marlin]

maybe check their repo to see if any other people have solutions (the extra models repo for the comfyui plugin)

[covert bough]

prompt: a blue, furry fluff, butterfly wing ears

[steep marlin]

I said it over in the sd3 channel, but if you guys want to do a pixart sigma+sd3 workflow, you can share the t5 between them. The extra models plugin author is working on a node to allow you to also use the triple clip node as well(since pixart doesn't use clipG/L). That way, if you have the big all in one SD3 model, you can also pull off it's clip for sigma.

But even using just the standalone t5 as the clip for SD3, it still works great

[loud apex]

If your using SDXL clip + t5 for SD3, which node do you use to combine the clips?

[steep marlin]

There's a triple clip node.

But if you're trying to pull off an sdxl model's clip for G/L and then append a separate t5 to it, I don't think there's a way yet. I tried it earlier and was running into issues(think I tried clipmerge or something)

[gray rampart]

hi, when i use "sd3Medium.vu4B.safetensors", comfyui has a error :"no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.",whitch model i need download

[steep marlin]

show a screenshot of it

[steep marlin]

no i mean of the graph

[gray rampart]

this?

[steep marlin]

oh that just means that the model is a version without them baked into the model

https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders

download at least clip L and G and put them in your models/clip folder

[gray rampart]

clip_g.safetensors and clip_l.safetensors?

thanks

[steep marlin]

and then use to load them in comfy

make sure to load them both, in this SS, i didnt change one to L

and if you want to also use T5 with it, download a version and use the triple clip node

[gray rampart]

ok,thanks a lot.i try it first ,thanks again

[steep marlin]

np

[ancient bloom]

I keep getting this error trying to use stable diffusion comfy AI for Krita's AI generative Ai. I am using Stability Matrix and have installed all the Models needed but it stillk says i am missing it

[covert bough]

krita doesn't have SD 3 support yet. SD 3 is a different network, krita will need to add code to let it run

[ancient bloom]

sorry I dont understand? Does this mean i cant use it?

[covert bough]

not yet

[ancient bloom]

alright then, i will use the local one rather than this remote server

thank you

[mystic plaza]

is there any node for defocus in comfui ?

[livid gust]

Blur by depth perhaps..?

[mystic plaza]

like i have created a bg through ai and need the bg to be defocus

[marble mirage]

what are the stable cascade requirements? tried in colab and failed t4 15gb vram. maybe i should try forcing fp16, increase RAM in the colab? or would i really need more VRAM?

[mystic plaza]

@sly sundial so a doubt in the workflow you created and sent me can i add a defocus image into it like a bg replacement without generation

[sly sundial]

Yes, just cutout the gen part and replace it with the load image.

[mystic plaza]

so now what do i need here bro @sly sundial

[sly sundial]

Follow where the image was being connected to when it was being gen'd and attach the load image node to the same places.

[mystic plaza]

??? @sly sundial

[sly sundial]

Why have you disconnected the mask as well?

I don't have the flow, or time at the moment.

Work it out...best way to learn 🙂

[mystic plaza]

the mask is not disconnected its saved in my load image

[sly sundial]

If you say so

[hardy tree]

@storm folio i see you added some stable audio support. is the support ready for actual generation. if it is, is there any basic workflow you can share on how to use it? ❤️

[storm folio]

t5_base.safetensors (goes in ComfyUI/models/clip): https://huggingface.co/google-t5/t5-base/blob/main/model.safetensors

stable audio open (goes in ComfyUI/models/checkpoints): https://huggingface.co/stabilityai/stable-audio-open-1.0/tree/main

[hardy tree]

ah ty 🙂

[storm folio]

it's only basic support for now but it should work

[hardy tree]

@storm folio at the decoding stage, im getting error: "weight_norm_fwd_first_dim_kernel" not implemented for 'BFloat16'

[storm folio]

update pytorch

[hardy tree]

ah

[storm folio]

how old is your pytorch?

[hardy tree]

well let me check

torch-2.1.2+cu118

[storm folio]

yeah that's pretty old, 2.3 is the stable now and 2.4 is going to be released soon

[hardy tree]

but is 2.3 available with 11.8?

[storm folio]

as far as I know it is

[hardy tree]

kk il update

actually how do you recommend i update it?

[storm folio]

are you on the standalone?

[hardy tree]

ah i found the link: pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu118

[storm folio]

or manual?

[hardy tree]

im using currently the portable comfy

with the python embed folder

[storm folio]

on the standalone you can just run the update/update_comfyui_with_python_dependencies.bat

[hardy tree]

i think i know how to update it, will let you know if i can then run the workflow

i hate python dependencies LOL.. cause i had to update xformers too, but then of course its not the "latest version" that works with 2.3...
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.3.0+cu121 with CUDA 1201 (you have 2.3.0+cu118)
Python 3.10.11 (you have 3.10.9)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
Memory-efficient attention, SwiGLU, sparse and more won't be available.
Set XFORMERS_MORE_DETAILS=1 for more details
xformers version: 0.0.26.post1

and also python version too? ugh

maybe i need to downgrade to some earlier xformers

[storm folio]

don't worry about xformers

just try it

[hardy tree]

kk

ok it generates nicely, stereo too 😮 btw is it always going to be 47 seconds? but now how to fix xformers 😦

[storm folio]

for xformers I think if you just update it it's going to work

yeah I need to add some more controls on the empty audio latent node

this is just the first basic implementation that can match the reference one

[hardy tree]

but i did update it, latest version of xformers, doesnt work with torch 2.3 and also 11.8 cuda and python version.
can you put also maybe a node that can preview audio instead of saving? cause i dont want to always save or go to the
output folder to check the audio

[storm folio]

yeah that's coming

oh you might need to switch to cu121 for xformers

but on pytorch 2.3 perf with vs without xformers should be pretty much the same

[hardy tree]

so now i have to install torch 2.3 with 12.1? ugh

[storm folio]

at least it is on my setup

[hardy tree]

kk

btw is audio encoding technically working? cause that would be cool if we can do like img2img but with audio

[craggy basin]

While failing utterly to get Hunyuan to work within Comfy, I noticed Comfy also supported things such as Llama3, supposedly quite easily. Is there a simple workflow and step-by-step I can use to try it?

[blazing rampart]

Anybody have any resources for advanced prompting techniques specifically in regards to SD?

All the videos/articles I can find are for normies, I'm looking for something that explains key technical and artistic terms (camera type, image POV, various styles, etc.)

[storm folio]

Yes it might work

Didn't test audio to audio

But vaeencode works for audio

[hardy tree]

@storm folio btw i fixed my xformers too, the problem was when i did just pip install xformers, but that doesnt install the correct cuda version at all anyway, so i had to go here: https://download.pytorch.org/whl/xformers/ and grab the correct wheel to install for my pc: xformers-0.0.26.post1+cu118-cp310-cp310-win_amd64.whl and now it is using xformers again! 😮

[storm folio]

I see so they do have some wheels with that cuda

[hardy tree]

yes lucky for me :3 cause i didnt want to complicate myself with installing either different python or torch again.
im gonna try the audio encoder

and im assuming the input has to be in stereo?

ok im getting some results nicely, but im assuming i have to change the denoise strength

hmm yea i mean technically it works, denoise lower and higher than 0.5 seems to produce bad results, so i stick to 0.5, this works with blank prompt, seems to recognize instruments nicely and also if i guide it with positive prompt, i can improve the result as well, so yea, this does work and i used stereo input and it matched the length of the input audio, so if you input 30 seconds, it will output 30 seconds

nice

now we just need audio preview node, and some control for the text to audio like duration length, etc

i didnt try the negative prompt, but im assuming it works too, just dont see a lot of point to it idk... cause you specify enough in the positive prompt and it will give you what you requested, not even sure what would i negate, anyway :3

[fading jewel]

@storm folio when you get a chance can you tell me why this is being done? it seems we're copying the conds to the GPU every single step? and also creating a fresh tensor out of them?

[storm folio]

That's because they can change each step

[cloud badger]

fyi, if you are on Windows versions of torch above 2.1.2 aren't compiled with Flash Attention, so it is likely to run slower.

[storm folio]

That's not really true when I tested

[hardy tree]

it runs nice for me, like fast

[cloud badger]

If I run any version above 2.1.2 I get errors saying that it wasn't compiled with flash attention

And it does run slower over longer workflows

[hardy tree]

i dont have those errors

[cloud badger]

You only get it once one the first run, then it doesn't come back up for some reason.
Unless they've very recently fixed it.

[hardy tree]

i mean i do have flash attention installed separately as a library

[cloud badger]

I'll try it again now with 2.3.1

Yeah still doing it in 2.3.1

  out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)```

And it's definitely a good bit slower when this happens vs 2.1.2 that doesn't have this error.

[cloud badger]

Did you not get the error? Or did you just find that it wasn't slower.

[hardy tree]

for me personally, i never got that error and its not slow

[cloud badger]

Not sure why you wouldn't get the error on Windows. It's the same binary, so it's not going to have been compiled with Flash Attention

[hardy tree]

what's your version of python?

[cloud badger]

3.11.9

I tried 3.12.4 but it breaks some stuff

[hardy tree]

ah cause i have 3.10

but idk if that is the reason

[cloud badger]

I don't think it would be because the error is specifically saying that it's not compiled with Flash Attention.

There's a few posts on the respective githubs but they aren't getting much traction.

[hardy tree]

i heard some people having problems with more newer python versions anyway, most people still stick to 3.10

[cloud badger]

Mines been perfectly fine on 3.11.9. No issues.

It's just any versions of torch higher than 2.1.2 give the error that it wasn't compiled with flash attention

[hardy tree]

can you somehow test by creating a virtual environment or something where you have 3.10 and try to use it, just to see if it gives the error?

[cloud badger]

I can, but I don't see it making a difference

[hardy tree]

are you using xformers?

[cloud badger]

no

[hardy tree]

can you try with xformers

[cloud badger]

Shouldn't need it when you are using sdp

[hardy tree]

yea but thats my point tho... cause your error comes from sdp

[cloud badger]

Yes because I'm trying to use SDP. Using xformers isn't going to prove anything because it's using something different

[hardy tree]

so you want to use sdp and not xformers?

[cloud badger]

Yes

[hardy tree]

is there any specific reason as to why?

[cloud badger]

It's generally better. There's a reason it's not in the requirements for Comfy anymore

[hardy tree]

hmm

[cloud badger]

There's an issue about it on the Comfyui page, but it's not a very good issue, there isn't actually any good information, except for a person moaning - https://github.com/comfyanonymous/ComfyUI/issues/3363

[storm folio]

I get the warning but the versions I tested were not slower than xformers

[cloud badger]

I can test it again, but when I tested it before a workflow that was 100 seconds on 2.1.2 was 120 seconds on 2.3.0 consistently

[fading jewel]

these comparisons are blocking because timesteps are on GPU

[cloud badger]

Ok fair enough then. Either they've fixed something, or perhaps something else was going on when I tested before, as there is no difference in speed now.

Initial SD3 Run - 49.02 seconds
After model loaded:
Run 1 - 12.75 s
Run 2 - 12.85 s
Run 3 - 12.68 s

SD1.5 Long Workflow:
Initial Run - 62.51 s
After Model Loaded:
Run 1 - 36.50 s
Run 2 - 36.83 s
Run 3 - 36.89s


Torch 2.1.2:
Initial SD3 Run - 48.55 seconds
After model loaded:
Run 1 - 12.69 s
Run 2 - 12.72 s
Run 3 - 12.72 s

SD1.5 Long Workflow:
Initial Run - 62.39 s
After model loaded:
Run 1 - 37.15 s
Run 2 - 37.47 s
Run 3 - 37.28 s

[fading jewel]

?????

oh wait yeah cpu rng

[severe pine]

where can I get this?
clip missing: ['text_projection.weight']

[steep marlin]

Newer torch uses a newer version of flash attention and its whl has to be compiled manually for Windows since they don't package it. It takes over an hour to compile on an i5 13600kf and then usually doesn't even work lol

If you go to the flash attention repo you'll see more about it

[hardy tree]

just curious, what made you go look through the code? :3

[cloud badger]

Yeah I read about it and I did compile a version of Torch with it successfully, but I don't know what I'm doing and it was missing DLL files that it needs for Comfy to have it run.

There doesn't seem to actually be a difference in speed anyway, so it doesn't matter.

[fading jewel]

mainly I think it started when I finally got my SD training script to have uninterrupted 100% usage, and after seeing that that is possible I realized that this means that if that isn't happening then something is almost certainly wrong somewhere

[steep marlin]

Well comfy is probably having the same issues I was having using a simple standalone diffuser script. Even though it succeeded in building the wheel, there's still some other compatibility issues in the libraries somewhere. Maybe try checking diffusers and transformers to see if anyone figured it out. It could just be something lame like it only works with some specific build of pytorch+python

[low heart]

Hi guys, can someone pls gimmi a tip or trick to get these last 2 nodes to install in ComfyUI

I have spent a full day installing nodes to get this one workflow working but these last 2 nodes are just impossible, i cant crack the install / setup process to get them to work.

i have them placed in the same folder as all the other custom_nodes

however, these 2 nodes did not have a requirments.txt installation file

[uncut dawn]

hello, how to add --lowvram to ComfyUI on Pinokio? https://pinokio.computer/

i just need to block usage spills from VRAM to system memory

[storm folio]

that's a setting in the nvidia driver

[steep marlin]

You can do it globally if you don't feel like pecking through folders to find Python.exe. the setting in Nvidia control panel is called CUDA fallback I think. You set it to prefer no sysmem fallback. But if you OOM on vram, it will spit errors

[brazen turret]

is there some plugin I am missing or something in order to enable live preview in ksampler? I am watching a youtube video that shows them using advanced ksampler but as it's processing it's showing a preview of the image for every step. How do I enable this

[steep marlin]

that's the setting in case you have a hard time finding it, make sure to click apply after changing it and restart comfyui

[errant ledge]

人

[stoic gulch]

https://i.imgur.com/BdaQ15X.png on windows 11 with py 3.11 using comfyui zluda... doubt it works but it loaded

comfyui is so f-n cool thank u guys

[lavish stone]

guys here is a way to add T5 to sdxl by i'm not a programmer so i don't know how could we do it : https://github.com/AIGText/Glyph-ByT5

[cursive ridge]

Help? I don't get it, I've downloaded hundreds of loras and checkpoints from civitae before and 99% of them worked wonderfully (and the others, the art just sucked), but ever since SD3, I get nonstop errors from them ALL, even after I select the correct models to fill in the nodes.

[stoic gulch]

they dont work with sd3

[cursive ridge]

They seem to be working for others though?? So I figure my settings are off, or something

[stoic gulch]

sdxl works with sdxl, 1.5 with 1.5, sd3 with sd3.. there is like 1 or 2 vae, and i doubt many people are making their own checkpoints and loras right now thats teh debate with the community license

i believe

https://i.imgur.com/NnWCyWC.png i think this is all you really get

mine are customized but its just combining nodes

[cursive ridge]

Hmmm, only 1-2 vae, here I've been putting in the vae I already have, oops!! Which vae actually work properly with SD3? In case that's the problem

[stoic gulch]

on civitai, when you search for sd3, they come up

[cursive ridge]

Thank you 🙂

[stoic gulch]

tbh, once again, i think this is where a lto of the debate lies. civit AI is not relly going to work with sd3 license

i dont know if anyone from civit will come out and say that, but this seems like war

🍿

[cursive ridge]

So download everything while I can, got it 😉

Clearly my prompting needs work ROFL
But at least I got ONE SD3 civitae model to work (it only crapped out on the same aspect at the end)

[stoic gulch]

https://civitai.com/models/515311/sd3-cotton-doll

[wispy kelp]

Are these troll loras? or everyone rushing to be the one to get the first stuff for sd3 out

[stoic gulch]

i think its more of "what *can * we do"

can i buy stock in stability?

yes.

[wispy kelp]

oh theyre public?

[stoic gulch]

$20

[cursive ridge]

I haven't tried that one, does it actually work?

[stoic gulch]

shhiiiiii

[cursive ridge]

The preview image scared me away from even looking at that one

[stoic gulch]

no clue im trying to get 3d to work. i just got audio pumping. then i gotta optimize workflows, then i gotta pretend im a visual artist for hire

[wispy kelp]

RealvisSD3 juggernautSD3 pls

[stoic gulch]

i'd apprciate some zavy thank you

[wispy kelp]

and no sdxl/sd3 merges - real deal! 😄

funny you say that zavy is my fav one now, best model for sdxl

[stoic gulch]

zavy is the goat

[cursive ridge]

Are you saying they do work, or they don't?

[wispy kelp]

i think theres a sdxl/sd3 merge already up

[stoic gulch]

no people are making sdxl sd3 hybrid checkopoints and loras but its not based on sd3, its just retrofitted, iiuc

[wispy kelp]

https://civitai.com/models/511463/afros-sd3-experimental-sdxl-merge

[cursive ridge]

I had Alex recommend I try the merge idea, so I'm tryng the Civitae ones, one by one, and not having any luck so far 😦

[cursive ridge]

This is one of the ones that is giving me neverending errors, what am I most likely doing wrong?

[wispy kelp]

idk havent tried any of this yet. feels like troll checkpoint and loras at this point for now...

ppl rushing to get anything "sd3" related out the gate

[stoic gulch]

i'm so new to all of this so im not in a rush. just praying for zavy

[cursive ridge]

zavy?

[stoic gulch]

zavy makes some really nice loras. zavy cinematic*, the one with two girls kissing, kind started me on my journey

https://civitai.com/models/295530?modelVersionId=332071 this is one of my favorites.

[cursive ridge]

I like his style, I"m testing it out on civitae. 🙂

I really should have tested the SD3 ones out before downloading lol

[stoic gulch]

sd3 is fairly particular. not a lot of room for messing around. its more about prompting

also, as i said, i dont know what i am talking about

[wispy kelp]

also the clip cant digest as many words (toekns?) as sdxl

my monster styles with over 8000 words go in fine into sdxl models but sd gives me bs like memory error

i have like a genereic negative body style with about 6000 words

and it does help a tiny bit even with sdxl models, not a smuch as with sd15 models but still

[stoic gulch]

have you split the text inputs?

[wispy kelp]

no

[stoic gulch]

if be so curious so see this

[wispy kelp]

idk know i just use the "sdxl style" node

[stoic gulch]

i am using ollama to reprompt my prompts, and its nice

and i have it split into 3

clip l+g, negative, and text

[wispy kelp]

so simple prompts are a no go?

it likes long natural language short stories

[stoic gulch]

you dont want any terminology that is contradictory

because then it tries to do both and then erase it the next step

you can see it in the preview soemtimes fiddling between two subject ideas

my text prompt was talking aobut zebras and my clip prompt was talking aobut the hulk, i got alike a hulk zebra hybrid

but each preview it was one or the other, it was weird

its thinking

https://i.imgur.com/0cxQE7e.jpeg is like a furry's dream, he's supposed to be dissintegrating into a black hole

[cursive ridge]

Every workflow I've gotten so far was awesome, as were the ones I downloaded from the SD3 github, however, as soon as I got to civitae... lol
It's usually the other way around though

Well if they are trolling, no problem, because I'm posting my resulting images on their civitae model pages 😄

[stoic gulch]

it text heavy for sure. it would be cool if there was an instruct variant so there is some contextual history

[cursive ridge]

I"m generally a lazy prompter and do short ones, now I feel extra lazy. Wow, those are some intricate prompts you must do 😄
That said, I wonder if my prompting is too short for these models.

[wispy kelp]

throw me the bets most realistic workflow for sd3 you found so far 😉

[stoic gulch]

i have created this articifially with ollama, so llama3 is a proxy context interpretter for sd3

[wispy kelp]

I am lazy too hence I added the most frequent negative and positive promts to a style file

[cursive ridge]

That's a good idea!

[stoic gulch]

tbh i think what people are trying to accomplish with loras and vae should be available with better wording and interpretability of sd3 and future generations

[cursive ridge]

Here's a successful one. Actually msot are successful, until I hit civitea lol

It's Biden and Puten, in case you aren't from the USA..

[uncut dawn]

😍 thank you so much for your time, this might actually solve my issue, will test it more

[cursive ridge]

I finally got this one to work, perhaps it was my using random (not SD3) vae that was the problem

[fading mural]

hella new, fresh out of the womb
how to insert more than 1 lora here if its a "chain"

[lusty pewter]

Literally just add another one in sequence.

[fading mural]

how

[lusty pewter]

In the most basic way possible.

[fading mural]

helpful thanks

[lusty pewter]

Add another node, connect model to model and clip to clip.

You cannot get more simple than that.

[cursive ridge]

You can copy/paste it. Or it as well as any connecting bits 🙂

or right click and add node, add lora

then connect the matching colour dots

[lavish stone]

use lora stackers

[coral siren]

Does anybody know, when using an input image in Comfy how to define a resolution, regardless of the input image res?

[lavish stone]

Depends what model you are using. I change the target resultuion in the Sdxl conditioner

[coral siren]

What node is it? There are several regarding conditioning, I would like to use it as well with Cascade btw

[lavish stone]

Sdxl advanced conditioner

Go download my workflow in civitai, and look the nodes in the Sdxl section

Search for my handle

[stoic gulch]

requesting deletion no real photos allowed

[cursive ridge]

Oops, thought swf ones were, sorry about that

[merry ermine]

Apologies if this has already been answered. I am trying to use SD3 with Comfyui and I can get it to select the model itself, but when I try to add the text encoders, I get undefined, as shown in the screen shot above. Any ideas what I am missing? The model and the text encoders are all in the same folder and all appear in the menu for choosing the checkpoint, but it's like the menu is empty for text encoders and I am struggling mightily to figure out just what I am missing. Thanks.

[marble mirage]

encoders should be in model/clip not model/checkpoint

[red dragon]

hello

can anyone help me with comfyui

I am getting an error

[cursive ridge]

Trying to remeber now, but try them in the clips folder

[stoic gulch]

i was trolling i was sayin git looked real sorry

[cursive ridge]

Oh LOL, can you tell that perhaps I've had mods tell me to remove NSFW deep fakes before? ROFL

[merry ermine]

Thanks for the help. That fixed that specific problem at least. Still waiting to see if I can get an image.

[stoic gulch]

its a joke.. there was a real flamingo** photo that won an AI contest

so i keep saying im banning people for real photos, but it reall just means i liked it

[cursive ridge]

Oh I get it now (had just woken up when I read that). I posted ones which were too real before, so had to take them down due to the deep fake thing lololol, so I easily fell for it LOL

[cursive ridge]

Nah the real one of Putin and Biden couldn't be made with SD3 😉

[cursive ridge]

Which error?

[red dragon]

Cannot import C:\ComfyUI\ComfyUI\custom_nodes\ComfyUI-MimicBrush module for custom nodes: No module named 'cv2'

[stoic gulch]

lol right i forgot was it was

[cursive ridge]

Have you tried (using the comfy manager) importing all missing nodes?

That works for me most of the time anyways

[red dragon]

It is saying
No Results

[merry ermine]

Well, it got me further than before... I was at least able to get a black square. But when I used the one labeled t5xxl_fp8_e4m3fn.safetensors, it got crabby with me and gave me this error (screenshot), so I tried sd3_medium_incl_clips_t5xxlfp16.safetensors for the T5xxl model and it didn't give me the error but it just gave me a black screen. Not sure what to do.

[red dragon]

@cursive ridge No Results

[cursive ridge]

Is the model that is listed in the workflow you are using, the exact same name as the model you are using? If not, just click on the one you want to use (that fixed most of my similar problems today.

THe other thing which has fixed most of my SD3 problems is me installing the correct Vae, instead of using my random SD2 ones

[red dragon]

Do install the repo using manager instead of cmd ?

[cursive ridge]

either works, just manager is easier sometimes.
For your errors, hopefully someone who actually knows what they are doing, and who gets less errors than you do, will come by shortly :>

[cursive ridge]

Did you also install any missing nodes via comfy manager?
Are you using SD3 specific Vae?

(those 2 things get rid of most of my screens full of red text)

Also, just correcting everyting that is outlined in red in the workflow tends to get rid of most errors

[merry ermine]

I am still trying to find an SD3 Specific VAE now that you mentioned that .. it was going to be the next thing I tried. I got all the nodes installed I think. I will look for the VAE and try that if I can find it.

[sly sundial]

It's baked into the models.

[merry ermine]

Good to know. If I could just find a loader that let me use the checkpoint that includes the model and all the clips at the same time, and had an output for the clips .. that would be cool.

[cursive ridge]

Search civitae for SD3 Vae, there are 2 🙂

[merry ermine]

Went on CivitAI and found a new workflow and a VAE so I am going to try a few things. Maybe I will also take a break and do something else for awhile. Thanks so much for your help 🙂

[cursive ridge]

I discovered a new comfyui thing today 🙂

[sly sundial]

You know there are example workflows for this? https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/comfy_example_workflows

[craggy basin]

@merry ermine Download this image and drag and drop to your ComfyUI. This is a simple workflow. Then place the three text encoders, I, G, and fp16 in the Clip folder and the plain SD3_medium model in the Checkpoints.

Before starting, go to the Update folder and double-click Update ComfyUI

[merry ermine]

I did something to work finally but it was with a workflow I found on CivitAi. It lets me use the model with the text encoders included ... when I try to use the indiviual text encoders I get an error. But at least I got a dog. Such as he is.
And yes, I did update Comfy first.

Thanks everyone - I at least have a work flow that works so I will play around with that. 🙂

[craggy basin]

One or all those steps will solve any issues

[merry ermine]

Prompt was just "dog"

[marble mirage]

module errors usually means python is missing something. cv2 is computer vision pip install opencv-python

[red dragon]

I already done that

but still giving me the same error

[marble mirage]

idk unless you did a system wide python install of it and comfyui is using an isolated venv that you didnt install to

[coral siren]

I'm actually looking for a cascade solution, thought there might be a general node for all models

[hardy tree]

@storm folio just want to confirm, is the new taesd3 from here: https://github.com/madebyollin/taesd ?

[storm folio]

yes

[hardy tree]

cool ty

[lavish stone]

i don't know if sdxl conditioner works for cascade, you could try it

[coral siren]

at least with Standard Cascade there is an individual Conditioner node, yet just that one, so I don´t know

we´ll see

[coral siren]

Found a solution (working for all models). Simply put a TransformImageResizeAbsolute directly behind the Load image one and I can simply set the according Resolution there:

Thank you for your help anyway

[lavish stone]

ah i did not know you wanted to resize the image, in your case i decode the image with vae then encode it simply. I keep the original resolution. If i want to resize i do it before in photoshop

[coral siren]

yes, like said, when swapping inputimages constantly on the fly I would like to have that process automated for "whatsoever" image resolution I put in. Too much effort comparibly otherwise 🙂

[red dragon]

hello

I am getting this error

ImportError: numpy.core.multiarray failed to import

how to fix i t

[coral siren]

btw I wouldn´t have mind if the outputresolution would simply have been altered, without actually resizing the image itself, then it´s enough the way it works now for starters 🙂

[stoic gulch]

can i load workflows into my current workflow?

[coral siren]

don´t know, yet you can simply select/mark one complete workflow and copy paste it over into a new one for a workaround

[brazen turret]

can someone come join me in https://discord.com/channels/1002292111942635562/1005530764886155264 and help me get HiRes Fix working properly, it WAS working last night and now it's totally screwed

[craggy basin]

Has anyone set up Lumina in ComfyUI?

[craggy basin]

Lumina

[steep marlin]

Looks good! Is that with the SFT version(supervised fine-tune)?

[hard marsh]

That Lil guy is so cute!

[grand pike]

Hello lovely people. I recently upgraded my PC with everything but a new graphics card. My comfyui install is on its on seperate hard drive. Would I need to re install comfyu ui?

[stoic gulch]

prob not if its in a venv. the rerequisites and drivers might be annoying

[grand pike]

Venv?

[bleak coral]

Has anyone tried MusePose in ComfyUI? I'm encountering a very strange error.

[stoic gulch]

feeling productive https://i.imgur.com/lA2KYc1.png

[silent knoll]

Hi, where is the best place to hire comfyui devs?

[stoic gulch]

is ther a pytorch 2.3.0 i can download or a xformers that works with 2.3.1? the website didnt work im trying to compile from source now

[distant pendant]

just install from https://pytorch.org/get-started/previous-versions/#v230
compiling torch is a nuclear option

[stoic gulch]

yes i was typing pip install pip install i didnt realize it was blocking the install

i keep getting this when i'm swapping between torch versions and i always forget how i fix it. OSError: [WinError 126] The specified module could not be found. Error loading "P:\ComfyUI-ZLUDA\.venv\Lib\site-packages\torch\lib\caffe2_nvrtc.dll" or one of its dependencies.

[stoic gulch]

this means you're not installing cu118/cu12x

[cursive ridge]

@storm folio might know some

[mystic plaza]

in comfyui we can mention from which direction the light rays are falling from right

[sly sundial]

The UI has no impact on this. It would depend on the interactions between your prompt, the model and the workflow you use.

Mostly the first two.

[silent knoll]

Thanks!

[fading mural]

can you load lora by saying lora:loraname

in prompt?

[mystic plaza]

i see

[sly sundial]

If the prompt node supports it you can.

[mystic plaza]

its just that i am trying to match original image lighting with the background i am generating

[fading mural]

how do i know if it does?

[sly sundial]

Try it, or read the repo details This one does https://github.com/ltdrdata/ComfyUI-Impact-Pack?tab=readme-ov-file#:~:text=ImpactWildcardEncode

[brazen turret]

btw this doesn't work in ComfyUI

when you load the lora you can adjust the model and clip weight before clip to your prompts. That's how lora is applied in comfy ui

I'm using ComfyRoll in order to load a LoRA stack and then apply that stack

[hardy tree]

interesting lora names there

[brazen turret]

yea too bad half of them destroy generation

[verbal sky]

Fun times for all - new video just dropped, please enjoy and SHARE with a billion people so I can overtake Mr. Beast...muhahah... video tutorial topics include Boaty, Updated Loader, Image Switching, Dynamic FX, Photopea layer save/retrieval within Comfy - https://youtu.be/wDJm0qkgIOE

[stoic gulch]

will comfy run on falkon browser? im running out of vrams

https://i.imgur.com/2XSBL0h.png

[stoic gulch]

https://i.imgur.com/JuV7QYM.png getting black boxes when i am trying tro generate videos

https://i.imgur.com/GNYa6NL.png

[stoic agate]

hey guys

is there a way to do multiple instant id on more than one face in the image?

something like multi ID?

[cursive ridge]

So the whole SD3/SDXL merge thing, can you also use loras with that?

Btw, anyone have a workflow for it? I haven't found one that doesn't have dozens of errors, usually about untitled nodes missing, even after I install missing custom nodes.

[soft pagoda]

you can use attention masks https://www.youtube.com/watch?v=4jq6VQHyXjg&t=92s

[stoic agate]

thank you

[remote ore]

Hi Everyone. Here is the link to the DeepFuze Repo: https://github.com/SamKhoze/ComfyUI-DeepFuze/tree/main

Please note that currently DeepFuze is optimized for Mac MPS+CPU, and Windows only supports CPU. If you are testing on Windows, make sure to choose your device as CPU on the nodes. We are working on enabling CUDA, and it will be released in a few days.

The playback nodes will be added shortly, so you will be able to hear the playback of the cloned voices. For now, it will merge to the lipsync, or you can connect the output to a play audio node.

Best Regards,

[bleak crypt]

@thorny lily Um.

[eager escarp]

use the exclamation emoji, it triggers an alert to the mods