Is there any way to do basic maths or conditional loops?

[Extraltodeus]

Like loop for X times in the k-sampler while changing the seed every time, or have a node that would perform simple maths? Or a router that would route somewhere if condition else somewhere else? These feats would be more than interesting but I'm not finding them.

[GalaxyTimeMachine]

Perhaps this will help? https://github.com/pythongosssss/ComfyUI-Custom-Scripts/blob/main/py/math_expression.py

[Extraltodeus]

That is nice indeed! thank you! Still needs a way to loop however.

[ntdviet]

@GalaxyTimeMachine please share the name of the custom node you use to get these rectangular lines, thanks

[GalaxyTimeMachine]

Isn't this in your ComfyUI settings? I thought it had been added there now.

[ntdviet]

@GalaxyTimeMachine oh shoot, I didn't catch up with the update, since when?...
Thanks btw

[Extraltodeus]

I ended up doing a lot of custom nodes in the end, including a way to save conditioning which can be then use as some sort of text embedding, also two clip interrogator nodes :)

https://github.com/Extraltodeus/CustomComfyUINodes

[ltdrdata]

Can the configuration of the repo be organized into a form that can be managed at once by creating __init__.py? In the current setup, since it is configured as it is, updates and other support cannot be provided as individual file units from the ComfyUI-Manager.

[Extraltodeus]

The ComfyUI-Manager?
I've dwelved into the nodes but I don't really know what you are talking about.

[ltdrdata]

The ComfyUI-Manager?
I've dwelved into the nodes but I don't really know what you are talking about.

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

This is the ComfyUI-Manager.
I am updating the DB periodically.

[ntdviet]

The noise injector does not inject anything if the noise latent is not at the same shape. Check your terminal if it is not warning you about it. If you try one of the workflows included in the shared image in the discussion with city96 you should probably be able to get an even higher quality.

Ah ok, so I can not inject 1056x1056 noise into 1024x1024 latents? good to know thanks. It didn't throw any warning for me. Thought I could cheat for more details by having a higher resolution noise map for smaller latents.

I used your guy's workflows to compare perlin noise with regular leftover noise and it was pretty disappointing for my SD1.5 checkpoint, totally blurred. The version with leftover noise got it pretty decent, though not perfect as you already stated.

Then I got surprised by the model pic, and yes, for model closeups, perlin noise performs pretty well.
In this kind of extreme pics, it outperforms normal noise by a lot :))

The hires-fix of faces is clearly better and more consistent with perlin

So why is it shit for buildings but good for characters? Because of the checkpoint I used? Can I have yours? Couldn't find the ckpt you used in your workflow

[Extraltodeus]

@city96 In my experience you always have to use the model used to generate the image to get the right sigma.

Also, I think it would be best to start a new discussion topic here on the main ComfyUI repo related to all the noise experiments. That way we can collect everything centrally instead of having it spread out over multiple issues/discussions/repos.

Agreed! The issue open on your repo might be a good place unless you have other thoughts.

[Extraltodeus]

@ntdviet but here again you are doing the same mistake of not using the same resolution! That's why you result is blurry. This is because your second pass does not actually have noise.

[ntdviet]

@Extraltodeus oh damn, ok, sorry for the error, I thought I checked everything. But then why is it working for the girls :))
Edit: sorry I'm the idiot, I got the upscale correctly for the girls, x2

[city96]

@Extraltodeus

My personal vote is for this current repo (the main ComfyUI one), since that's the most likely to get traction/other people joining in.

The issue on my repo could work, but I prefer discussions since you can have multiple top level comments at the same time for different things.

[ltdrdata]

Pretty nice! What do you need me to do? I wouldn't recommand to add each and every of my nodes tho. The most interesting might be the clip interrogator. The oobabooba is an adapted copy of another one so to make it work with the same name webui for the llama models for example so I don't know if it would be appreciated globally since it's relatively not from me. The rest of it is more or less a treasure cove. There are good and half-assed things.

This depends on the developer of the node. They can choose to provide useful nodes and miscellaneous nodes separately in a separate pack, or they can offer options in the configuration to enable or disable certain functionalities.

In my case, I provide three separate extensions based on the nature of the nodes. Users can receive the specific set of nodes/extension they desire.

On the other hand, for legacy nodes with messy dependencies, I offer them in a disabled state, and provide users with a configuration option to selectively enable them if they choose to.

You can choose how you want to provide it to the users, depending on your preference. However, I would suggest offering it in a form that can be easily installed using git clone, including __init__.py. This way, it can be easily tracked for updates when you make changes in the future.

[Extraltodeus]

I've done one that is worth sharing, I've included the init file too!
https://github.com/Extraltodeus/noise_latent_perlinpinpin