Nodes Browser
ComfyDeploy: How SD-Advanced-Noise works in ComfyUI?
What is SD-Advanced-Noise?
Nodes: LatentGaussianNoise, MathEncode. An experimental custom node that generates latent noise directly by utilizing the linear characteristics of the latent space.
How to install it in ComfyDeploy?
Head over to the machine page
- Click on the "Create a new machine" button
- Select the
Editbuild steps - Add a new step -> Custom Node
- Search for
SD-Advanced-Noiseand select it - Close the build step dialig and then click on the "Save" button to rebuild the machine
Colored Noise / Advanced Noise / Latent Noise Experiments
It is only fair that I shout out that all the other noise related repos such ash ComfyUI_Noise, noise_latent_perlinpinpin and comfy-plasma. I think the WAS node suite also has quite a few noise-related nodes.
While messing around with the stable diffusion VAE, I noticed the latent space behaves mostly linearly. I figured it should be possible to generate latent noise directly if I map out the per channel limits and give the whole thing an offset.
This repo also has some decoders, but those are a proof of concept tier at best. TAESD is better in every way.
LatentGaussianNoise
I'm not sure if the way I coded this even makes sense, or if it's even gaussian noise. It's just torch.random with a bunch of stuff like scaling/per channel random/etc.
Linearity / linear_encoder
This is the simplest encoder. It uses the fact that a change in the RGB channels creates a (mostly?) linear change in the 4 latents channels (which I just called A/B/C/D since I couldn't find any info about them).
The next step would be to plot at a higher precision and fit them onto a polynomial.