KSamplers with variation seed (aka Noise Hijack)

This node pack is retired

Thanks to noise being elevated to a proper datatype in Comfy, these nodes have been replaced with the much simpler, more flexible, and more reliable NoiseTools. Go get them there!

This node pack is retired - below is for historical interest only!

Custom nodes that replace KSampler and KSampler Advanced (or hijack others - see below) and allow you to generate small variations in the initial noise.

Conceptually, noise = random_based_on(seed) is replaced with noise = random_based_on(variation_seed) * x + random_based_on(seed) * (1-x) for some weight x. x=0 generates an identical image to the normal nodes, small values of x generate similar images.

x=0	x=0.02	x=0.04	x=0.06
x=0.08	x=0.1	x=0.12	x=0.14

Updates

1.1

• Main variation node now uses ContextManager to ensure noise function is always released
• Hijack nodes better at checking for inconsistent use
• Fixed saved workflows!

Usage

Once installed, Download this example workflow and check it works. Note that the variations seem to behave differently with different samplers and schedulers, so you might need to play around with those settings.

You can use the new nodes (found under sampling as KSampler with Variations and KSampler Advanced with Variations - for use with other KSamplers, see below) just like you use the original KSampler nodes, with

• seed is the seed for the original image
• control_after_generated is set to fixed by default
• variation_seed is the alternative seed
• variation_weight is the weight - typically quite small (try 0.1)

Batch inputs

If the batch size is 1, the node will generate a single image with the specified variation_weight.

If the batch size is larger than 1, the node will generate a set of images all based on the same original noise. The first image will have variation_weight of zero, the second will have variation_weight as set in the node, the third will have variation_weight of 2x the value set in the node, the fourth will have 3x etc..

Note that the way some samplers work means that batches don't really work like you'd expect. So if you have issues with batches, try other samplers!

Or download a workflow using any of the images in the example at the top.

Other KSamplers

There are two custom nodes under noise/hijack. Place the Hijack node directly before your KSampler, and the Unhijack node directly after it. Set control_after_generate to fixed. Like this, but with your KSampler:

SDXL Note

If you use the refiner, the noise is added by the first KSampler, so that's where you do the variations. Using the variations node on the second KSampler (where add_noise is disabled) doesn't do anything.

Installation

Use the ComfyUI Manager and search for chrisgoringe or variation.

Manual install:

cd [Comfy Install]/custom_nodes
git clone https://github.com/chrisgoringe/cg-noise.git
git clone https://github.com/chrisgoringe/cg-custom-core.git cg_custom_core

Then restart ComfyUI and reload the webpage.

Update

Use the ComfyUI Manager

Manual:

cd [Comfy Install]/custom_nodes/cg-noise
git pull
cd [Comfy Install]/custom_nodes/cg_custom_core
git pull

Then restart ComfyUI and reload the webpage.