ModelNoiseScale (Ray)

ModelNoiseScale (Ray) is a tiny node with a specific job: rescale the noise that gets injected at the start of sampling. It's the kind of utility that looks pointless until the day a model comes out with a noise-scaling quirk and this becomes the exact lever that fixes it.

Here's the context. Flow-matching models (Flux, LTX, AuraFlow, and their cousins) don't all initialize a generation the same way - some expect a "noise scale" baked into how much noise their initial latent gets. When ComfyUI's model config for a checkpoint gets the scaling slightly wrong, you see it as either generations that start from too much chaos (wild composition, things melting) or from too little (everything looks pre-baked and stiff). The noise_scale value multiplies that initial noise - 1.0 is the neutral default, above 1 adds more chaos, below 1 calms the start down.

Mechanically the node works by copying the model's existing sampling object, preserving its shift and multiplier parameters, and swapping in the new noise scale. That detail matters more than it sounds: it means you can adjust noise scale without clobbering a shift you set elsewhere - the two levers compose instead of fighting. It's a port of ComfyUI's stock ModelNoiseScale, adapted to the raylight distributed pipeline.

The inputs

• noise_scale - float, default 1.0, range 0 to 64. 0 effectively disables initial noise scaling; values above 1 are for when the model starts too clean.

That's the whole node. ray_actors in, ray_actors out, sits in the chain before your Ray sampler.

When it actually matters

If you're running a model whose recommended workflow includes a ModelNoiseScale node - some checkpoints' official workflows literally ship one - you need the raylight equivalent so the setting survives the trip into distributed mode. That's the real use case for this node, and it's why the pack bothers porting something so small: a multi-GPU run has no patience for a missing one-knob patch that quietly changes results.

Other than that: leave it at 1.0. It's a fix-it node, not a tuning toy. If you don't have a specific reason to touch it, touching it will just add a variable you don't need. Install is the standard pack route - ComfyUI Manager, or clone + pip install -r requirements.txt - and it needs no model files of its own.