SplitSigmasDenoise_motorway_edition

Inside a custom-sampling workflow, sigmas are the denoising schedule - the list of noise levels the sampler walks through from heavy noise down to clean image. SplitSigmasDenoise cuts that list in two at a denoise fraction, so you can run one pass on the noisy half and a different pass on the clean half. This is the Motorway version: the sigmas come out of the pipe by key, and the two halves go back in under keys.

The classic use is the two-pass setup people build by hand: denoise from 1.0 down to 0.3 with one sampler or model, then hand the latent to a second sampler for the 0.3→0 tail, often at higher resolution or with a different scheduler. Doing it with SplitSigmasDenoise means you're explicitly feeding the second stage the remaining sigmas rather than approximating with a denoise slider - which is the difference between knowing your math and hoping your math works.

What to set

• INPUT_sigmas_key (default sigmas) - the Motorway key holding the SIGMAS from your scheduler (a BasicScheduler or KSamplerSelect+SamplerCustom chain).
• denoise (default 1, 0–1) - the cut point as a fraction of the schedule. 0.5 splits the list in half; 0.3 keeps the first 30% of the noisy steps on the "high" side.
• OUTPUT_high_sigmas_key (default high_sigmas) - where the noisy first-half sigmas land.
• OUTPUT_low_sigmas_key (default low_sigmas) - where the clean second-half sigmas land.

Output is the MOTORWAY with both halves stored. Feed high_sigmas to a SamplerCustom for pass one, then low_sigmas to another for pass two. Note the outputs are sigmas, not latents - the latent is just the thing you keep passing along the wire between the two sampler stages.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/agilly1989/ComfyUI_agilly1989_motorway

Restart; it's under agilly1989 Nodes → Motorway-ed → sampling → custom_sampling → sigmas. ComfyUI Manager installs the pack too. No pip dependencies.

Where people get burned

Key errors are the same everywhere in this pack: 'MotorwayClass' object has no attribute 'hash_' means the sigmas key isn't in the Motorway - check the spelling, it's case-sensitive, and make sure your scheduler ran before this node.

The logic trap is subtler: denoise is a fraction, not a step count, and it's relative to whatever schedule your scheduler produced. A long schedule splits at a different absolute step than a short one, so the same 0.3 gives you different behavior depending on the scheduler and step count upstream. If the second pass behaves differently than you expect between runs, that's the first thing to check. And as with the whole pack: it's beta, actively changed, and motorway-ed nodes have come and gone across releases - an update hiding the node isn't your fault.