☁️BizyAir SplitSigmas

A sigma schedule is just the list of noise levels a sampler steps down through, start to finish. Most of the time you hand that whole list to one KSampler and never think about it again. But a growing number of workflows need to hand off half the schedule to one model and the rest to another - Wan 2.2's two-pass architecture is the clearest example, where a high-noise model handles motion and composition for the early, noisy steps, then a separate low-noise model takes over for the later steps that refine detail. SplitSigmas is the node that cuts a schedule into exactly two pieces so you can route them to two different samplers.

It's a small, honest utility node, and worth calling out explicitly: unlike most of this pack, it isn't a cloud API call. Its input and output types are plain SIGMAS, not one of BizyAir's BIZYAIR_* types, which means it's just local schedule math - no API key needed to get this one working, even though the rest of the pack expects one.

How it works

You give it a sigma schedule (from something like a scheduler node upstream) and a step index. It splits the list at that index and hands you back two separate schedules: everything before the cut, and everything after. Feed the first to your first sampling pass and the second to a resume pass on a different model or with different settings, and the two passes together cover the exact same ground the original single schedule would have - just split across two stages.

Inputs and outputs

• sigmas - the schedule you're splitting, type SIGMAS.
• step - an integer from 0 to 10000, defaulting to 0, which is the index to split at. If your schedule has 20 entries and you set step to 8, you get the first 8 in one output and the remaining 12 in the other.

It returns two outputs: high_sigmas (the earlier, noisier portion of the schedule - the part closer to pure noise) and low_sigmas (the later portion, closer to the clean image). Feed high_sigmas into your first-pass sampler and low_sigmas into the second-pass sampler that picks up where it left off.

If you'd rather split by how much denoising has happened rather than a raw step count, look at BizyAir SplitSigmasDenoise instead - same idea, different cut point.

Installing BizyAir

Even though this specific node doesn't need the cloud, it ships inside the same pack as everything else, so installation is identical. Easiest is ComfyUI Manager: search "BizyAir," install. Or clone it directly:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

then restart ComfyUI. Comfy-CLI users can run comfy node install bizyair. You'll still want to set an API key at some point if you plan to use any of the pack's other (cloud-backed) nodes, but this one specifically will run without it.

Common issues

The most common mistake is picking a step that doesn't match your intended split - if you want a roughly 50/50 handoff between two models but your schedule has 30 steps and you set step to 8, you've actually given the first model way less of the job than intended. Count your total steps before choosing a split point, or better, work backward from the fraction you want (a third, a half) times your total step count.

The other trap is architectural: this only makes sense if you actually have two samplers set up to consume the two halves. If you're not doing a two-model or two-pass workflow, you don't need this node at all - just run your full schedule through one KSampler.