LC Split Sigma Scheduler

If LC Split Sigmas (Advanced) is the BYO-curves version of two-pass sampling, this is the turnkey one. LC Split Sigma Scheduler takes a model, a scheduler, a step count, a swap point, and a denoise - and builds both halves of the split sigma schedule itself. Model in, sigmas_high and sigmas_low out, no separate scheduler nodes required.

The design is aimed squarely at SamplerCustomAdvanced-style consumers: it's built to hand that node a complete two-model, two-curve schedule in one shot. The primary Model produces the high-sigma (first-pass) schedule; the optional 2nd model produces the low-sigma (second-pass) schedule; and if you leave the second model unconnected, it falls back to the primary - so the node works as a plain single-model scheduler too, and only becomes dual-pass when you plug a second model in.

The inputs

• Model - primary model; builds the high-sigma, first-pass schedule.
• scheduler - the schedule type for both curves (the 9 standard choices; simple is the default).
• total_steps - full schedule length (default 40).
• step_swap - handoff index: high = 0..swap, low = swap..end. If ≥ total steps, the split is ignored and the first schedule runs to completion. Note this one is an input only - it's not re-exported as a socket.
• denoise - strength for building the full schedule (1.0 = full).
• 2nd model (optional) - the low-sigma pass model; falls back to Model.

Outputs: sigmas_high and sigmas_low.

How it works

Under the hood it calls ComfyUI's own calculate_sigmas on each model's sampling space (the denoise trim follows the standard steps/denoise re-expansion trick). Because each model contributes its own curve, the two halves can have genuinely different characters if the models differ. For wiring: you pair sigmas_high with Model and sigmas_low with the 2nd model in the downstream sampler. The LC123 README pairs this with LC Sampler Configure for the settings side - one node owns the numbers, this one owns the schedule.

Install

Part of LC123 - ComfyUI Manager (search "LC123") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes

Restart. No extra pip packages; MIT licensed. Cloud ComfyUI hosts (RunningHub etc.) need LC123 requested or a local install.

The honest take

This is the better on-ramp to two-pass sampling than the Advanced variant: fewer moving parts, fewer ways to miswire. But two-pass is still a technique you should reach for deliberately - for most single-model work (especially on distilled, CFG-1 models), one scheduler and one curve is all you need. When you do want a refiner-style finish, though, this is the cleanest way to build it.