ExtendIntermediateSigmas

ExtendIntermediateSigmas takes a SIGMAS schedule and inserts extra, interpolated steps between the existing ones - but only within a sigma range you choose. It's a refinement tool for noise schedules: your sampler keeps the schedule's overall shape, just with more resolution inside a specific window. Search it under "interpolate sigmas," which is exactly what it does.

It ships with ComfyUI core (comfy_extras.nodes_custom_sampler), and it's a recent arrival - and it has one of the more unlikely origin stories in the node registry: in June 2026 it became famous as the native-node trick for getting around Ideogram 4's safety filter.

How it works

A sigma schedule is a descending list of numbers. This node walks each adjacent pair and, for pairs whose values fall inside [end_at_sigma, start_at_sigma], inserts steps intermediate values interpolated between them. Pairs outside the range are left untouched.

• sigmas - the schedule to extend, e.g. from BasicScheduler or Ideogram 4 Scheduler.
• steps - how many extra steps to insert per gap (default 2).
• start_at_sigma - the top of the range. Default −1 means "no upper bound" (the code maps a negative value to infinity).
• end_at_sigma - the bottom of the range (default 12).
• spacing - the interpolation shape: linear, cosine, or sine. Cosmetic unless you're chasing a specific density.

Output is a longer SIGMAS tensor, still descending, same start and end. Wire it into SamplerCustomAdvanced exactly where you had the original.

The Ideogram 4 story

This is the part people actually Google this node for. Ideogram 4 trains a safety filter into its weights that triggers on a sudden drop in the initial sigma - the first big step from pure noise. The community (credit to silveroxides of the Chroma group) found that if you extend the early part of the schedule with intermediate sigmas, you remove the sudden drop, and the filter stops firing - "The sudden drop in initial sigma triggers the safety, that can be removed by removing the sudden drop" (thread 1tw6gmq, +202). The node then spread far beyond Ideogram as the standard "smooth out the front of a schedule" tool.

That same superpower is the general use case: when a model misbehaves on a schedule that jumps too fast at the noisy end, inserting intermediate steps softens the jump. It's a cheap fix that costs extra steps but no new machinery.

Where people get burned

• It doesn't change where sampling starts or ends. The first and last sigmas are preserved; only the middle gets denser. If you need to trim the schedule instead, that's SplitSigmas territory.
• Range direction confusion. start_at_sigma is the larger (noisier) bound, end_at_sigma the smaller. Get them backwards and the node quietly does nothing.
• It's not a quality guarantee. The Ideogram filter bypass is model-specific, and commenters in the same thread reported it didn't clear the filter for them. Treat it as a well-known trick, not a law.
• Step-count side effects. Inserting steps changes total sampling cost. On a model that needs exactly N steps (distilled models), extra intermediate sigmas can drift off the trained schedule - check whether the model expects a fixed step count before you extend.

If a schedule jumps too fast anywhere in its middle, this is the native node for inserting room to breathe.