TK3R Sigmas Interpolate

The scheduler you pick decides when denoising effort happens - whether the middle steps do the heavy lifting (Karras) or the work is spread evenly (linear). Sometimes the perfect schedule for your model sits in between two presets, and there's no stock node for "halfway between Karras and linear." TK3R Sigmas Interpolate is that node: feed it two sigma schedules with the same step count and it averages each pair, giving you a schedule that's exactly the midpoint of both.

The author's README says "just for fun," and it genuinely is a simple utility - but "for fun" undersells it a bit. The flow-matching era made schedule choice a real dial again, and the KB's own sampler lore is full of people hunting for the balanced schedule their model wants. This gives you a literal dial between two of them.

How it works

It takes sigmas_1 and sigmas_2 - both SIGMAS type, so they can come straight from scheduler nodes or from a custom schedule - and computes (sigmas_1 + sigmas_2) / 2.0 elementwise. One subtlety worth knowing: the two schedules must have the same step count. If they don't, the node logs an error and returns sigmas_1 unchanged rather than doing something clever. So match your step counts first; that's the entire failure mode.

Inputs and output

• sigmas_1 - first sigma schedule.
• sigmas_2 - second sigma schedule.
• sigmas - the elementwise average, same length as the inputs, which you wire into whatever consumes a sigma schedule (a sampler node, or a schedule-modifying node downstream).

There's no blend ratio - it's a fixed 50/50. If you want a weighted blend you'd chain it or bake the weighting in elsewhere; that's not this node's job.

Installing

Part of the TK3R Extensions pack:

• ComfyUI Manager → search "TK3R Extensions" → Install, then restart.
• Or cd ComfyUI/custom_nodes && git clone https://github.com/TK3R/ComfyUI_TK3R_Ext and restart.

Only dependency is rich; nothing to download.

A realistic use

Grab a Karras-scheduled sigma set and a linear one at the same step count, blend them, and see if your model likes the middle. If you're on a flow-matching model that "dislikes aggressive redistribution of denoising effort," a midpoint between an aggressive schedule and a conservative one is often a cheap win. And when the lengths don't match, the node tells you in the console - check your step counts rather than assuming it silently did something. The biggest practical gotcha is just that: it's an averaging node, not a resampler, so it will never fix a step-count mismatch for you.