Conditioning Sigma Gate

Some prompts are only useful for part of the denoise. The classic case: a "composition" prompt that should be loud while the latent is noisy and mostly silent once the frame's decided - or a detail booster you only want in the back half, so it refines without dragging composition around. Conditioning Sigma Gate is the on-ramp for that: it takes any CONDITIONING wire and restricts it to a window of sampling.

It's a gate, not a schedule. You pick where it opens and closes, and outside that window the conditioning is simply inactive. Think of it as "this prompt is only allowed to speak between step X and step Y."

Percent vs. sigma

Two ways to set the window, chosen with the mode combo:

• denoise percent - the plain 0..1 language (start 0 = first step, end 1 = last). No model needed, which makes this the mode you'll actually use for day-to-day work.
• sigma - real sigma values instead. This is the technically correct view of the scheduler's noise ladder, but a conditioning can only store a percent range, so sigma has to be converted. Wire a model in for an exact conversion using its own schedule, or a sigmas schedule as an approximate lookup table. The README's honest caveat: it's exact for normal schedulers, soft for Karras/exponential, because those are non-linear in percent.

So for a typical workflow you'd ignore sigma entirely: mode = denoise percent, start = 0, end = 0.6, and your prompt now governs only the first 60% of the denoise. If you want a negative-conditioning tweak that only shows up at the end, flip it: start 0.6, end 1.

Why it stacks safely

The bit that makes it safe to chain is that the gate intersects any range already on the conditioning rather than overwriting it. If a previous node already tagged the wire with a window, the gate narrows it instead of clobbering it. That means you can stack gates - one that trims the front, one that trims the back - and they compose cleanly. The README notes it also plays nice with the pack's Conditioning Blend ramp instead of fighting it.

Mechanically this is one of the pack's few nodes that touches sampling math rather than just text, but it still does it the ComfyUI way: it rewrites the start_percent/end_percent metadata on the conditioning, so there's no model patching and no slowdown. Drop it on any wire - a CLIP encode, a Fusion encode, anything upstream - and the one output goes back into the graph like nothing happened.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzExperimental

or search "Nynxz" in ComfyUI Manager, then restart. No dependencies, no model downloads, just a reasonably current ComfyUI - these nodes use the newer V3 node API, so update ComfyUI first if you see missing nodes. Node ids and defaults can shift between commits; that's the price of "experimental."