FlipSigmas

FlipSigmas does exactly one thing: reverses a sigma schedule. High-to-low becomes low-to-high. That sounds trivial until you realize how much of the "trick" workflow space is built on it.

A normal denoising run takes a latent from noisy to clean - schedule runs high to low. Flip that schedule and you get the reverse: a pass that starts from a nearly-clean latent and adds structure upward. That shape is useful in a surprising number of places, and rather than hand-building inverted schedules by hand, this node gives you the flip as a one-input operation.

How it works

The implementation is honest about being simple: it reverses the tensor's order, then handles one edge case - if the schedule starts at 0 after flipping (which is where it ends for a normal schedule), it nudges that 0 up to 0.0001. A literal zero as the first sigma would mean "start from a perfectly clean latent," which breaks the sampling math; the tiny epsilon keeps the solver honest without meaningfully changing the result.

The input and output

• sigmas (SIGMAS) - any schedule.
• SIGMAS output - the same schedule, reversed, with the epsilon guard applied.

It's a pure data transform. Wire it between a scheduler and a sampler and you've inverted the pass.

What people actually use it for

• Upscale / refinement passes. Some img2img-style refiners run a "low-to-high" schedule so the pass starts gentle and finishes with a full-strength noise push - the opposite polarity from normal denoising, which is exactly what certain texture-pass recipes call for.
• Building a noising stage by hand. Pair a flipped schedule with AddNoise and you control noise intensity as a curve rather than a fixed level.
• Normalizing shared workflows. When a downloaded workflow needs a schedule pointing the "other way" and you don't want to hand-type sigmas, this is the one-click fix.

Common issues

• The 0.0001 surprise. If you inspect the output and see a tiny value where you expected a clean 0, that's the guard working. It's harmless, but it trips people up when they're comparing schedules numerically.
• Expecting it to preserve a nice curve. Reversing a Karras curve doesn't give you a "good" low-to-high schedule - it gives you a mirrored one. Fine for tricks, wrong if you actually wanted a purpose-designed ascending schedule (reach for ManualSigmas).
• Using it as a substitute for denoise. They're different tools. denoise truncates a schedule; FlipSigmas reverses one. Conflating the two is a classic source of "why does my img2img look like this" confusion.

It's the kind of node that looks pointless on the shelf and essential the first time a workflow needs its polarity inverted - two minutes of "wait, I need the schedule backwards" and suddenly you're glad it ships in core.