Flip Flux Sigmas

InFluxFlipSigmas does one small job and does it carefully: it takes a SIGMAS tensor, reverses the order, and guards the edge case where the first value after flipping is zero. Flip the sigmas and an inversion workflow is walking the schedule backwards; skip the guard and you can run straight into a division-by-zero a step later.

Why the flip matters

Normal sampling goes from high noise to low noise - sigmas descending from ~1 toward 0. Inversion is the opposite direction: you start from a clean latent and add noise back in, so the schedule has to run up. The standard way to express that in ComfyUI is to flip the sigma array so the sampler consumes it low-to-high. Stock ComfyUI actually has a FlipSigmas node for exactly this, and you'll see it in the pack's example RF-Edit and Fireflow workflows. So why does this node exist?

The difference is the guard. After a flip, the first sigma is whatever was the last value of the original schedule - and the end of a sampler's schedule is frequently 0.0 (the "clean" step). Downstream code that normalizes or divides by sigma can choke on that. InFluxFlipSigmas checks: if the flipped head is 0, it bumps it to 1e-3, a tiny nonzero value that keeps the math stable without meaningfully changing the schedule.

Inputs and outputs

• sigmas (SIGMAS) in - your schedule, typically from a BasicScheduler or one of the pack's model-sampling nodes.
• SIGMAS out - the reversed schedule with the zero-guard applied.

You place it between the scheduler and the sampler whose direction you want inverted. In the RF-Edit pattern, the forward sampler runs down the flipped schedule while the reverse sampler runs back up it, which is how the same latent gets encoded and then decoded around an edit.

Installation

Same as every node in this pack: ComfyUI Manager → search "ComfyUI-Fluxtapoz" → install → restart, or clone https://github.com/logtd/ComfyUI-Fluxtapoz into custom_nodes. No model files, no extra dependencies beyond einshape.

Common issues

Honestly, this node is hard to misuse - it's one input, one output, deterministic. The two things to know: if your workflow already has a stock FlipSigmas and it works, you don't need to swap it; this is the variant that handles the zero edge case defensively. And if your inversion produces NaNs or infinities, check whether some flip is happening at all - an unflipped schedule feeding an inversion sampler is a silent failure that looks like garbage output, not an error message.