Mochi Prepare Sigmas

This is the most forgettable node in the pack and also the one that decides whether your Mochi edit looks clean or slightly dusty. Mochi Prepare Sigmas takes the sigma schedule from the Mochi Sigma Schedule node (the one that lives in kijai's ComfyUI-MochiWrapper) and does exactly one thing to it: if the schedule doesn't already end at exactly 0.0, it appends a 0.0 to the list. That's the entire node.

Why one appended number matters

Sigmas are the noise levels a sampler walks down, and a schedule that stops before true zero stops denoising early - your output keeps a layer of residual noise baked in. Visually that's grainy, hazy video that never quite snaps into focus. By forcing the last sigma to zero, this node guarantees the run goes all the way to a clean latent. It's the kind of defensive fix you'd write yourself after your third gray, soft result, except here the author wrote it for everyone.

The tooltip on the input says it plainly: "Override sigma schedule and steps." One SIGMAS in, one SIGMAS out. There's genuinely nothing else to configure, and that's the point - it's a pass-through that patches the schedule, nothing more.

Where it goes in the workflow

The pack's RF-Inversion edit workflow uses this node twice, and the two uses treat the output differently:

• Unsampling pass (video → noise): prepare sigmas, then flip them before they hit the sampler, so the schedule runs from clean video toward noise. The flip happens in a separate node (the core ComfyUI FlipSigmas node).
• Resampling pass (noise → edited video): prepare sigmas and feed them in unflipped.

The easiest way to mess this up is to reuse one prepared schedule for both passes and forget the flip is only for the unsampling side. When people say "the edit came out blurry/noisy," it's usually this node (or its placement) that's wrong, not the model.

Install

There's no Mochi support in this pack itself - it builds on the wrapper, so install that first:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-MochiEdit

No extra Python dependencies (there's no requirements.txt), just a restart. Same as every node in this pack, if anything errors after an update, git pull / update the wrapper first - the two repos were merged by their authors and drift between them is the usual breakage. And don't expect this pack to carry you into 2026: it's a November 2024 specialty tool for Mochi, a model the community largely left behind for LTX and Wan. But for its niche it does one job, and this little sigma patcher is quietly doing it well.