Lumina-Image 2.0 Sigma Scheduler

Lumina-Image 2.0 is the Alpha-VLLM text-to-image model whose whole scheduling story is one number: a fixed unit-flow ratio of 6.0, at 50 steps. That's it. No dynamic shift, no resolution-dependent curve - a fixed ratio, pinned. The Lumina 2 Sigma Scheduler builds exactly that schedule, and its entire job is to make sure you get it right instead of approximating it through a generic scheduler.

You'd think a one-number schedule would be trivial in any workflow, and mostly it is - right up until you realize that ComfyUI's generic schedulers aren't designed to hit an exact fixed ratio, and that getting 6.0 wrong by even a little visibly changes how denoising effort is distributed between composition and detail. This node removes the approximation. It's the smallest node in the pack's image-scheduler family, and it's honest about it.

Inputs

• mode - one choice, Official Fixed (6.0). The single-option dropdown looks silly until you remember the pack's rule: explicit selection, no silent fallback. It's the author's way of saying "you're choosing this on purpose."
• steps - default 50, the official count.
• strict_source - keep on to stay pinned to the official recipe.
• already_shifted - set true only if a model-sampling shift is already applied upstream in your workflow.
• start_step / end_step - slice.

Outputs are sigmas and schedule_info. sigmas goes straight to your custom-sampling path; schedule_info gives you the recipe and any warnings.

Scope, in the author's own words

The pack's compatibility notes are unusually precise here: support is limited to the original Alpha-VLLM text-to-image schedule with fixed ratio 6.0 and 50 steps. Video, mGPT, editing paths, dynamic shifts, execution, and quality claims are all explicitly excluded. So if you're running a Lumina video model or some fine-tune that does its own shift, this node is not aimed at you - treat its output as a reference point, not a prescription.

Installing

One pack, all nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart, search Sigmax. ComfyUI Manager lists it as "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies, no model downloads - the node emits sigmas, not weights.

Where it bites

The standard pack gotcha: feed sigmas directly to the sampler and don't stack a second scheduler or time shift, or you'll be applying 6.0 twice. And the scope gotcha: don't assume this one fixed schedule generalizes to every Lumina variant. For the original text-to-image model at 50 steps with CFG around 4, this is the schedule. For anything else, it's a starting point at best.