MiniMax H3 Sigma Scheduler

MiniMax H3 is the 33B omni-modal video model that opened weights in August 2026 and does video and synchronized stereo audio out of one context. Video of that size needs a schedule with actual opinions - it's a flow-matching model with distinct video and audio shift values - and the MiniMax H3 Sigma Scheduler is Sigmax's attempt to make that explicit without letting you hang yourself. It's the most feature-dense node in the pack, so let's separate the solid part from the experimental part.

The solid part: pick H3 Base FL2VA or H3 Base Ref2VA in variant, keep scheduler on its default h3_endpoint, and the node builds the pure endpoint-inclusive H3 video schedule. That default path needs no MODEL at all - it's dependency-free, just sigmas for the base recipe with the standard 12.0 video / 3.0 audio shift values.

The experimental part, three layers deep:

• scheduler - a ten-choice menu. h3_endpoint is the safe default. The other nine (simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal) are ComfyUI-native compatibility lanes that require an already-shifted H3 MODEL wired into the optional model input. The README is emphatic: connect the model only after an upstream MiniMaxH3SigmaShift node, keep the video/audio shifts aligned (12.0/3.0 for Base, 6.0/3.0 for the 768p recipe), and don't add a separate BasicScheduler. These are compatibility options, not MiniMax recommendations.
• turbo - a selector for four source-qualified community lightx2v Turbo recipes (e.g. h3.fl2va.lightx2v-turbo-4-v0.1-544p at 4 steps, h3.fl2va.lightx2v-turbo-4-v1.0-768p at 4 steps / 1344×768 with the 6.0/3.0 shifts). The node only constructs recipe-owned sigmas and a readiness receipt - it does not load or patch the LoRA. You load the matching lightx2v Turbo LoRA separately in the workflow. The tooltip's own words: "Experimental - community, unsupported."
• recipe_id - a legacy alias for turbo. Prefer turbo.

Inputs and outputs in one pass

variant (FL2VA/Ref2VA), steps (default 20), start_step/end_step for slicing, plus the optional turbo, recipe_id, scheduler, and model. Outputs: sigmas and schedule_info. Base video/audio shifts are 12.0/3.0; the 768p FL2VA turbo recipe uses 6.0/3.0.

Installing

Standard Sigmax install:

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

Restart, search Sigmax. ComfyUI Manager has "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies, no model downloads.

Where people get burned

The #1 support case, per the README's own troubleshooting table: you switch scheduler off h3_endpoint and the node asks for a MODEL or reports a shift mismatch. That's expected - the native lanes need the already-shifted model, the correct shift values for the chosen recipe, and no second scheduler. The second trap is the turbo selector: people assume selecting a recipe loads the LoRA. It doesn't. Recipe sigmas without the matching LoRA is just an unusual schedule. And remember the H3 community license geofences the weights out of the US, EU, UK, and South Korea - the schedule node runs anywhere, but the model you pair it with may not be licensed where you are.