MiniMax H3 AV Sigma Tail Subdivision / 音画尾段细分 (Advanced)

The tail of a flow-matching schedule is where the details get settled - and also where most samplers rush through. MiniMaxH3AVSigmaTailSubdivisionT8Advanced is the experiment that says "fine, keep every sigma knot you already have, and add a few more points right at the end, in base-flow time." Every input knot is preserved; extra_substeps decides how many extra full H3 audio-video DiT calls you're willing to buy. Zero is an exact pass-through, so this node is safe to drop into a workflow even if you're not sure you'll use it.

Where its sibling MiniMaxH3AVSigmaSameNFERedistributionT8Advanced moves existing points without adding calls, this one adds calls - so the first decision is whether you have the budget. actual_nfe output tells you exactly what you bought: original knots plus the extra substeps. Default mode is report_only, and the tooltip on extra_substeps is refreshingly plain: "Total additional full H3 audio-video DiT calls. Zero is an exact pass-through."

The inputs

• sigmas in, sigmas out - drop it between scheduler and sampler.
• extra_substeps (0) - how many extra calls to insert. Start at 0, read the report, then try 1–3.
• range_mode / tail_intervals (1) - where the new points go. tail_intervals picks how many original intervals, counted back from the zero-sigma tail, get subdivided. Default 1 = only the very last interval gets refined, which is the conservative reading of "tail."
• start_progress (0.75) / end_progress (1.0) - an explicit alternative window if you don't want to count intervals.
• spacing (base_time_linear) - how the inserted points are spread in base-flow time.
• shift_video (12) / shift_audio (3) - the dual-clock shifts, matching the pack's standard defaults.
• accept_turbo_schedule_ood - the same honest gate as the sibling node: 8-step Turbo schedules weren't trained with intermediate points, so you have to explicitly accept the out-of-distribution territory before the apply mode will run.

It reports both clocks - video and audio - which matters because H3's audio clock (40Hz) and video clock (24fps) can resolve the tail differently. If the audio feels unresolved but video is fine, this node's per-clock reporting is how you'd confirm which one actually got the extra attention.

Install

Standard T8mars pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart ComfyUI. It's in T8/MiniMax H3/Quality/Experimental, no extra dependencies, no model files. Try it on a short clip where you can hear the difference: 22 frames, extra_substeps 2–3, and A/B against your normal schedule. If the tail cleanup is real for your content, you've found a quality lever that costs a couple of extra forward passes instead of a whole re-run at higher NFE.