MiniMax H3 AV Tail Detail Schedule / 音画尾段细化 (Advanced)

The gentlest of the three schedule experiments in this pack. While its siblings move sigma points around or subdivide the tail in base-flow time, MiniMaxH3AVTailDetailScheduleT8Advanced does one specific, contained thing: it subdivides only the final H3 dual-clock interval. Default extra_tail_steps of 1 means one extra joint audio-video forward pass right at the end of sampling. That's it. No random noise injected, no wholesale rescheduling - just a finer descent into zero at the exact moment the last details form.

Why bother with such a small change? Because the final interval of a flow-matching sampler is where the last bit of structure snaps into place, and a single extra step there can be worth more than three steps scattered through the middle. The node's description spells out the mechanism for the useful case: extra_tail_steps of 3 adds a gradual 75% / 50% / 25% descent to zero - three progressively smaller hops into the endpoint instead of one big jump. That's a smoothing of the end of the trajectory, and it's the difference between "almost settled" and "settled" on motion micro-jitter and audio texture.

The inputs

• sigmas in, sigmas out - drop-in between scheduler and sampler.
• extra_tail_steps (1) - 0 is an exact pass-through; the node's own default adds one call; 3 gives the gradual descent.
• spacing (video_sigma_linear) - how the inserted tail steps are spaced. The default spaces them in video-sigma space.
• shift_video (12) / shift_audio (3) - the dual-clock shifts, same defaults as the rest of the family.
• profile (turbo_standard8) - the pack's standard 8-step Turbo baseline; all its Turbo dual-clock tests use it.

Outputs are sigmas, actual_nfe (so you can confirm the extra calls landed), and report_json. There's no out-of-distribution consent flag here - unlike its siblings - because the default only adds a step at the existing end of the schedule rather than inserting untrained intermediate times, which is a much safer intervention on a Turbo schedule.

The honest take

This is the one schedule experiment I'd actually reach for first, purely because it's the least invasive. One extra step at the end costs one forward pass and can't derail your schedule the way a full redistribution can. Run it, compare actual_nfe against what you expect, and A/B on a 22-frame clip where tail jitter is visible. If it helps, it helps cheaply.

It lives in T8/MiniMax H3/Quality/Experimental, and it doesn't need anything beyond the pack itself:

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

Restart ComfyUI, no extra dependencies, no model files. Schedule math only - which is exactly why it's the cheapest quality experiment in the pack to try.