MiniMax H3 Long Video Tail/Second Pass Plan (Advanced EXP/T8)

Long videos in the T8 pack aren't one giant generation - MiniMax H3 samples short clips (the pack's in-node loops use a 124-frame render window with a 22-frame context) and chains them segment by segment. That works great until you look at a segment and wish the tail of the shot had more refinement, or you want a low-sigma pass over the finished piece. That's what this node exists for: it plugs into the two in-node long-video runners and lets you plan an extra tail-subdivision or a manual low-sigma second pass per segment, without touching the loop itself.

The design principle is the pack's favorite trick, done right: it's append-only and disconnectable. The node's mode defaults to disabled, and if you disconnect it entirely, the long-video runners behave exactly as they always did - the author's own docs promise the old in-loop values stay bit-identical. So this is pure opt-in, which is exactly how an experimental node should behave. The bundled workflow is 04-long-video/2026-08-30_H3_In_Node_Long_Video_Prompt_Relay_EAV_Manual_Second_Pass_Advanced_EXP.json.

How it works

You pick one of three modes, and the node emits a sampling_plan object the long-video runner consumes when building its per-segment sigma schedule:

• disabled - the plan exists but changes nothing. The default, and the safe place to leave it.
• tail_subdivide - subdivides only the final H3 interval of the main trajectory into extra steps (extra_tail_steps, 0–8). Every inserted tail point is one more full joint audio/video forward pass, so the NFE grows, but nothing else about the main pass changes.
• manual_second_pass - after each segment's main sampling completes, a separate ordinary sampler runs again over the segment using your manually-specified sigmas. The default table is 0.5, 0.412, 0.350, 0 - the same low-sigma route the pack uses in its identity-preserve experiments.

Both modes keep the joint AV dual-clock contract: audio isn't frozen or re-noised on its own, it's sampled together with the video on the H3 dual clock, same as the main pass. The two passes share one bounded preview-cache state, Prompt Relay stays global, and Enhance-A-Video only audits the main pass - the pack deliberately doesn't reuse old EAV runtime errors on the second pass.

The inputs that matter

Only two, really:

• extra_tail_steps - how many extra steps to squeeze into the tail (only used by tail_subdivide). One, the default, is a reasonable starting point; more steps cost real compute.
• manual_sigmas - the comma-separated schedule for manual_second_pass. It must strictly descend and end at exactly zero; the node validates it and will refuse nonsense. tail_spacing (video_sigma_linear default, video_sigma_cosine, base_flow_linear) picks how the inserted tail steps are spaced.

The two outputs are sampling_plan, which feeds the runner, and report_json - the parsed schedule, mode, and how many extra NFE each pass adds, so you can see the actual cost before running.

Installing it

Same pack, same install: ComfyUI Manager → search "MiniMax H3 Audio T8", or

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

then restart. No model files and no extra pip packages for this node - requirements.txt is intentionally empty, with base dependencies supplied by ComfyUI. The usual T8 warning still applies: update ComfyUI core, frontend and Manager together and fully restart, because this pack leans on newer core APIs (comfy_api.latest, comfy.ldm.minimax) and a stale core makes every T8 node fail at import.

Where people get burned

• Treating a second pass as a free quality boost. The pack's scientific boundary is blunt: both modes change the joint video/audio trajectory and do not promise more detail or unchanged speech. Extra steps means extra forwards, and the tail can drift. Watch the full result and listen - the author is explicit that visual review plus listening are the human gate.
• Forgetting it only works in the two in-node runners. This plan feeds the in-node long-video loops (the In_Node_Long_Video_... workflows), not the Background/Accepted route or a generic sampler. If you wire it into the wrong workflow, nothing happens.
• Typing a manual schedule that isn't strictly descending. The validator rejects flat or ascending sigmas and anything not ending in zero. Copy the default 0.5, 0.412, 0.350, 0 and edit carefully.

If you've ever wished one segment of a long H3 render had a cleaner tail without re-rolling the whole chain, this is a tidy answer: plan it, review it, and disconnect it the day it stops earning its keep.