MiniMax H3 Output Trim (T8)

Every H3 render ends with the same awkward negotiation: the model gives you 124 frames at 24fps (5.167 seconds) or 362 frames, and what you actually wanted was a clean 5-second clip starting at a specific point. MiniMaxH3OutputTrimT8 is the unglamorous node that closes that gap - it takes decoded IMAGE frames plus an optional AUDIO track and trims both to the same start and duration, so your picture and sound come out of the same cut instead of you hand-editing them into alignment afterwards.

It's one of the rare stable (non-experimental) nodes in this pack, and it's deliberately simple: a couple of seconds-based inputs, no sampling math, no latent games. The idea is you can think in seconds - "I want 3.2 seconds starting at 0.8" - rather than doing frame arithmetic and then separately trimming audio in another tool.

The inputs that matter

• frames - the decoded IMAGE batch (this sits after AV Decode, not on the latent).
• start_seconds - where the trim begins, default 0.
• duration_seconds - how long the output is, default 5.
• fps - 24 by default. This is the field that bites people: give it the wrong frame rate and the seconds→frame conversion is wrong, so your "5 seconds" becomes 5 seconds of the wrong frames.
• audio - optional. Leave it disconnected for silent clips; connect it and it's trimmed to the same window, keeping AV sync.

Outputs: trimmed frames, trimmed audio, and a report_json with the exact trim math.

When you'd use it

Two places. First, after a Long Video continuation: the Continuation Concat node tells you it left a hidden tail frame to trim after decode (trim_tail_frames_after_decode), and this is the natural tool for that final crop. Second, as the last mile of a duration-planner workflow - the pack's Studio/Duration Planner metadata flows into it, and the seconds-based cut replaces a pile of manual frame math. If you've got a clip where the audio starts a hair late or you only want the good middle section, this is the "trim in ffmpeg but without leaving ComfyUI" move, and it keeps the two clocks honest.

It's not a sampler, not an upscaler, and it won't fix a bad render - if the first three seconds are a garbage ramp-in, trimming them away is exactly what it's for, but it can't un-ruin the rest. It's the boring utility you stop noticing once it's in the graph, which is a compliment.

Install

Part of the MiniMax H3 Audio T8 pack:

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

or search "MiniMax H3 Audio T8" in ComfyUI Manager and restart. No extra pip dependencies, no model files to fetch. The quickest way to see it working is the basic generation workflows under examples/workflows/01-basic-generation - drop it between AV Decode and your save node, set your start and duration, and read the report JSON if the cut lands somewhere unexpected. When the frames and audio don't line up, it's almost always the fps field disagreeing with the render's real frame rate.