MiniMax H3 Duration Planner (T8)

Here's a sentence that will save you a crashed ComfyUI session: you cannot pick any number of frames for H3. The model's latent grid accepts specific frame counts - the README calls out 22, 124, 362 - and "5 seconds sounds right" is exactly how you get an OOM or a refused latent. MiniMaxH3DurationPlannerT8 is the pack's translator: you give it a scene's start time, duration, and optional warmup/cooldown padding, and it hands back the valid length plus the exact slice-and-trim bookkeeping you need to render that window. It's pure math - no sampling, no audio loaded, just a timing plan.

How it works

The key insight is that H3 runs video at 24fps on a grid and audio at 40Hz, so a scene that "is 5 seconds" needs a render length that snaps to a valid frame count, then a final trim to the real duration. The planner separates three things so you don't have to:

• render duration - what you actually sample (the padded, grid-aligned window),
• source slice - where in the source audio/video to pull from (source_slice_start_seconds / source_slice_duration_seconds),
• final trim - what you cut after the fact to land exactly on the scene (final_trim_start_seconds / final_duration_seconds).

Inputs: scene_start_seconds and scene_duration_seconds are the scene; warmup_seconds / cooldown_seconds add context padding before and after (great for a beat of room tone); ensure_minimum_context (default true) refuses to produce a render window too small for H3 to work with; source_duration_seconds defaults to 0 = "unknown" - the sibling MiniMaxH3AudioWindowT8 reads it straight from your AUDIO if you need that.

Outputs you'll actually wire up

length (the INT frame count - this is what goes into your sampler's length/frame input), render_duration_seconds, the source-slice pair, the final-trim pair, prompt_timing_note (a string hint like "render includes 0.5s cooldown, trim to 5.0s"), and report_json. In an audio-driven workflow this feeds the Conditioning and the Audio Window so your prompt's timing note and the rendered frames agree on paper before you spend a minute sampling.

Installing it

From the MiniMax H3 Audio T8 pack (T8mars/T8star, GPL-3.0). Manager search MiniMax H3 Audio T8, or:

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

restart. No forced pip deps; models are yours to place. H3 weights' Community License excludes the US, EU, UK, South Korea.

The gotcha

The most common mistake is treating scene_duration_seconds as the output length. It isn't - it's the scene duration, and the planner may render longer and trim, or snap up/down to a valid grid count. Read the prompt_timing_note and the final-trim outputs before you assume your render is exactly 5.000 seconds. And if you're feeding this into an audio window, remember source_duration_seconds = 0 means "unknown," so the plan assumes you'll connect a real source or it stays conservative about context. This node is a planner, not a magician - but it's the difference between "why did that OOM" and "why did that trim perfectly."