MiniMax H3 LongMedia Planner

Single-prompt video generation tops out around 10–15 seconds, and that's not because the hardware gives up - it's because one prompt can't hold a story. MiniMaxH3LongMediaPlanner is the node that turns your one-shot into a multi-clip movie: a plan with a prompt, duration and optional seed per clip, wired straight into the Long Media Setup. It's the most user-facing node in the pack, and the one where "open" actually means "open a dialogue box and write a storyboard."

Connect its clip_plan output to the Setup, and the Setup automatically switches to MultiClip mode. The frontend renders editable clip cards - one per segment - while the actual card state lives in a hidden clips_json STRING widget that you don't edit by hand.

The inputs that matter

• global_prompt - shared by every clip, joined at runtime. This is where your immutable stuff goes: subject identity, wardrobe, environment, overall visual language, persistent camera grammar. Put it here once, not repeated in every card.
• multiclip_prompt - the structured import source. You can type a whole multi-shot script as clip_1: / clip_2: or shot_1: / shot_2: blocks and have it import into the editable cards. Saves enormous time on longer projects.
• multiclip_auto_import (default false) - auto-import the structured source into the cards whenever it changes.
• multiclip_import_request and multiclip_last_import_source - internal plumbing: a one-shot manual import trigger for dynamic sources, and a snapshot so auto-import never overwrites cards you've edited unless the source actually changed.

Outputs: clip_plan (H3_LONGMEDIA_CLIP_PLAN), clip_count (INT), requested_seconds (FLOAT - the sum of all clip durations), and report (STRING, a JSON summary). The planner needs at least two clips or it refuses to build.

Prompting it like it's a film, not a chat

The pack's own prompting guide has the sharpest advice, and it applies to this node directly:

• Keep identity/wardrobe/environment and "who is speaking" in the global prompt.
• Put only the new visible action of that interval in each card: "She walks toward camera through the battlefield. Medium-wide tracking shot."
• Use temporal glue words - continues, keeps, without stopping, same direction - so the model treats each card as a continuation, not a fresh scene.
• Don't start every card with "A new scene..." or "A woman appears..." - that's how you reset continuity between clips.

That last one is where beginners burn themselves: each card is a beat in one continuous shot, not a new video.

Authority and install

One nuance from the README: the Planner is authoritative only when workflow_mode = multiclip. If clip_plan stays connected while you've picked another mode, the Setup ignores it - a guard so a connected Planner can't silently override your workflow. Install is the standard one-pack dance, no extra pip deps:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-MiniMax-H3-LongMedia

Restart ComfyUI, or ComfyUI Manager → "MiniMax H3 LongMedia". The H3 checkpoint and VAEs go in ComfyUI's model folders.

Gotchas

• Two clips minimum, or it errors - that's by design; a plan of one isn't a MultiClip plan.
• Keep ComfyUI Dynamic VRAM enabled; the low-VRAM strategy depends on it.
• The H3 open weights are excluded from the US, EU, UK and South Korea under the MiniMax H3 Community License - H3 leads the open-weights image-to-video arena right now, but that lead comes with a licensing asterisk you should read before building on it.

The Planner is where LongMedia stops being "one video clip" and becomes "a film." Global prompt for the through-line, cards for the beats, and the engine handles the handoff between them.