MiniMax H3 Prompt Merge (Plan v2)

Somewhere in every Plan v2 workflow, the many moving pieces - a Project Setup, a couple of Image References, three Shots with dialogue - have to stop being a graph and become a single H3 prompt string that MiniMax will actually understand. That's the entire job of Prompt Merge, and it's the node that makes this pack feel like a compiler instead of a pile of widgets.

The name undersells it. This node doesn't just concatenate text. It takes the whole ordered plan and assigns every <Subject N>, <Picture N>, <Video N>, and <Audio N> label deterministically, numbers the speaker IDs in actual playback order, decides which H3 task type (T2VA, I2VA, L2VA, FL2VA, Ref2VA) the project needs, orders the native media routes, and validates the timing. No model is called - it's pure Python doing deterministic bookkeeping, which is exactly what you want when a language model later gets its hands on your prose.

The inputs and outputs that matter

It has exactly one required input: h3_plan, connected from the final Shot, Dialogue Event, or setup node in your chain. There's one optional widget, prompt_style, with two choices:

• Full structured prompt - the verbose, guide-compliant template. Good when you're learning or when you want to read every section.
• Compact low-token prompt - the same locked semantics, less repeated prose. If you're paying for tokens or hitting limits, this is the one.

Outputs worth knowing:

• h3_prompt - the finished, deterministic H3 prompt. Three or six sections depending on whether your plan has references.
• plan_context - the typed, compiled plan object. This is the thing every later node (enhancer, review gate, Apply Reference Plan) consumes. Prompt and plan travel as a matched pair.
• rewrite_request - a self-contained prose-only enhancement request for the structured enhancer, with the semantic fields locked so an LLM can't move your references around.
• problems_report - readiness, inventory, native routes, timing, and warnings. Read this before you queue anything; it tells you why your plan isn't valid long before MiniMax does.
• h3_length - the native frame count from Project Setup, forwarded.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart ComfyUI, look under MiniMax H3 → Plan v2. No extra Python dependencies, and it doesn't need any model downloads - the merge step is pure logic. You do need a ComfyUI build with native MiniMax H3 support for the downstream sampling to work.

Common issues

• Prompt and plan_context out of sync. This is the classic failure. h3_prompt and plan_context are a matched pair, and if you rewire one without the other - say, connect an old prompt to a new plan - the Apply Reference Plan node will reject the pair. Always take both from the same node.
• Media placed after the first Shot. The pack enforces "all references before the first Shot." If your problems report complains about inventory, that's usually it.
• Timing errors. Shot 1 must begin at 0.000, no gaps or overlaps. The compiler validates ranges and tells you exactly which shot is wrong.

One honest caveat: this pack is brand new and barely has community miles on it, so you're ahead of the troubleshooting threads either way. The good news is the node is verbose - its error messages read like a checklist, not a stack trace.