MiniMax H3 Structured Prompt Enhancer (Plan v2)

Writing a good H3 prompt is two different jobs. One is structural - correct labels, valid timing, references routed properly - and the compiler already does that. The other is descriptive: taking a bare scene and adding concrete blocking, lighting, motion, and synchronized sound so the model has something rich to work with. Structured Prompt Enhancer (Plan v2) is the node that delegates the second job to a Qwen LLM while keeping the first locked down hard.

The cleverness is in the lock. Qwen gets your compiled plan and returns JSON prose - but it never writes an H3 prompt. The system prompt is explicit: no section names, no [Shot N] markers, no cut timestamps, no dialogue-tag markup, no speaker IDs. Python validates the JSON, reconstructs the full H3 document, and restores every structural field itself. If Qwen tries to renumber your Subjects or move your references, the change is rejected, not silently accepted.

The two modes

• Intent-locked expansion (default) - your original prose stays verbatim; Qwen appends only compatible production detail. This is the safe one.
• Creative expansion - Qwen may rewrite the editable prose and add presentation choices. References, timing, dialogue, shot order, and routes stay locked either way.

Inputs that matter

• clip - a complete generation-capable Qwen3-VL or Qwen3.5 CLIP, or MiniMax H3's 50-layer conditioning CLIP with the Generation Tail Loader connected. This is an LLM-capable ComfyUI CLIP, not an OpenAI vision CLIP.
• plan_context - from Prompt Merge.
• system_prompt - editable prose behavior; blank uses the built-in production-detail pass.
• visual_analysis - off by default (text metadata only), and the tooltip calls it "safest and lowest-memory." When on, it attaches sampled image/video frames as evidence. Note the sharp rule: audio is never inferred from waveform data; audio meaning comes only from your Audio Reference metadata and transcripts.
• max_new_tokens, temperature, top_k, top_p, min_p, repetition_penalty, presence_penalty, seed - the standard sampler set. max_new_tokens (default 1200) is the one to raise for many long shots.
• thinking - off is recommended for direct schema-following output.

Outputs

The ones that matter: enhanced_prompt (the rebuilt, validated final prompt), editable_prose (validated JSON for Apply Structured Prose), enhanced_plan_context (recompiled locked plan matching the enhanced prompt), and enhancer_report (mode, generation, visual-analysis, fallback, and validation status).

Install and prerequisites

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

The pack itself is dependency-free, but this node is the exception to the "nothing to download" rule: it needs a real Qwen3-VL/Qwen3.5 CLIP, or the MiniMax H3 conditioning CLIP plus the generation tail file (see the Generation Tail Loader article). Both are model downloads, not pack dependencies.

Gotchas

• Fallback is not failure. If Qwen returns empty, repetitive, or invalid output, the node falls back to the deterministic compiler draft and tells you in enhancer_report. Your prompt doesn't vanish - you just get the un-enhanced version.
• No invented story facts. The system prompt forbids adding characters, props, locations, dialogue, or lyrics not in the supplied prose. If your shots are vague, the enhancement is thin; feed it real material.
• Keep clip and tail consistent. Complete Qwen3-VL doesn't need the tail; the truncated 50-layer conditioning CLIP does. Mixing them up either wastes the tail or errors.

One honest take: if your shots are already detailed, you can skip this node entirely and go Prompt Merge → Apply Reference Plan. The enhancer pays off when you're drafting thin and want production-grade prose without hand-writing it. That's the audience it's built for.