JZL - 🎬 剧本与镜头处理器

This is the engine room of the JZL MiniMax pack. JZL - 🎬 剧本与镜头处理器 (Script & Shot Processor) takes a rough story idea and, in a single LLM call, turns it into a complete segmented shooting script for MiniMax H3: N segments, each with its own H3 prompt plus scene, video, and audio dispatch instructions. That's the "one-click 漫剧 (comic-drama) creation" pitch the whole pack is built around, and this node is where the magic happens.

The reason it exists is the blank-page problem: a video pipeline is only as good as the per-segment prompts you feed it, and writing sixteen structured H3 prompts by hand is brutal. Let an LLM do the decomposition - story → shots → prompts - in one pass, then let the rest of the pack route those prompts into the actual generation nodes.

How it works

Pick a backend, feed a story, get a structured document out. The output is the "four-section" format the downstream nodes expect: [SHOT_START] blocks each containing ===H3_PROMPT=== (with a multi-field body like detailed_description), plus ===SCENE_INSTRUCTION===, ===VIDEO_INSTRUCTION===, and ===AUDIO_INSTRUCTION=== dispatch sections. The dispatch instructions are slot lists - 角色:孙悟空, 场景:花果山, 道具:金箍棒 - that the dispatcher nodes later turn into actual reference-image assignments.

The LLM backend is your choice:

• 本地模型 [local] - runs a GGUF model through a llama-server subprocess (llama.cpp build b10436) that the pack installs for you. No more llama-cpp-python: process-isolated, runs, exits, and releases VRAM. The README's Qwen-class defaults (up to 256K context, temperature 0.6) are tuned so the model keeps the [SHOT_START] format strict instead of free-associating.
• 在线API [api] - wire the api_config input from the pack's 🌐 API 设置 node and it reads the saved config automatically.

Inputs that matter

• story_input + story_name - your story text, and the folder name it gets saved under.
• mode - 拆解模式 (Decompose) breaks an existing story into shots; 生成模式 (Generate) writes a story from scratch.
• story_style - 19 genre presets (热血战斗 to 谍战风云), each packing tone, color direction, and rhythm guidance into the system prompt.
• segment_count / segment_duration - 4/6/9/12/16/20/24 segments, each 4–15s. Note the tooltip on the model loader: longer scripts need more context (6 segments ≈ 12K tokens, 56 ≈ 128K), so raise n_ctx/max_tokens as you scale.
• ref_image_intro / ref_video_intro / ref_audio_intro - declare your reference assets in 角色A = 孙悟空 style lines. The node only counts elements you declared, which keeps the LLM's invented characters out of the statistical table and the slot map. This is the anti-hallucination trick and it matters.
• enable_scene / enable_props / enable_video / enable_audio - which dispatch instructions get emitted. Flip them off to shrink the output.
• preference - camera-language preferences from the pack's 🎯 MiniMax H3 偏好设置 node.
• use_custom_rule + custom_rule_path - replace the built-in segmentation rules by pasting text or pointing at a .txt/.py rule file.

Outputs

• 剧本输出 (STRING) - the full structured script. This is what feeds the ✨ 提示词增强 node and then the 📋 分段处理中心.
• BUS (JZL_H3_BUS) - a hidden side-channel carrying the model/API/preference/style config so downstream nodes silently inherit the same backend. Don't ignore it: the PromptEnhancer needs this wire to know which LLM to reuse.

Install and the one gotcha that bites everyone

Pack install: ComfyUI Manager → search JZL or ComfyUI-JZL-MiniMax-H3, or git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3 into custom_nodes/, restart. You need a recent ComfyUI.

The gotcha: local mode will not work until you install the llama runtime. Run this once from the pack directory:

cd ComfyUI/custom_nodes/ComfyUI-JZL-MiniMax-H3
python install_runtime.py          # Windows portable: ..\..\..\python_embeded\python.exe install_runtime.py

--dry-run previews the detection, --list-backends shows what your machine can run. Until the runtime's runtime_config.json exists, local mode fails at model load - it's the single most common reason this node "doesn't work" on a fresh install. And if local generation suddenly returns stale config when you switch from API to local, that's cache: the node intentionally re-runs when you change backend, style, or segment count, so let it regenerate.

Also remember the model you're scripting for: H3's local weights are geofenced by the MiniMax H3 Community License out of the US, EU, UK and South Korea - if that's you, the API path is the one to use.