Multishot Advance AI Scenario Director

Writing the H3 contract is the slow part. This node offloads it.

MiniMax H3 is a 33B omni-modal video model - 4–15 second clips at 24 FPS with native synced audio - and the Multishot Advance pack chains those clips into long scenes. The catch: H3 doesn't take a plain sentence. It wants a structured contract, with subject_definitions, a soundscape, per-clip [Shot N] blocks separated by ---, and (if you're doing dialogue) exact <d>[English] ...</d> tags. Great output, but the typing is brutal when you're iterating on a five-clip story.

That's what the Multishot Advance AI Scenario Director is for. It's the optional, explicitly experimental authoring node in this pack: you describe the story in plain language, and it hands back structured English fields ready to drop into the pack's Prompt Builder. It doesn't render a single frame - it's a text generator sitting before everything else. And the nice part: it's fully local. The source calls your LLM over urllib to http://127.0.0.1:11434 (Ollama) or your LM Studio / llama.cpp server. No API key, no cloud, no prompts leaving your machine - the same reason the community reaches for local LLMs for this kind of prompt-authoring job at all.

How it works

You give it scenario - the story, subjects, actions, environment, what each reference Picture should contribute. Optionally wire in up to nine reference_images, three reference_video_*, and three reference_audio_*. If analyze_visual_references is on and you've set a vision_model, it base64-encodes each Picture (and contact-sheet frames of each Video) and sends them to a multimodal model for analysis before the chat model writes anything.

The chat model then produces a structured response, which the node renders into one copyable block: subject name and description, retention focus, global/style/soundscape/music prompts, and the full shot script with [Shot N] camera cuts and --- scene separators. prompt_output is the single output, a STRING, and it's written so the Pack's Prompt Builder can eat it directly.

The inputs that actually matter

• scenario - the primary input. It's an error to leave this and scene_briefs empty.
• chat_model - the exact name of an installed local instruction model, or type auto for discovery. The greyed placeholder (gemma3:12b) is a hint, never a submitted default.
• scene_count and duration_per_scene - how many --- clips to produce, and how long each (H3's 4–15s range).
• dialogue_language - spoken dialogue only. Prompt prose is always English; this just forces <d>[Turkish] ...</d> style tags without translating the words.
• vision_model - needed only if you want Pictures/Videos analyzed. qwen3.5:9b-class multimodal models do the job.

Wiring it in

AI Scenario Director.prompt_output -> Prompt Builder.external_prompt_input

First run imports the generated block into the Prompt Builder's editable fields and stops downstream. Review it, fix what the LLM got wrong, then run again to continue to the sampler. That stop is deliberate - this node is not feature-complete, and the author wants you to look before you render.

Installing

ComfyUI Manager → search "ComfyUI-H3-Multishot-Advance", or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/KursatAs/ComfyUI-H3-Multishot-Advance

Restart ComfyUI. There's no requirements.txt - the Director is pure stdlib, so no pip installs. But it assumes you already run a local LLM server. Install Ollama (or LM Studio / llama.cpp), ollama pull a chat model, and keep it running before you queue. The pack itself doesn't bundle H3 weights; load those through the pack's Multi Model Loader, and grab ComfyUI-GGUF if you use quantized H3 checkpoints.

Gotchas

• It's experimental. Quality and structured-output reliability vary by model - treat the output as a first draft.
• scene_briefs (one brief per --- block) must match scene_count or the node errors out.
• Audio isn't transcribed. If an audio ref is a voice-timbre source, name its owner in reference_audio_notes - the Director won't guess.
• thinking only activates for Ollama models that advertise native thinking; unsupported models just skip it.

Like any local-LLM helper, it's a timesaver, not an oracle. Use it to beat writer's block on the H3 contract - then edit like you mean it.