H3 Story Director — LLM Model (API)

H3 Story Director - LLM Model (API) is the answer to the obvious objection to the main Director: "I don't want to route my story planning through OpenRouter." This variant runs the exact same multimodal planning, validation, and plan compilation - but instead of calling OpenRouter, it takes an LLMMODEL on a socket, and you get to decide what that model is. A local Gemma through Ollama, a datacenter endpoint through the pack's own H3 LLM Model (API) node, a YALLM provider - anything that emits an LLMMODEL output works.

It's a clean swap. The api_key, model, reasoning, and timeout_seconds inputs disappear (no longer relevant when you're not calling OpenRouter), and in their place is the llm_model connection. Everything else - the scene schema, the reference tagging, the validation, the outputs - stays identical.

How it works

The node hands your story idea, reference images, and optional VHS source-video batch to whatever model you've wired in, still using the strict structured-output schema. The important part: it still performs all final plan validation locally. Incomplete JSON, missing scenes, duplicate IDs, underspecified scene prompts, unassigned <Picture n> tags - all rejected before any MiniMax render starts, regardless of which model produced them. With a YALLM model that doesn't expose structured-output parameters, the node supplies the exact JSON Schema in the system instruction and rejects malformed responses on its end. The model can be sloppy; the node is not.

The inputs that matter

• llm_model - the one required input you can't get wrong. Accepts the pack's H3 Ollama Model (Local), its H3 LLM Model (API), or the LLMMODEL output from YALLM's LLM Model (API) / LLM Provider (API) nodes.
• director_profile - this one does the work of matching your model. OpenRouter keeps the compact scene schema. Gemma switches to a stricter structured worksheet - explicit-action checklist, duration-aware beats, physical performance, camera, sound, final state - which is the profile to pair with a local Gemma via Ollama. The profile doesn't select the model; it shapes what the node asks for.
• story_idea - optional as always; leave empty for full creative control.

The rest is the standard Director surface: scene_count, scene_duration_seconds, steps, draft_only, genre, language, motion_style, director_mode, video_sample_frames, four reference images, source_video.

Outputs

Identical to the main Director: plan_json (into MiniMax H3 Context Loop Plan), scene_prompt, mode_prompt, story_bible, synopsis, validation, usage_stats, credits_remaining, source_video_analysis. No usage stats from OpenRouter here - those come from your connected model's service, if it reports them.

Install

ComfyUI Manager → search Vision Prompt Assistant, or:

cd ComfyUI/custom_nodes
git clone https://github.com/elgalardi/ComfyUI-VisionPromptAssistant

Restart, wire in an LLMMODEL, and go. This variant is the one to reach for when you want the Director's structure without paying per OpenRouter request - the local Ollama path makes the whole planning stage free per call after the model is loaded. The tradeoff is plain: a 12B local Gemma is not writing at frontier quality, but it's writing for a fraction of the cost, and the local validation catches most of the damage a weaker model would otherwise do. Start with draft_only on either way.