Select H3 Scene

Generate H3 Prompts hands you a whole chapter's worth of scene prompts at once. Select H3 Scene is the node that lets you pick exactly one - chapter 3, scene 7, say - and pull out everything you need to actually run that single shot through MiniMax H3 Reference to Video. It's the seam where this text pipeline finally touches the thing that renders moving pictures, and it exists to stop you from getting the attachment order wrong.

Feed it the prompts manifest from Generate H3 Prompts, set chapter_index and scene_index (both 1-based - yes, it trips people up on the first try), and it walks the manifest to find that scene. Index out of range or a scene that was skipped, and it raises a clear error listing the indexes that are available. This is a pure selection node: no LLM calls, no file writing, instant.

The outputs tell you exactly what to attach

Five outputs, and the last four are really one story told four ways:

• h3_scene (type MINIMAX_H3_SCENE) - a structured object with chapter id, scene index, title, the prompt, the ordered picture/audio IDs, and the bindings. The one you'd keep for scripting.
• prompt - the finished H3 prompt text, ready to paste into whatever reference-to-video node or API call you use.
• image_asset_ids_in_h3_order - the list of picture asset IDs, one per line, in the exact order the prompt's <Picture N> labels use them.
• audio_asset_ids_in_h3_order - same deal for <Audio N>.
• binding_json - the full structured binding: each <Subject N>, the entities behind it, and every picture and audio asset attached to that subject.

That repeated emphasis on order is the entire point of this node. MiniMax H3's subject/picture/audio labels are request-local - a scene references a handful of your hundreds of global assets under compact local numbering - and if you attach PIC_CHAR_001_BACK_VIEW where the prompt expects the face front, Elena renders from the back in a close dialogue shot. The node's own docstring is blunt about the division of labor: it returns IDs in H3's expected order, and it "never fabricates media from the textual briefs." You render the PNGs and WAVs from the reference-asset prompts back in the consolidation stage, and you attach them in exactly this order when you fire the generation.

Where it fits

This is a real workflow's last text node: LM Studio Configuration → Extract/Load → Consolidate/Load → Generate H3 Prompts → Select H3 Scene, and from here the h3_scene (or just the prompt string plus the ordered asset IDs) feeds a MiniMax H3 Reference-to-Video setup in ComfyUI. Given H3's launch reception - day-zero ComfyUI support and the community immediately praising its image-to-video and multi-reference behavior - this kind of disciplined, per-scene reference management is what keeps a long-form project from dissolving into a cast of lookalikes. The pack ships no install surprises here beyond the shared one: the API key still comes from ComfyUI Settings, and the heavy Qwen model still lives in LM Studio.