Load Chapter Catalogs

The heaviest, slowest part of the MiniMax H3 Novel pack is extraction - dozens of sequential local-LLM calls per chapter to turn prose into structured character/location/object catalogs. Load Chapter Catalogs is the node that lets you skip it. It reads previously saved *_references.json files back off disk and hands them to the pipeline exactly as if Extract Chapter References had just produced them, which means you can tweak consolidation settings, re-run, and iterate without paying for the extraction again.

One input, validated hard

It has one input: catalog_path. Point it at a folder containing *_references.json files (it scans them in natural sort order and deliberately excludes consolidated_references.json), or at a single such file saved by Extract Chapter References. The tooltip is accurate about both cases. It validates each file has the required chapter_id and schema_version fields - so an old v1-era file from before the pipeline's schema bump will fail loudly rather than silently feeding Consolidate References garbage. That's a feature: garbage in would otherwise mean a corrupted cast list downstream.

What it plugs into

Two outputs, both identical in type to what the extractor returns, which is the whole trick: chapter_catalogs (type MINIMAX_CHAPTERS) wires straight into Consolidate References, and catalog_summary gives you the quick per-chapter character/location/object count so you can confirm you loaded what you think you loaded. Since it produces the same types as Extract Chapter References, you can even swap the two nodes in a workflow without rewiring anything downstream - keep one as the "recompute" path and the other as the "reload" path.

It's a thin, honest utility and it doesn't try to be more: no LLM config input, no model calls, nothing to tune. The one thing it needs from you is a path that exists - a typo or a folder that contains no matching files and it errors immediately rather than returning an empty catalog. Install is the shared pack story: it's one of the seven nodes in yuyuki/minimax_h3_novel_pipeline, so clone that repo into ComfyUI/custom_nodes or grab it from ComfyUI Manager by searching "MiniMax H3 Novel Pipeline", and restart. There's no model download for this pack at all; the only external dependency is a Qwen model running in LM Studio for the nodes that actually talk to an LLM - this one never does.