Parse Structured Music LLM Output

Anyone who's wired a chat LLM into a generation graph knows the pain: the model gives you exactly what you asked for, wrapped in "Here is your enhanced prompt!" and a closing paragraph, and you have to scrape the good parts out of the chat-shaped noise. That's failure mode one from the local-LLM playbook, and it's exactly what this node exists to kill. Parse Structured Music LLM Output takes the assistant's text, extracts the [Caption], [Lyrics], [Title], and [Image_Prompt] sections, validates that the required music sections are present, and emits each one as a clean STRING - plus per-song seeds and provenance.

The bundled system prompt demands the canonical [Caption] → [Lyrics] → [Title] → [Image_Prompt] order, but the parser is deliberately order-tolerant: an LLM that shuffles the sections won't nuke an otherwise usable result. What it won't tolerate is missing or empty required sections - it raises an error rather than silently generating a song with no caption. Silent failure is worse than a loud one here.

Inputs

• structured_llm_output - the full assistant text from your LLM node.
• song_count - how many song variants to emit from this source (1–100). More songs = more downstream generation time, obviously.
• seed_mode - random_each_song (fresh seed per variant) or increment_from_base (reproducible sequential seeds from base_seed). For reproducible batches, increment is the one you want.
• base_seed - only meaningful in increment mode.
• user_prompt - stored with the parsed result for provenance.
• source_name_override - replaces the auto-derived source identifier when non-empty.
• fallback_title - used only when no usable [Title] survives parsing; it never replaces a valid LLM title.

Outputs

Every output is a list (is_list), so ComfyUI maps them across your batch: caption, lyrics, title, image_prompt, source_name, generation_seed, run_index, variant_count, source_path, prompt_origin, and prompt_provenance_json. Wire caption + lyrics into the MiniMax Music 3 node, image_prompt into the artwork branch, and the provenance outputs into MiniMax Song Metadata.

Installing it

Pack install via ComfyUI Manager (search "MiniMax Music Production Toolkit") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/ComfyUI-MiniMax-Music-Production-Toolkit.git
cd ComfyUI-MiniMax-Music-Production-Toolkit
python -m pip install -r requirements.txt

Dependencies: scipy, soundfile, imageio-ffmpeg, mutagen, Pillow. Restart ComfyUI and hard-refresh once.

Gotchas

If you get a hard parse error, the LLM probably skipped a required section or mangled the tags - the README's honest caveat applies: "the example LLM stage depends on the quality and instruction-following ability of the GGUF model you choose." A small or heavily-quantized model that truncates its output mid-[Lyrics] will trip this node on purpose. And the cache-buster rule applies upstream: if the LLM output looks identical every run, the LLM node is caching - wire in the LLM Session ID node, not the parser.