SongBloom Lyric Generator

Most people don't want to hand-write SongBloom's tag format - they want to describe a song and have the lyrics appear. That's this node. You give it a prompt ("write a nostalgic song about youth"), a style, a mood, a language, and it returns ready-to-sing lyrics with all the [verse] / [chorus] / [intro] structure already in place.

The interesting bit is that it has two entirely different brains, and you should know which one you're using. By default use_deepseek is on: it builds a system prompt from your style/mood/language choices and sends the request to DeepSeek's API, then runs the result through format cleanup and validation. That's the quality path - a real LLM writing real lyrics. But it needs an API key, which is where people stall. Flip use_deepseek off and it falls back to a built-in template assembler: it picks section-by-section fill-in lines from a local library keyed to your language and style. No network, no key, no cost - and the lyrics are predictable and generic, because that's what template generators are.

Inputs that matter

The required row is all vibe-setting:

• text_prompt - what the song is about.
• style - 12 options (pop, rock, ballad, folk, electronic, jazz, classical, hip-hop, country, blues, r&b, reggae).
• mood - 12 options (happy, sad, nostalgic, energetic, peaceful, romantic, melancholic, angry, hopeful, mysterious, dramatic, playful).
• language - chinese, english, mixed, japanese, korean, spanish, french.
• structure - simple / standard / complex / custom.
• length - short → very_long.
• format_style - traditional / modern / minimal / detailed (author's tooltip: "lyric format style").

The optional row is where the DeepSeek config lives - use_deepseek, deepseek_api_key, deepseek_model (only DeepSeek-V3, internally mapped to deepseek-chat), deepseek_system_prompt (the default is a full Chinese prompt instructing strict format output - you can override it), plus custom_tags, theme_keywords, user_requirements, rhyme_scheme, line_length, repetition_style, emotional_intensity, and target_audience. The tooltips on those are the author's own descriptions - rhyme patterns like aabb/abab, target audience, etc. The extra tuning fields mostly matter when DeepSeek is on, since they get folded into the context.

Output is a single generated_lyrics (STRING) - ready for SongBloomLyricProcessor or straight into the generator.

The API key question

DeepSeek keys are resolved in a priority order: node input → DEEPSEEK_API_KEY env var → DeepseekAP-config.json in the pack root → deepseek.api_key in config.yaml. If no key is found anywhere, the DeepSeek path returns the raw prompt with a message saying so. For $5-ish you'll get a long way on DeepSeek-chat, but note the obvious: this is a paid external API bolted onto a local-generation pack, so it inherits every downside - you need network, you're sending lyrics to a third party, and a 400/401 usually means a bad key.

Install

Same pack install: search "SongBloom_ComfyUI" in Manager, or

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/SongBloom_ComfyUI
cd SongBloom_ComfyUI
pip install -r requirements.txt

Restart, and the node is under SongBloom/Lyrics. If your songs keep coming out garbled, remember the English-vs-Chinese bias in the model itself - the lyrics this node writes may be flawless while SongBloom still sings them wrong.