MiniMax Music

Yes, it's odd that a vision-language node pack has a music generator in it. And yes, you'll use it anyway, because MiniMax's music models are genuinely good at turning a text prompt into a full song, and having that inside ComfyUI means your audio pipeline stays in one graph. This node calls MiniMax's music API - it generates tracks from music-3.0 / music-2.6 (plus free tiers of each), or does voice/cover-style generation with music-cover if you feed it an existing song.

The name is the one trap to watch: "MiniMax" here is the Chinese AI company, not anything to do with Min-Max text processing in your workflow. And unlike the pack's local nodes, this one does call a real API, so it needs a real key: MINIMAX_API_KEY, read only from the ComfyUI server environment. No key in the workflow, ever. If you haven't set it, the node raises a clear error telling you to set it and restart.

The mechanism is straightforward - it builds a request to one of two fixed regional endpoints and decodes the response into playable audio. You pick region (global_en for the international API, cn_zh for the China endpoint), a model, and then the creative inputs: prompt (what the music should sound like), lyrics (optional text to sing), is_instrumental to skip vocals, and lyrics_optimizer to have MiniMax clean up your lyrics before setting them. aigc_watermark is sent only to the cn_zh endpoint - it's an AI-content flag, so don't expect it to do anything on the global side. Output comes back either as a url or as hex-encoded bytes (output_format), and audio_format gives you mp3, wav, or raw pcm with your choice of sample_rate (up to 44100) and bitrate (up to 256 kbps).

For covers, you supply the source material through the optional audio_url or audio_base64 inputs and a cover_feature_id. stream toggles streaming responses, and timeout_seconds defaults to a generous 600 - music generation is slow, don't shrink it.

The three outputs wire straight into the pack's audio tools: audio is a real AUDIO socket you can connect to Play Music or Save Audio nodes, wave_form is the visualization preview, and sample_rate is there so downstream nodes know what they're working with.

Installing it

No special install beyond the pack itself - ComfyUI Manager (search "ComfyUI VLM nodes"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes
python -m pip install -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements.txt

Then set MINIMAX_API_KEY in the environment that launches ComfyUI and restart. The README points at the official global and China docs for account and content requirements - worth checking, because free tiers and content rules differ by region.

Gotchas

• Key not picked up → it's an env issue, not the node. Set it where ComfyUI is launched, then restart.
• Free tiers (music-3.0-free) exist in the dropdown and are fine for testing, but expect slower/lower-quality output and rate limits.
• The cn_zh vs global_en split is real - pick the region that matches your account. Wrong region = auth or availability errors.
• This node is marked as an output node, so it doesn't need a downstream consumer to execute, but you'll almost always want to connect audio to a Play/Save node so you can actually hear your work.