Magnific Generate Music

Generate music from a prompt and get it back as a ComfyUI AUDIO object you can combine with your video. That's the pitch, and the interesting part isn't the music - it's that the result arrives as a socket rather than a file you have to hunt down and drag in.

This is a cloud node. It calls Magnific, it costs money per call, and it does nothing offline. If you want music you own outright and can generate without a meter running, local options exist and our own KB spends a chapter on them - ACE-Step is the open answer to Suno, and it runs in ComfyUI. Use this one when Magnific is already in your stack or when you specifically want the models on their menu.

What you set

prompt is a multiline text box - describe the music. model is a ten-entry list, which is really five products with both their display names and their raw slugs (Google Lyria, Lyria 3 Short, Lyria 3 Long, ElevenLabs Music, ElevenLabs Music v2, plus google-lyria, google-lyria-3, google-lyria-3-pro, elevenlabs-music-generation, elevenlabs-music-generation-v2). Google Lyria is the default. Pick one spelling and stick to it in a shared workflow.

duration_seconds runs 10 to 300, and here's the tooltip you actually need: only Lyria 3 Long and ElevenLabs honour it - the rest are fixed at 30 seconds. Set 120 with a 30-second model and you get 30 seconds, without a complaint. That's the single most common surprise with this node. instrumental asks for no vocals.

seed exists, but not the way you're expecting: music generation takes no seed, so changing it only forces a re-run. Don't build a "reproduce this exact track" workflow on it.

folder comes from Magnific Save To; without it, results go to your Personal project.

Outputs are audio, creation_identifier and metadata. metadata is JSON - send it to Magnific Metadata (unpack) if you want the model and prompt as separate values.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Or install "CustomNodePacks" from ComfyUI Manager and restart. The Magnific nodes are ported into this pack from Magnific's vendor pack magnific-comfyui 0.7.0, so one install covers them; don't also install the vendor zip. Sign in via the Magnific menu or the Magnific Save To node - credentials land in ~/.magnific/comfyui_auth.json, never in the graph.

The dependency people trip on

This is one of the pack's two audio nodes, and audio needs a decoder. If your ComfyUI build has no audio decoder, the node tells you to install PyAV in ComfyUI's Python environment and restart:

pip install av

Run it in the same interpreter ComfyUI uses; the portable install has its own. Also keep the pack's requirements.txt away from a working install - it lists opencv-python, scipy and safetensors as real requirements and a lot of commented-out extras, and the README explicitly warns that a blind pip install -r can overwrite ComfyUI's bundled torch, numpy and Pillow.

Where it fits

The natural pairing is video: generate a clip, generate a score for its duration, combine. It's also the cheapest way to stop a silent render feeling unfinished. Just size the prompt to the runtime you can actually get - a 30-second fixed-length model against a 5-second clip means you're throwing 25 seconds away, and against a 60-second clip means you're looping.

One caveat that isn't the node's fault: music generated from a text prompt is a different job from sound that matches picture. For foley and synced effects, you're better off with a local sound model and a bit of timing discipline than asking a song generator to hit a cut.