FL MiniMax Music 3 Audio VAE Loader

This is the node you'd skip if you're only generating music, and the one you need the moment you want to feed audio back in. The name is nearly identical to the VAE that ships with the FL MiniMax Music 3 Loader, but the "Audio" in the title is doing real work: it loads the full MiniMax Music 3 DAV - the waveform encoder and decoder - rather than just the decoder side that turns generated latents into sound.

The distinction is exactly where people get confused. The plain Music 3 Loader gives you a VAE that can decode latents to audio, which is all a text-to-music workflow needs. This node pulls the full encoder/decoder pair from the SimpleTuner/MiniMax-Music-3-Encoder repo (a ~306 MB file, sha256-verified and resumable like the rest of the pack's downloads). With it, you can run VAE Encode Audio: feed it 44.1 kHz mono or stereo audio and it converts real sound into native Music 3 latents.

Why does that matter? Because once audio exists as latents, it's in the same space as generated audio - which is what opens up audio-conditioned music workflows, and it's the same encoder the SimpleTuner training backend uses on the data side. If you ever want to take an existing track and have Music 3 do something with it rather than dream it up from a prompt, this is the piece that gets you there.

What it does

No inputs, one output:

• vae - the full MiniMax Music 3 DAV, ready for VAE Encode Audio or decoding.

It downloads the encoder artifact on its first execution (again, only when the node actually runs), with progress on the node. That's the whole interface. It's a thin node by design - the value is that the pack fetches, verifies, and wires a model file you'd otherwise be hunting for on HuggingFace yourself.

Install

Same as the rest of the pack - ComfyUI Manager, search FL MiniMax Music 3, or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxMusic3.git
cd ComfyUI-FL-MiniMaxMusic3
pip install -r requirements.txt

Restart ComfyUI. You also need a current ComfyUI that actually supports MiniMax Music 3, since the node depends on the model type being registered in the core.

Realistic expectations

If your entire workflow is "prompt → song," you don't need this node - the regular loader's VAE covers you, and this is a ~306 MB download for nothing. Reach for it when audio is an input to your graph (encoding existing tracks), or when you want the same encoder/decoder setup on the inference side that the training side uses. And know that encoding is the slower direction: it's the full waveform pass, not the lightweight decode. Keep your source at 44.1 kHz and mono/stereo and it does what it says on the tin.