Music Stage

Music Stage is the entry point of ComfyTV's audio story - the node where you type "lo-fi, jazz piano, rainy, 90bpm" and get actual music back, entirely on your own machine. The name is a hint: this is the generate step in the flow, and the generation engine behind it is a local ComfyUI workflow you already have the pieces for. The pack itself adds no API, no cloud, no key. The name is not a lie.

Under the hood, the workflow combo picks an audio-generation backend, and the shipped default is ACE-Step v1 Song - the open-weights text-to-music model that the community's shorthand for "local Suno," and the one the KB notes is genuinely strong on instrumentals while vocals remain the weak suit. That's not ComfyTV's fault, it's the model's honest limit; for instrumental beds and scores it's excellent, for pop vocals it's a work in progress.

The inputs that matter:

• main_prompt - free-form style tags: genre, mood, instruments, BPM. This is your whole song brief.
• lyrics - leave empty for instrumental. The tooltip is the rule: non-empty triggers vocal generation. Want a song with vocals? Paste lyrics here and it goes down the vocal path.
• duration_s - 1–240 seconds, default 30.
• bpm, timesignature, keyscale, language - the musical controls (ACE-Step 1.5-style: key like "C major", time signature in beats per bar, lyrics language code).

Output is a single audio track, which auto-spawns an Audio Picker so you can keep the good takes from a batch - the classic generate → pick → edit loop of the pack.

Now the part that trips people up: this node is not free to run, it's free to install. The model files are yours to fetch. The ACE-Step workflow needs ace_step_v1_3.5b.safetensors in ComfyUI/models/checkpoints/ (it's a checkpoint-format model, ComfyUI core can load it - no extra custom nodes). The docs also list MiniMax Music 3 as an alternative backend with its own diffusion/VAE/text-encoder files under ComfyUI/models/. First Run with an un-downloaded model fails with a missing-model error in the log; grab the file, drop it in the right folder, re-run. It's the one node in this batch of 20 where "install" means download a multi-GB model, not just clone a repo.

Install the pack itself the usual way: ComfyUI Manager search "ComfyTV", or git clone https://github.com/jtydhr88/ComfyTV into custom_nodes/, then a full backend restart (Desktop/macOS: clone by absolute path into the running instance). ComfyTV declares zero extra pip dependencies; the heavy lifting here is done by ComfyUI's own ACE-Step support. If the audio is thin, don't fight the model - push through the AudioFX suite (Stereo widen, Saturate, Reverb) the way this pack is designed, and mix it into something that sounds finished.