ACE-Step Text2music Gen Params

This is the creative heart of the ACE-Step pack. ACE-Step Text2music Gen Params is where you describe the music, write the lyrics, and decide whether you're making a track from scratch, covering an existing song, or regenerating a slice of one. Everything you configure here gets packed into a gen_params bundle that feeds the Text2music Server node - this node builds the instructions, the Server node executes them.

The one decision that shapes everything else is sample_mode. OFF (the default) means you're in manual mode: you write the caption (style description) and lyrics yourself. ON means you type a single sentence into sample_query - "a funk rock song with groovy bass and punchy drums" is the default - and an LLM writes the caption, lyrics, bpm, key, duration, and time signature for you, then hands the whole package to the generator. If you want a quick sketch and don't care about the details, sample mode gets you there in one prompt.

The fields that matter

• vocal_language - 51 options, default en. Set the singing language; unknown is there if you want to let the model guess.
• caption - the style/genre description. The default is a hilariously detailed disco-funk spec; you'll replace it, but it's a good example of how much texture ACE-Step responds to.
• lyrics - with [verse], [chorus] section tags. Leave it empty for an instrumental. In manual mode the pack treats empty lyrics as instrumental automatically.
• auto - ON (default) lets the language model decide bpm, key, duration, and time signature. OFF reveals bpm, key (e.g. C major), duration (seconds, -1 = model decides), and time_signature so you can pin them down.

The other controls only do anything when you feed it audio, which is what makes this node more than a text-to-music form:

• src_audio (AUDIO) - connect a source song and the task auto-switches to cover. Tune remix_strength (1.0 = preserve the original) and cover_strength (noise injection, 0 = clean cover). Heads up from the community: ACE-Step deliberately doesn't preserve the original melody on covers - a choice to dodge copyright trouble with commercial songs, not a bug.
• is_repaint - with a source connected, this flips to repaint mode: set repaint_start / repaint_end in seconds to regenerate just that time range while keeping the rest of the track.
• refer_audio (AUDIO) - reference audio that guides style and timbre, usable in any mode. It's guidance, not a source.
• audio_codes - paste tokens from an Audio Codes node to use a previous generation's internal representation as the source instead of raw audio.

How it actually works

Under the hood the node classifies the task before bundling: repaint if is_repaint is on with a source, cover if any source or codes are present, otherwise plain text2music. When auto is on it zeroes out the manual bpm/key/duration so the server's LLM fills them in. The audio inputs get encoded to WAV and embedded in the request as multimodal content - the cloud API, and your local ACE-Step server, accept audio directly in the prompt.

Install

Ships in the ACE-Step-ComfyUI pack: ComfyUI Manager → search ACE-Step-ComfyUI, or

cd ComfyUI/custom_nodes
git clone https://github.com/ace-step/ACE-Step-ComfyUI.git
cd ACE-Step-ComfyUI
pip install -r requirements.txt

then restart ComfyUI. Dependencies are just torch, numpy, and requests - all the real compute happens server-side.

Common issues

The classic beginner trap: you toggle sample_mode ON, then try to edit caption/lyrics and wonder why nothing changes. In sample mode only sample_query, vocal_language, and is_instrumental are honored - everything else is LLM-decided. And if you're in manual mode expecting a song but get an instrumental, you probably left lyrics empty, which the pack reads as "no vocals." One more: remember the gen_params output has to reach the Server node's gen_params input - this node does nothing visible until that wire is connected and you hit Queue.