ACE-Step Understand

Understand is the "what am I actually listening to?" node. You feed it an audio file and it hands back a full breakdown: a caption describing the style, the lyrics, bpm, keyscale, duration, and a formatted analysis_text. In one pass it answers the questions you'd normally spend an evening trying to pin down by ear, and - here's the practical payoff - the outputs wire directly into the generation nodes as a spec. It's the analysis half of a clone loop: Understand a song you like, feed its caption/BPM/key into Text to Music, and generate in that same direction.

The interesting mechanic is that it doesn't just listen with a classifier. The node encodes your audio into ACE-Step's semantic "codes" using the DiT's audio encoder, then runs the language model's understand_music over those codes. That's why config_path is in the required inputs even though no music is being generated - the encoder needs a real model to produce the codes. The target_duration input tells the model what length to reference while analyzing.

The inputs that matter

• audio - the track to analyze.
• config_path - the model config used for encoding (acestep-v15-turbo default).
• target_duration - reference length for analysis, 30s default.
• language - a hint about the vocal language if you know it; helps the LLM transcribe.
• temperature - the tooltip's advice is worth quoting: lower (0.0-0.3) is more precise, "try 0.1 for better accuracy." Default 0.3.
• top_k / top_p / repetition_penalty - the usual sampling controls; the defaults already lean toward faithfulness.

Outputs: analysis_text (a formatted summary string), caption, duration (FLOAT), bpm (STRING), keyscale, lyrics.

Install

Shared pack install - nothing extra for this node beyond the standard model stack:

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5 && pip install -e .
cd ComfyUI/custom_nodes
git clone https://github.com/kana112233/ComfyUI-kaola-ace-step.git
cd ComfyUI-kaola-ace-step && pip install -r requirements.txt
huggingface-cli download ACE-Step/Ace-Step1.5 --local-dir ComfyUI/models/Ace-Step1.5

Restart ComfyUI after installing.

Where people get burned

Two realistic frustrations. First, the output is only as good as the encoder and LLM's judgment - BPM and key on a complex, tempo-shifting track can be off, and the README itself positions Understand as a starting analysis, not sheet music. Second, note that bpm comes back as a STRING here (unlike Create Sample's INT), so if you're wiring it into a node that expects a number, you'll need a conversion or a text node in between. For quickly capturing the direction of a track - the caption and tags - it's excellent, and that's the workflow worth building around it.