Add Vocals to Music · V6

If you're starting fresh in 2026, this is the node in the Add Vocals family you should be reaching for. KIE's docs label V5 and V5.5 as discontinued, which makes V6 the live tier - better vocals, richer detail, and a 5000-character lyrics budget instead of the older caps.

The job is unchanged. You have an instrumental - maybe a local ACE-Step render, maybe a stem, maybe something you wrote - and you want a sung topline on it. VOCALS are precisely where the local stack stops competing: ACE-Step's instrumental output is genuinely good and its lyrics and vocals are the known weak spot. Suno's vocal is the reason this stays an API call.

How it works

The pack generates its node classes from KIE's official API documentation at startup and splits each documentation page into one node per published model version. So V6 is baked into this class - there's no model selector, and the · V6 in the display name is the only sign of the tier. The node builds the documented request body, submits it to KIE's task endpoint (envelope model ai-music-api/add-vocals, inner model V6), polls KIE's record-info endpoint every couple of seconds with exponential backoff until the task reports success, then downloads the mp3 and returns a native ComfyUI AUDIO. Failures raise with the task id and the provider's error message.

The inputs that matter

Five are required:

• upload_url - the instrumental, as a URL KIE can fetch. Per the docs, source audio plus title, negative_tags and style is enough to generate.
• prompt - the direction for the performance; it doubles as lyrics only when lyrics is empty.
• title, negative_tags, style - required, and they steer the result heavily. On V6 the docs cap style at 1000 characters and lyrics at 5000.

Optional: vocal_gender (m/f; the docs admit it only improves the odds), style_weight (0–1, default 0.61), weirdness_constraint (default 0.72), audio_weight (how strongly the source audio pulls), lyrics, variety (0–4, default 1 - 0 holds the style tightly, 4 chases ideas), timeout_seconds (1200) and callback_url (leave empty for normal ComfyUI use).

Outputs: audio (wire into Save Audio or whatever comes next), url, all_urls_json, task_id, raw_json, plus credits_consumed and credits_left as floats - both also printed on the node after a run.

Getting your instrumental in

upload_url is a plain string widget; you can't connect an AUDIO to it directly. Use KIE • Upload Audio from KIE Next/Utility/Media, which takes an AUDIO socket and returns a public URL string, then right-click upload_url here, pick Convert widget to input, and wire it in. That converts the node into something a fully local generate → upload → sing-on-it chain can drive.

Installing

ComfyUI Manager, search "KIE Nodes Next". The README still describes the registry listing as prepared rather than published, so expect the clone route to be the one that works:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart, then set the key once under Settings → KIE.ai Nodes Next → Connection. It's validated before it overwrites a working key, only a masked fingerprint reaches the UI, and nothing about the key goes into the workflow JSON. Headless installs can set KIE_API_KEY, which takes priority. Dependencies: requests, Pillow, numpy, PyYAML - that's it.

Where people get burned

• Source audio the API can't reach. Local file paths don't work as upload_url. Upload first, every time.
• prompt winning over lyrics in your head, but not in the payload. It's the other way round: lyrics takes priority.
• No key configured raises an explicit error naming the settings page.
• Mixed-release installs are refused on startup. If the pack's version and its generated-node build disagree, delete the folder and reinstall cleanly.
• Credits per attempt. This is a metered call, not a sampler. credits_consumed / credits_left exist so you can see the price of iteration; change one field at a time.
• The 8-minute family limit applies to upload-based source audio.

One workflow note that saves money: run the node once with everything at defaults and just a couple of words in lyrics. If the vocal character is wrong at that point, no amount of style_weight tuning will fix it - you're on the wrong tier for the sound you want, and that's worth learning for the price of a single call.