Generate Music · V6_MINI

The one you probably want first

If you're going to learn a single node from this family, learn this one. KIE_Next_V6_MINI_e5ffe3de is prompt-to-song on KIE's V6_MINI tier, and unlike its Cover/Extend/Add-Vocals siblings it doesn't need source audio - just a description, a style and a title, and it also accepts image, audio and video references as real ComfyUI sockets.

Where the other nodes in this pack shine is stitching one specific job onto audio you already have. This one is the workhorse underneath: a song from nothing, in a graph, so the output can flow straight into a video pipeline instead of living in a browser tab. That's the argument for API music nodes in the first place - generating the track is only half the job, and having it already inside ComfyUI as an AUDIO object is the other half. Local models cover a lot of this ground free; clean vocals and lyric phrasing are still the hosted models' advantage.

The mode matrix you have to understand

Two required booleans decide almost everything, and the tooltip lays it out:

• custom_mode = true (the default) means you're in control: style and title carry the song, and the extra weights apply. instrumental then splits it - true needs only style and title; false additionally needs prompt, which is sung verbatim as your lyrics.
• custom_mode = false is the simplified mode: prompt becomes a description and lyrics get written for you. Then instrumental true needs style + prompt, and false needs style, prompt and lyrics, where lyrics acts as an attachment alongside the description in prompt.

Most of the "why did it ignore my lyrics" and "why is my description being sung" confusion is just these two toggles disagreeing with your expectations. There's one automatic behaviour worth knowing: image references are only legal when custom mode is off, and the pack detects that combination and switches custom_mode to false for you locally, dropping the now-invalid custom-only fields before the request goes out. It prints a line about it rather than failing remotely, which is the right call.

The rest of the inputs

style, title, negative_tags - the genre, the label, and the exclusion list. duration - 10 to 360 seconds, and only valid in custom mode. persona_id + persona_model - apply a saved Suno style or voice persona. style_weight, weirdness_constraint, audio_weight, all 0.65 - how strictly the style holds, how experimental it gets, and how the audio features are weighted (that last one does nothing on vocal-free output, per the tooltip). variety - 0 to 4, from "same style twice" to "unreasonably varied." vocal_gender - a probability nudge only, in custom mode only.

And the references: image_urls takes a normal IMAGE socket (batches welcome), while audio_1/audio_2/audio_3 and video_1/video_2/video_3 each take one native audio or video input. Anything you connect gets uploaded to KIE automatically and swapped in as a URL - no copy-pasting links.

Outputs

audio - the finished track, native ComfyUI audio. url - the first result URL. all_urls_json - every URL in the response, and your route to the second take, since only the first reaches the socket. task_id - the job handle. raw_json - the full provider payload, which is where you look up the track id that Extend Music needs. credits_consumed and credits_left - the spend, also drawn on the node after the run.

Install

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

Restart ComfyUI. Dependencies are requests, Pillow, numpy and PyYAML - no model files, nothing to download, zero VRAM. Windows users have the one-click installer in the repo's dist/ folder. ComfyUI Manager support is the pack's stated direction (registry metadata ships under publisher felipeosilva); the clone is the route that works today.

Then paste your KIE.ai API key once in Settings → KIE.ai Nodes Next → Connection - validated before it replaces a working key, shown only as a masked fingerprint. Servers can use KIE_API_KEY, which takes priority.

Failures worth expecting

The node waits for the job by polling, 2.5 seconds rising to about 15, until timeout_seconds (1200 by default) expires. If it times out, the task usually isn't dead - check the task_id with KIE • Wait for Task before resubmitting and paying for a second run. Provider failures surface with KIE's reason in raw_json; content-filter rejections on explicit lyrics are the common one. Empty-reference combinations get rejected upstream too: in non-custom mode you need at least one of the reference inputs, style, or lyrics, and in custom mode at least style, lyrics or negative tags.