Generate Music · V6

This is the current-tier song generator: text in, a full track out, with vocals that hold up. If you've looked at the · V5 and · V5.5 versions of this node, here's the short version - KIE's docs mark both of those discontinued. V6 is the live tier, with a 5000-character lyrics budget, a 1000-character style budget, and materially better vocals.

The wider context hasn't changed. Local music generation means ACE-Step: free, 4GB-friendly, genuinely good at instrumental music, and the known weak spot exactly where you'd expect - vocals and lyrics. Suno's vocal is the reason this remains an API call for anything sung. So the split is: local for beds, loops and cheap iteration; this node when the singing is the point. Every call costs KIE credits and sends your prompt out.

Two siblings are worth knowing about before you commit: the pack also ships separate nodes for V6_WILD (the docs' description: pushing creative boundaries for bolder, more distinctive expression) and V6_MINI (lightweight and fast, balancing quality and speed). Same endpoint, different tier - and since they're separate classes, you can wire both and compare for the price of two calls.

How it works

The pack builds its node library from KIE's official API documentation at startup, one node per documented model version, which is why there's no model dropdown here - V6 is baked into the class. The node assembles the documented request body, submits it to KIE's task endpoint (envelope model ai-music-api/generate, inner model V6), then polls KIE's record-info endpoint every couple of seconds with exponential backoff until the task reports success. On success it downloads the audio result and returns it as a native ComfyUI AUDIO; on failure it raises with the task id and the provider's message.

The inputs that matter

Three required, and they define your mode:

• custom_mode - default true (detailed control). Set it false for the simplified path.
• instrumental - default true. The docs' matrix: in custom mode with this true, only style and title are required; with it false, style, title and prompt are required and prompt is sung as the exact lyrics.
• prompt - exact lyrics in custom mode, a loose brief in non-custom mode where lyrics are written for you (3000-character cap there).

Optional, mostly custom-mode-only: style, title (80 characters, required in custom mode), negative_tags, vocal_gender (m/f; the docs only claim it raises the odds), style_weight, weirdness_constraint, audio_weight (docs: unsupported when there are no vocals), persona_id / persona_model, duration (10–360s, valid on this tier), variety (0–4), lyrics (5000-character cap, takes priority over prompt).

Then the reference sockets, which are the nice part: image_urls takes an IMAGE, and audio_1/2/3 plus video_1/2/3 take native AUDIO and VIDEO inputs. The pack uploads them and passes URLs for you - no plumbing.

Outputs: audio, url, all_urls_json (the endpoint's example response contains two tracks; this is where you keep both), task_id, raw_json, credits_consumed, credits_left.

The trap the pack already handles

Attach an IMAGE to image_urls and leave custom_mode at its default and KIE rejects the request - after uploading your file. The pack intercepts it: on Generate Music it forces non-custom mode locally and strips the fields that aren't valid there (duration, style, title, negative_tags, the weight and vocal fields, persona_id, instrumental), logging exactly what it removed. Sensible, but it means the submitted payload isn't literally what your widgets say. Read the console.

Installing

ComfyUI Manager, search "KIE Nodes Next" - the README describes registry publication as prepared rather than shipped, so the copy-paste route is what works today:

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 at Settings → KIE.ai Nodes Next → Connection. It's validated before replacing a working key, only a masked fingerprint is shown to the frontend, and KIE_API_KEY takes priority on headless installs. Four small dependencies; no torch, no checkpoints, no VRAM.

Where people get burned

• Mode mismatch. Custom mode, instrumental true, lyrics filled in → no lyrics. The matrix is not decorative.
• Two tracks per request. audio is the first; the rest live in all_urls_json.
• No key configured raises an error naming the settings path.
• Mixed-release installs are refused at import. If version and generated-node build disagree, delete the folder and reinstall - don't patch by hand.
• Credits per iteration. credits_consumed and credits_left print on the node; watch them while tuning, because this is the expensive node in the family.

Run it once with defaults, then change a single field per run. You're not sampling here, you're paying - and the fastest way to burn credits is changing style, lyrics and duration together and learning nothing.