Extend Music · V6_WILD

The short version

KIE_Next_V6_WILD_688eeb10 continues a Suno track by its audio_id, running the V6_WILD model - the version KIE pitches as pushing "creative boundaries for bolder, more distinctive musical expression." It's the same mechanism as its V6_MINI sibling, with a model that's more likely to take your track somewhere you didn't plan.

Two nodes in this pack look almost identical and are worth telling apart before you wire anything. Extend Music continues a track that lives on Suno's side by id. Upload And Extend Audio continues a file you upload by URL. Both exist because those are genuinely different starting points, not because the author got bored.

WILD is the right pick when the original was a springboard rather than a finished thing. Instrumentals, jams, a two-bar loop you want turned into a piece of music. For "extend my finished pop song by eight bars, don't embarrass me," you want one of the tamer versions.

Mechanism

The node builds the extend payload, hands it to KIE's task API, and then waits. Polling starts at 2.5 seconds and backs off toward 15 until the job resolves or timeout_seconds (default 1200, max 7200) expires. When the task reports success the audio is downloaded and decoded into ComfyUI's AUDIO dict. The V6_WILD version string rides inside the request payload while the outer envelope uses KIE's market name for the extend endpoint - the split that pre-v0.4.10 builds got wrong, so if submissions fail on model grounds, update and restart before hunting elsewhere.

No local model, no VRAM. Your ids and prompts go to KIE, and you pay in credits, which the node prints on itself once the run finishes.

Inputs

• audio_id - required. Take it from a previous Suno generation's track entry, not its task id. That mix-up is the most common reason a run comes back as a provider-side failure.
• prompt - lyrics for the new material. Required widget, though the source id alone is technically enough.
• continue_at - where the extension begins, in seconds, default 60. It has to be inside the original track's length. Leaving a little room before the end gives the model something to blend with and hides the join.
• instrumental - optional, defaults to false, which is the opposite default to the Upload And Extend nodes. Flip it if you're extending an instrumental, and note that when it's true the docs say not to pass lyrics, prompt or vocal_gender.
• style and negative_tags - optional steering for the continuation. The docs suggest matching the original's style; with WILD I'd suggest matching it in style and letting weirdness_constraint do the wandering instead.
• lyrics - explicit lyrics, which take priority over prompt as lyrics. Structure tags like [Verse] and [Chorus] are the expected shape.
• style_weight, weirdness_constraint, audio_weight - 0.65 each by default. audio_weight is your continuity control.
• variety - 0 to 4; the tooltip's own summary of 4 is "unreasonably varied," which is either a feature or a warning.
• vocal_gender, persona_id + persona_model - voice nudging and saved style/voice personas.

Outputs and what to do with them

audio is the extended track - native ComfyUI audio, so Save Audio, a mux node, or a downstream video pipeline all just work. url is the primary result URL and all_urls_json holds every URL the response contained; with multiple takes returned per request, reading the JSON is how you get at the alternate. task_id is the handle you'll want if you plan to chain a second extension. raw_json is the full payload for when something reads oddly, and credits_consumed / credits_left tell you what it cost.

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 when it's done. The only dependencies are requests, Pillow, numpy and PyYAML; there are no weights, no model downloads, no VRAM budget. There's a Windows one-click installer in the repo's dist/ folder, and ComfyUI Manager support is the direction the project is heading (registry metadata ships under publisher felipeosilva), but the manual clone is what works today.

Then paste your KIE.ai key once in Settings → KIE.ai Nodes Next → Connection. The pack validates it against KIE before overwriting a working key, masks it to a fingerprint, and shows a connected status. Servers can use KIE_API_KEY instead, which takes priority.

Caveats

The honest one is that extension is the least seamless thing these models do - you're asking for a new section that continues a performance it never heard as a whole. Overlapping continue_at a few seconds before the end is the practical fix. Beyond that: a bad audio_id fails at the provider, timed-out jobs are usually still running upstream (poll task_id with KIE • Task Status before resubmitting), and raw_json is where the provider's stated reason lives when something gets rejected.