Add Instrumental to Music · V6_WILD

What this node is

Give KIE_Next_V6_WILD_56abd8db an audio URL and it comes back with an instrumental rendition of that track, performed by Suno's V6_WILD model. KIE's blurb for that version - "bolder, more distinctive musical expression" - is the reason to choose it here: where the tamer versions tend to reproduce your arrangement faithfully minus the singer, WILD is more willing to reinterpret.

That makes it the right tool when what you want is a different instrumental, not a karaoke backing. Think a score rebuild for a video cut, a bed for narration that shouldn't sound like the vocal track's leftovers, or an experiment where you hand it a busy mix and see what it picks out.

If you want your song back with the vocals surgically removed and nothing else changed, that's a stem-separation job, not this one - and the pack has a node for that too.

How the node works

It assembles the request, submits it to KIE's task queue, then polls for completion: every 2.5 seconds at first, backing off toward 15, all bounded by timeout_seconds (default 1200, up to 7200). When the job lands, the returned audio is downloaded and decoded into ComfyUI's AUDIO object. V6_WILD is carried inside the request body while KIE's market model name carries the outer envelope - the split that used to be wrong in builds before v0.4.10, which is worth knowing if you see model-level rejections on an old install.

As with every node in this pack, the model runs on KIE's side. Your upload leaves the machine, KIE's servers must be able to fetch it, and the run debits your KIE credit balance - reported back on the node once it finishes.

Required fields, and the one that isn't called what you expect

• upload_url - the source track, as a public URL. Local file? Route it through KIE • Upload Audio and paste the result.
• title - required, cosmetic, 100 characters-ish.
• negative_tags - required, the exclusion list. Keep it tight ("heavy metal, fast drums" is the shipped default) and don't use it to describe what you do want.
• tags - required, and this is where the desired style goes. Note that the sibling Add Vocals nodes call this field style; this endpoint documents it as tags. Same job, different key - copy-pasting a workflow between the two families is where you'll notice.

Optional controls: lyrics, allowed by the schema but not what you want populated on an instrumental job. vocal_gender, likewise present for shared-plumbing reasons. style_weight (0.61) - how strictly the tags are honoured; weirdness_constraint (0.72) - how far it's allowed to wander, and the WILD-specific dial that matters most; audio_weight (0.65) - how much of the source's character survives, which is exactly the "faithful instrumental vs. reinterpretation" axis. variety (0–4) decides how different the two takes you get back are from each other.

Outputs

audio is the instrumental as native ComfyUI audio - Save Audio, mux it under a video, or feed it into another music node. url is the primary result URL; all_urls_json is the full list and is where the second take lives, since the socket only carries the first. task_id identifies the job, raw_json is the complete provider response, and credits_consumed / credits_left show the 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 afterwards. Dependencies are requests, Pillow, numpy, PyYAML - nothing to download beyond that, no GPU load, no model files. A one-click Windows installer lives in the repo's dist/ folder, and ComfyUI Manager support is the plan (registry metadata is present under publisher felipeosilva), though the git clone remains the install that always works.

Set the API key once in Settings → KIE.ai Nodes Next → Connection. The pack validates it against KIE before overwriting a saved key and never displays the full key in the UI. Headless servers can set KIE_API_KEY, which takes priority over the saved one.

Where it goes sideways

WILD means WILD: if the output is unrecognisable as your track, that's the model working as advertised - lower weirdness_constraint and raise audio_weight to pull it back. Vocals showing up in the "instrumental" usually trace to something in tags, lyrics or style describing a singer. Failed jobs print a reason into raw_json; unfetchable URLs are the most common cause. And a timeout is not a failure - the task often completes upstream, so check task_id with KIE • Task Status or KIE • Wait for Task before resubmitting and paying twice.