Generate Mashup Music · V4.5PLUS

Same endpoint as the V4.5 mashup node, different model version. V4.5PLUS is the "richer sound, max 8 min" rung of that ladder, and for a genre blend it's a reasonable pick inside the legacy family - bigger, denser results than V4.5, without moving up to the V6 nodes that KIE now treats as current.

If you're building a music bed for a render, this is the kind of node that saves you from hunting royalty-free loops: two existing tracks you like, one hybrid that matches your mood, delivered as a normal AUDIO object you can pipe anywhere.

Mechanism

Pure API call. The node assembles a POST to KIE.ai, tagged with the transport model ai-music-api/mashup and V4_5PLUS as the model version inside the payload, then polls the task until it resolves or timeout_seconds (default 1200 - twenty minutes) runs out. Nothing loads onto your GPU; the only local work is decoding the returned MP3 into ComfyUI's audio format, using the torchaudio (or PyAV) that ComfyUI already ships with.

Inputs you cannot skip

upload_url_list is required and wants a JSON array of public URLs - this endpoint fetches the audio itself:

[
  "https://example.com/a.mp3",
  "https://example.com/b.mp3"
]

Only the first two entries are used. style, title and prompt are required widgets; style is the real steering wheel, prompt doubles as lyrics only if lyrics is empty (lyrics wins otherwise), and title lands in filenames downstream so don't leave it as "Relaxing Piano" unless that's the plan.

Beyond that: variety (0 = strict single style, 1 = default balance, 2 = clearly distinct takes, 3 = bold exploration, 4 = maximum, wildly different), the 0–1 dialects style_weight / weirdness_constraint / audio_weight, and vocal_gender - which the docs describe honestly as only increasing the probability of the requested voice, not guaranteeing it. duration and persona_model are scoped to the V5.5/V6 family and are inert here. persona_id defaults to the placeholder string persona_123; empty optional strings are dropped before submission, so clearing it is safe.

Outputs

audio is the blend. all_urls_json holds every URL the response returned - mashups come back with multiple takes and audio is just the first, so this is worth wiring to a text display node if you want to compare them. url, task_id and raw_json cover linking and forensics, and raw_json is the only honest way to see what the provider actually did with your two source tracks. credits_consumed and credits_left are floats, and the node prints "credits spent / credits left" on its face after every run.

Setup

Manager → search the title ComfyUI-KIE-Nodes-Next; or:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Restart ComfyUI afterwards. It needs requests, Pillow, numpy and PyYAML - no weights, no model directory, no build step. Then add your key in Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key; the pack validates it against KIE before replacing a stored key, keeps it out of the workflow, and shows a masked fingerprint. Servers can export KIE_API_KEY instead and skip the settings dance.

The boring but important caveats

Legacy version. KIE's parameter docs list V4.5PLUS (Discontinued) with V4, V4.5, V4.5ALL, V5 and V5.5, and offer V6/V6_MINI/V6_WILD as the live family. The node exists because the docs still publish the version; if KIE drops it, the node will start erroring, and the fix is to switch to the V6 sibling rather than debug anything.

Not a remixer. You can't keep track A's drums and track B's vocal - that's stem separation plus a DAW. This blends whole songs, and the output is whatever the model decides a blend of those two styles sounds like.

Metered per attempt. Nothing gets previewed before you spend. Running variety: 3 a dozen times to browse ideas is a real expense for what is usually a background decision.

Copyright, not code. Feeding two released songs in and publishing the blend is a licensing question. The node doesn't care; your distributor might.