YinChao Reference Music

Drop in a clip, get back a whole new track in the same vibe. That's the whole trick of YinChao Reference Music: instead of describing a sound with words and hoping the model gets it, you hand it audio and say "like this." It's the pack's most creative node and the one where billing gets interesting, because the platform charges you by how closely the output mimics your reference.

What it actually is

The reference/cover-generation node of the YinChao (音潮) pack - an official wrapper for the YinChao platform's paid music API. Nothing runs locally. Your input AUDIO is encoded to MP3, uploaded, and the platform generates a new song conditioned on it. The result comes back as native ComfyUI AUDIO, ready to wire into anything else in your graph. Pricing is roughly ¥0.22 (about $0.03) per song, but with the twist that similarity - your main control - is a billed dial, not a free setting.

The pack is new and has zero community track record yet, so treat the output quality as unproven. What is proven is the plumbing: the node handles upload, async submit, polling, download, and decode entirely internally.

The inputs that matter

• audio - required. Your reference clip as native AUDIO. Keep it under the 10 MB upload cap and mono or stereo.
• similarity - a combo with exactly four stops: 0.2, 0.8, 1.3, 1.5, defaulting to 0.8. This is the fidelity dial, and since the platform bills by similarity, higher isn't always better - it's pricier. Start at 0.8, and only push toward 1.5 when you've confirmed 0.8 isn't giving you what you want.
• prompt - optional multiline. Style notes layered on top of the reference, for when "like this" isn't quite the whole instruction.
• lyric - optional multiline. Give it words for a true cover; leave it blank for an instrumental-style rework.

One thing to know: the generation model is pinned internally to v3.5 for reference tasks, same as the extend node. You don't get a model choice here.

How it works

Encode to MP3 (PyAV, 192 kbps), upload as a reference file, submit an async reference task, poll every 3 seconds for up to 10 minutes, download, decode. The song count is hard-coded to n=1 so a run can't silently generate two songs and charge you twice. The one thing the node won't do is quietly degrade your input: hit the 10 MB cap and it fails with a clear error instead of re-encoding your reference into something worse.

What comes out

Five outputs, same as the other song nodes: audio (native AUDIO), title, lyric, audio_url (direct hosted MP3 link), and task_id.

Install

Same pack as the other three YinChao nodes. Search "YinChao Music" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/yinhcao/ComfyUI-YinChao
python -m pip install -r requirements.txt   # requests, av>=12, numpy

Restart ComfyUI. PyAV bundles its own codecs, so no system ffmpeg install, and the pack needs a current ComfyUI (V3 schema).

Grab an API key at platform.yinchaoyongxian.com and set it under Settings → YinChao → API Key - it's deliberately not a node input, so it never lands in workflow JSON. Headless setups use the env var:

export YINCHAO_API_KEY="your-api-key"

Common issues

• 10 MB upload cap. At 192 kbps that's roughly seven minutes of reference audio; anything longer errors clearly.
• Mono or stereo only - other channel layouts are rejected.
• Similarity bills you. A 1.5-similarity run costs more than a 0.2 one. Test cheap, then commit.
• Your reference audio leaves the machine and passes through the platform's content moderation, so this isn't the node for anything you don't want on a server you don't control.
• Missing key / no balance / moderation rejection all fail loudly - the pack's house style is explicit errors, never a fake-empty success.