Image to Song Reference YuE
Writes lyrics from your image and sings them in the style of a reference song you upload, following its vocal style, its instrumental feel or both through YuE's in-context learning.
Overview
This app takes a picture that sets what the song is about and a reference recording that sets how it sounds. Lyrics are drafted from the picture, then YuE, the HKUST and M-A-P music model, sings them with its in-context learning (ICL) checkpoint, which writes new songs in the style of the audio it is given.
The reference is first split into vocals and backing, and the clone type picks which part steers the result:
- vocal: the reference singing guides the new vocals over a fresh accompaniment.
- instrumental: the reference backing guides the arrangement under new vocals.
- both: vocals and backing guide the song together, for the closest match.
What you can build
- Parody tools that turn a user's photo into a song in the style of a track they picked.
- Jingles that echo a brand's existing music while singing about a product photo.
Inputs
- An image, uploaded or linked.
- A reference song, as a file or URL. The default is a speech clip, so test with music.
- Clone type: vocal (default), instrumental or both.
- Required style fields for genre, mood, instrument, gender and timbre, picked from YuE tags such as K-pop, cello, nostalgic or husky vocal.
- A seed (default 42).
Outputs
The finished MP3 mix, its vocal and instrumental stems, the lyrics as plain text, and the two WAV files the reference was separated into, which show how cleanly it was split.
Recommended settings
The app uses only the first 30 seconds of the reference, the length the YuE authors recommend for ICL, so start the file at the part you want it to follow. The authors report better musicality when that part is a chorus.
Limitations
- There is no lyrics field; the words always come from the image.
- The authors note that a reference raises the share of good results but narrows variety, since output stays close to it.
- The stage-1 model is YuE's English ICL checkpoint.
Safety & compliance
The YuE weights carry an Apache 2.0 license. The authors leave plagiarism checks to the creator and suggest marking published tracks as AI-generated, so check results against the reference. Use a real singer's recording as a vocal reference only with that person's consent.
API quick start
Run image-to-song-with-reference-YuE with a single API call.
{
"inputAudio": "https://your-cdn.com/input.mp3",
"audio_prompt_type": "vocal",
"genre": "Pop",
"instrument": "guitar"
}curl -X POST "https://api.wiro.ai/v1/Run/wiro/image-to-song-with-reference-yue" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"inputAudio": "https://your-cdn.com/input.mp3",
"audio_prompt_type": "vocal",
"genre": "Pop",
"instrument": "guitar"
}
JSONOther models you might like
Explore similar music generation models.



