Introduction

You already know YuE2 plans a symbolic ABC score before it sings. Now the community just opened the pocket wider: hum a melody for 10 to 30 seconds, drop style + lyrics, and get a finished song that keeps your melody, builds structure around it, and keeps going long after your hum stops.

Mothersuperior released the YuE2-hum-to-song (“humsong”) adapter on Hugging Face, and Kytra (@sin_ceriously) shipped a dedicated ComfyUI node pack plus demo video. This isn’t a new base model—it’s the missing bridge from your hummed hook into YuE2’s planner and decoder.

How Hum-to-Song Actually Works

YuE2 writes songs in two stages: an AR planner drafts an ABC score then 25 Hz semantic tokens; a NAR decoder turns tokens into VAE latents and 48 kHz audio. Melody is decided in the planner—so a hum has to enter in two places:

  • Score continuation (no training) — SheetSage2 (melody-only) turns the hum into YuE2’s two-voice ABC. The prompt leaves the score open (no ABC_END), so the planner keeps writing sections, transitions, and an ending in the same key/meter/range. On their tests the hummed phrase recurred as the hook two to four times later. A ~20 s hum can become a 2–3 minute song.
  • Prosody adapter (trained) — the decoder never sees raw hum timbre. Pitch is tracked (pYIN), turned into a filtered sine carrier, VAE-encoded, and injected at decoder layers 0, 7, 14, and 21 via LoRA + projections. That preserves how you hummed it—timing and contour—without locking your mic timbre.

Melody modes: continue from hum, hum only (close the score after your bars), or ignore hum (planner writes its own score; adapter can still shape phrasing). Inference also exposes hum influence via classifier-free guidance on the hum channel.

Training notes from the card: ~3,163 self-supervised pairs from separated vocal stems, partial-hum rules so “hum a bit, get a whole song” is the default, combined release as a rank-96 LoRA against stock YuE2-3B. License inherits CC BY-NC 4.0—non-commercial only.

Pro Tip

Hum something voice-like into a clean mic—pure sine tones won’t transcribe. Treat the hum as the opening and the hook, not the total length; the planner decides where the song ends. Dial hum influence carefully; the adapter is subtle by design and shines most in the first hummed section.

How to Get Started with YuE2 Hum-to-Song

  1. Grab base YuE2-3B plus the adapter weights from Hugging Face (hum_adapter_v1_combined.safetensors for stock YuE2, or the ComfyUI-layout safetensors).
  2. Scripts path: hum_continue.py for open-score continuation, then infer_hum.py for prosody rendering to FLAC.
  3. Or load Kytra’s ComfyUI HumSong pack: Loader, Hum Input (file or mic), Hum to Song, HumSong Output (score view highlights hummed bars).
  4. Watch the demo, then iterate style/lyrics/seed with a short hum.

Model: Mothersuperior/YuE2-hum-to-song · ComfyUI suite: ComfyUI-FS_Audio_Suite · Release post:

· Demo: YouTube · Base: m-a-p/YuE2-3B

Conclusion

Hum-to-song is the YuE2 community drop that finally lets you own the melody before the bounce—score continuation for what you hummed, prosody adapter for how you hummed it. Open weights, ComfyUI nodes, CC BY-NC caveats on the base.

Hum a hook, open the score, and let that unfiltered motion finish the song.

—Anabel ♡