Engines: YuE2, HeartMuLa, ACE-Step
Weavy Studio runs three open music models. What each is good at, what it reads, its licence, and when to pick it.
Pick the engine at the top of the Create panel. Lyrics, title, language, seed, versions and length work the same everywhere; the extra options change per engine.
| YuE2 | HeartMuLa | ACE-Step | |
|---|---|---|---|
| Model | YuE2-3B (M-A-P) | HeartMuLa-RL-oss-3B + HeartCodec | ACE-Step 1.5 XL + LM planner |
| How it works | writes a melody-and-chord score (ABC), then sings it | autoregressive tokens, RL-tuned to follow style tags | an LM plans the song, a diffusion model renders it |
| Reads the style as | a free description | short tags (male vocal, rock, harmonica) | a caption |
| Extra controls | composition mode, CFG, ABC score, two samplers | CFG, temperature, top-k | artist voice (LoRA + reference record), planner on/off, steps |
| Licence | CC BY-NC 4.0 | Apache 2.0 | MIT |
YuE2
The most controllable: see and edit the score it sings, tune both samplers. Clear diction, Polish included. Weights are non-commercial for companies (details).
HeartMuLa
Commercial-friendly. It ignores long phrases, so the studio shortens your style into tags automatically: the voice first (male vocal / female vocal), each phrase cut to three words, BPM and key dropped. The preview under More options shows exactly what is sent.
ACE-Step
Commercial-friendly and the only engine that can keep an act's own voice: choose a voice such as Luka and the studio loads that artist's voice LoRA on the ACE server and sends one of their records as the reference voice. Versions of ACE jobs run one after another, because the voice adapter is shared on the server.
From the catalogue
⤓ From catalogue (top of the Create panel) searches the WeavyAudio catalogue. Picking a track fills in its title, artist, language, lyrics and style. With ACE-Step the track also becomes the source:
- Cover keeps its melody, rhythm, chords and arrangement and re-renders it in your style and voice (cover strength 1.0 = to the letter; lower gives the style and voice more room). The length follows the original.
- Reference voice takes its vocal timbre and mix as the reference instead.
Both work together with an artist voice (LoRA) — e.g. Luka singing a cover of a catalogue song.