All settings explained
Every Weavy Studio / YuE2 control: composition mode, style strength (CFG), takes, seed, length, ABC score and per-stage sampling (temperature, top-p, top-k, repetition penalty).
Open More options in the Create panel. Defaults are tuned; change one thing at a time.
Composition
- Full plan (default) – the model first writes a score: melody and chords in ABC notation, then sings over it. Most stable structure.
- Melody – plans only the melody line; the accompaniment is free.
- No plan – straight from lyrics and style. Looser, sometimes more surprising.
Style strength (CFG)
Classifier-free guidance: how strictly the audio follows the style and lyrics. 1.0 = off, 2.0 = default (cleaner vocals, closer to the style), above 3 can sound stiff.
Versions
How many takes to render (1–8). Each uses its own GPU slot, so 3 takes take about as long as 1. Picking the best of several is the single most effective quality lever.
Seed
Empty = random. The same seed with the same settings reproduces the same song. Lock gives every take the same seed (useful to compare settings).
Max length
Upper bound of the song in seconds (30–360). The audio stage produces about 25 tokens per second; the limit caps that stage.
Score (ABC)
Paste your own melody and chords in ABC notation to have them sung (Full plan / Melody modes). Every finished version shows the score it used under Details → Score – copy it, edit a bar, render again.
Sampling – two stages
Generation has two token stages, each with its own sampler:
| Setting | Score stage (default) | Audio stage (default) | Effect |
|---|---|---|---|
| Temperature | 0.70 | 1.00 | randomness; lower = safer, flatter |
| Top-p | 0.90 | 0.95 | keep the most likely tokens up to this probability mass |
| Top-k | 30 | 100 | keep at most k candidates |
| Repetition penalty | 1.005 | 1.20 | discourage loops |
| Penalty window | 100 | 50 | how far back the penalty looks |
Changed values are marked with a dot (•). ↺ Defaults restores everything.