YuE2 Field Report 4: No Sequential Production Path
No audio input, no stem output, and an identical supplied score yields a backing correlation of +0.133. Capability matrix and measurements.
TL;DR — Sequential production is unavailable in YuE2. No audio input, no stem output, and an identical supplied score yields a backing stem correlating at +0.133 against a +0.95 same-performance reference.
Fourth field report on local YuE2. Question tested: can a beat be fixed first and a melody added afterwards, as in a conventional production workflow. Reporting date: 21 September 2026.
Capability matrix
Unavailable:
- fixing a beat and singing over it — no audio input exists
- backing-only render — empty lyrics yields wordless singing
- backing reuse across takes — identical scores correlate at 0.133
- stem output — one mixed file per render, no code path emits parts
Available:
- exact melody and harmony control — score round-trips byte-identical
- passage-level edits — measured surgical in an earlier test
Test 1 — empty lyrics field
Common guidance: leave lyrics empty to obtain an instrumental. Measured against the same song generated with lyrics, after source separation:
- vocal-stem RMS 0.0934 (empty) vs 0.0872 (with lyrics)
- vocal share of total energy 38% vs 41%
- score still carried 371 vocal notes
An empty lyrics field yields wordless singing, not a backing track. The guidance appears to describe the browser demo rather than the Python runtime.
Test 2 — score inheritance
- A (lyrics empty) — 116 BPM, 175.9 s score, 371 vocal notes, 85 chords, sha
0f5984530f9d - B (A score + lyrics) — 116 BPM, 175.9 s, 371 notes, 85 chords, sha
0f5984530f9d - C (control, no score) — 116 BPM, 136.6 s, 254 notes, 64 chords, sha
8b7d4a491867
Transfer is exact: identical SHA-256, project comparison tool returns match: true, zero differences. Composition is fully under control.
Test 3 — does an identical score reproduce the backing
- A vs B, identical score — +0.133
- A vs C, different scores — +0.027
- B vs C, different scores — +0.036
- calibration, same latents through two decoders — +0.954 to +0.991
It does not. Backing band energy also diverges: 20-60 Hz falls 23.9% → 18.4%, 60-120 Hz rises 37.9% → 45.8%. Kick-to-bass weighting is redistributed between renders of the same composition.
Documented cause: "Editing style, lyrics or ABC requires new semantic generation and synthesis. Cached latents are reusable when changing only the decoder, not when changing the music." Behaviour matches the documentation exactly.
Scope of score control
Fixed by abc=: pitches, durations, chord symbols, meter, tempo, section boundaries — verified byte-identical.
Not fixed: instrument balance, groove, mix, performance.
Prior measured result remains valid: a 32-note edit in a supplied score moved the edited window +1.8 semitones with control windows identical to the hertz. Score-level control is precise; it operates on the composition layer only.
Available workarounds
- Post-production — render twice, separate, recombine. Editing rather than generation; quality bounded by separation artefacts.
- Different model — one accepting audio conditioning or emitting stems. The YuE2 runtime contains no code path for either.
Reference: measurements
Raw figures behind the tables. Method and the fixed/not-fixed breakdown: Axortex.
The numbers
vocal-stem RMS A (no lyrics) 0.0934 C (lyrics) 0.0872
vocal share of energy A 38% C 41%
backing-stem correlation
A vs B (identical score) +0.133
A vs C (different scores) +0.027
B vs C (different scores) +0.036
same latents, different decoder +0.954 to +0.991 <- what 'same performance' looks like
backing band energy, 20-60 / 60-120 / 120-250 / 250-2k / 2k+ Hz
A 23.9 37.9 26.1 8.5 3.6
B 18.4 45.8 19.8 12.0 4.0
The kick-versus-bass balance moves by five to eight percentage points between two renders of the same composition. That is not a re-mix of one performance; it is a second performance.
References: generation-and-covers · generation · repo.
Image: William Braid White, sectional illustration of player piano interior action, 1909, Public domain, via Wikimedia Commons.
← Back to all posts