All tracks below are matched to the source prompt's first-second RMS. This removes HyperGANStrument's 10-16 dB raw-level advantage on the two reported weak-attack rows. The metric table still uses unnormalized model outputs.
| Metric (9 rows) | Current adaptive flow | Exact energy anchor | HyperGANStrument |
|---|---|---|---|
| Full-note LSD, lower is better | 3.4221 dB | 3.3845 dB | 5.5337 dB |
| RPA@50c, higher is better | 0.9374 | 0.9374 | 0.8900 |
| Prompt attack/sustain error, lower is better | 1.2274 dB | 0.5586 dB | 3.6520 dB |
| R3 frame-energy error, lower is better | 1.5667 dB | 0.0020 dB | 8.4826 dB |
| Onset-only target LSD, lower is better | 5.9132 dB | 5.9462 dB | 8.3430 dB |
The energy anchor improves prompt dynamics and full-note fidelity without changing pitch accuracy. Its mean onset-only target score is 0.033 dB worse, so listening remains the deciding gate.
HyperGANStrument was 10.60 dB louder in the previous raw panel. Compare attack shape here at matched level.
flowstrument-energy-anchor-keyboard-004-minus4.html
HyperGANStrument was 16.28 dB louder in the previous raw panel. Check impact definition and decay after level matching.
flowstrument-energy-anchor-mallet-056-plus18.html
Check that the energy anchor retains the attack and superior sustain already heard in the current flow.
flowstrument-energy-anchor-string-056-plus3.html
flowstrument-energy-anchor-keyboard-004-minus1.html