What we did. For eight emotion pairs (e.g. Amusement + Sadness, Fear + Anger) we activate both LoRA adapters at once and compare four settings: each emotion alone (100%), both at 100%, both at 50%. Because a single generation is noisy, for every setting we draw 16 samples with different seeds and keep the best by A_z+B_z. We show the top-2 of 8 and the top-2 of 16 so you can hear how much drawing more seeds helps.
What the numbers mean. A_z / B_z = EmoNet intensities of the two emotions in the same clip; we want both high. blend / genu = naturalness proxies.
What we found. Both emotions co-occur (A_z>0.5 ∧ B_z>0.5 in the best-of-16) in 1/8 pairs. Drawing 16 seeds instead of 8 raises the best combined A_z+B_z by +0.24 on average — i.e. more seeds noticeably help surface a clip where both emotions land. Take-away: two emotion LoRAs are composable, but the co-occurrence is seed-sensitive, so best-of-N sampling is the practical way to get a clip that carries both; high-arousal emotions (Anger) still tend to dominate a quieter partner. Per-pair audio below.