Adds sync_strength (0.0–3.0, default 1.0) to PrismAudioSampler.
The scale is applied post-conditioner (after Sync_MLP) to the conditioning
tensor before it enters the DiT. Since CFG always uses zeros as the null
sync embedding, this cleanly scales the sync guidance signal:
effective_sync_guidance = cfg_scale * (sync_strength * cond - 0)
Higher values tighten temporal audio-video alignment; 0.0 disables sync
guidance entirely (audio conditioned only by video + text features).
Not applied in T2A mode where sync is replaced by the learned empty_sync_feat.
Also logs sync temporal coverage vs audio target duration, with a warning
when they differ by more than 0.5s (stale or mismatched features).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ethanfel
a315093743