concepts
Continuity
How Cadenza keeps shot five looking like shot one: locked references and a render feedback loop.
Last updated 2026-07-10
The drift problem
Generating shots independently re-rolls the dice each time: the character's face shifts, the wardrobe changes, the light moves. Regenerating to fix it is where budgets die.
Locked references
When a sequence starts, Cadenza locks one reference per dimension: character, style, and lighting. Every shot resolves the same locked set, so shot one and shot N share the same reference identity. Brand product images can be pinned as references too, and pinned brand references are never overwritten.
The feedback edge
When an early shot renders well, its frame is fed back to refresh the character and style references while keeping the same reference identity, so lineage is unbroken and later shots inherit the proven look.
Multimodal encoding
References are attached per model: reference-driven models receive image references with per-aspect influence weighting, others receive start frames. The encoder decides the mechanics; the guarantee stays the same.
Note: The no-drift guarantee is tested under load: concurrent sequences each keep their own locked reference set, with no cross-sequence bleed.