Layer 0101
Scene Planning Layer
An LLM analyzes the campaign brief and produces a structured shot plan — a JSON breakdown of each scene: duration, setting, mood, framing. The "director" layer, ensuring clips tell a coherent story, not a disconnected set of AI generations.
Brief → JSONStructured shot planNarrative coherence
Layer 0202
Consistency Anchor Layer
The core of the system. A detailed style prompt describing exact visual treatment — lighting, color grade, camera style — is prepended to every clip's generation call. A reference image is passed into the video model's image-to-video mode, so the product and model appearance are anchored, not reimagined, each time.
Style promptImage-to-video anchoringZero drift
Layer 0303
Generation Layer
Individual 6–8 second clips generate in parallel via Google Veo (primary) and Kling (fallback, noted for strong reference-consistency). A background job queue handles many clips simultaneously, making multi-clip outputs practical within a reasonable wait time.
Veo · primaryKling · fallbackParallel job queue
Layer 0404
Assembly Layer
ffmpeg stitches clips together, applies a consistent color grade so every clip reads as one production, and mixes in audio — delivering a final cut that feels authored, not assembled.
ffmpeg stitchUnified color gradeAudio mix