Maintain human-in-the-loop for brand alignment while prioritizing high-quality reference images and using LLM-based evaluation to accelerate iteration; design the system to tackle compound visual consistency challenges across scenes.

The system slashes production time from weeks to hours and scales across multiple customer segments by leveraging AWS services (SageMaker, Bedrock, Lambda, Step Functions, ECR, ECS) and advanced AI models (Anthropic Claude Sonnet 3.7, Wan 2.1/2.2, Segment Anything, Nova for frame selection).

Readers can replicate the approach by exploring Amazon Bedrock and SageMaker and by reaching out to the AWS Generative AI Innovation Center for collaboration.

The creative ideation pipeline unfolds in three stages: (1) generate customer segments using Bedrock/Claude to define personas; (2) produce 4–6 high-divergence concepts via creative briefs; (3) refine storyboards with stochastic feature sampling and human-in-the-loop review to ensure brand alignment and precise audiovisual specs.

AI-generated ads achieved higher narrative coherence and a 25% gain in originality versus the existing library, delivering a 15–30 second spot in about 12–15 minutes on ml.p4d.24xlarge SageMaker instances, with multi-GPU sharding sustaining per-scene efficiency.

The architecture comprises four layers: data/storage (S3, ECR); processing/orchestration (Lambda, Step Functions); GPU compute (multi-GPU SageMaker with tensor parallelism and FSDP, TorchServe, ECS for speech synthesis); and a React/Cognito-based user interface for review and approval.

The video pipeline covers five modalities (text, image, video, audio, overlays) using curated models (Wan 2.1 for reference-to-video and text-to-video; Sesame AI Lab for speech), with an LLM-driven quality judge that measures narrative adherence, visual quality, and brand compliance and auto-regenerates when needed.

Bark.com and AWS collaborated to build a scalable AI-powered video-generation pipeline designed to produce personalized social content rapidly without sacrificing quality or brand consistency.

To sustain visual consistency across scenes, the project employs semantic consistency via element extraction, blueprint generation, and prompt transformation, plus a visual-consistency framework using reference image propagation (Nova frame analysis, Segment Anything for segmentation, reference propagation for continuity).

Ablation studies confirm the importance of reference image propagation and hierarchical scene planning for engagement and visual coherence, underscoring the role of narrative-element extraction in maintaining consistency.