How JoyCastle Accelerated a 100k+ Ad Asset Library with Amazon Nova Multimodal Embeddings

Background

JoyCastle, a global mobile‑game developer, maintains a rapidly expanding advertising asset library (over 100,000 videos and images) used for user‑acquisition campaigns. Manual tagging and keyword‑only search became bottlenecks, causing creative teams to spend excessive time locating suitable assets.

Challenge

High cost and inconsistency of manual labeling.

Keyword search lacks semantic understanding, failing on natural‑language queries such as “character being flicked by a finger”.

Cross‑modal retrieval (e.g., image‑to‑video) is not supported.

Low reuse rate; many high‑quality assets remain buried in folder hierarchies.

Solution Overview

In collaboration with AWS Game Industry Solutions, JoyCastle built an AI‑powered asset management system based on Amazon Nova Multimodal Embeddings . The model maps text, image, video, and audio into a single semantic vector space, enabling unified vector search and automatic video segmentation that matches the typical 5‑second clip needed for ad creatives.

Key Technical Advantages

Segmented Embedding: videos are automatically split into 1‑30 s clips, each with its own embedding, eliminating intermediate conversion steps.

Four embedding dimensions (3072, 1024, 384, 256) allow trade‑offs between accuracy and cost.

Supports both synchronous (real‑time) and asynchronous (batch) processing.

System Architecture

The pipeline runs on AWS services:

Amazon Bedrock – provides access to the Nova model.

Amazon OpenSearch Service – vector database with HNSW index and millisecond‑level KNN retrieval.

Amazon Lambda – serverless compute for embedding generation and search requests.

Amazon S3 – stores raw assets and generated embeddings.

Amazon SQS + DynamoDB – task orchestration and state tracking for asynchronous jobs.

Two core workflows are illustrated in the figures below.

Figure 1: Asset ingestion flow.

Figure 2: Asset retrieval flow.

Core Capabilities

Multimodal input & cross‑modal search: text‑to‑video/image/audio, image‑to‑video, video‑to‑video, audio‑to‑audio.

Semantic video segmentation returns precise timestamps and similarity scores for each clip.

Flexible vector dimension selection to balance precision and cost.

Sync mode for low‑latency queries; async mode for large‑scale batch embedding.

Practical Deployment

A sample repository provides a one‑click deployment script:

# Clone the sample repo
git clone https://github.com/aws-samples/sample-multimodal-embedding-models
cd sample-multimodal-embedding-models

# Deploy
./deploy_model.sh nova-segmented

Prerequisites: an AWS account with Bedrock access to the Nova model, AWS CLI v2, Node.js 18+, CDK v2, and Python 3.11.

Evaluation

Testing on 170 game‑ad assets (130 videos + 40 images) produced the following results:

96.7 % recall success rate – the target content was retrieved.

73.3 % high‑precision recall – the target appeared in the top‑2 results.

Cross‑language capability: Chinese query scored 78.2 % vs. English 89.3 % (3072‑dim embedding).

Best‑Practice Recommendations

For advertising‑asset scenarios, use the SEGMENTED_EMBEDDING mode with a 5‑second segment length and the 1024‑dim embedding, which offers the best accuracy‑cost balance for most workloads.

Conclusion

The AI‑driven pipeline eliminates manual labeling, supports natural‑language and cross‑modal queries, and dramatically improves creative productivity. It also opens the path to automated slicing, intelligent stitching, and closed‑loop performance testing, redefining how game advertisers generate and manage creative assets.