Magic Memories
Generative AI

Project Objectives

1. Model and Approach Identification:
Evaluate different GenAI models to enhance captured images, considering background removal, object detection, re-lighting, style transfer, upscaling, face swap, and more.
2. Cost and Performance Analysis:
Assess hosting and inference costs, evaluate workflow efficiency, and identify models with balanced performance for realistic outputs without excessive hallucinations.
3. Creative Direction Integration:
Collaborate with creative directors to ensure AI-enhanced images align with artistic and brand standards.
4. Hosting Feasibility:
Compare the use of open-source models with private, proprietary ones, considering options for both API and self-hosted solutions.
5. User Interface (UI) Adjustments:
Propose necessary UI changes to facilitate on-demand GenAI functionalities for a seamless user experience.

Methodology

DWS used a structured approach, starting with model identification and selection based on performance metrics, complexity, and suitability for the client’s use cases.

The research assessed the following:

• Background Removal and Object Detection: By leveraging models like U-Net and YOLO, the team evaluated precision in isolating main subjects and identifying objects within images.
• Style Transfer and Story Creation: Neural style transfer was tested to apply artistic styles, while image sequence generation was considered for storytelling, providing an engaging user experience.
• Stickerization and Depth Effects: Unique capabilities like sticker creation and depth/parallax effects added layers of personalization and engagement.

Each GenAI component underwent a SWOT analysis to determine its strategic value. This led to an understanding of which models offered the best performance-cost balance and ensured the visual quality aligned with Magic Memories' brand.

Implementation and Testing

• Testing API Endpoints: Five API endpoints were created to test different GenAI functionalities, allowing Magic Memories to send an image and receive enhanced outputs. This infrastructure enabled seamless integration and testing across various use cases.
• Dataset Creation: A test dataset of source images was developed to evaluate the performance of each GenAI workflow, focusing on maintaining image accuracy, avoiding visual inconsistencies, and ensuring a cohesive look across generated images.
• Creative Director Collaboration: Creative directors were engaged in prompt engineering, curation, and iterative refinement of AI-generated content, bridging the gap between technical capabilities and artistic standards.

Results and Recommendations

DWS provided a comprehensive final report outlining findings and recommendations, including:

• Model Recommendations: Suggestions for the optimal GenAI models for different tasks, emphasizing the cost-effective performance of open-source models and the precision of private ones for high-detail tasks.
• Cost and Infrastructure: A detailed breakdown of hosting costs, emphasizing the balance between GPU performance and resource allocation for high-demand scenarios.
• UI Modifications: Suggested UI changes to accommodate on-demand GenAI functionalities, enhancing user engagement through intuitive access to personalized image enhancements.
• Creative Workflow Integration: Recommendations for incorporating creative directors in the GenAI workflow to maintain brand consistency and creative vision across AI-enhanced images.

Conclusion

The partnership enabled Magic Memories to explore cost-effective, high-quality GenAI options that could enhance customer engagement by transforming captured images into memorable, shareable moments. By establishing a foundation for GenAI-driven workflows, the case study illustrated a successful framework for merging AI technology with creative vision, creating a dynamic experience for customers and laying the groundwork for future innovation.