PetGenAI: Automated Pet Portrait Generation

Overview & Goal

A printing factory wanted to offer a scalable personalization service where customers could upload photos of their pets, and designers would create unique artwork based on those images. The existing process was slow and labor-intensive, requiring manual editing for every order. The goal was clear: automate and scale the design pipeline so a single designer could handle up to 10x more orders without compromising quality.

Challenges

Building a scalable personalization pipeline involved multiple layers of complexity:

• High-Volume User Media: Handling potentially many gigabytes of customer-uploaded images daily.
• Scalability: Managing workloads that varied from steady daily demand to sharp traffic spikes around promotions.
• Consistency of Generated Artwork: Ensuring the pet subject remained recognizable across multiple styles and post-processing steps.
• Integration with Existing Systems: Seamlessly connecting to the factory's CRM.
• Hybrid Model Hosting: Balancing local GPU workloads with external API calls to manage costs while keeping latency low.

Solution

We delivered a distributed ML-powered service that automated image processing, design generation, and delivery into the factory's workflow.

Key System Components

Design Decisions

• Stable Diffusion chosen for its adaptability and open ecosystem of fine-tuning and LORAs.
• AWS ECS ensured GPU scaling and reliable batch inference for large order sets.
• AWS Lambda + SQS offloaded sudden surges without over-provisioning costly GPU instances.
• Pipeline Modularity: Each stage (segmentation, inpainting, style application) isolated for debugging, replacement, and scaling.
• Adobe API integration preserved final professional quality and compatibility with existing print workflows.

This project demonstrates end-to-end ML engineering: from model fine-tuning to production deployment, handling real-world scalability challenges while maintaining quality standards.