Project Overview (Executive Summary)

Photozig, Inc. is a leading technology provider specializing in digital solutions for health and media production. They needed a powerful, efficient way to refine image-generation models for diverse event categories. Avahi created a solution that leveraged AWS services and integrated LoRA adapters into the SDXL model, enabling Photozig to optimize image generation for multiple use cases. The outcome delivered faster development cycles, streamlined data handling, and a merged SDXL model that elevates Photozig’s image fine-tuning capabilities.

About the Customer

Photozig, Inc. develops advanced digital applications, focusing on improving health and media workflows through cutting-edge technology. With a vision to enhance creative production, Photozig serves clients seeking to leverage modern AI solutions for superior image and video generation.

The Problem (Customer Challenge)

Photozig had a specific need to fine-tune their image-generation workflows for different events—ranging from birthdays to family gatherings. Existing solutions required repetitive training and manual intervention whenever they wanted to adapt their base model for each new scenario. This process slowed the overall development cycle and hindered the rapid rollout of new features.


If left unaddressed, Photozig would continue to spend valuable time duplicating efforts, testing separate model versions, and wrestling with inconsistent outputs. The lack of a unified pipeline stood in the way of delivering timely enhancements to customers and capitalizing on new business opportunities in media production.

Why AWS

Photozig recognized AWS as the ideal platform due to its robust suite of AI and machine learning services, global infrastructure, and cost-effective scalability. Amazon SageMaker in particular offered easy setup for training advanced AI models and streamlined workflows for data management.


In addition, AWS’s flexible storage services, high-performance compute, and well-established architecture patterns ensured Photozig could quickly iterate on new features while maintaining reliability and security

Why Photozig Chose Avahi

Avahi’s deep technical expertise with AWS and proven track record in generative AI projects made it the perfect partner to guide Photozig’s image-fine-tuning initiative. Avahi’s in-house AI architects and data scientists provided the specialized skill set needed to implement LoRA adapters for the SDXL model.


Because Avahi is an advanced AWS partner, Photozig saw the advantage of leveraging best practices and tailored solutions that would scale efficiently within AWS. Avahi’s structured approach, clear communication, and ability to deliver tangible results in a tight timeframe further solidified this partnership.

Solution

Avahi devised an end-to-end AI workflow that began with collecting Photozig’s event-specific images in Amazon S3. Once uploaded, the data was used to train LoRA adapters via Amazon SageMaker, targeting categories such as birthdays and family reunions. This approach allowed each adapter to refine the SDXL base model in its respective domain. After training, Avahi merged the adapters into a single model, preserving each adapter’s unique enhancements while preventing duplication of effort. Testing and validation of the unified SDXL model took place on Amazon SageMaker notebook instances, ensuring it met
expected accuracy benchmarks.

To make the final model accessible, Avahi deployed it on Amazon EC2 GPU instances for real-time inferencing. Amazon API Gateway was introduced to manage the APIs, enabling quick integrations for any client-facing application or backend process. Throughout the engagement, Avahi followed a clear schedule to maintain scope, regularly shared progress updates, and iterated on feedback from Photozig’s team.

By consolidating all model variants and data pipelines, Photozig now benefits from a streamlined workflow that supports multiple categories in a single environment—cutting down on the effort needed to iterate on new concepts in the future.

Key Deliverables

• Trained LoRA adapters for multiple event categories and integrated them with the SDXL model
• Created a streamlined data pipeline for image-generation requests
• Set up an API layer for easy integration via Amazon API Gateway
• Validated the merged model with testing and quality assurance
• Deployed a GPU-backed inference environment on Amazon EC2
• Provided basic documentation detailing the solution architecture and usage

Project Impact

By combining LoRA adapters with SDXL on AWS, Photozig gained a flexible, cost-efficient way to produce finely tuned AI models without duplicating training efforts. The enhanced model can handle diverse user requests, accelerating the rollout of new features for media production.

• Completed training for five distinct event categories within a three-week window
• Delivered a single merged model capable of handling all categories in one workflow
• Reduced the complexity of maintaining multiple models across separate pipelines