What started as "just implement four image generation use cases" turned into a journey of technical nightmares, stakeholder frustration, and seemingly impossible challenges. We will share how our expectations of "just use a GenAI model" crashed into reality when dealing with product visualization, semantic editing, and brand consistency. Experience our rollercoaster of CUDA dependency issues, model limitations, and how persistence (and some creative problem-solving) finally got us there.

Target Audience: Anyone interested in implementing AI projects
Prerequisites: Basic understanding of AI concepts and cloud architecture is helpful
Level: Advanced

Extended Abstract:
As consultants at Slalom, a human-centered business and technology consultancy, we work with leaders who expect more. This session tells the story of our journey implementing four different image generation and editing use cases - and how we learned the hard way that "just use GenAI" isn't quite that simple.

Our story unfolds in three acts:

Act 1: Expectations

Our confident pitch: "These are standard use cases, we will just use existing models!"

Why we thought combining different models would be straightforward

Our initial approach to hand product visualization and brand consistency

How we planned to implement semantic editing

Act 2: Disappointments

When we discovered why single models couldn't handle our specific requirements

CUDA dependency hell: When nothing works as documented

The growing frustration as each use case revealed new technical challenges

Resource management nightmares and exploding costs

When semantic editing produced hilariously wrong results

Act 3: Unexpected Wins

How persistence and creative problem-solving helped us overcome obstacles

The solutions we found when standard approaches failed

Building stakeholder trust through transparency about challenges

What actually worked in the end (and why it was different for each use case)