Building an AI personalized story app is not about generating images

The hard part of creating an AI app for personalized stories is not generating an image. It is everything else.

A few months ago I started building Buklea, a personal project that began with a very simple idea: creating personalized stories for my children using artificial intelligence. On paper, it seemed easy: the user uploads a photo, writes a few details and the AI generates an illustrated story where the child becomes the protagonist.

The reality is quite different. Today there are models capable of generating impressive images with very little effort, so the challenge is no longer just getting a beautiful illustration. The real challenge is building a product that works consistently, safely, at scale and with a viable cost structure.

While developing Buklea, I ran into several interesting problems that I think are common to many products based on generative AI.

1. Consistency is much harder than it looks

If a story has 20 pages, it is not enough for each illustration to be good in isolation. The character still needs to look like the child in every scene, the clothing has to remain consistent, important objects need to appear correctly and the settings should evolve in a logical way.

All of this has to happen while the character changes pose, expression, lighting or camera angle. Current models have improved enormously, but they still struggle when they need to maintain coherence across long sequences.

In fact, much of the work is not about generating images. It is about designing systems that help models remember what they need to preserve from one page to the next. This topic deserves a full article of its own.

2. Privacy becomes critical when you work with children

One of the areas that made me think the most was the management of reference images. To get good results, the system usually needs to use photographs of the child, and that raises important questions from the very beginning:

• Where are those photographs stored?
• Who can access them?
• How do we prevent them from appearing in another context?
• How do we allow models to use them without exposing them to other users?

When we are talking about children, privacy stops being just another product feature and becomes one of the core requirements of the system. The technical architecture has to be designed with that in mind from day one, not added later as an improvised layer of protection.

3. Generating images is expensive

There is a common perception that AI has made content creation dramatically cheaper, and that is partly true. But when you try to build a commercial product, new questions appear.

A story may require dozens of images. Each image has a cost, each regeneration has a cost, tests have a cost and errors have a cost too. Suddenly, the business model becomes strongly constrained by the technology being used.

Very often the challenge is not generating better images, but making sure the experience remains profitable for both the user and the product.

4. Product design matters more than infrastructure

As engineers, we often fall into the temptation of designing complex architectures too early: advanced CDNs, multiple cache layers, distributed systems or queues for everything. Those are interesting problems, but they are not always the problems that need to be solved first.

In small projects or validation-stage products, the priority is usually somewhere else. The important question is not whether your system can generate one million stories. The important question is whether someone actually wants to use it.

While building Buklea, I have tried to constantly remind myself that the goal is not to build the perfect architecture, but to learn as much as possible with the least possible complexity.

What surprised me the most

Perhaps the most interesting part of the whole project is that the hardest challenges were not the ones I expected at the beginning. I thought the main problem would be generating the illustrations, but that has been the easy part.

The difficult part has been everything around those illustrations: consistency, privacy, costs, user experience and product design. And I suspect this happens in many generative AI projects.

In the next articles, I will go deeper into some of these topics and share the solutions I have been exploring while building Buklea.