How MamaPapa's AI storybooks actually got built

MamaPapa's AI storybook feature generated its first story on March 29, 2026, and it was the app's first real bet on generative AI: turn the photos a family already has saved into a personalized, illustrated children's story, with their own kid as the main character. It's the feature DreamShots later grew out of, and the pipeline behind it shaped a lot of how MamaPapa builds AI features today.

Images first, text second

The obvious way to generate an illustrated story is to write the text, then generate an image for each page to match. MamaPapa's pipeline runs the opposite order: images first, text written to match afterward. That ordering exists specifically for consistency. A child's appearance, an art style, a scene's mood, all of it is easier to keep coherent across a multi-page story if the images are settled first and the narration is written to fit what's actually on the page, rather than generating text first and hoping the illustrations match it page by page.

Moving off a self-hosted model

Storybook generation didn't stay on one image-generation setup for long. An early version ran on a self-hosted SDXL model via Modal, the same GPU-hosting approach FloraFever's plant-identification pipeline uses for its own generative layer. That got replaced with hosted image providers offering tiered quality levels, letting a family choose between a faster, cheaper generation and a slower, higher-quality one depending on what a given story needs. Owning the inference stack makes sense for a high-volume, low-complexity task; for something as visually demanding as a full storybook's worth of consistent illustrations, a hosted provider with real quality tiers turned out to be the better trade.

Built multilingual from early on

Story language support landed within days of the feature's initial launch, not as a later addition. A family whose default app language is Danish gets a story written in Danish, not an English story with translated UI around it. That's a meaningfully different engineering problem than translating buttons and labels: the AI generation step itself has to produce narrative text in the right language, not just have its surrounding interface localized.

From one feature to a hub

Storybooks didn't stay a single, standalone feature. It went through a significant rework a couple of months after launch, covered separately in how MamaPapa unified its billing, and more recently it became the anchor of a broader "Creations" hub that DreamShots now shares alongside it, rather than two disconnected AI features living in different corners of the app. That consolidation matters for the same reason the swappable generation backend does: MamaPapa's generative features are being built as a growing family of related capabilities sharing infrastructure and a UI home, not as one-off bolt-ons shipped independently of each other.

Storybooks was the first proof that a personalized, AI-generated feature could work for a product whose entire premise is protecting family photos from exactly the kind of broad AI analysis MamaPapa deliberately avoids elsewhere. Getting that balance right on the first attempt is a lot of why the features that followed it could move faster.

More on what MamaPapa does today is on its product page.