FreshPrints

PROJECT OVERVIEW

The Problem

While auditing I started noticing something important in Mixpanel. Nothing dramatic at first—just a small pattern that kept repeating. When I spoke with a few of our users, the story became clearer. They come in with all kinds of creative plans, but our image library isn't always ready for what they have in mind. So when they scroll through the options and don't find anything that fits, the excitement fades. That initial spark disappears—and many users simply don't return.

Collaboration with

Instead of clicking “customise this” and entering the editor, more users were skipping that step and going straight to “request a mockup.” It turned into a quick escape whenever they couldn't find the visuals they needed. What seems like a small shift on the surface created a much bigger ripple behind the scenes: a growing stack of manual mockup requests for our template design team.

This wasn't a usability issue. It was a confidence problem. Users didn't trust that they could create something good from scratch.

Strategic Question

How might we reduce the effort required to go from intent → usable design?

We began with deep discovery sessions, including stakeholder interviews, competitor audits, and real user shadowing. These insights revealed major friction points in engagement.

The Solution

By creating an AI image generation feature, we can supply our users with any images they need—and inspire them to keep creating. And our users loved the result after it launched in April 2025.

The Process

As AI image generation rapidly developed in quality, we hypothesised that it could solve one of our most important problems—supplying every user with any image they could imagine.

We knew that images are critical to the success of a user's design, as users who experienced a 0 result search while looking for content were 4.4% less likely to return.

The Problem with Prompting

I jumped into fast research on other existing AI image tools, and saw users often struggled with describing what they wanted to create.

To combat this issue, I worked with engineers and our visual designers to train “Style selectors” that would allow our users to pick a style via thumbnails with visual examples, removing the need to learn a complete lexicon to prompt.

We explored different thumbnail and messaging for styles. Initially we did not provide a “no style” option, until testing showed that some users really wanted an option to let the product choose the best style for them.

The Launch

Text to Image launched on April 4th 2025, with a quarter of a million images published in the first month.

But after launch…retention trended downwards by 4%

Testing led to the hypothesis this was because we had initially built the feature as a mobile app to save development time. We refined our discovery experience—plus added new and improved styles, and removed a limit to free usage.

We included an “AI Image Generator” entry point when users open the Design tab to look for photos.

We also welcomed DALL-E to our ecosystem. We can't wait to see where it goes from here!