How AI illustrations from your photos actually work (behind the scenes)

The most common question we get from new users is some version of: how is my kid (or my mom, or my dog) actually showing up on the page? It looks like the same character from front to back. Is the model just trained on my photos? The answer is more interesting than that, and it explains why the book takes the time it takes to generate.

Do the illustrations actually look like your subject?

Short answer: yes, when the system is given good references and is allowed to do its serial-grounding trick across the book. Long answer is the rest of this post — including the boring privacy bit, which comes first because people reasonably ask.

We do not train on your photos

First, the boring but important part. We never train a model on your photos. The reference photos you upload are sent to Google Gemini as inputs at generation time, used to make the illustrations for this one book, and then they sit in your account in your private Supabase storage. They are not pooled into anyone else's book and they are not used to fine-tune any model we run.

What we do is more like character consistency than character training. The references are sitting next to the prompt, every time we ask the AI to draw a page, so the AI keeps redrawing the same character.

Page 1 is the canonical character sheet

When you generate a book in quality mode (the default for character-driven stories), page 1 is special. It is generated first, on its own, before any of the other pages exist. The prompt for that page is heavier than the others — it tells the AI to establish the character in a clear, full-body, well-lit pose, treating page 1 as the canonical look of your subject.

Then, for page 2, we send the AI three things: your reference photos, page 1 as an image, and the page-2 text. Page 3 sees: references, page 1, page 2, and the page-3 text. Page 4 sees: references, page 1, page 3, and the page-4 text. And so on. Page 1 is in every prompt. The immediately-previous page is in every prompt. That is the trick that keeps your subject from turning into a different person by page four.

This is also why quality mode is serial — page 3 cannot start until page 2 is done, because page 2 is part of page 3's prompt. It is slower than firing them off in parallel. It is also the difference between a real keepsake and a book where the curly-haired kid becomes a different kid partway through.

Fast vs. quality mode

There is also a fast mode. It fans every page out in parallel using only the reference photos — no page 1 grounding, no previous-page grounding. It is much faster (a couple of minutes instead of more like ten) and the trade-off is real: character consistency drops. Hair and skin tone tend to drift. The face shape varies. It looks like the same person if you squint.

We use fast mode for stories where a recognizable likeness matters less, and we use it as the fallback when quality mode keeps timing out. For books that star a specific person or pet we default to quality.

What the AI gets right

Within a single book, when the references are good, the AI is genuinely good at: keeping hair, skin tone, and coloring consistent, keeping body shape consistent, keeping a face readable across pages, putting the subject into varied poses (sitting, walking, sleeping, looking up at the sky), and matching the time of day and weather described in the text.

It is also genuinely good at composition. Pages where the character is the subject get framed like an illustration would be framed — the figure near the rule-of-thirds intersection, not stuck in the middle of every page.

What the AI gets wrong

Honest list, because pretending otherwise wastes everyone's time:

• Eye color. The model is unreliable about heterochromia and unusual eye colors. If your subject has one blue eye and one brown, expect to use the AI Assistant on the pages where it matters.
• Fine details. Specific marks — a particular freckle pattern, a scar, a pet's heart-shaped white patch — drift. The model gets the general look but specific shapes are not stable.
• Subtle features. A clear, distinctive look reads correctly. Very subtle distinctions settle into whatever the model is most confident drawing.
• Hands. If you ask for a page where someone is holding something or someone, the hands can come out wrong. We try to crop those tightly. Sometimes one slips.
• Counting. If you write a prompt with 'three kids,' you may get two or four. Specific counts above two are unreliable.

These are the failure modes the AI Assistant was built for. You do not have to live with them.

Fixing pages with the AI Assistant

Inside the Studio, every page has an AI Assistant box. You can tell it what to change in plain English. The system reads your request, decides whether you are asking to change the text, the illustration, or both, and dispatches only what is needed.

Requests that work well are specific and small. 'Make her eyes blue.' 'Change this scene to morning.' 'Move the figure to the right side of the frame.' 'Rewrite this page in a warmer tone.' Requests that work less well are vague and large. 'Make this better.' 'Try again.' 'I do not like this.'

When in doubt, ask for one specific thing at a time. The Assistant is a precise tool, not a do-over button.

Why we tell you all this

Because AI products that pretend they are magic age badly. The mechanics are not actually that mystical. We are sending Gemini your reference photos and a careful prompt, page after page, keeping the previous output in context so the character holds. We made the system as good as we know how to make it. It is still going to have a bad day on the occasional page. The Assistant exists for those pages.

Understanding what is happening under the hood makes you a better collaborator with the system. You will write better prompts. You will pick better reference photos. You will spot the kinds of errors the model is prone to, and you will fix them quickly. The book you end up with will be better than the book you would have gotten by treating it as a black box.