Will it actually look like my child?
The only honest way to answer that is to show it - on the cases that usually go wrong, not the ones that always go right.
Eight cases, each one a feature that image generators are known to erase. For every case we wrote down what had to survive before running it, then ran the same reference photograph through our ordinary pipeline twice, independently, and judged both outputs against that list. Every image here is the first output of its pass - nothing was generated repeatedly and then picked from, and the second pass is published whether it agrees with the first or not. Three cases were re-run on 24 August after we changed the prompt to fix what the first run exposed; their pre-fix pictures are still on the page underneath the new ones, because "we fixed it" with only the good result showing is a claim rather than evidence. Of 27 things that had to survive, 22 held on both passes, 1 was lost in at least one, and 3 came out differently between the two passes - which is the number we would most like to be zero.
Deep brown skin, tightly coiled hair
age 5This is the case we got wrong ourselves. Until late July our cover prompt carried no likeness clause, and a boy with deep brown skin and close-cropped hair came back lighter-skinned with wavy hair - while his interior pages were correct. It is now the first case we check.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| Skin tone still deep brown | Held | Held |
| Hair still very short and tightly coiled | Held | Held |
| The gap in his front teeth is still there | Held | Held |
Fine, fair, barely-there hair
age 1The hardest case there is, because a near-bald baby is what image generators most want to "correct". A parent reviewing a rival product described exactly that: their one-year-old came back with long thick black hair, when she barely has any and it is blonde.
What to notice: Re-run after the prompt was changed to say that an absence is part of the likeness. It helped and it did not solve it: both passes still give her more hair than the reference has, though less than before, and pass 2 also drifted the colour from pale blonde towards strawberry - a criterion that had held on the earlier run. The catastrophic version, a full head of thick dark hair on a blonde baby, has never happened here. The mild version still does.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| Still almost bald | Partly | Partly |
| What hair there is stays fine and fair | Held | Partly |
| Eyes still blue | Held | Held |
| Still reads as a baby, not a toddler | Held | Held |
Straight black hair with a fringe
age 7A fringe is a specific shape, and it is the sort of detail that gets averaged away. This case also tests the collapse a reviewer described when two sisters who look nothing alike came back as "the same girl with very subtle differences".
What to notice: The one difference between the passes is the hair colour: pass 2 is black, pass 1 reads as a warm dark brown. Side by side it is obvious; on its own you would not notice. This is the kind of drift two passes exist to surface.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| The fringe still sits above the eyebrows | Held | Held |
| Hair still straight and black, not waved or browned | Partly | Held |
| Recognisably the same child in both passes | Held | Held |
Curly ginger hair and freckles
age 4Freckles and red hair are the first things smoothed into something more average. If they are still there twelve chapters later, the identity lock is doing its job.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| Freckles still on the cheeks | Held | Held |
| Hair still ginger and curly | Held | Held |
| Eyes still green | Held | Held |
Round glasses
age 6Glasses are an object rather than a feature, and objects are what a generator adds and removes most freely - the same failure mode as a hair clip or a hearing aid. Either they are there, the same shape, or they are not.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| The glasses are still there | Held | Held |
| Frames still round, not rectangular | Held | Held |
| Hair still light brown and straight | Held | Held |
Long braids with beads
age 8A second dark-skinned case with a completely different hair structure to the first, because "we got dark skin right once" is not the same claim as "we carry hair structure". Braids also have a length and a count, which is harder to fake than a colour.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| Braids still long, not shortened | Held | Held |
| Beads still on the ends | Held | Held |
| Skin tone unchanged | Held | Held |
| Still reads as eight, not five | Held | Held |
A visible hearing aid
age 5The case that matters most to the family it belongs to, and the one a generator is most likely to tidy away. A child who wears a hearing aid and does not see it in her own book has been told something about herself. This is the case that failed.
What to notice: This is the case that failed, and it is the case the prompt fix was written for. BEFORE: the device vanished completely from one portrait out of two, and from both story scenes. AFTER, with a worn item named as part of the person and the side of the head fixed: it is there in both portraits, clearly, on the right ear - and it is there in the first story scene too, with the hair tucked back so it can be seen. It is still missing from the second scene, where the pose turns that ear away from the camera. That is better than deleting it and it is not what we asked for, so it stays recorded as lost. The old pictures are below the new ones.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| The hearing aid is still there | Held | Held |
| Still visible once she is in a story scene | Held | Lost |
| Hair still short and dark | Held | Held |
A baby with no hair
6 monthsThe hardest case in the set and the most commercially relevant: baby books are the single biggest thing people search for and find us with. A six-month-old has no hair, no teeth and few fixed features to hold on to, so this is where a generator is most likely to draw a toddler instead.
What to notice: Re-run after the prompt fix, and the two misses came out differently. The skin tone is now right on both passes - before, he came back a shade lighter and warmer than his photograph. The hair did not fully resolve: both passes still give him a short covering where the reference has none, closer than before but not none. He is still recognisably this baby and still a baby. The pre-fix pictures are below, and the odd arm attachment in the earlier scene is in them - it is the reason nothing goes to a printer until you have seen it and said yes.
| What had to survive | Pass 1 | Pass 2 |
|---|---|---|
| Still no hair | Partly | Partly |
| Skin tone unchanged | Held | Held |
| Still reads as an infant, not a toddler | Held | Held |
| Eyes still dark brown | Held | Held |
How the likeness is held across twelve chapters
The photo is used once, to draw a single reference character. Every later illustration - the cover and all twelve chapter pictures - is generated from that character, not from the photo again. That is the reason the child does not drift page to page, and it is also why the photo itself is deleted as soon as the character exists.
It is also why each book stars one child. A single face can be held steady; two faces start to blend into each other, which is the failure behind a reviewer’s complaint that two sisters who look nothing alike came back as “the same girl with very subtle differences”. Siblings get a book each.
How this was done, and what it cannot tell you
The criteria came first. For each case we wrote down the two to four things that had to survive before generating anything, phrased so that “held” or “lost” is a matter of looking rather than taste - a fringe is either above the eyebrows or it is not. A criterion invented after seeing the picture is not a test, it is a caption.
The same reference, twice. Each case was run through the pipeline two times independently from one reference photograph, and both results are on this page. That is there to answer the obvious objection - that we generated until it worked - because an inconsistency between the passes is visible to you rather than only to us. Where the two passes disagree, the table says so.
It is our own pipeline, unchanged. The same prompts, models and settings that draw a real customer’s book, called the same way. Nothing was hand-retouched, nothing was upscaled, and no image was regenerated and swapped in.
What this cannot tell you. Three things, and they are real. The reference portraits are generated rather than photographs of real children, so this measures whether the pipeline carries a face, not how it copes with a badly lit phone snap of a wriggling toddler. We judged the results ourselves - there is no blind rater, and you should check the pictures against the criteria yourself rather than take our word for the verdicts. And eight cases with two passes each is a small sample: it is enough to show what the pipeline does with hard features, and not enough to put a percentage on your book.
The one test that settles it for your child costs nothing: upload their photograph and read the cover and first chapter free. If the child on the page is not recognisably them, close the tab.
Common questions
Will the book actually look like my child?
That is what this page exists to let you judge before you spend anything. The illustration keeps skin tone, hair colour, hair texture, eye colour and face shape from the photo you upload, and holds them across all twelve chapters. You see the cover and first chapter free, so you can decide on the likeness before paying.
What makes a good photo to upload?
One clear photo of the face, looking towards the camera, in daylight, with nobody else in frame. A phone photo is ideal. Avoid sunglasses, hats that hide the hairline, heavy shadow, and screenshots of a screen - a screenshot loses the detail the illustration is built from.
Does it work for every skin tone and hair type?
That is what the cases on this page test, and they were chosen because they are the ones image generators most often get wrong: deep brown skin with tightly coiled hair, long braids with beads, a baby with no hair at all, a fringe, freckles, glasses. Skin tone and hair structure held on both passes in every case except the two babies, where the skin came back slightly lighter on one of them and both gained a little hair. We publish the hard cases rather than the easy ones, and the results rather than the good results.
What did it get wrong?
Three things, all on this page. A baby with almost no hair came back with a little more hair than the reference has - short, fine and the right colour, but more of it - and that happened on both passes and on both babies we tested. The darker baby also came back a shade lighter and warmer than his photograph. And a hearing aid was dropped altogether on one pass out of two, and is missing from the story scene in both. The pattern is that the pipeline is good at carrying what is there and less reliable about preserving an absence or a small worn object. Both are now written into the prompt itself - a worn item is named as part of the person, with the side of the head stated, and an absence is named as part of the likeness. The pictures on this page are from the run before that change, and they stay up until we re-run the study and publish the new ones. We would rather show a fixed failure than a page that never had one.
Are these real children?
No, and deliberately. The reference portraits are generated, not photographs of real children, so that no real child’s face is published on a page about how we handle children’s photos. Everything after the reference is our normal production pipeline, unchanged.
What if the likeness is not right when I see my book?
Tell us before you approve it and we will regenerate it - as many times as it takes, and nothing is printed until you say yes. Because each book is made for one named child we fix rather than refund, and we do not put a limit on fixing.
Why does each book star only one child?
Each book has one hero whose name is on the cover - that is a choice about whose story it is, not a limit. We used to say the illustration blends two children together; we tested that across 88 illustrations at two, three and four people and found no blending and no swapped faces, so the family stories now hold up to four. Siblings and twins still get a book each.
Last updated:
Judge it on your own child’s photo - the first chapter is free.
Browse the collection