How CLIP Teaches AI What 'Cheerful' Means | MyOwnChildbook
Type “a cheerful child in a yellow raincoat, classic picture-book style” into an AI image generator and within seconds an illustration appears that shows exactly that. Nowhere in the code is there a line that literally translates “cheerful” into raised mouth corners. Yet it works, time after time. Edwin, data engineer at MyOwnChildbook, explains which model makes this possible, and why it also explains why we deliberately avoid naming specific illustrators in our own style choices.
The model that ties language and image together
The key is CLIP (Contrastive Language-Image Pre-training), developed by OpenAI and described in the paper “Learning Transferable Visual Models From Natural Language Supervision” (Radford et al., 2021). CLIP was trained on roughly 400 million image-caption pairs from the open internet. During that training, the model does not learn a dictionary, but a geometric space: every image and every piece of text is converted into a long string of numbers, an “embedding”. Images and text that belong together conceptually end up close together in that space.
Edwin: “The easiest way to picture it is a map with thousands of dimensions instead of two. ‘Cheerful child’ and a photo of a laughing child sit close together on that map. ‘Rainy day’ and a photo of a wet street too. The model was never given a definition of ‘cheerful’ - it just saw millions of examples where that word belonged with that kind of image.”
Why style descriptions work without naming anyone
This property of CLIP-like models explains a concrete choice we make at MyOwnChildbook. In our style options, we deliberately use descriptions like “classic 19th-century fairy tale illustrations” or “modern children’s book illustration”, and never the name of a specific living illustrator or a well-known animation studio.
Part of that is a safety choice: platforms such as OpenAI train safety filters that react sharply to the combination of a specific artist name with a child context. But it is also a technical observation. Because the internet is full of text that describes styles in terms of era, technique, and tradition (“watercolour”, “nineteenth-century”, “hand-drawn line work”), the embedding model has learned to recognise those style concepts just as sharply as individual proper names. A descriptive phrase often activates a style cluster that is just as strong, and sometimes more broadly applicable, than a single name would.
Where embeddings hit their limits
This kind of model is not magic, and certainly not an exact science. Embeddings are statistical associations, not understanding in the human sense. For rare combinations - a style description that rarely appears alongside a particular subject in the training data - the result becomes less stable and less predictable.
More importantly: if you are looking for an exact, pixel-for-pixel imitation of one specific illustrator’s work, no CLIP-guided system will reliably deliver that. Embeddings capture the mood and the style family, not the precise brushwork of an individual. For anyone genuinely after a replica of one artist, a human illustrator with that specific portfolio is the only reliable route, not AI.
The alternative: fine-tuning instead of describing
There is a different technical approach from the one we take. Instead of relying on the knowledge a model like CLIP already has about styles, you can fine-tune a small additional model (techniques such as DreamBooth or LoRA) on a handful of example images of exactly one desired style. That can turn out more precise for that one narrow style, but it is also slower, more expensive per style option, and less flexible if you - like us - want to offer dozens of different styles side by side. Both approaches have their place; we choose descriptive prompts because breadth and speed matter more to us than exactly cloning one specific look. Our explanation of the full gpt-image-2 pipeline shows how that choice relates to the rest of the generation process, including how we maintain character consistency across multiple pages.
What this means for the end result
Edwin, also a father of two: “My kids do not notice the exact line thickness of an illustration. They react to the mood: is it warm, is it recognisable, does it feel like a real picture book? That is precisely the level at which CLIP-like embeddings operate - mood and style family, not microscopic precision. As it happens, that is also the level at which a child judges an illustration.”
The fact that the photo you upload is then processed neatly and temporarily is a separate layer of the pipeline. You can read more about that in our explanation of what happens to an uploaded photo of your child.
The mathematics behind an AI illustration is invisible to whoever opens the book. But the choices that follow from it - descriptive style language instead of a name, speed over exact imitation - do determine what a page ends up looking like. Technology and taste meet right there.