How We Train a Custom AI Model on Your Child's Photos
When you upload photos to DreamDraft, something remarkable happens behind the scenes: we train a unique AI model that learns what your child looks like. Not a generic “brown-haired child” — your specific child, with all their unique features. Here's how it works, explained without the jargon.
The Short Version
You upload 5-10 photos. Our AI studies them, learning the specific combination of features that makes your child look like them — the shape of their eyes, the texture of their hair, the way they smile. Then it uses that knowledge to generate new, original illustrations of your child in any scene, any pose, any setting. The result: a storybook character that parents and kids instantly recognize.
What is LoRA? (The Non-Technical Version)
LoRA stands for Low-Rank Adaptation. In plain language, it's a technique that lets us teach an existing AI model about a new concept — in this case, what your child looks like — without rebuilding the entire model from scratch.
Think of it like this: imagine a talented artist who has spent years learning to draw children's book illustrations. They already know how to draw faces, bodies, expressions, backgrounds, and lighting. LoRA is like showing that artist a handful of photos and saying, “Now draw this specific child.” The artist doesn't need to relearn how to draw — they just need to learn what makes this child unique.
That's exactly what LoRA does with our AI. It takes a powerful illustration model that already knows how to create beautiful art, and fine-tunes a small portion of it to understand your child's specific appearance. The training takes about 2-3 minutes and produces a model that can generate your child in any scene.
Why 5 Photos Are Enough
You might wonder: how can an AI learn what someone looks like from just 5 photos? The answer is that our base model already understands faces, expressions, and human features at a deep level. It has been trained on millions of images. What it doesn't know is the specific combination of features that make your child unique.
Five photos from different angles and with different expressions give the model enough data points to capture that uniqueness. We recommend:
- Front-facing photos — clear view of facial features
- Side angles — profile shape, ear position, jawline
- Different expressions — smiling, serious, laughing
- Various lighting — indoor, outdoor, natural light
- Everyday photos — school pictures, candids, selfies all work great
More photos (up to 10) can improve results, but 5 good ones are plenty. You don't need professional photography — your camera roll is perfect.
From Photos to Illustrations
Once the model is trained, here's what happens when we generate an illustration:
- Scene description — Our story AI provides a description of what should happen on this page (e.g., “Emma discovers a glowing treasure map in the garden”)
- Image generation — The trained model creates an illustration of your child in that exact scene, maintaining their likeness while adapting pose, expression, and setting
- Consistency check — Because every illustration uses the same trained model, your child looks like the same person across all pages — different poses, same kid
This is fundamentally different from services that paste a photo onto a pre-drawn template. Our illustrations are generated from scratch, with your child's features woven into the art itself.
What About Character Consistency?
This is the hard problem that avatar-based services sidestep entirely. When you use pre-drawn templates, consistency is easy — it's the same drawing on every page. But the character doesn't look like your child.
With AI-generated illustrations, achieving consistency while maintaining likeness is genuinely difficult. Our LoRA approach solves this because the model has internalized what your child looks like. Whether the character is running, sleeping, flying on a dragon, or sitting in a classroom — the model generates the same child. Their curls stay curly. Their freckles stay in place. Their smile is their smile.
What Happens to Your Photos?
We take privacy seriously, especially when it comes to children's photos:
- Photos are encrypted in transit and at rest
- They are used solely to train your personalized model
- Photos are automatically deleted after 30 days
- We never share, sell, or use your photos for any other purpose
- DreamDraft is fully COPPA compliant
Your trained model is also private — it can only be accessed through your account and is used exclusively for your book.
Try It Yourself
The best way to understand the technology is to see it in action. Creating and previewing a book is completely free — you can upload photos, train a model, and see the results without entering a credit card.