Training a generative neural network on children’s drawings.

Each of us can recall how the world felt differently in childhood—as if it existed within its own universe. We expressed all these ideas and dreams through play and drawing. A child’s drawing is an expression of sincerity and an endearing carelessness, as if one could capture a moment of happiness within those strokes and lines. To preserve this atmosphere, we conceived the idea of transferring these joyful imperfections into a generative neural network, allowing everyone to immerse themselves in a moment of true happiness.

To train the model, we gathered children’s drawings in various styles, along with different images in that style (buildings, people, animals, plants—everything that surrounds us in everyday life).

Kids' drawings possess a distinctive style—vivid masks, asymmetric composition, high-contrast colors, and cross-hatching lines of varying thickness. It is important to note that children do not create a direct, precise image in their work, but its subject matter can be deduced through characteristic features of the silhouette and form. This generative model has been able to adopt such techniques, creating images using these same recognizable methods.

1. Data collection and dataset preparation
2. Environment setup
3. Training configuration

4. Automatic Image Captioning (analyzes each image in the dataset; generates a text description (caption); example: for a kids' drawing of a rocket → «a colorful drawing of a rocket»)

5. Configuring Hugging Face Access (connecting the token and saving it to the cache)

6. Course Launch
7. Model Saving and Loading