Researchers from MIT, among others, have created a website with an AI that generates a fairytale with images and text, based on input from the user.
Creating fairytales and creative texts using AI is a challenge. According to the researchers behind the fairytale AI, the biggest challenge is that a fairytale needs to have an unexpected direction and a coherent feeling and tone. Several existing machine learning models have the ability to create fairytales, but all have their shortcomings, according to the researchers behind the new model.
In a study, the researchers combined strengths in different AI models for text generation. They also designed a model to retrieve images in existing databases based on the content of a text.
To train the model, texts from Reddit and existing children's books were used. The Reddit texts were cleared of negative messages and offensive words. A total of 300,000 texts were included in the training. The researchers fine-tuned the model several times based on the results' degree of creativity, positivity in the messages, diversity in sentences, readability, and how "fairytale-like" the texts were.
The researchers also created a model to calculate the difference between the
images that were selected for the saga to get the highest possible feeling of coherence.
The platform was available online for a while under the name "Fairytailor," and the user could choose between creating a fairytale through "autocomplete" or "high quality auto-complete." In the latter option, several different fairytales were created coupled with a ranking based on different parameters. All that was needed from the user was a small snippet of the text.
To evaluate the platform, people were asked to submit fairytales generated by the AI and rate the works. The rating was based on three different parameters: how coherent it was, the degree of creativity, and the degree of clarity. The highest score was 5 points.
This is the average end result: Clarity: 3.4, creativity: 3.7, coherence: 3.5.
Read more here: https://github.com/EdenBD/MultiModalStory-demo