Problem
Build an application that meaningfully leverages generative AI.
Constraints
Had to use an existing dataset of drawings to train the GAN and CNN models.
Hypothesis
The Google Quick Draw dataset, with 50 million images across 345 categories, would provide sufficient diversity to train both recognition and generation models effectively.
Process
- Trained GAN models to generate images similar to the drawings in the dataset based on a given word
- Trained CNN models to classify images in the dataset for drawing recognition
- Awarded points to player and AI based on how quickly the word was guessed
- Added gradual reveal of the word's letters as time decreased
Solution
A game with a custom-trained AI opponent that can both recognize player drawings and generate its own.
Metrics
- 3 rounds of guessing and 3 rounds of drawing per game session
- No noticeable lag when AI is drawing or guessing
- Trained on 50 million images across 345 categories from Google Quick Draw dataset
Evidence
Lessons Learned
Learned how to train GAN and CNN models on image datasets for image classification and generation.