DeepPixelMonster final

Final slides

Transcript

1. Deep Convolutional Generative Adversarial Networks (DCGANs) for Creating Pixel Art Lawrence Du
2. Creating art for mobile apps takes skill and money
3. Problem: The Costs of Making Art for Indie/Mobile Games Total cost of adding one new character: $109,000 Artwork cost: $30,000
4. Problem: The Costs of Making Art for Indie/Mobile Games Total cost of adding one new character: $109,000 Artwork cost: $30,000
5. Deep Learning for AI assisted art generation Fake pictures of bedrooms created by deep neural network
6. One franchise built around pixel art... Pokémon Red (released 1998)
7. Using Pokémon as a training set • 64 x 64 x 3 pixels = 12,228 features per image • 700+ Unique Pokémon • 76,000+ animation frames
8. Feature engineering with images • database used for classifying appropriate images • 231 Pokémon selected (19,231 frames) • Expansion of training set size: ◦ Brightness ◦ Hue ◦ Contrast ◦ Left-right transformations.
9. DEMO www.DeepPixelMonster.com (TRY IT YOURSELF)
10. Implementation • Written using • 30+ neural architecture combinations tested. • Training takes 20-30 minutes on GPU • Front-end Python • Amazon AWS
11. Implementation DISCRIMINATOR GENERATOR
12. Network Balancing DISCRIMINATOR LOSS GENERATOR LOSS
15. • LINE messenger - $238 million on stickers in 2015 • Easier to send a sticker than to write Chinese characters (kanji) Additional use cases for AI pixel art
16. Lawrence Du larrydu88@gmail.com PhD Biological Sciences (UC San Diego)
17. Solution: Use Deep Learning for AI assisted art generation AI generated music album covers