Conference Presentation, Keynote, Product Demonstration
Designing Characters with Deep Learning: Spellbrush (W18) - YC Gaming Tech Talks 2020
- SpellRush intends to scale content creation without becoming a massive studio by developing deep learning tools for the art industry, with the immediate goal of creating images indistinguishable from those drawn by professional illustrators.
- The company's AI tools are designed to generate characters in under two seconds, a task estimated to require professional illustrators between two and fifteen hours, enabling the production of hundreds of characters or "endless" variations simultaneously.
- Technology capabilities include idempotent generator and discriminator programs that eliminate latent noise, allowing control over character expressions, colors, and illustration styles, with style transfer tasks requiring minimal time compared to the hours needed by human artists.
- Training models currently requires seven to ten days with individual costs ranging from $3,000 to $4,000; to mitigate high operational expenses, SpellRush built a DIY supercomputer running at approximately 60 cents per hour versus roughly $10 per hour for cloud spot instances.
- The company trains its networks on publicly available internet data, primarily starting with anime aesthetics, addressing dataset skew where female characters outnumber males at a ratio of approximately 1:6 and people of color comprise less than 3% of available illustrations.
- Active research efforts include improving the generation of darker skin tones and male characters beyond real-world data frequencies, as well as pursuing automated animation techniques using Live2D, Spine workflows, and super resolution.
- Operational strategy involves building the world's first AI-illustrated game internally while maintaining a lean team of five people with plans to hire a sixth person, specifically targeting roles for a 2D animator, motion designer, real-time VFX artists, and an AI research intern in the coming winter.
- An existing older model named "waifu labs.com" is currently online, allowing user interaction for character selection and customization, serving as a foundation for the company's broader asset pipeline development.