Designing Characters with Deep Learning: Spellbrush (W18) - YC Gaming Tech Talks 2020

TL;DR

Spellbrush uses generative adversarial networks (GANs) to generate high-quality anime-style character portraits in under two seconds, reducing art production time from hours. The company trained on ~10 million internet images, found severe demographic skew (1:6 female-to-male, <3% dark skin tones), and manually corrected the generator to improve representation.

Key insights

• On AAA game budgets, art accounts for 50–70% of total production cost; AI can compress a 2–15 hour human illustration task into sub-two seconds and generate hundreds of variations in the same time.
• GANs consist of a generator network (learns to draw) and a discriminator network (learns to distinguish real from fake art); both are trained adversarially by backpropagating the discriminator’s corre
• After training, the latent noise can be manipulated to control the generator’s output: same character in different expressions, colors, or illustration styles.