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

TL;DR

Spell Rush, a company utilizing AI tools for art, has developed a deep learning model that can create high-quality character illustrations in seconds, significantly reducing production time and cost.

Transcript

my name is corey uh i'm the ceo at spell rush and i'm here to talk to you today about uh designing characters with deep learning so um we're small rush uh we're a yc company as well uh we're building deep learning tools for art and artists what exactly does this mean art is hard so my co-founder is a professional artist but we're a fairly small tea... Read More

Key Insights

• 💭 Art is becoming increasingly more expensive and difficult to scale, but AI tools can help in the art production pipeline and save time and costs.
• 💡 AI tools can generate illustrations on par with what professional illustrators can create, but at a much faster rate, allowing for the creation of hundreds of characters in the same time frame.
• 🔍 The use of generative adversarial networks (GANs) in AI technology enables the generation of high-quality images that are indistinguishable from real data sets.
• 🖌️ The AI tools can generate multiple variations of characters, including different expressions, colors, and illustration styles, all of which would typically take a trained artist hours to achieve.
• 📚 The training of AI models involves crawling publicly available images from the internet, with a focus on the anime aesthetic which has the most data available.
• 🌍 Efforts have been made to improve the generation of darker skin tones and male characters to better represent real-world diversity within the AI-generated illustrations.
• 🖥️ The training of AI models can be complex and expensive, but building a DIY supercomputer in-house helps reduce costs and allows for more efficient training.
• 💼 The company is working on creating the world's first AI-illustrated game and is hiring artists, animators, motion designers, and AI research interns to join their small team.

Questions & Answers

Q: How does Spell Rush's AI tool compare to illustrations created by human artists?

Spell Rush's AI tool has demonstrated the ability to generate character illustrations that are indistinguishable from those created by human illustrators in terms of quality. The tool can produce high-quality illustrations in seconds, whereas it would take a professional artist hours to create the same content.

Q: What is the advantage of using deep learning tools for art production?

Deep learning tools, such as Spell Rush's AI model, offer several advantages for art production. Firstly, they significantly reduce production time, allowing artists to generate more content in less time. Secondly, these tools can handle large-scale projects without the need for a massive studio or budget. Lastly, they provide the ability to create variations and customize characters easily, saving artists from repetitive tasks.

Q: How does Spell Rush train their AI model?

Spell Rush trains their AI model using generative adversarial networks (GANs). They have a generator network that learns to draw art and a discriminator network that learns to differentiate between real and generated art. By repeatedly feeding images to both networks and updating their weights based on feedback, the AI model improves its ability to produce high-quality illustrations.

Q: How does Spell Rush address representation in their AI-generated illustrations?

Spell Rush acknowledges the imbalance in representation within their dataset, with a higher number of female characters and limited representation of people of color. To address this, they have put effort into improving the AI model's generation of darker skin tones and male characters. While their dataset does not represent population statistics accurately, they aim to create more diverse and inclusive art through their AI tools.

Q: What is the cost and infrastructure required for training Spell Rush's AI models?

Training Spell Rush's AI models is a complex and expensive process. They have built their own mini supercomputer with a cluster of computers, custom language, and top-tier GPUs. Training each individual model takes about seven to ten days and costs around $3,000 to $4,000. However, their in-house infrastructure reduces the running cost to approximately 60 cents per hour.

Q: What are the future plans for Spell Rush?

Spell Rush is currently working on developing the world's first AI-illustrated game. They are a small team looking to expand their workforce and are hiring artists, animators, motion designers, and AI research interns. They are utilizing their AI tools and expertise to create innovative and visually appealing gaming experiences.

Summary & Key Takeaways

• Spell Rush is a small startup that aims to solve the challenge of scaling up art production in the gaming industry by using AI tools.

• They have created a deep learning model that can generate character illustrations that are on par with professional artists in a fraction of the time.

• The model is trained using generative adversarial networks (GANs) and has the ability to create variations in expressions, colors, and styles for the same character.