Spellbrush Combines Generative Diffusion Models and Turn-Based Tactics Games
Spellbrush is a San Francisco-based AI studio that builds generative image generation models and video games. Founded in 2016, the 30-person company is best known for creating niji・journey, an illustration engine built specifically for anime aesthetics. The company is now applying its diffusion research directly into game design, developing mobile turn-based tactics titles that integrate custom generative models into gameplay systems.
Founding and Engineering Background
Spellbrush was founded by CEO Cory Li, Creative Director Ruwen Liu, and CTO Haitao Mao. Li previously co-founded Benchling after studying bioengineering at MIT, while Liu also studied at MIT before leading visual direction at the company. Spellbrush went through the Y Combinator W18 batch.
To power its machine learning research while keeping operational costs down, the team built an in-house GPU supercomputer running SLURM. In a presentation for Y Combinator, Li noted that training GAN models on public cloud infrastructure cost several thousand dollars per run, leading the startup to assemble its own cluster of Titan RTX GPUs in its office to reduce training costs to approximately 60 cents per hour. The company's backend stack uses Elixir for real-time services, Python with PyTorch and JAX for model training, and C# via Unity for game systems.
Generative Models and Game Production
The studio's technology path expanded from early generative adversarial networks to large-scale diffusion models. Spellbrush released Waifu Labs in 2019 to showcase character generation, followed by Aeromancer in 2022, an early game experiment with real-time model integration. Spellbrush also collaborated on the Niji Model v5 for Midjourney, as reported by Nerd's Chalk.
In a talk on Founders You Should Know, Li stated that niji・journey generates roughly one million images per day. The studio maintains offices in downtown San Francisco and Akihabara, Tokyo, as it develops its next title, Project Crown. According to revenue estimates published by GetLatka, Spellbrush generated an estimated $5 million in revenue in 2024.
Current Open Roles
Spellbrush is currently filling five technical roles across machine learning research, software engineering, and game development:
- The
AI Researcher (Anime)($75,000–$200,000) focuses on designing and training large-scale image diffusion models using PyTorch or JAX to support niji・journey and game assets. - The
AI Infrastructure Engineer($75,000–$200,000) builds machine learning infrastructure and inference systems across distributed GPU and TPU clusters. - The
LLM Engineercrafts prompts and structures language model output to power character interactions and world-building for a mobile turn-based tactics game. - The
Frontend Engineer (Anime Games)($180,000) builds single-page applications and web interfaces in React and TypeScript to link generative models directly with creative tools. - The
Anime Games - Unity Engineer($60,000–$180,000) writes C# code in Unity to connect generative model pipelines to tactical game logic.
Open Roles Summary
| Role | Location | Work Mode | Salary |
|---|---|---|---|
LLM Engineer |
San Francisco, CA | Onsite | Not listed |
AI Infrastructure Engineer |
Tokyo, JP / San Francisco, CA / Remote | Remote | $75,000 – $200,000 |
Frontend Engineer (Anime Games) |
San Francisco, CA / Tokyo, JP / Remote | Remote | $180,000 |
AI Researcher (Anime) |
Tokyo, JP / San Francisco, CA / Remote | Remote | $75,000 – $200,000 |
Anime Games - Unity Engineer |
San Francisco, CA / Tokyo, JP / Remote | Remote | $60,000 – $180,000 |



