Game-focused generative AI company Anchornode said on Monday it will jointly develop with South Korean AI company Upstage a benchmark to assess AI capabilities to understand and inspect game asset production specifications. The two companies plan to build evaluation questions and a scoring system based on requirements from game production sites and release the output as open source within the year.

The project will be 추진 as a collaboration task under the "DIP-Spot Open Innovation in Dongjak" programme hosted by the Seoul Business Agency (SBA), the company said. Upstage is taking part as a demand company and selected Anchornode as a collaboration partner.

Anchornode said game graphic assets such as characters, icons, backgrounds and UI must meet project-specific production specifications, including size, reference points, transparency processing and engine setting values. It said such specifications and inspection criteria vary by project, and some remain in developer experience and work practices. For AI to play a larger role in game production, it needs not only the ability to generate images but also the ability to understand production requirements and verify whether outputs meet those conditions.

Anchornode will define the specifications that AI must understand and verify based on knowledge from game production sites, and will design evaluation questions and ground-truth data. Upstage will provide the Sola model and API, as well as benchmark design know-how. In a preliminary check conducted by Upstage, "Sola Pro 4" read a Unity engine asset settings file and carried out a task to write the specifications that must be complied with as a list of rules in a prescribed format.

The two companies will use the benchmark development as a starting point to make concrete cooperation between language models and game-focused AI. They plan to lay the groundwork to assess AI understanding and inspection capabilities for game production specifications and, based on that, explore the possibility of expanding cooperation, including linking with game development tools.

Anchornode CEO Jae-ho Won (원재호) said for AI to move beyond making images well and be used in actual game production, it must be able to understand project-required specifications and check whether outputs meet those conditions. He said Anchornode will connect its on-the-ground game production knowledge with Upstage's language model technology to create specific evaluation criteria for what AI must understand and perform in game production.