Kazuma Kaneko recently announced a new game featuring procedural image generation. He claims that the model was trained on a set of images bootstrapped from his artwork. The typical method for creating style-specific image models involves fine-tuning general-purpose models that have been trained on very large corpuses.
This market resolves YES if any version of the system shown in marketing or used at game launch was trained on content that was not either drawn by Kaneko or bootstrapped from a model starting with only Kaneko's work as input.
Update 2025-07-02 (PST) (AI summary of creator comment): The creator has pointed to an article showing the AI generating images similar to other copyrighted material, and stated this seems like conclusive evidence for a YES resolution.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ22 | |
| 2 | Ṁ16 | |
| 3 | Ṁ9 | |
| 4 | Ṁ0 |
https://automaton-media.com/en/news/ai-in-kazuma-kanekos-new-game-seemingly-outputs-images-similar-to-disney-characters-and-other-copyrighted-material-raising-concerns/ This seems like pretty conclusive evidence that it was trained on outside work. Does the NO-holder @khang2009 want to offer a rebuttal?
@ProjectVictory That’s exactly the situation that I figured, and would resolve YES since it was at some point trained on non-Kaneko (or bootstrapped) content.