China is hot on the heels of the US in the AI race.
Weeks after Google had released Genie 3, its world-creation model that created game environments on the fly, China’s Tencent has released a similar model of its own. Called Hunyuan-GameCraft, the model can generate interactive game videos. Unlike Google, Hunyuan-GameCraft is open-source, and the authors have given details on how they’ve created the model.
“Recent advances in diffusion-based and controllable video generation have enabled high-quality and temporally coherent video synthesis, laying the groundwork for immersive interactive gaming experiences,” Hunyuan-GameCraft’s authors said on its launch page. “However, current methods face limitations in dynamics, generality, long-term consistency, and efficiency, which limit the ability to create various gameplay videos. To address these gaps, we introduce Hunyuan-GameCraft, a novel framework for high-dynamic interactive video generation in game environments,” they added.
Like with Genie 3, the game world can be manipulated with the keyboard, with users using arrow keys to move around. Hunyuan-GameCraft creates videos based on these movements, and then recursively uses the created videos to create new ones. The model is trained on a large-scale dataset comprising over one million gameplay recordings across over 100 AAA games.
The model even preserves history, like Google had shown with Genie 3. Google had shown off how one could paint a wall in the game, pan away from it, and when one panned back, the paint was still there. Hunyan-GameCraft seems to be able to similarly preserve history. “For immersive game experiences, 3D consistency and scene coherence are essential for interactive video modeling of game scenes. Thanks to hybrid history condition, Hunyuan-GameCraft effectively preserves the original scene information after significant movement,” the authors say. The Hunyuan-GameCraft model recommends 80GB of VRAM for the full version, while the quantized 13B model can run on consumer-grade GPUs like the 4090.
While it remains to be seen if Hunyuan-GameCraft’s performance can match that of Genie 3, the release of the model shows that Google doesn’t quite as much of a lead as people thought when Genie 3 was released. Genie 3 was the first world-creation model that really caught the attention of the public, and even led many to wonder if its release was a proof that we ourselves were living in a simulation. But Chinese labs are able to release models with the same premise, showing that Genie 3 might not necessarily be the outlier that everyone thought when it was released. And China is once again flexing its capabilities in AI, showing that it now has competent models to do everything from generating text, code, images, videos, and even whole new worlds.