Tencent introduced the AI system Hunyuan-GameCraft , which creates interactive videos from individual images. Unlike regular video generators, Hunyuan-GameCraft allows the user to control the camera in real-time using the WASD keys or arrow keys. This enables free movement in generated scenes and viewing them from different angles.
The system operates on the open model HunyuanVideo and supports movement forward, backward, left, right, up, and down, as well as camera rotation left and right. Each key movement is translated into numerical values understood by the video generator. The AI considers the duration of key presses, affecting the speed of movement in the scene.
To improve video quality during long sessions, Tencent applied the Hybrid History-Conditioned Training method. The model creates each new video fragment gradually, relying on previous frames. This ensures the video remains smooth and responsive to user actions without losing quality.
Hunyuan-GameCraft was trained on over a million gameplay recordings from a hundred well-known games, including Assassin’s Creed, Red Dead Redemption, and Cyberpunk 2077. Additionally, the team created three thousand motion sequences based on 3D objects. During testing, the system demonstrated more precise control and better image quality compared to other AI for cameras.
GameCraft operates at speeds of up to 6.6 frames per second in real-time and responds to user actions in less than five seconds. The entire code and model weights are available on GitHub, and a web demo is being prepared for launch. The system complements the list of AI for creating interactive worlds, which already includes products from Google DeepMind and Skywork.
