Hunyuan GameCraft

Hunyuan GameCraft is an open-source interactive game video generation framework developed by the Tencent Hunyuan research team and released in August 2025. The model generates dynamic, playable video sequences from a single scene reference image and action inputs — keyboard signals (W, A, S, D, arrow keys) and mouse movements — that control character navigation and camera direction within the generated environment. Its core architectural innovation is a hybrid history-conditioned training strategy that preserves scene information across extended sequences as the user navigates, preventing the visual drift and inconsistency that affected earlier interactive video generation models. The framework also generates dynamic environmental content including moving clouds, rain, snow, and water flow, rather than static scene backdrops. Hunyuan GameCraft is intended for game developers who want to prototype environment concepts and character interactions without committing to full 3D modeling and rendering workflows, AI researchers studying interactive generative video models, and technical creators who want to experiment with keyboard-controlled scene exploration. The model was trained on a dataset of over one million gameplay recordings spanning more than 100 AAA titles, which gives it broad coverage of game visual styles from realistic environments to stylized and pixelated aesthetics. See top alternatives. Outputs from Hunyuan GameCraft are compatible with standard game engine workflows — generated video sequences can serve as visual references or concept validation for environments built in Unity or Unreal Engine. The full-precision model requires approximately 80GB of GPU VRAM for optimal inference, which places it outside the range of most consumer setups. A quantized 13-billion parameter version, available through Hugging Face, can run on consumer-grade GPUs including the RTX 4090, making the model accessible to individual developers and researchers with high-end consumer hardware. PCM distillation reduces inference steps and computational overhead while maintaining temporal consistency across long video sequences. The model does not generate real-time playable output — it generates video sequences that simulate gameplay, which can be used for prototyping and visualization but not as a drop-in replacement for real-time game engine rendering in production titles View alternatives.