What is Hunyuan-GameCraft
Hunyuan-GameCraft is a highly dynamic interactive gamevideo generationframework jointly launched by the Tencent Hunyuan team and Huazhong University of Science and Technology. Based on unifying keyboard and mouse inputs into a shared camera representation space, it achieves fine-grained motion control and supports complex interactive inputs. The framework introduces a hybrid history-conditioned training strategy, which can autoregressively extend video sequences, preserve game scene information, and ensure long-term temporal consistency. Based on model distillation technology, Hunyuan-GameCraft significantly improves inference speed and is suitable for real-time deployment in complex interactive environments. The model is trained on a large-scale AAA game dataset, demonstrating excellent visual fidelity, realism, and action controllability, significantly outperforming existing models.
Main Features of Hunyuan-GameCraft
- Highly dynamic interactive video generation: Generates highly dynamic interactive game video content from a single image and corresponding prompt, supporting real-time user control based on keyboard and mouse inputs.
- Fine-grained motion control: Unifies standard keyboard and mouse inputs into a shared camera representation space, supporting complex and fine-grained interactive inputs, such as speed and angle control.
- Long-term video generation: Supports generating long-term coherent video sequences, preserves historical scene information, and avoids scene collapse.
- Real-time interaction: Significantly improves inference speed, reduces latency, supports real-time interaction, and enhances user experience.
- High visual fidelity: Trained on a large-scale AAA game dataset, the generated videos have high visual fidelity and realism, and are applicable to a variety of game scenes and artistic styles.
Technical Principles of Hunyuan-GameCraft
- Unified action representation: Maps keyboard and mouse inputs (such as W, A, S, D, arrow keys, etc.) into a continuous camera representation space, and encodes the input camera trajectory into feature vectors based on a lightweight action encoder, achieving smooth action interpolation.
- Hybrid history-conditioned training strategy: Based on historical context integration and mask indicators, autoregressively extends video sequences. In each autoregressive step, historical denoising blocks are used as conditions to guide the denoising process of new noisy latent variables, preserving scene information and solving the problem of error accumulation in autoregressive generation.
- Model distillation: Implements model distillation based on the Phased Consistency Model (PCM), distilling the original diffusion process and classifier-free guidance into a compact eight-step consistency model, significantly improving inference speed and reducing computational overhead.
- Large-scale dataset training: Trained on more than one million game recording data entries from over 100 AAA games, ensuring broad coverage and diversity of the model. Fine-tuned with a carefully annotated synthetic dataset to improve accuracy and control capability.
Project Links of Hunyuan-GameCraft
- Project Website：https://hunyuan-gamecraft.github.io/
- GitHub Repository：https://github.com/Tencent-Hunyuan/Hunyuan-GameCraft-1.0
- HuggingFace Model Library：https://huggingface.co/tencent/Hunyuan-GameCraft-1.0
- arXiv Technical Paper：https://arxiv.org/pdf/2506.17201
Official Examples of Hunyuan-GameCraft
Single-action scene, natural screen dynamics, the windmill rotates naturally
- Prompt：A picturesque village scene featuring quaint houses, a windmill, lush greenery, and a serene mountain backdrop under a bright blue sky.
Chinese: A picturesque rural scene, with quaint cottages, windmills, lush greenery, and a serene mountain background under a bright blue sky.
Multi-action scenes, allowing simultaneous perspective changes and walking
- Prompt：A sunlit courtyard features white adobe buildings with arched doorways and windows, surrounded by lush greenery and palm trees, creating a serene Mediterranean ambiance.
- Chinese: A sunlit courtyard, white adobe buildings with arched porches and windows, surrounded by lush greenery and palm trees, creating a tranquil Mediterranean atmosphere.
Historical consistency, no deformation after moving the camera away and back
- Prompt：A medieval stone castle stands tall under a dark sky, its glowing windows contrasting with the surrounding snow-covered landscape.
- Chinese: A medieval stone castle stands tall under a dark sky, its glowing windows contrasting sharply with the surrounding snow-covered landscape.
Application scenarios of Hunyuan-GameCraft
- Game video generation: Rapidly generate game trailers, demo videos, and in-game cutscenes, helping game developers validate game concepts and designs in the early stages.
- Game testing: Automatically generate game scenes and interactions, used to test game performance and player experience, reducing the workload of manual testing.
- Game content expansion: Generate new levels, scenes, and interactive content for existing games, extending the game's lifecycle and increasing player engagement.
- Interactive video content: Generate interactive video content for video platforms and social media, where users input commands to control the video's development, providing a brand-new viewing experience.
- Virtual reality (VR) and augmented reality (AR): Generate immersive interactive content in VR and AR applications, enhancing users' sense of immersion and participation.