miHoYo AI Paper Leaked: Anti-Entropy Team's New Method Skyrockets Animation Generation Efficiency

A recent AI research paper from miHoYo's Anti-Entropy team has drawn wide attention in the gaming and AI fields. Titled 'Efficient Controllable Motion Generation via Diffusion Models for Game Characters,' the paper proposes a new framework for efficiently generating controllable character animation using diffusion models. Compared with traditional motion capture or hand keyframing, the method can significantly reduce animation production costs while maintaining motion fluidity and diversity.

The Anti-Entropy team is miHoYo's internal core department focused on AI technology R&D, and has previously applied some AI-assisted tools in Honkai: Star Rail and Genshin Impact. The newly revealed paper shows they designed a generative architecture based on conditional diffusion models that can produce animation sequences consistent with physical laws and character settings from text descriptions or motion tags. Enter 'sudden turn while running' or 'elegantly swinging a sword,' for example, and the model outputs the corresponding skeletal animation data.

The technology's core innovation is introducing a balance between 'controllability' and 'efficiency.' Traditional diffusion models often need tens of seconds or even minutes to generate high-quality animation, but the Anti-Entropy team compressed generation time to real-time levels by optimizing sampling steps and network architecture—making dynamic motion generation inside game engines possible. They also proposed a 'pose guidance' mechanism letting developers constrain generated results with a few keyframes, preventing the AI from producing eerie motions that miss expectations.

As an industry trend, AI character animation is becoming a hot spot in game development. Unity's Muse tools could already generate basic animation from natural language, but miHoYo's work focuses more on high-precision, interactive game characters. For open-world and massively multiplayer online games, AI-driven animation generation means NPCs can have richer body language instead of relying on preset limited motion libraries. This not only boosts immersion but dramatically reduces the art team's workload.

Deployment still faces challenges, however. First is the model's adaptability to specific styles (such as miHoYo's signature Japanese-style cel shading); second is how to avoid 'jitter' or 'clipping' problems in AI-generated motion. The Anti-Entropy team says in the paper they will further optimize through mixture-of-experts models and reinforcement learning.

Looking ahead, miHoYo may integrate this technology into flagship projects such as its unannounced Project SH (an urban open world). As AI animation tools become open source, small-to-mid teams may also use similar solutions to rapidly iterate game prototypes. Within the next two years, AI can be expected to become an indispensable part of the character animation pipeline—and miHoYo stands at the forefront of this transformation.