miHoYo’s Anti-Entropy Team Unveils 10x Faster 3D Character Animation Model

In a newly surfaced AI paper, miHoYo’s Anti-Entropy team has proposed a new character animation generation model called Motion Diffusion Transformer. Based on a text description or simple motion input, the model can generate high-quality, multi-style 3D character animation sequences within seconds—more than 10 times the efficiency of traditional methods. The breakthrough quickly drew intense attention from game developers and AI practitioners alike.

The paper details how the model combines diffusion probabilistic models with a Transformer architecture to achieve efficient modeling of complex motion sequences. Unlike traditional animation pipelines that depend on large amounts of hand-keyframed work or motion capture data, the Anti-Entropy team’s model only needs natural-language instructions such as “jump while running” or “turn around and swing a sword” to automatically generate continuous, lifelike motion sequences. It also supports fine-grained control over motion style, rhythm, and intensity, giving game characters unprecedented expressiveness.

For the game industry, the potential impact is profound. In open-world and MMO games, character animation production often consumes enormous amounts of labor and time. miHoYo itself has accumulated vast animation assets across titles like Genshin Impact and Honkai: Star Rail, and applying the new model could compress content production cycles from weeks to hours, dramatically improving development efficiency. The technology may also drive more dynamic, personalized NPC behavior generation, making game worlds more immersive.

Notably, the Anti-Entropy team’s achievement is not an isolated event. In recent years, tech giants such as NVIDIA and Meta, along with numerous AI startups, have all been exploring generative AI approaches to character animation. NVIDIA’s Avatar Cloud Engine, for example, already delivers real-time voice-driven facial animation, while the Anti-Entropy team’s new model pushes the boundaries of full-body motion generation further. It is easy to see 2025 as the pivotal turning point when AI-driven game animation moves from the lab into large-scale commercial use.

Real-world deployment still faces challenges, though. Ensuring the physical plausibility of generated animation and the coherence of multi-character interactions, integrating seamlessly with existing game engines (such as Unity and Unreal Engine), and avoiding the “uncanny valley” effect in generated content are all problems miHoYo and the industry must keep working on. In addition, defining copyright for AI-generated content and re-dividing roles within teams will be issues developers have to confront head-on.

Looking ahead, with continued investment from pioneers like the Anti-Entropy team, we have every reason to believe AI character animation will no longer be the exclusive domain of a few big studios—it will gradually open up to small and mid-sized teams and even indie developers. When that happens, every character in a game could have unique, real-time evolving motion, and players will step into virtual worlds that feel more vivid and alive than ever.