miHoYo AI Paper Revealed: How the Anti-Entropy Team Revolutionizes Character Animation Generation

miHoYo, a globally leading game developer, recently saw its Anti-Entropy team publish a heavyweight AI paper titled 'Efficient Character Animation Generation via Diffusion Models.' The paper details a new method for rapidly generating high-quality character animation using diffusion models, drawing wide attention from game developers and AI researchers alike.

In traditional game development, producing character animation is extremely time-consuming and expensive. Animators must hand-craft keyframes for every motion—from walking, running, and jumping to complex combat combos—then perform fine adjustment and blending. For open-world games with huge character rosters like Genshin Impact and Honkai: Star Rail, the volume of animation work grows exponentially. The Anti-Entropy team's paper aims precisely at this pain point.

The paper's core technology is applying diffusion models to animation generation. Diffusion models have achieved enormous success in image and video generation in recent years; their basic principle is to progressively corrupt training data by adding noise, then learn the reverse denoising process to generate new data. The Anti-Entropy team innovatively adapted this framework to skeletal animation data. They built a Transformer-based architecture capable of handling joint rotation and position data across time series. Trained on massive motion capture data and in-game animation data, the model learned to understand the intrinsic relationships and physical laws between different motions.

Specifically, the method's input can be a simple text description ('a character sprinting forward rapidly') or a rough reference animation. The model outputs a smooth, natural, physically plausible animation sequence. More importantly, it supports conditional generation—specifying a character's initial pose, movement speed, or emotional state—for fine-grained control. This gives game designers unprecedented flexibility to rapidly iterate different motion designs without waiting for animators to craft them frame by frame.

The technology's potential impact is profound. First, it promises to dramatically shorten development cycles: a complex combat animation sequence that once took weeks may now be generated in minutes. Second, it can significantly reduce animation production costs—especially for indie studios and small teams, where AI animation generation tools let them create character expressiveness rivaling AAA blockbusters with limited resources.

miHoYo is not the only company exploring AI animation generation. Engine giants like Unity and Unreal Engine are also actively integrating AI tools. Unity's Muse toolkit already includes AI-based animation generation, and Unreal Engine's MetaHuman framework uses machine learning to simplify character creation. What sets miHoYo's paper apart, however, is its focus on optimizing for real-time in-game interaction scenarios. The animation it generates is not only high quality but computationally light, deployable directly in a game's runtime environment—meaning NPC animation can adapt dynamically to player behavior, creating more natural and immersive interaction.

Looking ahead, AI's application in game animation will no longer be limited to generation. We may see AI-driven procedural animation systems that adjust a character's body posture and reactions in real time based on in-game physics collisions, environmental changes, and character state. Combined with natural language processing, players may one day issue complex motion commands to characters by voice, with AI instantly generating the corresponding animation in response.

Of course, the technology faces challenges. How do we ensure copyright ownership of AI-generated animation? How do we prevent models from producing bizarre motions that violate physical common sense or a game's art style? These are questions developers and the legal community must explore together. What is undeniable is that the Anti-Entropy team's paper marks AI's gradual evolution from an auxiliary tool into a core driving force of game content creation—heralding an important shift in the game development paradigm.