Inside miHoYo’s AI Paper: Dynamic Character Animation Generation Gets 10x Faster

In the game industry, character animation has always been a core pain point in development. Traditional hand-crafting or motion capture pipelines are slow and laborious, while AI-generated animation has struggled with incoherent motion and inconsistent style. An AI paper from miHoYo’s Anti-Entropy team is now generating buzz: it proposes a dynamic character animation generation framework combining diffusion models with physical constraints, achieving breakthroughs in both efficiency and quality.

The method proposed in the paper is called PhysDiff, short for Physical Diffusion for Controllable Character Animation. Its core idea is combining a diffusion model with a physics simulator: the diffusion model first generates a preliminary motion sequence, then a physics engine (such as PhysX) applies constraint corrections to ensure physical plausibility—no foot sliding, no body clipping. Compared with traditional methods, PhysDiff boosts generation speed 10x while preserving motion diversity, and it requires no large amounts of hand-labeled data.

The Anti-Entropy team presents extensive experimental comparisons in the paper: PhysDiff-generated running, jumping, and attacking motions significantly outperform existing state-of-the-art models such as MotionDiffuse and MDM in temporal coherence and style consistency. More importantly, its computational resource consumption drops by about 60%, meaning ordinary developers or small teams could plausibly run it on local hardware.

For the game industry, the application prospects are broad. First, it can drastically simplify the character animation pipeline: complex motion sequences that once took weeks could now be finished in hours. Second, PhysDiff supports conditional control—developers can input a simple text description (such as “light-footed running” or “heavy sword swing”) or a reference video and get the corresponding animation, sharply reducing communication costs between art and design teams.

In addition, the method is especially friendly to indie game developers. Unity and Unreal Engine have begun integrating AI animation plugins, but most rely on cloud APIs with high latency and cost. If PhysDiff is open-sourced or integrated into editors, indie teams could gain animation generation capabilities rivaling AAA studios, right on their local machines.

The technology does face challenges, though. The coupling of the physics simulator with the diffusion model still produces occasional “jitter,” especially during complex terrain interactions. The miHoYo team proposes using a “temporary smoothing filter” to mitigate it in the paper, but the problem is not fully solved. Style transfer precision also has room to improve, particularly for non-humanoid characters such as quadrupeds and mechanical beings.

Looking ahead, competition in AI character animation will revolve around “controllability” and “real-time performance.” miHoYo’s PhysDiff sets a new benchmark for the industry, and more game companies are expected to follow suit with similar tools in the second half of the year. Engine makers like Unity and Unreal may also accelerate the integration of such technology, pushing game development into an “AI-native” era.

For players, this means richer, more vivid game worlds. NPC movement will no longer feel stiff, and character combat animation will be smoother and more natural. This paper from miHoYo may well be a key to the next generation of game experiences.