Chinese scientists’ AI tool generates, edits lifelike human actions in 3D motion
Being able to realistically and accurately recreate human motions could be extremely useful in several fields, as it can help developers create more lifelike characters in video games and animations, enhance immersive experiences in virtual reality (VR), and improve the quality of training videos in areas like healthcare, sports, and emergency response.
In a bid to achieve this, researchers from Peking University’s Institute for Artificial Intelligence in Beijing and the State Key Laboratory of General AI have unveiled new AI models that can simplify a range of human motions for digital characters and avatars.
AI model for realistic motion synthesis
The researchers’ approach to generating human motion, presented at this year’s Conference on Computer Vision and Pattern Recognition (CVPR 2025), combines a data augmentation technique called MotionCutMix with a diffusion model known as MotionReFit, Tech Xplore writes.
According to Yixin Zhu, senior author of the paper, while motion generation had made great strides, the ability to edit existing human motions was still severely lacking. In creative fields like game development, animation, and digital art, professionals often work by refining and modifying existing content instead of creating everything from the ground up.
Therefore, Zhu and his team set out to build a system capable of editing any human motion using simple written instructions, without needing task-specific details or body part labels. The system is designed to handle both spatial edits—focusing on specific body parts—and temporal edits, which adjust movement over time.
It also generalizes well across different scenarios, even when trained with limited annotated data. To make this possible, the researchers introduced MotionCutMix, a simple yet effective training method that teaches AI to edit 3D human motions based on text input. Similar to how chefs combine ingredients to create various dishes, MotionCutMix generates diverse training examples by blending body parts from different motion sequences.
Smooth transitions between body parts for realistic animations
The learning approach developed by the researchers allows the selection of specific body parts, such as arms, legs, or torso, from one motion sequence and blends them with parts from another sequence.
Instead of creating a jarring transition between movements, MotionCutMix gradually smooths the boundaries between body parts, resulting in more natural, fluid motions. For each blended motion, a new training example is generated, consisting of the original motion, the edited version, and a text instruction that describes the change.
Previous methods for generating human motions typically used fixed datasets, often consisting of annotated videos of people moving in different ways. In contrast, MotionCutMix generates new training samples on the fly, enabling learning from large libraries of motion data without the need for manual annotation.
This approach is especially useful since much of the content available online is not annotated and thus cannot be utilized by other methods. The new framework also allows editing both the specific movement of body parts and the style of those movements.
Thus, MotionCutMix requires far fewer annotated examples to achieve strong results, generating potentially millions of training variations from a small set of labeled examples. By training on diverse combinations of body parts and motions, the model learns to handle a wider range of editing requests.
Despite the increased complexity of training examples, the process remains efficient, with soft masking and body part coordination ensuring smoother, more natural edited motions without awkward transitions or unrealistic movements, the researchers noted.
Source: Interesting Engineering
Lithium-ion batteries offer longer lifespan, fast charging with nanoscale engineering
Chinese scientists’ AI tool generates, edits lifelike human actions in 3D motion