A domestic research team has developed an artificial intelligence technology that allows a 3D character to mimic the movements in a photo by inputting just one photo. This technology is expected to significantly lower production costs and entry barriers across the metaverse, game, and animation industries, as it can implement natural 3D movements without the need for multi-angle shooting data or complex motion capture equipment, which were essential in the existing 3D content production process.

A research team led by Professor Kyung-Don Joo of the UNIST Graduate School of Artificial Intelligence announced on the 25th that they have developed an AI model called 'DeformSplat (Rigidity-aware 3D Gaussian Deformation)' that deforms the posture of a 3D character generated based on 3D Gaussian Splatting without distortion.

3D Gaussian splatting is a cutting-edge AI technology that reconstructs 3D objects from 2D images, and is rapidly gaining attention in the graphics field.

On the other hand, the existing method required images or continuous video data taken from multiple angles to move the character, and had limitations in that if there was insufficient data, shape distortion occurred, such as abnormal bending of arms and legs.

The Deformsplat developed by the research team can naturally change the pose of a 3D character with just one photo.

The results of the experiment showed that when the character performed movements such as raising his arms or twisting his body, the proportions did not collapse not only from the front but also from the side and back, and the joints hardly ever stretched like rubber.

The core of this technology is two things: △Gaussian-to-Pixel Matching and △Rigid Part Segmentation.

Gaussian pixel matching connects the Gaussian points that make up a 3D character with the pixels in the photo, accurately conveying the pose information contained in the photo in 3D. />
Rigid body segmentation technology automatically finds and binds rigid structures that must be deformed during movement, thereby maintaining the shape of robots, dolls, and characters without distortion.

Professor Joo Kyung-don explained, “Existing technologies have the problem that when trying to move a 3D object with just a single photo, the shape is greatly damaged.” He added, “This research is a technology that allows AI to identify the structural characteristics of an object on its own, distinguish the area that acts as a skeleton, and generate natural movements.”

He then emphasized, “This will significantly lower the threshold for 3D content production, which previously required specialized equipment and expensive production personnel.”

This research result was accepted as a paper at SIGGRAPH ASIA 2025. SIGGRAPH ASIA, the world's most prestigious computer graphics and interactive technology conference hosted by the ACM, has been recognized for the technical perfection and innovativeness of the research.

This year's conference was held in Hong Kong from December 15th to 18th.

This research was conducted with the support of the Ministry of Science and ICT's Institute of Information and Communications Technology Planning and Evaluation (IITP) and UNIST Graduate School of Artificial Intelligence.