An artificial intelligence (AI) technology has been developed that can restore a three-dimensional (3D) shape of a pet dog from just a single photo and even create animations. You can meet an avatar that looks just like my dog in the virtual space.
Ulsan National Institute of Science and Technology (UNIST) announced on the 24th that a research team led by Professor Joo Kyung-don developed an AI model called "DogRecon (Dog Reconstruction)" that can create a movable 3D avatar from just one photo of a dog.
Dogs have different body shapes depending on their breed, and due to their four-legged walking characteristics, it's often challenging to capture their joints, making them difficult animals for 3D restoration. Particularly, restoring a 3D shape from a single two-dimensional photo often results in some areas being inaccurately generated or distorted due to a lack of information.
The researchers explained that DogRecon applies a dog-specific statistical model to capture differences in body shape and posture by breed, and it can restore parts that are hidden by generating various angled images automatically with generative AI. It also precisely reproduces the dog's curved body shape and fur texture based on a video restoration technique that expresses a point in three-dimensional space softly rather than sharply.
In performance tests, DogRecon created a 3D avatar of a dog that was natural and accurate, similar to existing video-based technologies, even from just a single photo. Previous models often depicted dogs in unrealistic forms, such as having an extended torso or awkwardly bent joints when the dog was in a seated position, and features like ears, tails, and fur were often clustered inaccurately.
The researchers noted that DogRecon can generate movements for the avatar with just text input, allowing it to be utilized in virtual reality (VR) and augmented reality (AR).
Researcher Cho Kyung-soo, the first author, said, "As more than a quarter of all households have pets, I wanted to extend the 3D restoration technology that has been developed with humans in mind to pets. This will be a tool that allows anyone to implement their pet in a digital space and make it move on its own."
Professor Joo Kyung-don said, "This research is a significant achievement that combines generative AI with 3D restoration technology to create models of pets that resemble reality. I look forward to expanding it to various animals or personalized avatars in the future."
This research result was published in the June 2 issue of the International Journal of Computer Vision, which is the leading journal in the field of computer vision. Demonstration videos can be found on the project website (https://vision3d-lab.github.io/dogrecon/).
References
International Journal of Computer Vision (2025), DOI: https://doi.org/10.1007/s11263-025-02485-5