A single photograph taken with a 360° camera can now be used to generate a 3D indoor space, including its walls, floors, and furniture. By generating the structural layout and individual objects separately, the technology enables furniture to be moved or replaced and offers potential applications in digital twins and virtual environments for robot learning.
Professor Chang D. Yoo’s team in our department, in collaboration with NAVER LABS, has developed InSpace, a framework that jointly generates an indoor space’s structural layout and individual objects from a single 360° image. The work was selected for a Spotlight presentation at ECCV 2026, a leading international conference in computer vision.
This achievement follows the team’s one Oral paper and one Spotlight paper at ICML 2026, bringing its total to three Oral or Spotlight papers at ICML and ECCV this year. The respective selection rates were approximately 0.7% of submissions for ICML Oral, 2.2% for ICML Spotlight, and 1.57% for ECCV Spotlight. These selections highlight the team’s research contributions across machine learning and computer vision.
(left): the output is decomposed into the structural layout and the individual assets >
Previous image-to-3D methods have largely focused on generating individual objects visible in a photograph. Creating an indoor environment, however, also requires modeling walls, floors, and the spatial relationships between these structures and the objects within them. Without this structural context, generated furniture may float above the floor or be positioned inconsistently with its surroundings.
To address this challenge, the team combined the broad visual coverage of 360° imagery with a generation framework that explicitly considers spatial structure. While an ordinary photograph captures only part of a room, a 360° image provides richer information about the surrounding environment. InSpace uses this information to generate the room’s overall structure together with its individual objects.
The input is represented in equirectangular projection (ERP), which maps a spherical 360° view onto a rectangular image, much as a globe is unfolded into a world map.
A key strength of InSpace is its ability to generate an entire indoor scene while keeping individual objects separately editable. Furniture is produced as separate 3D meshes, allowing objects to be repositioned or replaced. Potential applications include indoor digital twins, virtual environments for robot and humanoid navigation and manipulation training, interior design simulations, and AR/VR content creation.
Professor Yoo said, “This research broadens the target of 3D generative models from objects to spaces,” adding, “A technology that turns a single photograph into a usable 3D space could be a key to resolving the data bottleneck that robots and AI face in learning real-world spaces.”
Gwanhyeong Koo, a Ph.D. candidate at KAIST, is the first author of the paper. ECCV 2026 selected 164 of its 10,473 submitted papers for Spotlight presentations, corresponding to approximately 1.57% of submissions.
This research was conducted in collaboration with NAVER LABS, and was supported by the Korea Planning & Evaluation Institute of Industrial Technology (KEIT) (Project No. RS-2024-00417108) and by the Institute for Information & Communications Technology Planning & Evaluation (IITP) (Project No. RS-2021-II211381, “Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments”).
- Paper title: InSpace: Structure-Aware 3D Indoor Scene Generation from a Single 360° Image
- Link to original paper: https://arxiv.org/abs/2607.03990
- Project page: https://kookie12.github.io/InSpace-Project-Page/
