SceneReGen builds 3D scenes from a single image, generating objects in a shared coordinate system

On August 25, 2026, the paper SceneReGen: Generative Reconstruction of 3D Scenes from a Single Image was published on arXiv. The authors propose a method for generative reconstruction of 3D scenes from a single image: instead of reconstructing the scene as a whole, SceneReGen generates individual objects and assembles them in a common coordinate system aligned with the observation. The key idea is selective pose factorization: the object's orientation is encoded directly in the generated mesh, while translation and scale are estimated separately from scene features. On a subset of 3D-FUTURE, the method shows better results in scene-level CD, scene-level F-Score, and 3D bounding-box IoU among the considered methods. The authors have not released the code.

What it means

SceneReGen solves a problem that none of the tools in our reference currently addresses. OpenLRM, Hunyuan3D 2.1, and TRELLIS are generators of single 3D models from one image (image-to-3d). They create one object in a centered, scale-normalized coordinate system. SceneReGen works one level above: it takes a scene image and instance masks as input and outputs multiple objects arranged in a common scene. The gap between object generation and scene reconstruction is exactly the problem the authors call the representation gap. In our fact sheets, this gap is directly visible: for all three tools, the task is listed as image-to-3d, not scene reconstruction. None of them accepts instance masks or assembles multiple objects into a common scene. SceneReGen has no practical application yet: the code is not published, the weights are not released, and the developer does not specify VRAM requirements or export formats. This is a research work that shows a direction — generative assembly of scenes from ready-made assets — but does not provide a ready-made tool for comparison with what is already in the CPU3D reference. If the authors publish the code and weights, SceneReGen could become the first tool in our reference for generative scene reconstruction rather than single models. For now, it is a reason to follow the development of the topic, not a replacement for existing generators.