SpatialCrafter turns a single image into an explorable 3D scene via a generative proxy

Researchers presented SpatialCrafter, a two-stage framework for generating explorable 3D scenes from a single image. The paper was published on arXiv on August 27, 2026. Instead of directly using a video diffusion model with incomplete signals such as sparse point clouds or 2D panoramas, the authors first build a global 3D proxy via the PaSS Flow module, then refine the scene's appearance with a Generative Deferred Refiner. For training, a hybrid dataset of 115 thousand scenes was assembled. The authors have not released the code. More details in the arXiv paper.

What it means

SpatialCrafter tackles the image-to-scene task, i.e., generating an entire scene rather than a single object. This sets it apart from the tools in our reference, which work with individual 3D models. OpenLRM handles image-to-3d: from one image, it produces a single object in OBJ format. The fact sheet mentions nothing about scenes or proxy geometry. Hunyuan3D 2.1 also generates a single 3D asset from an image, with an emphasis on PBR materials. Explorable scenes and generative proxies are not mentioned in its description. TRELLIS supports image-to-3d and text-to-3d, outputting a textured mesh in GLB. Working with scenes from a single image is not stated in the fact sheet. For now, SpatialCrafter remains a research effort without open source code, so it cannot be practically compared with available tools. If the code appears, this could expand the AI for 3D section toward scene generation rather than just individual models.