Thinking in Blender: Staged Executable Inverse Graphics with Vision-Language Models
2026-06-01 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors explore using vision-language models to turn a single image into a 3D editable scene using Blender code, without needing special 3D data or extra views. They created a method called SEIG, which breaks down the reconstruction into steps like geometry and lighting to improve accuracy. Their tests show that doing this step-by-step helps get better 3D results. They also demonstrate how these editable scenes can be used for various creative tasks.
Inverse graphicsVision-language models3D reconstructionBlenderScene compositionGeometryMaterialsLightingExecutable codeTask decomposition
Authors
Guangzhao He, Rundong Luo, Wei-Chiu Ma, Hadar Averbuch-Elor
Abstract
Inverse graphics is a longstanding and highly underconstrained problem that seeks to reconstruct images as editable 3D scenes which can be rendered, relit, and manipulated. In this work, we investigate whether pretrained vision-language models (VLMs) can perform executable inverse graphics directly from a single image by reconstructing a scene as an editable Blender program, without relying on specialized 2D or 3D foundation models, differentiable rendering, or multi-view supervision. We introduce Staged Executable Inverse Graphics (SEIG), an agentic framework that reconstructs a 3D scene from a single image by progressively refining scene factors including geometry, materials, composition, and lighting directly in executable Blender code space. We evaluate our framework across diverse scenes using a range of reconstruction metrics spanning pixel-level, perceptual, and semantic fidelity. Our experiments show that staged reconstruction substantially improves reconstruction fidelity, highlighting the importance of task decomposition for executable inverse graphics with general-purpose VLMs. Finally, we showcase various downstream applications enabled by the reconstructed editable Blender scenes.