Title: Thinking in Blender: Staged Executable Inverse Graphics with Vision-Language Models

Abstract:

Reconstructing images into editable 3D environments capable of rendering, relighting, and manipulation constitutes a classic inverse graphics challenge, one that has long suffered from significant underconstraint. This study explores the potential of pretrained vision-language models (VLMs) to execute inverse graphics tasks directly from single images. By generating editable Blender scripts, our approach operates without the need for specialized 2D or 3D foundation models, differentiable rendering techniques, or multi-view data supervision.

We present Staged Executable Inverse Graphics (SEIG), an agentic framework designed to rebuild 3D scenes from individual images. SEIG achieves this by iteratively refining critical scene attributes—such as geometry, materials, composition, and lighting—within the executable Blender code space. To assess performance, we tested the framework across a variety of scenes, employing metrics that evaluate pixel-level accuracy, perceptual quality, and semantic consistency.

Our results demonstrate that a staged reconstruction process significantly enhances fidelity, underscoring the value of decomposing tasks when applying general-purpose VLMs to executable inverse graphics. Furthermore, we illustrate several downstream applications made possible by the resulting editable Blender scenes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC