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

arXiv: https://arxiv.org/abs/2606.02580 · HTML: https://mlsanity.com/paper/2606.02580/

## Key facts

- Cited by 10 papers in the map (data through 14 Oct 2026).
- Placed in VLM benchmarks and visual reasoning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 23 of its 46 references that are in the map.
- First posted on arXiv 1 Jun 2026.

## Placed in

- Multimodal LLMs / vision-language models / [VLM benchmarks and visual reasoning](https://mlsanity.com/mllm/vlm-benchmarks-and-visual-reasoning/)

## Builds on (most cited first, 23 of 23)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [3D Gaussian Splatting for Real-Time Radiance Field Rendering](https://mlsanity.com/paper/2308.04079/) (2023)
- [CAT3D: Create Anything in 3D with Multi-View Diffusion Models](https://mlsanity.com/paper/2405.10314/) (2024)
- [Extracting Triangular 3D Models, Materials, and Lighting From Images](https://mlsanity.com/paper/2111.12503/) (2021)
- [PhySG: Inverse Rendering with Spherical Gaussians for Physics-based Material Editing and Relighting](https://mlsanity.com/paper/2104.00674/) (2021)
- [SAM 3D: 3Dfy Anything in Images](https://mlsanity.com/paper/2511.16624/) (2025)
- [Learning Object-Compositional Neural Radiance Field for Editable Scene Rendering](https://mlsanity.com/paper/2109.01847/) (2021)
- [GS-IR: 3D Gaussian Splatting for Inverse Rendering](https://mlsanity.com/paper/2311.16473/) (2023)
- [CSGNet: Neural Shape Parser for Constructive Solid Geometry](https://mlsanity.com/paper/1712.08290/) (2017)
- [Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model](https://mlsanity.com/paper/2410.13882/) (2024)
- [Deep 3D Capture: Geometry and Reflectance from Sparse Multi-View Images](https://mlsanity.com/paper/2003.12642/) (2020)
- [Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning](https://mlsanity.com/paper/2601.11109/) (2026)
- [VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space](https://mlsanity.com/paper/2508.19247/) (2025)
- [Non-rigid Point Cloud Registration with Neural Deformation Pyramid](https://mlsanity.com/paper/2205.12796/) (2022)
- [Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives](https://mlsanity.com/paper/2307.05473/) (2023)
- [The Scene Language: Representing Scenes with Programs, Words, and Embeddings](https://mlsanity.com/paper/2410.16770/) (2024)
- [Dynamic LiDAR Re-simulation using Compositional Neural Fields](https://mlsanity.com/paper/2312.05247/) (2023)
- [Generating Physically Stable and Buildable Brick Structures from Text](https://mlsanity.com/paper/2505.05469/) (2025)
- [IR3D-Bench: Evaluating Vision-Language Model Scene Understanding as Agentic Inverse Rendering](https://mlsanity.com/paper/2506.23329/) (2025)
- [MeshCoder: LLM-Powered Structured Mesh Code Generation from Point Clouds](https://mlsanity.com/paper/2508.14879/) (2025)
- [3DAxiesPrompts: Unleashing the 3D Spatial Task Capabilities of GPT-4V](https://mlsanity.com/paper/2312.09738/) (2023)
- [Volumetric Disentanglement for 3D Scene Manipulation](https://mlsanity.com/paper/2206.02776/) (2022)
- [Unsupervised Discovery of Object-Centric Neural Fields](https://mlsanity.com/paper/2402.07376/) (2024)

## Built on by (most cited first, 10 of 10)

- [Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs](https://mlsanity.com/paper/2609.11499/) (2026)
- [4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes](https://mlsanity.com/paper/2610.03715/) (2026)
- [Diffuse the object, keep its label: curating detector training data from a few unlabeled photographs via VLM-built 3D vegetation scenes](https://mlsanity.com/paper/2608.09691/) (2026)
- [3DHarnessBench: Probing Agentic 3D-to-Code Capabilities of Frontier Vision-Language Models](https://mlsanity.com/paper/2609.06535/) (2026)
- [LEGO-Anything: Coding Agents for 3D Scene Reconstruction](https://mlsanity.com/paper/2609.36380/) (2026)
- [AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents](https://mlsanity.com/paper/2609.40353/) (2026)
- [LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction](https://mlsanity.com/paper/2610.01863/) (2026)
- [VETR: Execution-Grounded Self-Improving Agent for Symbolic Graphics Reasoning](https://mlsanity.com/paper/2610.13098/) (2026)
- [CineRecon: Benchmarking Frontier Models for Agentic Cinematic Reconstruction](https://mlsanity.com/paper/2610.16577/) (2026)
- [PhysCodeGym: Benchmarking Coding Agents for Physically Verifiable 4D Reconstruction](https://mlsanity.com/paper/2610.17288/) (2026)

## Abstract (arXiv)

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.