Researchers build VIGA to turn 2D images into editable Blender 3D scenes

VIGA is a multimodal agent that uses Blender as a visual thinking engine. Feed it a single 2D image, and it reconstructs the scene into an editable, programmatic 3D environment complete with physics and interactions.

What it can do

Unlike standard 3D generators that spit out uneditable static meshes, VIGA builds procedural scenes you can actually work with inside Blender.

It can build scenes from scratch using basic geometric primitives like parametric cuboids and spheres. If you need detailed assets, it can also plug into existing tools like Meshy or SAM-3D, treating their outputs as a starting point before rigging up physical interactions and scene layouts.

Beyond 3D scenes, the researchers found that VIGA's approach naturally generalizes to other code-driven creative tasks, including 2D slide presentation design.

How it works

At the heart of VIGA is an analysis-by-synthesis loop. Instead of guessing the final 3D scene in one shot, the agent writes code, renders the scene in Blender, inspects the result against the original image, and iteratively refines its instructions.

This process relies on interleaved multimodal reasoning combined with an evolving contextual memory. By keeping track of previous editing steps and render mistakes, VIGA can debug object placements, materials, lighting, and camera angles step by step.

To put it to the test, the team introduced BlenderBench, a suite of 30 challenging tasks spanning camera adjustments, multi-step edits, and compositional modifications. Across these tests, VIGA delivered an average performance improvement of 124.70% over standard one-shot baselines like GPT-4o and Qwen3-VL-8B. It also clearly beat BlenderAlchemy, a memory-less variant of itself, and scored a 35.32% improvement on the single-step BlenderGym benchmark.

Availability and limits

The research paper is available on arXiv (2601.11109), with demo videos and task breakdowns hosted on the project website.

The system has been evaluated with both proprietary models like GPT-4o and open-weight models like Qwen3-VL-8B. While the technical details and benchmark are fully documented, anyone looking to run it locally should keep an eye on their project page for packaged code releases and setup guides.

Why it matters

For 3D artists and game developers, editability has always been the main bottleneck of AI generation. Getting a messy triangular blob is rarely useful in a real production pipeline. VIGA shows a much smarter alternative: let the AI write clean Blender Python code and use the render engine as its ground truth.

We think using a graphics engine as an interactive sandbox neatly solves the spatial reasoning gap that vision-language models struggle with. If you work in 3D pipelines or procedural generation, this agent-first workflow is definitely one to watch.