I’ve watched a lot of people try to turn flat images into usable 3D models, and honestly, most of the time it’s a slog. You wrestle with software, tweak meshes for hours, and still end up with something that looks off. So when I came across this experiment, I stopped scrolling immediately.
This one comes from a sharp AI professional who decided to test just how far the newest models can go. The creator handed Codex a set of images and asked GPT-6 Astra to build a full 3D scene from scratch. Not a rough blob, a detailed model with interior and exterior views. I think it’s one of the cleanest real-world examples of AI-native tool use I’ve seen this year.
The problem: images are flat, 3D is hard
Here’s the setup. The original poster had a set of reference images from architect Ar. June Chow, showing both the exterior and interior of a gallery concept. Beautiful shots, but static. Turning that kind of reference into an actual 3D model usually means opening heavy software, learning its quirks, and manually rebuilding every wall, light, and surface.
That’s the wall most people hit. The vision is there, but the execution eats your whole week. The expert wanted to know if an AI model could carry that load instead.
The solution: headless Blender + Python
This is the part that made me sit up. The creator didn’t ask Astra to click around inside Blender’s normal interface. Instead, Astra controlled Blender through Python scripts running headless in the background on the laptop.
If “headless” is new to you, here’s the simple version. Blender has a graphical interface, the buttons and panels humans use. Headless means skipping that entirely and driving the program straight through code. No mouse, no menus, just scripts telling Blender exactly what to build.
What I love is how the original poster framed it: this is a more AI-native way of using a tool. The model doesn’t need the human interface at all. It talks to the software the way software talks to software.
A few things stood out about the approach:
- No extra setup: the expert pointed out there was no need to configure a Blender MCP connection.
- Full pipeline in code: Astra could build, render, and export the entire scene by running Python scripts directly in Blender.
- Runs locally: the whole thing worked in the background on a personal laptop, not some giant cloud rig.
Why it matters: when an AI can drive a tool through its scripting layer instead of its buttons, it stops being a fancy assistant and starts being an operator. That’s a big shift in how we’ll use creative software.
The result: a working scene in one afternoon
So how long did all of this take? According to the person who shared it, one afternoon. That’s it.
And it gets better. Within the first hour after the initial prompt, Codex had already produced a working MVP. A functional first version, ready to look at. After that, the creator spent the rest of the time fine-tuning details and testing different approaches to sharpen the output.
I was genuinely impressed reading that timeline. An hour to a working model is the kind of speed that changes how you even plan a project.
The clever twist: a playable web explorer
The original poster also referenced a recent tutorial article from OpenAI and called it an absolutely inspiring read. But this savvy professional made one smart change of their own.
The OpenAI walkthrough leaned on Unreal Engine. Instead of going that route, the creator asked Astra to build a playable web explorer using Three.js. If you haven’t used it, Three.js is a JavaScript library that renders 3D graphics right in the browser. No install, no heavy engine, just open a link and walk through the scene.
That decision makes the whole thing shareable. Anyone with a browser could explore the gallery concept, which is a lot friendlier than needing a full game engine on your machine.
What you can take from this
Even if you’re not building 3D galleries, there are real lessons here worth stealing:
- Think headless: ask whether your AI can drive a tool through scripts instead of clicking the interface. It’s often faster and cleaner.
- Chase the MVP first: the creator got a working version in an hour, then refined. Rough and functional beats perfect and unfinished.
- Pick lightweight outputs: choosing Three.js over a heavy engine made the result instantly shareable. Match the tool to who needs to see it.
- Borrow and adapt: the expert learned from an existing tutorial, then swapped in a better fit for their own goals.
The bigger picture here is that AI is starting to operate our creative tools directly, not just suggest ideas about them. When a model can build, render, and export a full scene on its own, the gap between concept and finished work shrinks fast.
The original poster hinted there are more Astra tests coming, and I’ll be watching closely. If you want the full breakdown and the visuals of that gallery scene, go check out the complete LinkedIn post. It’s absolutely worth a look.