Build a Platform 9¾ LEGO Scene with GPT-6 Astra and V2Fun

Overview

This experiment recreates a Platform 9¾ LEGO-style scene using GPT-6 Astra as the planner and 3D agent, with V2Fun handling complex 3D asset generation. Instead of treating the reference as one text-to-3D task, the workflow breaks the scene into separate components and assembles them into a structured final asset.

Tools Used

• GPT-6 Astra: reference understanding, task planning, scene decomposition, spatial reasoning, and final assembly.
• V2Fun: generation of complex individual 3D components.
• Three.js / procedural code: simple, regular, or repeated geometry where direct code is more efficient.
• Visual reference: defines the target appearance and component relationships.
• GLB: final 3D output format.

Step 1 — Analyze the Reference

GPT-6 Astra first studies the complete Platform 9¾ scene and identifies its major visual elements. Front, side, rear, and isolated reference studies help clarify the structure before modeling begins.

Step 2 — Decompose the Scene

The scene is divided into independent assets: the platform, brick wall, canopy and pillars, Platform 9¾ sign, bench, luggage trolley, lamp, and smaller decorative props. This converts one complex generation task into several well-defined 3D tasks.

Step 3 — Generate Complex Components with V2Fun

Complex components are generated separately with V2Fun. This avoids using large amounts of procedural geometry code to approximate irregular or visually detailed assets. In the demonstrated V2Fun workflow, major components can return within tens of seconds; the total end-to-end workflow time was not benchmarked.

Step 4 — Assemble with GPT-6 Astra

After generation, GPT-6 Astra determines the scale, rotation, position, and spatial relationships of the individual assets. It reconstructs the scene by placing the wall on the platform, aligning the canopy and pillars, positioning the sign and bench, and arranging the smaller props.

Step 5 — Add Procedural Details

Not every element needs a foundation-model call. Simple and regular structures can be created or refined with Three.js or procedural code. The agent chooses whether each part should be generated, coded, reused, or assembled.

Step 6 — Final Assembly

The separate assets are combined into the final Platform 9¾ LEGO-style scene. The workflow preserves explicit component relationships instead of producing only a single opaque mesh.

Workflow

Reference → Understand → Decompose → Generate → Assemble → Refine → Export

Prompt

A fully assembled LEGO brick, with the whole brick building piece designed for the Three-Quarters Station in 9 from Harry Potter. You need walls and a three-quarter sign for 9, and you can have benches under the wall (like at a train station). Building block style.

Result

Final output: hogwarts-platform.glb
File size: 44.02 MB
Complex component generation: within tens of seconds in the demonstrated V2Fun workflow
Total workflow time: not benchmarked

This workflow shows how GPT-6 Astra can act as a 3D agent: it focuses on understanding, planning, and construction, while V2Fun focuses on complex geometry generation.