Bilawal Sidhu uses GPT-6 Astra to rebuild a scanned room as an editable Blender scene
Bilawal Sidhu demonstrates using GPT-6 Astra to turn a RealityScan room mesh into an editable Blender scene with isolated objects and procedural materials. He also demonstrates physics in the rebuilt scene and distinguishes it from point clouds or an unstructured triangle mesh.
TL;DR
- GPT-6 Astra rebuilt a scanned room as editable Blender objects, according to the creator's walkthrough.
- The output has a scene hierarchy resembling human-built 3D work, Sidhu explains in a follow-up.
- The demo includes physics: Sidhu switches off gravity and lets the reconstructed furniture float and collide.
Bilawal Sidhu started with roughly 400 photos he took in 2019 of his parents' old house. He says the reconstruction also replaced the scan textures with procedural materials, without downloading replacement assets.
From RealityScan to Blender
Sidhu describes four stages:
- Photogrammetry: RealityScan turns the photographs into a mesh.
- Reconstruction: GPT-6 Astra uses the scan as a reference to rebuild the room as a clean, editable Blender scene.
- Materials: The captured textures become procedural materials.
- Object separation: Individual objects are isolated and editable, enabling the physics demonstration.
The walkthrough does not include prompts, reconstruction time or a downloadable Blender scene.
Scene hierarchy and physics
Sidhu describes the result as a hierarchically decomposed scene, comparable to how a human would model it. He contrasts that with three earlier export options:
- Point clouds.
- Unstructured triangles, or “triangle soup.”
- Gaussians.
The editable hierarchy is the standout. The gravity sequence puts that object-level control to work.
Blender's rigid-body system moves and rotates mesh objects without deforming them. Rigid bodies must be explicitly added to objects, so object separation and physics configuration are distinct parts of the workflow.
Studios and simulation environments
Sidhu pitches reconstructing scans as a way to build reinforcement-learning environments, alongside applications in VFX, AR/VR and robotics. He argues that Opus and Astra subscriptions could lower the barrier previously imposed by 3D-authoring and simulation expertise.
He also urges existing studios to pivot toward this work, even joking about applying to Y Combinator.
Scale, topology and thin structures
Sidhu's replies add three details about geometry:
- Relative scale: He credits the real-life scan as a reference with keeping proportions consistent.
- Topology: He describes the rebuilt geometry as “pretty damn good”.
- Thin structures: He flags these as particularly difficult for traditional photogrammetry, a limitation of the capture stage.