GPT-6 Astra 3D: The Essentials

Last updated: September 21, 2026

GPT-6 Astra 3D turns 1 to 8 reference photos of a hard-surface object into a real editable 3D asset. Instead of one fused mesh, it reconstructs the object as separately named, selectable parts, each with clean topology and its own PBR material, so a prop can be recolored, swapped, or handed off to Blender or a game engine part by part. It is built for props and hard-surface subjects such as vehicles, weapons, tools, and furniture, not characters. Try it on the GPT-6 Astra 3D model page.

Reach for GPT-6 Astra 3D whenever the result needs to stay editable, that is, when individual parts must be selected, recolored, or swapped later.

How to Use the Model

How GPT-6 Astra 3D Works

Feed it 1 to 8 photos or renders of a single object. The first image sets the viewpoint the result is checked against; any extra images help the model resolve the sides that first photo does not show. Even a single clean reference photo is enough to get a fully separated, editable result.

a vintage brass diving helmet, riveted brass, thick glass viewports

Turntable of the result, built from a single reference photo at Maximum effort. It came back as 33 separate named materials over roughly 151,000 triangles, ready to recolor part by part. Open this asset in Scenario

Feeding Multiple Reference Views

When the object is more complex, add up to 8 photos. Use a front view and a rear (or rear three-quarter) view rather than two side profiles: the model checks the result against the first image, so the extra views should cover ground the first one cannot, not repeat it from another angle.

a compact all-terrain scout drone, matte gray and orange chassis, rugged tires

Turntable built from front and rear reference photos of the same design, capped at a 20,000 triangle game-ready budget with Maximum effort and 8 refinement passes. Open this asset in Scenario

Two independently generated reference photos of the same imagined object can differ slightly in proportions, since each comes from a separate generation. Real photos of one physical object, or renders from a single 3D scene, avoid that drift entirely.

Naming the Subject

The optional Subject field takes a short name for the object, for example "a tugboat". Leave it blank when the reference photo already shows one clean subject. Fill it in when the reference has clutter or shows more than one object, so the model knows which one to build. Every example in this article set it to a short, direct description, for example "a mid-century teak and brass side table" for the side table example below.

Matching Effort and Polycount to the Object

Modeling effort, refinement passes, and the polycount cap should track how complex the object actually is, not sit at the same setting for everything. Simple geometric shapes finish in minutes on Standard effort with a low polycount; mechanisms with small moving parts need Maximum effort, more refinement passes, and a higher polycount cap to hold on to their detail.

"a mid-century teak and brass side table". buildEffort: Standard, refineSteps: 4, faceBudget: 100,000. Open this asset in Scenario

"a cute rounded cartoon toy robot, glossy teal and yellow plastic". buildEffort: Standard, refineSteps: 4, faceBudget: 80,000. Open this asset in Scenario

"a cartoon-style toy food truck with a friendly face". buildEffort: High, refineSteps: 6, faceBudget: 150,000. Open this asset in Scenario

"a steampunk hand-cranked brass flashlight, gear mechanism, leather grip". buildEffort: Maximum, refineSteps: 8, faceBudget: 250,000. Open this asset in Scenario

The side table and toy robot are simple shapes and came back in a few minutes on Standard effort. The flashlight, with its exposed gear mechanism and braided leather grip, needed Maximum effort and 8 refinement passes to keep that detail intact.

Parameters

Five parameters control the reconstruction: what to build it from, what to call it, how hard to work at it, how many correction passes to allow, and how many triangles to deliver.

images

Required. 1 to 8 photos or renders of a single object. The first image sets the viewpoint the result is checked against; the rest help resolve the sides that first photo does not show. The scout drone example uses two views (front and rear); the diving helmet example reconstructs a full object from just one.

prompt

Optional. Name the object in a few words, for example "a tugboat". Leave it blank when the reference photo already shows a single clean subject; fill it in when the reference has clutter or shows more than one object, so the model knows which one to build. The side table example used "a mid-century teak and brass side table".

buildEffort

Standard, High, or Maximum, default Maximum. Match it to how complex the object actually is instead of leaving it maxed out for everything. The side table and toy robot examples used Standard and finished in a few minutes; the steampunk flashlight, with its exposed gear mechanism, needed Maximum to hold on to the small mechanical detail.

refineSteps

0 to 10 correction passes, default 7. The run stops on its own once the result stops improving, so a higher cap mostly matters for genuinely intricate subjects. The flashlight example used 8 passes to resolve its gears and braided leather grip; the side table needed only 4.

faceBudget

1,000 to 2,000,000 triangles, default 250,000. This is a hard cap applied at export, not a suggestion. The scout drone example was capped at 20,000 for a game-ready budget and still kept its wheels and sensor turret readable; the flashlight example used the full 250,000 to preserve its gear teeth and rivets.

Use Cases

  • Game props: build weapons, vehicles, and set dressing as editable meshes with parts already separated for material variants.

  • Product visualization: turn product photography into a 3D asset for interactive viewers or configurators.

  • Film and VFX previz: convert concept photos or physical maquettes of props into a scene-ready starting mesh.

  • E-commerce: give shoppers a rotatable 3D view of a real product straight from its catalog photos.

  • Education and museums: digitize physical objects or replicas into inspectable, part-labeled 3D models.

  • Tabletop and print-and-play: turn concept art or physical minis into printable terrain and prop kits.

Tips for Better Results

  1. Use front-and-back reference views, not two side profiles, for multi-image input. The model checks the result against the first image and needs the others to cover what that first view cannot show.

  2. Fill in the Subject field whenever the reference photo has clutter or more than one object, even though the field is optional.

  3. Match buildEffort and refineSteps to the object's real complexity. Simple geometric shapes finish in minutes on Standard; mechanisms with small moving parts benefit from Maximum and a higher refineSteps cap.

  4. Set faceBudget to your pipeline's real target up front. It is a hard export cap, so a game-ready asset and a hero-detail asset should use different values, not the same default.

  5. Shoot or generate reference photos with clean, even lighting and an uncluttered background. The model separates the object from its surroundings more reliably that way.

  6. Expect strong per-part material separation, often a couple dozen named materials, even from a single reference photo. That separation is what makes the result editable afterward.

  7. Budget extra refineSteps for fine structures like braided leather straps, thin crank handles, or rivets; they hold up well at higher settings.

Known Limitations

  • Independently generated reference photos of an imagined object can drift slightly in proportion between views, since each comes from a separate generation. Real photos of one physical object, or renders from a single 3D scene, avoid this.

  • Built for props and hard-surface subjects, not characters. Use a character-specific model for humanoid or creature subjects.

  • Surfaces never shown in any reference photo are inferred, not observed. Treat a fully hidden underside or interior as a best guess, not a measurement.

  • Maximum effort with a high refineSteps and faceBudget takes real time, on the order of 20 to 25 minutes per asset in testing. Plan batch runs accordingly.