Describe the subject, its structure, its materials and how large it really is. GPT-6 Astra builds a textured mesh, and the preview lets you orbit the GLB before you download it.
GPT-6 Astra · text to 3D
Text to 3D on Voor AI takes a written brief and returns a textured model rather than a picture. GPT-6 Astra is the model behind this route: describe the volumes, the materials and the room, declare how large the subject really is, and download a GLB you can orbit here, open in Blender, or drop straight into an engine. The spacecraft below is the published example for this model. It comes out as a closed hangar, so the viewer loads the same meshes with the roof unlinked and looks down into the room — otherwise every camera angle outside the building shows nothing but a wall. The link under the viewer is the complete file with the roof left on.
Prompt anatomy
Every layer in a text to 3D brief removes a guess the model would otherwise make for you. The order below matches how the published example was written, and each layer changes a different part of the file you get back — which is why editing the materials line alone will not fix a silhouette you dislike.
Name one thing and its job. “A compact uncrewed exploration spacecraft on three landing legs” gives the model a single object to close; two nouns of equal weight usually produce two half-objects fused together.
Walk the volumes in order — fuselage, wings, canopy, nozzles, panels, landing gear. Each clause you add is a decision the model no longer has to guess, and that is what keeps the hidden side consistent with the side you described.
Say the finish per part, not globally: brushed metal, off-white ceramic, dark rubber, smoked glass. Text to 3D bakes colour and roughness into the mesh, so a material you never mention arrives as generic grey.
A subject alone returns a floating prop. Naming the room — ribs, ceiling, gantry, carts, a closed door — makes the generator build a staged scene around it, scaled to the value you set in subject size.
Close the brief with what must not appear: no people, no weapons, no logos, no readable text. Negative clauses are the cheapest way to stop the model decorating a clean asset with plausible but unwanted detail.
A compact uncrewed exploration spacecraft resting on three landing legs inside a maintenance hangar. Build a coherent rounded fuselage, short swept wings, a sealed smoked-glass cockpit canopy, recessed engine nozzles, thin ceramic hull panels with visible seams, and landing gear firmly supporting the craft on the floor. Add a maintenance gantry, two equipment carts, organized cables and tool cases. Surround the craft with structural ribs, a high ceiling and a large closed hangar door. Use brushed metal, off-white ceramic, dark rubber and glass, with subtle surface wear and contrasting orange service panels. Keep the whole spacecraft visible and leave room around it for inspection. No people, weapons, brands, logos or readable text.
Subject size
Subject size is the longest dimension of the finished subject in metres, and it is the one setting on this route that changes the content of the output rather than its polish. A hangar built around a 12-metre craft has to invent more surrounding structure than a mug built at 0.1 metres, so the same brief at two sizes returns two different scenes. Move the slider before you generate and the credit estimate stays exactly where it is.
Smallest allowed subject. Detail is inferred, not measured.
Small props with a closed shell and one moving part.
Furniture and tools that need a believable footprint.
Vehicles, where wheel arches and glass area carry the read.
The published example on this page — a craft plus its staging.
Largest allowed subject. Expect architecture, not interiors.
What lands in your downloads
Text to 3D returns a single binary glTF file. Geometry, UV maps, textures and material assignments travel together, so there is no texture folder to relink and no axis conversion waiting on the other side. The numbers below are read from the published example file on this page, not from a specification sheet.
Import the GLB and every textured part arrives as its own object. Decimate for a game build, re-UV for a print pipeline, or keep the topology and use the file as a block-out for a hero asset. Named meshes make the selection step short: 92 of them in the published example instead of one welded blob.
The viewer at the top of this page is the same GLB a download gives you. Drop the file into a web page with a glTF viewer, into a product configurator, or into an AR session — no render farm, no lighting rig, because the shading information is baked into the textures.
Honest limits
None of these are reasons to avoid the route; they are reasons to write the exclusions line and to keep a second generation in the budget when the asset has to match something real. When the object does exist and its proportions matter, photograph it and switch to the sibling route instead — the comparison below is the short version of that decision.
The generator on this page is already set to GPT-6 Astra text to 3D; there is nothing to switch before you start.
Image to 3D accepts up to ten references in one run, so a three-quarter front and a rear view usually beat one perfect shot.
Both routes return the same GLB format and both sit on the 3D models page, so a finished asset can move between them without a conversion step. If you are building the reference first, generate the still with text to image and feed that file into image to 3D.
Text to 3D turns a written description into a textured three-dimensional model instead of a picture. On this page the prompt is the only required input: you describe the subject, its structure, its materials and its surroundings, and the result is a real mesh file you can orbit in the preview and download. GPT-6 Astra is the text to 3D model behind this route, and it returns GLB rather than a flat render.
One text to 3D generation costs 2,787 credits here, however long the prompt is. GPT-6 Astra is priced per run rather than per mesh or per polygon count, so the subject size field and the number of words you write do not change the price. The generator shows the same estimate next to the Create button before anything is charged.
A single binary glTF file with the extension .glb. It carries the geometry, the UVs, the baked textures and the material assignments in one container, which is why Blender, Unity, Unreal, Godot, three.js, model-viewer and macOS Preview all open it without a conversion step. The published example on this page is 2.7 MB across 92 meshes and 27 textures, and because the model closed a roof over the scene, the viewer removes that roof so the craft inside is visible — the download link under the viewer is the complete file with the roof left on.
The prompt field accepts up to 4,000 characters, and the model reads the whole thing. The examples that come back most usable name one subject, describe its main volumes in order, list the materials, then describe the surround and finish with explicit exclusions. A treated prompt of 80 to 150 words is usually better than a one-line noun, because unseen sides of the object have to be inferred from something.
Subject size sets the longest dimension of the finished subject in metres, from 0.05 to 20. It is a scale statement, not a quality dial: a 12-metre brief describes a hangar-sized scene, while 0.2 metres describes something you could hold. Telling the model the real size changes how much surrounding detail it invents, because a scene has to be proportioned around a known subject.
Closed, solid objects with a recognisable silhouette: appliances, tools, furniture, vehicles, containers, props, modular architecture, and staged interiors. The model is strongest when the prompt describes surfaces it can build in one pass, such as brushed metal, painted ceramic, rubber, glass and fabric. Quiet, evenly lit presentation also helps, because the textures are baked into the mesh rather than lit by a scene.
Four places show up in our runs. Very thin geometry such as wires, chain and fine lattice can arrive fused. Openings that depend on exact measurements, like a specific bolt pattern, are approximated rather than measured. Readable lettering and logos on a surface come back as texture noise, not text. And anything that only exists as a photograph — a specific person, a real storefront — is invented rather than reconstructed.
Yes, but the sensible split is to refine the brief rather than the mesh. Re-running with a sharper description of the part that disappointed you costs one more generation and keeps the file clean; pulling a GLB into Blender to rebuild the geometry usually costs more time than the rerun. Download the GLB, inspect it in a viewer, and change the prompt only for the elements you can name precisely.
Use image to 3D when the object already exists and its proportions matter, because photographs pin the silhouette and the material placement to something real. Stay with text to 3D when the object does not exist yet, when you are exploring a concept, or when you need a whole scene around the subject rather than a single prop. Both routes share the 3D picker, so you can compare one prompt against one photo without leaving Voor AI.
Name the subject and its volumes, say the materials per part, declare the real size in metres, and close with what must stay out. The credit estimate appears before GPT-6 Astra starts building.
Generate a 3D modelOptional cookies help us understand usage and measure ads. Essential cookies stay on so Voor works.