GPT-6 Astra users build editable Blender models from photos, plans, and scans

• Editable Blender scenes are the common thread in the early Astra demos. LinusEkenstam reported a claimed centimeter-accurate studio from rough ideas and a few photos in 11 minutes LinusEkenstam's studio-model post.
• The strongest reference packets combine measurements with imagery: LinusEkenstam supplied tape-measure figures, five photos and two panoramas LinusEkenstam's input breakdown, while LukeW used a 38-page plan-and-render PDF LukeW's cabin walkthrough.
• Photogrammetry is becoming an editable starting point rather than a final asset. bilawalsidhu used about 400 photos, a RealityScan mesh and Blender to clean a scanned home into a manipulable scene bilawalsidhu's editable-scene workflow.
• The physical version is already showing up in creator posts: LinusEkenstam said Astra used photos, measurements and a small manual drawing to make an STL replacement part, then sent it to a printer LinusEkenstam's replacement-part report.

OpenAI says on its GPT-6 Astra launch page that access is rolling out from a limited set of organizations to ChatGPT paid tiers and API platforms. An independent Blender starter guide publishes editable files and Python for Astra-assisted projects without requiring a Blender MCP connection. A separate published game workflow describes connecting Codex to Blender MCP, generating art references, then iterating against in-game screenshots at 60fps.

LinusEkenstam’s studio test is the clearest early example of image-plus-measurement reconstruction. He calls the result centimeter accurate, based on rough ideas and a handful of photos.

He later detailed the input packet in LinusEkenstam's follow-up:

• Floor-tile size, total room size, and the high and low ceiling points.
• Five photos, including two panoramas.
• IKEA kitchen identification and drawer-front names, which Astra used to fetch a matching texture image from IKEA’s site.

Floor plans with dimensions make the result exact, LinusEkenstam said in a later reply. He also described the agent switching between computer use and Blender MCP LinusEkenstam's MCP reply, rather than relying on one control path.

The final photoreal polish was separate from the scene construction: LinusEkenstam said the Cycles render passed through GPT Image 2.0 as a last step in LinusEkenstam's rendering reply. He plans to release the Blender file alongside a website Astra built for annotations and feedback LinusEkenstam's review-loop reply. The studio and bookshelf experiments used 2% of his monthly Codex allowance, by his count LinusEkenstam's usage estimate.

LukeW reported feeding Astra a 38-page PDF of cabin blueprints and renders, then getting a to-scale interior-and-exterior walkthrough in about 10 minutes.

His compact version of the request was simply, “these are the plans for a cabin I am building, make me a 3D model I can do a walk through with,” in LukeW's prompt example. He later described the delivered web artifact as 21 MB of HTML plus a 6 MB GLB file LukeW's file-size reply.

Higgsfield described a second architecture pipeline: one floor plan becomes scene code, then Blender builds and Cycles renders the location model.

The same account reported rebuilding New York in 3D in 20 minutes higgsfield_ai's New York demo and a one-shot Blender villa tour higgsfield_ai's villa post. These posts give useful output references, but none publishes a dimensional comparison against the underlying building data.

bilawalsidhu began with a photogrammetry model of a drawing room and asked Astra to rebuild it in Blender using the scan’s own textures and procedural shaders.

A later version of the experiment lays out the chain more fully.

1. About 400 2D photos, captured with a DSLR according to bilawalsidhu's capture reply, become a RealityScan photogrammetry mesh.
2. The mesh enters Blender as the classical reconstruction layer.
3. Astra cleans the scene into editable objects, retaining scan textures and creating procedural shaders where needed.
4. A critic compares the result with multiview imagery. bilawalsidhu described two loops of roughly 45 minutes in bilawalsidhu's critic-loop post, and later put the pace at about two rounds an hour bilawalsidhu's timing reply.

The source asset was an FBX with an 8K texture; bilawalsidhu said Astra initially used its diffuse map, with a normal-map pass planned for higher-frequency detail bilawalsidhu's texture reply. Close-up reference photos improve the result, he added in bilawalsidhu's detail-photo reply.

The key limitation is visible in the workflow itself. bilawalsidhu said the mesh is currently Astra’s ground truth, so errors in the scan can survive into the editable scene; he is testing whether querying the original imagery can narrow that gap bilawalsidhu's ground-truth reply.

LinusEkenstam’s replacement-part story is the most complete photos-to-fabrication chain in the evidence pool.

His reported sequence was:

1. Supply photos, measurements and the problem description for a discontinued toilet part.
2. Let Astra locate a manual with a small 2D drawing and combine it with the supplied references.
3. Generate the model in under two minutes.
4. Export an STL, send it to a printer in Sweden, and collect the printed part about 1.5 hours later.

LinusEkenstam said Astra works directly through Blender or Rhino and that his MCP setup can test tolerances and assembly order LinusEkenstam's manufacturing reply. He labeled the photo-to-print pattern “P2S” in LinusEkenstam's short reply.

A single photo also became a real-time 3D bookshelf, searchable and linked to Goodreads in LinusEkenstam’s demo.

He said the result was one-shot, with no follow-up conversation, in LinusEkenstam's one-shot reply. The linked Book Snapper app describes the less theatrical product layer: camera recognition of titles and authors, plus a personal library.

LinusEkenstam proposed a sturdier cataloging stack in a bookshelf follow-up: local OCR and ISBN lookup for identification, then 3D representations of each book’s cover, back and spine, enriched with edition and collection data.

MengTo reported using Astra to build a Three.js game, recreate its scenes in Blender and generate the trailer. He said the rocks, trees, monuments and characters were procedural code, keeping the game under 2 MB before image and music assets, while detail work still needed repeated pushing.

petergyang’s Astra, Blender and Godot tutorial covers four distinct prototypes:

• A Star Fox-style space shooter.
• A first-person shooter on a moving train.
• A StarCraft-style RTS level.
• A roguelike deck builder about AI founders.

MengTo separated the hard part from the asset pass: games are hard, while Blender is comparatively easy because it is not interactive MengTo's game-design reply. In the same project, he said all assets were generated rather than taken from a package MengTo's asset-source reply.

Output constraints matter. AIandDesign said an unconstrained attempt produced 500,000 polygons, then reported a version capped at 10,000 polygons with no visible loss of fidelity in AIandDesign's polygon-budget reply. Other posts show that same creator building a game environment from one image AIandDesign's one-image scene and adding slot machines to its first level AIandDesign's game-scene post. Quality can still vary within a run, om_patel5 observed, with strong reference-led assets and weaker structural details in the same post-apocalyptic game scene om_patel5's game analysis.

The examples now extend past architecture and game environments. higgsfield_ai showed a Blender camera that separates into 1,025 modeled objects.

• A robot built in a Blender session was presented as ready to animate by higgsfield_ai's robot post, while higgsfield_ai's Z3 recreation rebuilt Konrad Zuse’s programmable computer.
• higgsfield_ai also reported turning its office into an explorable Roblox environment higgsfield_ai's office-tour post.
• koldo2k asked Astra to build Diagon Alley in Blender and walk through the process with camera movement koldo2k's Diagon Alley post. JameslabiQ posted a Blender recreation of Craig Federighi JameslabiQ's recreation post, and gokayfem used Astra to show both the original and current Colosseum gokayfem's Colosseum post.
• A forensic-reconstruction prompt published by higgsfield_ai asks for open-source skull references, detailed Blender modeling, face reconstruction, then a Seedream photoreal render higgsfield_ai's skull workflow. The same account called its Houdini experiment an initial agentic-VFX test higgsfield_ai's Houdini post.