Two concepts. Two rulers. One rule. Espresso machine and rangefinder
Never build on a drawing.
Measure something that cannot drift.
Two GPT Image 2 concepts on Scenario became real Blender models, part by part, built by agent teams under one supervisor. One team measured a generated 3D mesh. The other measured the concept image itself. Pick an object, then drag across the picture.
The shared trap
Both builds started where most do: ask the image model for front, side and top views, then model from them. Both found the views could not be trusted, in two different ways.
Espresso machine · views drawn one at a time
Every full edit moved something else, and the views disagreed in 3D: the lever's side plate faced the camera in both the front and the side view. Dropped after about 40 minutes and 620 CU.
Rangefinder · hero and six views in one image
| Cross-view check | A | B | C |
|---|---|---|---|
| Height spread | 18.3 % | 14.5 % | 18.7 % |
| Body depth spread | 5.3 % | 32.4 % | 13.0 % |
| Side views mirrored right | no | yes | no |
One image keeps the style together, but a script that measured the six views against each other found no sheet consistent. Worse, the front view placed the windows, lens and dials about 22 mm left of where the hero shows them.
Two answers
Neither team built on what a generator drew. Each found a reference that holds still, measured it, and scored every gate through a camera fitted to the concept.
Ruler A · measure a generated mesh
The concept went to three image-to-3D models on Scenario at once (770 CU, 11 minutes). The most coherent mesh was scaled so the tower stands 0.40 m, rendered in fixed orthographic views at 0.3 mm a pixel, and used as a ruler. Agents measured it and were never allowed to copy, snap to or join its geometry. Where it disagreed with the concept, the concept won.
Ruler B · measure the concept image
No mesh generator was allowed, not even as a measuring proxy. The hero image is near axonometric: both ends of the body have the same height and the verticals stay vertical. That gave the camera directly (azimuth 29.9, elevation 13.6, 12.5 px per mm), and every visible feature was back-projected onto the plane it sits on and written to one millimetre file, LAYOUT.json.
Every gate, four checks
Select a row to show that gate above.
Side by side
| Lever espresso machine | Rangefinder camera | |
|---|---|---|
| Concept | GPT Image 2 on Scenario, one prompt, 1024 × 1536 | GPT Image 2 on Scenario, hero + six views in one 3840 × 2160 image |
| The ruler | Hunyuan 3D 3.1 Pro mesh, chosen from three | The concept image, through its solved camera |
| Overlap at blockout | 0.924 | 0.973 |
| Overlap, final | 0.945 | 0.967 |
| Builders | 3 agents (body, brew, lever) + 1 independent reviewer | 4 agents (body, lens, top controls, small fittings) |
| Parts | 67 | 73 |
| Modelling cage | 102k vertices, 203k edges | 136k vertices, 276k edges |
| Clock, gates | 3 h 15 | 2 h 33 |
| Agent time | 5 h 12 | 4 h 29 |
| Scenario credits | 1,403 CU | 205 CU (+114 CU for the film) |
| Look | Cycles, nine colour samples checked against the concept | Cycles, flocked lens interior, lights hidden from transmission |
Overlap is the silhouette IoU from each concept's own fitted camera, excess counted against it. The two numbers are not a contest: the objects, the concepts and the rules differ. The rangefinder build was not allowed to use a mesh generator at all; the espresso build used one only as a ruler.
The final models
Drag to orbit, pinch or hold ⌘/Ctrl and scroll to zoom. Mesh shows the real quad edges the agents modelled, over clay; Wireframe shows them alone.
Loading the model
The renders
Every image renders the rebuilt geometry: no base mesh, no generated image.
The rangefinder film
Five camera moves rendered from the final scene, the process screens, a Gemini 3.8 TTS narration and an ElevenLabs Music v2.5 bed, all on Scenario.
Make your own
The making-of walks through both builds with every gate and number. The agent skill gives your agent the same route through Scenario MCP for a concept of your own.