CAD Scene research method
How We Test Camera and Geometry Retention
A source-guided render should keep the view you chose. This method tests that claim without confusing a convincing image with an accurate one.
Published . Method owner: CAD Scene. Responsible reviewer: Vladimir Mindru, Principal Architect at Yellow Office architecture.
Fixed test inputThe short answer
Hold the source, prompt, Composition, model, settings, and output size fixed. Generate repeated outputs. Score camera, perspective, framing, and visible structure separately from restyle quality. One aligned output proves only that run.
What does render retention mean?
Geometry retention
The output keeps the source camera, horizon, vanishing points, crop, massing, openings, repeated elements, and object positions. Materials and light may change. Structural placement should not.
Restyle quality
The output also needs to realize the requested materials, light, atmosphere, and photographic finish. A near-copy of the clay source cannot receive a strong overall verdict.
A pixel-difference score cannot separate intended material changes from unwanted geometry drift. We compare the source and output on these two axes, then inspect the images directly.
The fixed fixture
The example input stresses two-point perspective, repeated windows, thin balcony slabs, and long vanishing lines. GPT Image 2 generated the synthetic clay scene. The accepted image stays fixed for every condition.
- Style
- Warm contemporary residential architecture
- Scene
- Six-story apartment block on a quiet urban street
- Lighting
- Late-afternoon sun with long soft shadows
- Materials
- Red brick, bronze frames, glass, concrete
- Entourage
- Pedestrians, street trees, parked bicycles
The sample
The current harness has six synthetic fixtures. Four test Enhance. Two test Edit. Optional reference-image conditions expose another source of drift.
Fixtures cover exteriors, an interior, a complex roofline, an aerial tower view, and two masked edits. They stress repeated openings, perspective, roof edges, mullions, and untouched regions.
Synthetic scenes isolate the method and avoid customer data. They do not replace a later benchmark on comparable CAD and BIM exports.
Run the same scene through six checks
- 01
Freeze the input
Use one accepted source image. Do not crop, resize, or clean it between conditions.
- 02
Freeze the condition
Keep the Composition, user prompt, model, native settings, reference state, and output size unchanged.
- 03
Repeat the render
Run five outputs per test cell by default. One favorable image is not a stable result.
- 04
Score two axes
Judge geometry retention and restyle quality separately. Geometry cannot win by leaving the source unfinished.
- 05
Check data integrity
Require at least 80% scored coverage. Missing models or failed cells block a verdict.
- 06
Review the contact sheet
An architect checks every source and output together before any prompt or product claim changes.
The 1–5 scoring rubric
An automated multimodal judge records both scores and a short note. Out-of-range scores fail the cell. Human review remains the final check.
| Score | Geometry | Restyle |
|---|---|---|
| 1 | Camera, perspective, crop, or structural positions clearly move. | The requested finish is absent, off-spec, or damages the scene. |
| 3 | The broad view holds, but visible edges, openings, or proportions drift. | The direction is present, with uneven materials, light, or blending. |
| 5 | Viewpoint, horizon, vanishing points, crop, and visible structure match. | The requested finish is complete and coherent without hiding the source. |
What can this method prove today?
It proves that CAD Scene has a fixed, inspectable way to test a prompt change. The pipeline has passed a live smoke test. The full five-run, multi-model evaluation is still pending.
No benchmark result is published on this page.
There is no accuracy percentage, winning model, competitor score, or universal geometry claim. Results will remain separate until comparable outputs, complete coverage, and architectural review exist.
Limits
- Synthetic fixtures do not represent every CAD or BIM export.
- Ordinal scores are review aids, not dimensional measurements.
- A multimodal judge can miss an architectural error.
- Enhance and Edit need separate interpretation.
- Model and provider behavior can change after the review date.
- A rendered image is not construction verification.
For workflow context, start with the screenshot-to-render guide. Prepare the input with the CAD screenshot capture checklist. Revit users can compare the browser and native-plugin paths. Product operation remains in the Enhance Help article.