What We Learned Reviewing 50 Historical SketchUp AI Renders

AI rendering claims are easy to write and hard to evaluate. “Preserves geometry” can mean anything from keeping the general room type to retaining the exact camera, openings, built-ins and object inventory.

We assembled 50 historical Maquete source-and-output pairs to see what a real production archive could tell us. The review was useful, but it also showed why historical work should not be presented as a controlled benchmark.

The short version: the set provides credible evidence of the recurring strengths and failure modes in source-aware rendering. It does not produce a defensible current-pipeline pass rate.

Executive summary

• The review used 50 existing SketchUp source views and their historical outputs. No images were generated for this report.
• Camera direction and the main architectural composition were often the most stable parts of an output.
• Small fixtures, loose furniture, material tone, pattern scale and reflective detail were more likely to vary.
• A visually strong image could still contain a meaningful design change.
• Close views exposed errors that were easy to miss in wide compositions.
• Because the outputs span different dates, workflows and pipeline versions, we are not publishing a single geometry-preservation percentage.
• The first-pass review was AI-assisted and has not been certified by an independent reviewer or by an architect.

What we reviewed

The eligible internal pool contained 383 completed renders whose recorded source application was SketchUp and whose source views belonged to Maquete co-founder and licensed architect Laura Consoni.

Repeated exports can make a production archive look more varied than it is. We therefore compared low-resolution structural image hashes and treated very close matches as near-duplicates. This reduced the pool from 383 source views to 55 structurally distinct views.

We then selected 50 using farthest-point sampling over those hashes. In plain language, each new selection was chosen because its visual structure was relatively different from the views already selected. That step reduced the chance of filling the review with consecutive exports of the same camera.

The resulting sample included interiors, close material and fixture views, and exteriors across Guided, Match View and other historical workflows. The source and shipped output for every fixture were placed side by side for visual inspection.

Why this is not a benchmark

A controlled benchmark would rerun every source through one fixed production pipeline, retain every first attempt, record each regeneration, and review all outputs using a predeclared scoring rubric.

This review does not do that. Its outputs were generated between late April and late July 2026 across different workflows and evolving versions of Maquete. Some stored automated scores were created by historical validators, but those scores were not used as human ground truth here.

We also did not reconstruct a complete attempt history for every fixture. That means the archive cannot reliably answer:

• the current first-attempt geometry pass rate;
• the current regeneration rate;
• mean attempts per accepted image;
• validator false-pass or false-reject rates;
• whether a later pipeline version would improve an older result.

Those questions require a controlled run. Publishing percentages from this mixed archive would imply a level of comparability the data does not have.

The review criteria

We looked at six areas:

1. Camera — viewpoint, crop, horizon and field of view.
2. Architectural geometry — walls, floors, ceilings, beams, openings, stairs, built-ins and fixed cabinetry.
3. Object inventory — scene-defining furniture, fittings and fixtures.
4. Materials — family, placement, tone, scale, texture and gloss.
5. Realism — whether the image reads coherently at normal presentation size.
6. Visible ambiguity — source elements that were difficult to interpret from the flattened viewport.

The SketchUp viewport was the comparison source. We did not inspect the underlying 3D model, construction drawings or as-built project, so this review cannot establish dimensional accuracy.

Finding 1: camera and main composition can survive while details change

Across many pairs, the output retained the source camera direction and the main arrangement of the room. A kitchen still read as the same kitchen; an exterior kept its principal massing; a joinery wall stayed in the same part of the frame.

That stability matters, but it can create false confidence. Once the overall image looks familiar, smaller changes become easier to overlook:

• a tap changes type;
• a chair disappears or shifts;
• cabinet divisions simplify;
• a pendant changes shape;
• a window or curtain gains a different proportion.

The practical lesson is to review the envelope and inventory separately. “It looks like the same room” is not a sufficient acceptance test.

Finding 2: simple close views are revealing

Several fixtures isolated sinks, worktops, taps or small areas of joinery. Their camera and dominant geometry were comparatively easy to compare because the source contained fewer competing elements.

They also made local errors conspicuous. A changed tap configuration, basin edge, countertop thickness or joint position occupies a larger share of the image in a close view.

Wide views are useful for atmosphere and overall composition. Close views are better tests of material assignment and fixture fidelity. A project presentation needs both.

Finding 3: material family is only the first level of consistency

An output may keep “timber” as timber and “stone” as stone while still changing the design intent.

The review repeatedly highlighted subtler questions:

• Did pale oak become orange or grey?
• Did a honed surface become glossy?
• Did fine stone movement become high-contrast veining?
• Did tile or plank scale change?
• Did vertical grain become horizontal?
• Did the material spread onto a surface where it did not belong?

This is why a useful material brief includes surface, location, family, colour, finish, scale and direction. See the step-by-step guide to keeping materials consistent across multiple AI render angles.

Finding 4: exteriors preserve a hierarchy, not every contextual fact

In the exterior pairs, the principal building mass often remained recognisable while landscaping, background buildings, cars, ground texture and sky changed more freely.

That may be acceptable for an early mood image. It is not acceptable when a neighbouring boundary, protected tree, access route or site level is a project fact.

The intended use must therefore be decided before rendering:

• Concept image: contextual variation may be tolerable.
• Client design review: building and site relationships need closer checking.
• Planning or sales material: verified context should come from authoritative survey, model and project information.

Photorealism does not turn invented context into evidence.

Finding 5: source ambiguity becomes output interpretation

The renderer sees a flattened image. Dark planes, missing edges, overlapping objects and unclear openings can have several plausible interpretations.

In the reviewed set, simple and clearly bounded source geometry was generally easier to compare. Visually dense scenes offered more opportunities for objects or edges to be simplified, combined or reinterpreted.

When several attempts misunderstand the same element, strengthen the source rather than extending the prompt indefinitely:

• separate adjacent surfaces with distinct colours;
• expose the edge of an opening;
• remove temporary geometry;
• avoid overlapping silhouettes;
• save the exact camera as a SketchUp Scene;
• make important fixed elements legible.

The SketchUp model pre-flight checklist covers this preparation in detail.

Finding 6: realism and fidelity are different scores

Many historical outputs looked substantially more photographic than their SketchUp sources. That does not automatically make them faithful.

A render can be:

• realistic and faithful;
• realistic but structurally wrong;
• faithful but visually underdeveloped;
• both unrealistic and inaccurate.

For architectural review, fidelity comes first. Check the camera, envelope, openings, fixed joinery and scene-defining objects before judging atmosphere, styling or polish.

What the review supports today

Limitations and conflicts

This is a Maquete-run review of Maquete outputs, not an independent study. The eligible source cohort belongs to a company co-founder and is not representative of every customer, building type or modelling style.

The sample-selection method improves visual diversity, but 50 of the 55 structurally distinct views still come from a limited body of work. Some related rooms and projects appear from different cameras.

The first-pass visual review was AI-assisted. It has not been independently reproduced, and Laura Consoni has not yet certified the fixture-level findings as an architect. Client-sensitive source/output pairs remain internal; the public Apartment LNC case study shows an approved seven-view project set without exposing private project records.

Most importantly, this report describes historical production evidence. It should not be read as a measurement of the current pipeline.

What we would need for a controlled benchmark

A future benchmark would:

1. freeze one production pipeline and record its versions;
2. rerun all 50 fixtures with their stored briefs;
3. keep every first attempt and regenerated result;
4. apply the same camera, geometry, inventory, material and realism rubric;
5. have the results reviewed independently, with an architect reviewing architectural cases;
6. publish counts, not only percentages;
7. include successes, failures and ambiguous cases;
8. release a privacy-safe result table that reproduces every headline number.

We have not run that study. It would incur generation cost and should be undertaken only when the question justifies the expense.

The useful conclusion

The historical archive is valuable because it replaces a vague promise with inspectable lessons:

• preserve and compare the source camera;
• review fixed architecture before atmosphere;
• treat object inventory as a separate fidelity problem;
• describe material placement and finish, not only material names;
• use close views to expose local drift;
• keep drawings, schedules and professional review authoritative.

Showing limitations is not an admission that the product is uniquely unreliable. It is an accurate description of generative architectural rendering—and a more useful basis for trust than an absolute guarantee.

Continue reading: Geometry preservation in AI rendering, why AI renders hallucinate furniture, and the Apartment LNC multi-view case study.