We measure whether AI staging tools move the walls

StagingBench Team — Wed, 04 Nov 2026 12:34:34 +0000

A virtual staging image can look convincing while changing the room it is supposed to show. Furniture attracts attention, but the useful evaluation question often sits behind it: did the walls, windows and doors stay where they were?

StagingBench is an independent benchmark of AI virtual staging tools built around that question. We use the same empty-room photos, select Modern style and keep the first output. Method v0.2 measures architecture drift after alignment. Lower drift means the room's architecture stayed put more closely.

The number needs context. Test settings, framing failures, alignment failures and the compare images all affect what a reader can conclude. Style Smarter's room-type pass provides a concrete example of reading those pieces together.

Keep the input and selection rule consistent

Using the same photos gives the comparison a common starting point. Selecting the same style also keeps the requested direction consistent. Neither choice makes every tool identical, but both make the test conditions easier to describe and inspect.

Keeping the first output defines another part of the experiment. If a tester generates repeatedly and publishes a favourite, the displayed result includes the tester's selection. Our published output records the first attempt under the stated settings.

This rule does not predict what a user could achieve after further attempts. It gives the benchmark a clear boundary. Someone reading the result knows which attempt the evidence represents, without needing to guess how many discarded outputs came before it.

What architecture drift tells you

Architecture drift asks how closely the output preserves the room's structure after alignment. Walls, windows and doors are the reference points that matter to this question. A lower value indicates closer preservation under the method.

The methodology explains method v0.2. When quoting a result, keep that method version attached. A drift value by itself should not be treated as a percentage of correctness, a guarantee of suitability for a listing or a complete judgment of the image.

Framing is another part of the evidence. A changed view can remove information from the original photograph. Alignment failures also deserve explicit attention: when a comparison does not align, readers need to see that limitation alongside the numerical summary.

Style Smarter: the settings behind the result

The Style Smarter room-type pass is dated 3 November 2026. It covers all 10 test rooms. We used Modern style, set the room type per photo, kept the first output and used the Pro download.

The room-type selections were:

Living: rooms 01, 02, 05, 07 and 10.
Bedroom: rooms 04 and 06.
Office: rooms 03 and 09.
Kitchen: room 08.

These details define the run. They help someone distinguish a result produced with room-specific settings from a result generated under different conditions. Any reuse of the numbers should preserve that context.

Style Smarter's founder provided the Pro account free. Stating that alongside the settings makes the access arrangement visible to anyone assessing the test.

Read both means with their coverage

Mean architecture drift was 20.5 across all 10 rooms and 21.5 across the 9 that aligned. Those are the reported summaries for this pass. Keep each value paired with its coverage when citing it.

The room-level notes add essential information: rooms 04 and 06 failed framing. Room 09 failed alignment, so its framing is unknown.

Unknown framing is an unresolved result. It cannot support a claim that the room passed framing, and it should not silently become a framing failure either. The recorded distinction matters when someone copies the data into a comparison, report or agent response.

The same care applies to the means. Quoting 20.5 without its coverage, or presenting 21.5 as though every room aligned, loses information the public result supplies. Readers should be able to trace a number back to the precise set it describes.

Inspect what the summary cannot settle

The Style Smarter result gives a practical reading sequence: review the settings, note the framing and alignment outcomes, and inspect the comparisons before deciding how much weight to give the mean.

Other published examples show why this matters. In room 05, Collov AI passed framing and had low drift, but removed the two steps to the bay. Pedra kept them. That is a change the automatic score misses. The images are available for Collov and Pedra.

Room 01 supplies two further checks. virtualstaging.ai's comparison shows a moved column and a changed soffit, with a framing failure. TINTY's comparison has a crop and framing failure, retaining roughly 89% of the frame.

These examples explain the limits of treating a summary as the whole evaluation. A small architectural detail can carry information about the physical space, while a crop can remove information from view.

Reuse the data without losing its meaning

The published results.json is available under CC BY 4.0. Builders and agents can use the numbers with attribution. A useful citation names StagingBench Team, links the data and methodology, and identifies the tool and tested pass.

Preserve failure notes and coverage when building derived comparisons. A tool tested across 10 rooms does not, by itself, establish that every tool or the entire benchmark edition is complete. Read coverage at the tool level and avoid inventing an overall ranking from partial evidence.

For an agent-generated answer, that means carrying qualifications into the answer itself. The distinction between all-room and aligned-room means should remain visible. So should a room whose framing is unknown.

Access, pricing and disclosure

The Style Smarter result above used Pro. Its current Free plan offers 5 renders per month at 1K, with no card required. Our Free room-01 PNG had no watermark. Paid Basic is $19 per month for 60 renders at 1K. Those plan details describe access; they should not be substituted for the settings of the measured Pro pass.

Some tools on StagingBench have affiliate links, disclosed on their individual pages and the disclosure page. Style Smarter has no affiliate relationship with StagingBench. Scores are never paid for. The complimentary Pro account remains a separate disclosure attached to this test.

A useful benchmark leaves the reader able to inspect its claim. For AI staging, that means keeping the original room in view, preserving the conditions behind each result and giving failures as much visibility as the average.

DEV Community: StagingBench Team