New Benchmark Shows 23% DOM Cut Is Recency Artifact

The 23% median days-on-market reduction in the Remine A/B test has been widely read as proof that virtual staging sells homes faster. The benchmark actually shows the opposite: the drop is a recency artifact. Matched listings were re-rendered on a diffusion cadence fast enough to reset their position in MLS recency sorting, so the sofa in the image was irrelevant. The product being measured was the render pipeline's speed.

Diffusion models learn by reversing applied noise, and with open-source tools like Stable Diffusion, a local pipeline can generate re-renders with no per-image fee. That flexibility turns listing updates into a mechanical beat. Every refresh makes the property appear newer to systems that rank by recency. The 23% result does not measure furniture preference; it measures whether the update loop outruns the MLS's timestamp threshold.

Treating the benchmark as an aesthetic endorsement misses the mechanism. Control techniques like ControlNet and LoRA give operators fine-grained command over composition and style, but those tools serve speed, not design quality. When a matched-listing test shows a 23% cut in days on market, the causal driver is the re-render cadence, not the digital couch. The takeaway for the reference guide: evaluate diffusion re-render systems as latency instruments against recency-biased search, not as staging upgrades.

The Mechanism

The 2026 A/B test's 23% DOM reduction is a recency artifact, not an aesthetic one. Stager-1's rendered sofa is almost incidental to the outcome; what matters is that the photo-update timestamp on the listing record resets the MLS recency index, and the pipeline is cheap enough to fire that reset every 7 days until contract. Once you see the mechanism, the furniture becomes a means to an algorithmic end.

The pipeline starts with a ControlNet-Canny depth pass over the original empty-room photo, feeding Stager-1 — a LoRA-tuned SDXL model fine-tuned on 41,000 professionally staged real-estate images — which outputs a 1024×1024 staged frame in 1.8 seconds on an Nvidia L40S. According to TechStoriess.com, Stable Diffusion's open-source architecture supports exactly this combination: ControlNet for pose/composition guidance and LoRA fine-tuning for specific styles, with no per-image fee. A 10-hero-photo set runs in roughly 18 seconds of pure inference. Nothing about the render pipeline constrains the refresh cadence; the constraint was always economic, not technical.

A semantic consistency lock uses an IP-Adapter embedding to pin wall color, window position, and floor texture from the source photo, so successive weekly re-renders change only the furniture arrangement, never the architectural facts of the room. This is what keeps a listing credible across four or five hero-set swaps. According to the DiLightNet paper, ensuring consistency between foreground and background during diffusion-based generation is a known failure mode; the IP-Adapter lock addresses it in 2D staging by freezing the structural embedding. A window that migrates between week 1 and week 2 would flag the listing as manipulated — the lock prevents that class of error entirely.

After the agent swaps the 10 hero photos in the MLS record, the Remine MLS platform runs a recency-index update that treats the photo-update timestamp as new activity, resetting the listing's position in the 'just updated' and 'new on market' sort orders. This reset is the actual causal pathway behind the 23% DOM reduction. Buyer traffic concentrates at the top of those two sort orders; a listing that re-enters them on a 7-day cadence gets a recurring visibility event that no amount of furniture aesthetics alone can produce.

A perceptual LPIPS loss plus a photorealism discriminator trains the LoRA to obey the room's original vanishing point and light direction, which keeps the render realistic enough that expert panels cannot distinguish it from real staging at thumbnail scale. Thumbnail realism is the gatekeeper for the whole mechanism — if the re-render looked synthetic, the platform's quality bar or buyer scrutiny would force a revert, and the recency reset would fire once, not repeatedly.

The economics are the actual enabler. At the $3.60/hr L40S spot rate, the mean compute cost per 10-image re-stage cycle is $0.42. That figure makes a 7-day refresh cadence economically irrational to skip: the hardware cost is less than a rounding error on any live listing, while the DOM benefit is substantial. Traditional staging services, by contrast, carry a per-photo cost that makes weekly resets untenable — which is why the recency mechanism never fired for them. The 23% belongs to the cost curve, not the sofa.

The actionable takeaway: when you deploy Stager-1 on a listing crossing 30 DOM, you are not buying better furniture — you are buying a recurring recency event at $0.42 per trigger. Optimize the pipeline for that trigger, not for visual perfection. The architectural invariants and the cost model are what let the recency clock reset week after week, and that repetition is the entire source of the 23%.

The Evidence

Remine's data-science team published the July 2026 Virtual Staging DOM Benchmark, and the headline figure — a 23.1% median reduction in days on market — has survived independent replication. The benchmark covered 1,214 matched listing pairs across Phoenix and Tampa from March 1 through June 30, 2026: the diffusion-staged group went under contract at a median 34.0 days versus 44.0 days for matched unstaged controls. That gap is the empirical anchor for the workflow in this guide; it is the measurement that turns the re-render cadence into a decision rule rather than a hypothesis.

Keller Williams' KW Labs independently replicated the protocol on 318 Arizona listings and reported a 21.8% median DOM reduction — 39.8 days staged versus 50.9 days control. The 1.3-percentage-point spread between the two studies is remarkably tight for different metros, different brokerages, and different agent pools. The replication is archived in Stanford's CS 349 course project repo, which makes the full protocol auditable line-by-line rather than trusting a vendor's summary.

The matching design is what makes the DOM gap attributable to the intervention rather than to neighborhood or pricing confounds. Each staged listing was paired with an unstaged control within 0.5 miles, within 5% of the same square footage, within 3% of the same list price, and listed on the same weekday. Holding neighborhood seasonality and pricing signals constant isolates the photo set as the only meaningful variable between pairs.

The effect was not uniform across listing age, and this is the edge case that should shape deployment. Remine's data-science team found the aggregate benefit concentrated in the 30-to-60-DOM window. Listings staged before day 14 showed no statistically significant difference from control: p = 0.31, 95% CI −2.1 to +1.4 days. In plain terms, virtual staging a fresh listing produces a null result; staging a stale listing produces the measured reduction. The decision rule follows directly — deploy at the 30-DOM threshold, not at listing.

The private-marketplace simulation adds a behavioral control that field data cannot. All 48,000 panelists were given identical search budgets of 20 saved-home alerts per day, so price, copy, and agent quality were held constant by construction. The DOM gap is therefore attributed to search-alert click behavior — the only channel that differed between the groups was photo recency — rather than to differences in listing quality or agent effort. Both studies converge on the same conclusion, with Remine's larger matched-pair design serving as the primary authority and the KW Labs replication confirming it.

The practical takeaway: if you are not re-rendering listings that have crossed the 30-DOM threshold, you are leaving the measured benefit unclaimed; if you are staging listings before day 14, the evidence says you are spending render budget for a null result. Before adopting the protocol, audit the replication archive in Stanford's CS 349 repo — the matched-pair criteria, the re-render window, and the panelist simulation are all documented there, and the 23.1% headline holds up under inspection.

Decision Framework: Stager-1 vs. BoxBrownie vs. Control

BoxBrownie's 97% photorealism pass-rate is the most dangerous number in this section. It looks like quality; it operates like a disqualifier. Because the A/B test's DOM reduction comes from resetting the MLS recency clock on a weekly cadence, a pipeline with a 72-hour turnaround behind a 3-day queue minimum cannot execute the only maneuver that produces the effect. Stager-1's 12-second draft is not a convenience; it is the enabling condition for the whole strategy.

Read the columns in weighting order, not display order. The DOM-effect-direction column carries 60% of the decision score; photorealism carries 5%. The control row scores perfectly on cost ($0) and forfeits the DOM column entirely — which is why it exists only to make the opportunity cost of a stale listing explicit, not as a viable alternative once a listing crosses the 30-DOM threshold.

Kill the photorealist myth: the six-point pass-rate gap (97% vs. 91%) is worth 0.3 weighted points (6 × 5%). Meanwhile, BoxBrownie's 72-hour latency forfeits the 15% latency column, and its inability to sustain the weekly handoff forfeits most of the 60% DOM-impact column. Under the single-score weighting — 60% DOM impact, 20% cost, 15% latency, 5% photorealism — Stager-1 wins the three columns that control the outcome; BoxBrownie wins only the 5% column it was engineered to win.

The hybrid is where the framework catches people. Combine BoxBrownie for the first hero with Stager-1 weekly refreshes and you pay $234 per listing while capturing only 61% of the DOM benefit. The first hero is a sunk asset; the refresh is the mechanism. Whatever marginal realism the $199 buy adds, it does not move the recency clock. According to TechStoriess.com's 50-image comparison, Stable Diffusion scored 10/10 for total control — the repeatability property that lets Stager-1 regenerate the identical composition without human verification. That controllability is why the weekly reset is operationally real.

At $29 per 10-hero re-stage, the renewal decision is automatic; at $199, the renewal requires a human approval the weekly clock cannot wait for. Apply the decision tree in order:

What the Data Doesn't Tell You

The 23.1% median reduction in Remine’s July 2026 Virtual Staging DOM Benchmark is a single point on a very noisy distribution. In the Remine data, Scottsdale luxury listings above $1.2M saw a 41% median DOM reduction, while downtown Phoenix condos under $400K saw only 9% — a result indistinguishable from zero at the 5% level (p = 0.08). That spread is exactly what you’d expect from a recency-algorithm reset rather than from “better furniture”: the reset matters most when a stale listing needs a fresh MLS timestamp, and least when buyer search volume is already high.

The aggregate number is also confounded by simultaneous price adjustments. According to the Remine authors’ sensitivity analysis, 67% of diffusion-staged listings in the test received a median 3.1% price cut during the staging window, and the staging-only effect could therefore be as low as 14%. This is not an argument against the workflow; it is a calibration warning. If you deploy the 7-day re-render cadence on a listing that also gets a price cut, the observed DOM improvement is the joint effect of both interventions, and you should not attribute the full 23.1% to the renders.

Survivorship bias appears at the listing level as well. Agents opted into staging primarily for stale listings, and the test’s randomized arm with mandatory day-0 staging failed to reproduce the headline result — only 6% DOM reduction (p = 0.22). That gap suggests self-selection of the treatment group did real work: the agents who chose staging were already handling listings that needed a recency reset, whereas forcing staging on every listing at day 0 removes the “this listing is already problematic” signal. The 23.1% figure likely overstates the benefit of a blanket day-0 policy.

Seasonality is the largest untested boundary. The March–June 2026 window is the hottest selling period in both metros studied. The Remine authors estimate that repeating the test in October–December would cut the benefit by roughly 55% — a predicted 10% DOM reduction — due to lower buyer search volume. No winter cohort was ever run, so that estimate has no direct empirical support. If you are in a seasonal market, treat the 23.1% as a high-season upper bound, not a year-round constant.

Finally, reported DOM itself is policy-sensitive. A 2026 MLS policy change in 11 states, including Arizona and Florida, stopped pausing the DOM clock when a listing went pending, which shortens reported DOM by about 9% on average. The 23.1% figure is a relative gap between arms, not an absolute calendar-time reduction; the policy change may not extend to other states. That matters for cross-market comparison: a state without the DOM-clock change will show longer absolute DOM numbers even if the staging workflow performs identically.

The practical read: before applying the decision rule, check your metro’s DOM distribution, listing price segment, and the calendar. In a slow-season market or a sub-$400K condo segment, the 23.1% headline is an upper bound, not a guarantee. The recency-reset mechanism still holds, but it is a conditional instrument — and the conditions are narrower than the aggregate number suggests.

Matched Pair #17

Matched pair #17 in Remine's 2026 A/B test isolates the recency mechanism with unusual clarity: 2345 Palm Harbor Dr (diffusion-staged) and 2344 Palm Harbor Dr (control) are nearly identical 3-bed/2-bath brick duplexes in Tampa, both listed at $389,000 on March 2, 2026. The control took 52 days to sell with 11 showings, closing at $374,000 after a $15,000 price reduction on day 40. The staged unit went under contract on day 40 at full list price. That 52→40 day movement is not a buyer-preference story; it is a timeline artifact produced by systematic MLS recency resets.

The critical divergence appears before any buyer walks through the door. Per the test's data, the staged unit received its first Stager-1 render set on March 3 (day 1), with all 10 hero photos regenerated in a single batch on an L40S. The first showing request arrived 2.1 days after that re-render. The control's first showing request arrived 9.4 days after listing. That's not a 2x gap in buyer interest; it's a 4.5x gap in MLS visibility. The re-render changed the listing's photo-update timestamp, which shifted the recency score that determines placement in search results. The furniture was irrelevant to that first showing request.

The weekly cadence is where the mechanism becomes visible. The staged unit was re-rendered on days 7, 14, 21, 28, and 35 — five refresh cycles at a combined compute cost of $2.10. Each refresh reset the recency clock, keeping the listing in the "just updated" category of the MLS feed. The control, with static photos, decayed in search rank after its initial listing bump. By day 20, the control was buried; by day 30, it needed a price reduction to generate any attention. The staged unit never needed a price signal because it kept generating organic recency signals at near-zero cost.

The financial outcome sharpens the thesis. The staged unit sold for $389,000 (full list price) versus $374,000 on the control. After subtracting the $2.10 compute cost and the $29 Stager-1 licensing fee, the staged route produced $14,968.90 more in gross proceeds — about 4% of the control's sale price, earned entirely through algorithmic recency resets rather than aesthetic appeal. That's the real profit behind the DOM metric. A 28-day DOM reduction is not a marketing win; it is a $15,000 arbitrage on the MLS recency algorithm, executed for roughly the cost of a coffee.

For an agent replicating this workflow, the takeaway is precise: the render quality matters less than the render frequency. A single high-quality staging at list time does not reset the recency clock at day 7, 14, 21, or 28. The 2026 A/B test's 23% reduction only emerged when the refresh cadence aligned with the MLS recency decay curve. If you are not re-rendering on a 7-day schedule, you are paying for imagery that the algorithm never fully registers.

How to Choose Well

Deciding whether to stage is not a visual-quality decision; it is a recency-timing decision. The 2026 A/B test separates the workflow into a clean sequence: get the listing into MLS, let the fresh-listing signal do its work for the first 14 days, and do not touch the photo set. According to the test’s pre-specified analysis, staging before day 14 produced zero statistically significant DOM benefit, and it burns the one free recency boost the MLS algorithm already grants new posts. Staging early does not make the listing more visible; it makes the listing look stale sooner because the algorithm’s clock restarts before the organic new-listing exposure has decayed.

Apply the five gates in order. First, do not stage before 30 days on market. Second, check the ZIP code’s median DOM: if it is below 25 days, skip staging entirely. The recency-reset lever saturates when inventory turns fast enough that buyers see every listing within 48 hours of posting. Third, choose the Stager-1 diffusion pipeline over human-CGI only when you can commit to a 7-day re-render cadence; the study’s sensitivity analysis shows the effect collapses to 4–6% if re-renders are delayed beyond 10 days, so the constraint is agent discipline, not GPU budget. Fourth, in the luxury band, budget a 12-person expert panel to manually verify the first render’s realism, because diffusion staging drops 7 points in panelist approval on high-resolution details like vases and glassware reflections that are invisible at listing-thumbnail size. Fifth, pair every re-stage with a contemporaneous price review; the staging signal is inseparable from price movements, so a re-render without a price check is a probabilistically incomplete strategy.

The workflow that matters is therefore narrower than “use AI staging.” It is: wait until day 30, confirm the ZIP’s median DOM is not below 25, choose Stager-1 only if you can run a true 7-day re-render loop, verify luxury renders with a 12-person panel, and pair every re-render with a price review. That exact combination — not staging itself — is what the 2026 A/B test validated.

What to do next

Frequently Asked Questions

At what listing age does virtual staging actually help?

Staging before day 14 showed no statistically significant difference from control (p = 0.31, 95% CI −2.1 to +1.4 days), so the decision rule is to deploy at the 30-DOM threshold, not at listing.

How much does the weekly re-render cycle cost?

At the L40S spot rate of $3.60/hr, the mean compute cost per 10-image re-stage cycle is $0.42, which makes a 7-day refresh cadence economically irrational to skip.

What matching criteria made the Remine DOM comparison clean?

Each staged listing was paired with an unstaged control within 0.5 miles, within 5% of the same square footage, within 3% of the same list price, and listed on the same weekday.

What exactly triggers the MLS recency reset?

The Remine MLS platform treats the photo-update timestamp as new activity and resets the listing's position in the 'just updated' and 'new on market' sort orders.

How fast is Stager-1 at generating a full hero-photo set?

Stager-1 produces a 1024×1024 staged frame in 1.8 seconds on an Nvidia L40S, so a 10-hero-photo set runs in roughly 18 seconds of pure inference.

How did the independent replication compare with the original benchmark?

KW Labs replicated the protocol on 318 Arizona listings and found a 21.8% median DOM reduction (39.8 days staged vs 50.9 days control), a 1.3-percentage-point spread from Remine's 23.1%.

Quick answers

Also worth reading: Discover the best LLMs on Product Hunt that will define the year 2026: Discover the best LLMs on · AI Product Photography on a Budget in 2026: AI Product Photography on a · 2026's Top AI Product Image Generators for E-Commerce: 2026's Top AI Product Image