Stable Generation: staging every angle of a room, consistently
The most requested staging feature we don't offer yet - stage one photo of a room, and every other photo of that room follows with the same furniture in the same places. Here's the honest story of building it: what we call “stable generation,” how close it is, and everything that still goes wrong.
- What it is: stage one photo of a room, approve it, and have every other angle re-stage itself with the same furniture in the same places — the staging feature agents ask us for most.
- Where we are: our best system gets it flawless about 20% of the time, and acceptable about 60%. Not good enough to sell — so we haven't.
- The rest of the field: we ran a blind, like-for-like head-to-head against the boldest competitor we could find. Our opinion: neither is good enough yet.
- What to do today: stage one strong photo per room with our single-photo virtual staging — it's excellent, and a better deal.
Want it the day it's ready? Get one email when we launch it →
Agents ask us for multi-angle staging almost every week — stage one photo of a room, and have every other angle of that room follow automatically. I'll be honest with you: it frustrates me that we still can't hand it to you, and I'm genuinely disappointed we haven't been able to follow through on something so many of you want.
What bothers me just as much is watching other companies advertise this feature as if it's a finished, solved thing. It leaves agents assuming we've fallen behind, or that we're somehow not capable of building what everyone else supposedly already has. So I want to be straight with you about what's really going on.
From everything we've tested — our own work and theirs — my honest read is that nobody can do this reliably yet. We've put an enormous amount into it, and as far as I can tell we've made as much progress as anyone in this space. The difference is simply that we're not willing to sell you an unfinished product, while — in our honest opinion — some competitors are charging for something that will most likely disappoint the person who buys it.
This article is our progress toward stable, multi-angle staging: what we're seeing across the field, what we're able to do versus others, and how what everyone charges compares — including the single-photo staging we're proud to offer right now. I'm not writing it to call anyone out or make anyone feel bad. I just want to show you what we see, honestly, and let you make up your own mind.
The workflow we recommend
Upload every photo of the room, exactly like you do today.
Stage a single photo of that room and get it looking exactly how you want.
Every other photo of that room is staged automatically - same furniture, same places.
The problem we're solving
You've shot a listing. A great room gets two photos, because one angle doesn't do it justice - so you start with two empty shots of the same room:
You stage the first and approve it. Now you want the second staged to match - same sofa, same chairs, same rug, in the same places. With every tool on the market today, what you'll usually get instead is a second, unrelated staging: different furniture, rearranged. Buyers flip between the two photos and it reads as fake.
What stable generation does
“Stable generation” is our name for keeping the staging stable as the camera moves. It reconstructs the room in 3D, computes where every approved piece must appear from the new camera position, and only lets the AI paint inside those spots. Here are all four photos together - your two originals on top, the staged versions directly below them:
The two staged photos (bottom row) carry the same sectional, leather chairs, marble table, rug and arc lamp, in the same places relative to the stone fireplace - seen from two completely different camera positions. The rooms themselves are untouched.
That's a real, unretouched result, verified by two independent automated judges. So why isn't there a launch button at the end of this post? Because today it comes out this clean about one room in five. We've gone from “it can't be done” to “OK about 60% of the time, genuinely well about 20%.” The rest of this post is the work behind that number.
The road to stable generation
We built and measured six versions on the way here - the LAI-Stable_Generation line (SG for short), each a different bet on how to keep furniture consistent across a camera move, and each with a nickname so it's easy to talk about. Here are the first five, in the order we built them; the sixth (3.4 “Keystone”) is our most advanced single model and gets its own section below.
† scored on a fresh held-out set. The last two bars are the blind head-to-head below — our best output and Edensign, judged blind on the seven rooms we both ran; that's a stricter test than our own development grading, which is why our bar there sits below SG 4.0. Full rubric in Appendix A.
How it works: Show the AI the approved photo and ask it to reproduce the same furniture from the second camera.
What worked: Almost nothing - which is the point. It's the honest floor to measure everything else against, and it's still what most tools on the market actually do.
Where it broke: It copies the approved angle's composition instead of re-imagining the room from the new camera. Independent academic work this year (the CVSBench benchmark) found every major model does this.
How it works: Generate both staged angles together in a single wide image, then cut them apart. Inside one generation the model keeps furniture consistent because it's drawing one scene.
What worked: Our first real passes - furniture stayed recognisably the same between the two halves.
Where it broke: Each half is lower resolution, and furniture still drifts between them: a cushion changes colour, a lamp changes shape.
How it works: Detect floor landmarks in both photos and use them to pin where each piece should land in the new view before the AI paints.
What worked: Cheap, and it nudged placement in the right direction.
Where it broke: Floor hints are too weak a leash - the model acknowledged them, then placed furniture wherever its instincts pulled.
How it works: Hand the approved staging to a video AI, have it 'dolly' the camera toward the second angle, grab the best frame, clean it up.
What worked: Genuinely fixed placement on rooms nothing else could handle - the video prior understands that a camera move preserves the scene.
Where it broke: Furniture details drift during the move, and on wide turns the video AI re-imagines the whole camera path. Also the most expensive approach by far.
How it works: Reconstruct the room's 3D geometry from the two photos, project every approved piece into the second camera so its position is computed rather than guessed, then have the AI repaint the rough projection.
What worked: Best furniture accuracy of any approach, and positions that are mathematically correct going in. This is the core of stable generation.
Where it broke: The AI, asked to repaint, still sometimes overrides the correct positions and slides furniture to face the camera.
The one deep-dive worth seeing is Parallax, because it's the foundation of everything above. Instead of asking the AI to imagine the new angle, it computes it: reconstruct the room in 3D, then project the approved furniture into the second camera. The projection is streaky, but every position in it is geometrically correct - and that's what the AI refines into a clean photo.
LAI-Stable_Generation-3.4 (Keystone): our most advanced model
Our most advanced single model, 3.4 “Keystone”, takes Parallax and adds a hard architecture lock: the AI may only paint inside the exact spots where furniture belongs, so the room's walls, windows and floors come back pixel-perfect by construction. On a fresh set of rooms it had never seen, it produced the cleanest results of the whole project - but still passed cleanly on only about 1 in 10.
Keystone taught us the decisive lesson: we ran it on two different AI vendors' models, and both fail the identical way - they rearrange furniture to face the camera, even inside the locked regions, even when handed the correct answer. That's not a bug in one product. It's where image models are in mid-2026. Furniture accuracy is solved. Obedience isn't.
What failure looks like
This one takes a moment to read, which is the point. The room is one long open living-and-dining space; the two photos face opposite ends of it. The front door in the second photo is real - it's right there in the empty shot.
The architecture is right; the furniture story is fiction. From the dining end the camera should see the dining set in the foreground and the back of the sofa. Instead the AI mirrored the whole arrangement to face the camera and pushed the dining set to the far wall. A buyer sees furniture teleporting - and wonders what else in the listing isn't real. That's exactly why we won't ship at 20%.
What it costs (the part nobody publishes)
Our measured raw AI cost per room, using the best models:
| Staging the first photo until you accept it | ~$0.25 |
| Each additional angle - Keystone (cheapest reliable) | $0.26 |
| Each additional angle - Dolly (highest quality) | up to $1.25 |
| A typical 3-photo room | around $1.20 |
A standard single staging image costs us roughly $0.10–$0.13. So at today's reliability, honest pricing would mean charging several times more per photo for results we'd reject two times out of five. That's not a product - it's a coin flip with a price tag.
So how does the rest of the field do?
Everything above is about our own work. But you'd be right to ask the obvious question: is multi-angle staging hard for us, or hard for everyone? The only honest way to answer that is to put our results next to other tools' — same rooms, same grading — so the rest of this post does exactly that. This is the first of these comparisons; we intend to run more, on other tools, and publish them here whether we come out ahead or not. We're starting with the boldest claim in the category.
What other tools claim - versus what their own pages say
Several tools advertise multi-angle consistency today. We compared each one's marketing to its own documentation and terms, and to independent evidence. Quotes are verbatim from their public pages as of July 21, 2026. One data point frames the rest: the biggest player in this space by volume shipped the feature as a two-photo beta for bedrooms and living rooms - and has since stopped advertising it. If this were easy, the biggest player by volume wouldn't be walking it back.
| Tool | Their marketing says | Their own pages also say | Independent evidence |
|---|---|---|---|
| Virtual Staging AI (biggest by volume, acquired by Zillow in 2024) | “Stage rooms from two perspectives with the same furniture and layout” with “consistent layout across views” (homepage, archived March 2026) | Their beta docs: limited to 2 photos, works “best with bed and living rooms, other room types are currently incompatible,” “photos from opposite angles will not work well,” and even for compatible pairs a good result “is not guaranteed” (archived docs) | As of July 21, 2026 the feature is no longer advertised or documented - both were removed between March and April 2026 (compare archives). The one hands-on beta review noted “a lamp switching sides or a book moving” (source) |
| Edensign | “Every Angle. One Consistent Style.” An AI “3D spatial map” places furniture “in consistent positions... across all camera angles” (source) | Their terms: output “may contain inaccuracies, artifacts, or unexpected results” and they do “not guarantee that Output will be accurate”; all credits non-refundable (source) | No Trustpilot, G2 or Capterra presence we could find. The main press article promoting it appears to be sponsored content (source). An affiliate review praises the consistency but notes failures on unusual layouts and strong shadows (source) |
| Pedra | “Pedra keeps your staging consistent across every angle” (source) | Their release notes describe a manual flow: select “Use furniture from another image,” choose your staged photo, regenerate (source) - the same reference-photo technique as our Echo model | No third-party test of the feature found; only their own testimonials page |
| Stager AI | “One Room. Multiple Angles. Seamless Staging” (source) | Workflow is limited to 2 photos of a room | 34 App Store ratings total; when a reviewer asked how to get the same furniture into multiple pictures, the developer replied the capability was still being worked on (source) |
| Collov AI | “Same pieces appear naturally from every angle” (source) | Same page: “our design team first creates a consistent base version” - a human service, minimum 3 images, from 20 credits per image, delivered “within 3 business days” | No independent hands-on test of the multi-angle feature that we could verify; a July 2026 third-party review lists the feature but reports no testing results or workflow detail (source) |
| ListingAI | We say: not yet. This post - success rates, failures and costs included - is our marketing. | ||
The tell
Here's the part worth sitting with. Even at our current best, we won't sell this to you, because one room in five isn't good enough. Meanwhile the feature is marketed hard across the industry, priced, and sold on non-refundable credits.
Yet the tools selling this don't all seem fully convinced by it themselves. Edensign, for one, markets multi-view heavily but won't put it in its own developer API — as of July 2026 the Edensign API stages one photo at a time, with no multi-view option, so the feature lives only inside its web app. And the biggest player by volume, Virtual Staging AI (owned by Zillow), went further, quietly removing its multi-view feature entirely between March and April 2026 (their homepage archives, before and after).
So ask yourself what it means when a company will charge you for a feature it won't build into its own developer platform - and when the biggest player by volume walked away from it altogether. We'd rather publish this progress report, warts and all, than sell you a coin flip.
Why we ran this - and when
We ran this benchmark in July 2026, for two reasons. First, you can't set an honest bar in a vacuum - we wanted to know how our own system stacks up against the best tool people can actually use today. Second, we kept seeing multi-view staging promoted with big promises - “stage any room in 15 seconds,” “one click, multiple views, delivered instantly” - and doubted they held up reliably on real listing photos.
So rather than guess, we tested it: we ran Edensign's own multi-view tool - it makes the boldest fully-automatic claim in the category - across our benchmark rooms, and graded the output with the exact rubric in Appendix A, the same one we hold ourselves to. One thing worth noting from their own pages: the ads say “any room in 15 seconds,” while Edensign's own multi-view page says a room takes 30-45 seconds and that the feature works “exceptionally well for bedrooms and living rooms.” We're publishing the head-to-head - same inputs, our system and theirs, side by side - so you can decide for yourself.
And if any vendor thinks our rooms are unfair, we'll run any two-angle set they nominate. The one rule: the sofa doesn't get to move.
Head to head: same rooms, both systems, graded the same way
We ran seven of our benchmark rooms through Edensign's own multi-view tool. To keep ourselves honest we first graded every result blind — two independent AI judges, told only that they were scoring “an AI staging tool”, never which one and never that one set was ours. Then we went back over every pair by hand, angle by angle, and that closer human read is what you see below — including marking down our own images that first looked clean.
The claim we were testing, from their paid Instagram ads (July 2026):
Edensign paid Instagram ads, captured July 2026 (note the “Ad” label). Their claim; below, what it produced on our rooms.
| Room | Camera turn | ListingAI | Edensign |
|---|---|---|---|
| Backyard patio | 168° | Acceptable, with errors | Critical errors |
| Dining room | 150° | Critical errors | Critical errors |
| Living room | 107° | Critical errors | No major errors |
| Bedroom | 33° | Critical errors | Critical errors |
| Kitchen | 119° | Acceptable, with errors | Critical errors |
| Living room | 81° | Critical errors | Critical errors |
| Entry & stairwell | 157° | Acceptable, with errors | Critical errors |
| Result (of 7) | 0 clean · 3 acceptable · 4 critical | 1 clean · 0 acceptable · 6 critical | |
Read that honestly. We went looking for the best competitor we could find — the one making the boldest fully-automatic claim in the category — ran the same rooms, and graded them the same way. The point isn't the scoreboard above; it's that the strongest tool out there, at this feature, still turns out work we wouldn't feel comfortable putting in front of your buyers — and on plenty of rooms, so do we. The room-by-room breakdown below shows exactly where each one holds up and where it breaks. That closer look even caught a flaw in our own pipeline — on the bedroom, a user had accepted our result, but a second look showed our generative variant had quietly re-angled the wall of the room, so it wasn't a faithful staging of the photo we were given. That is the actual state of the art, not a highlight reel. They sell it anyway. We won't.
Backyard patio · camera turns 168°
Dining room · camera turns 150°
Living room · camera turns 107°
Bedroom · camera turns 33°
Kitchen · camera turns 119°
Living room · camera turns 81°
Entry & stairwell · camera turns 157°
A transitional space that barely takes furniture, so this one is a soft result for both.
The honest summary
Across seven shared rooms, graded angle by angle:
- ListingAI never nailed both angles cleanly (0 of 7), but landed acceptable-or-better on 3 of 7. Where it slipped it was the second angle drifting — a different piece of furniture, a moved rug — and three times it altered the room itself, which we count as an automatic fail.
- Edensign produced one genuinely clean room (1 of 7) and fell short on the other six. Its misses were bigger: whole layouts and furniture sets changing between angles, and a patio surface swapped for synthetic turf.
- Neither is good enough to sell as reliable multi-angle staging. One tool is wrong less often; both are wrong too often.
That is exactly why, today, we recommend staging one strong photo of each room rather than chasing matching angles — and taking a few attempts per room until you have one you are happy with.
Method: seven rooms both systems accepted, each system's two output angles first judged blind for furniture match, placement and architecture by two AI graders (the second re-derives the camera geometry from landmarks before seeing the first's scores), then reviewed by hand angle by angle. Edensign produced three variations per room; we show the first. Same input photos for both, verified identical.
What has to be true before we launch
- It works on most rooms. Our bar is a clean match on well over half of typical listing photo sets, verified by the rubric below - not cherry-picked demos.
- We tell you upfront when your photos won't work. Stable generation already includes a check that reads the 3D geometry of your photos in about 12 seconds and knows whether two angles are compatible - before spending your credits, not after.
- The price is honest. You pay for matched angles that actually match.
The encouraging part: most of the machinery is done. Furniture accuracy - solved. Keeping the room itself untouched - solved. Knowing in advance which photo pairs will work - solved. The one remaining gap is the models' habit of rearranging furniture to face the camera, and model quality steps forward every few months. When a release closes that gap, stable generation is ready for it.
Want to know the day it launches?
Leave your email and we'll send exactly one message when stable generation goes live.
In the meantime: what we do stand behind
Multi-angle staging isn't ready — from anyone we've seen. But single-photo virtual staging absolutely is, we're proud of ours, and it happens to be a much better deal than what the competition charges. Our pricing is built around the fact that AI isn't perfect:
- You only pay for what you keep. 1 credit to generate a preview, 2 more only when you accept it — and if you're not happy, it's just 1 credit to try again. You never pay full price for a reject.
- $36 a month is 300 credits — about 100 staged images, roughly 36¢ each. And because rejects are cheap, even if you turned down half of everything we generated, you'd still keep about 75 images for that same $36 — around 48¢ each. (Here's exactly how credits work, and why the number varies →)
Against their own advertised prices (edensign.io/pricing, July 2026):
| Plan | Per month | Staged images | Per image |
|---|---|---|---|
| ListingAI Professional | $36 | ~100 (≈75 if you reject half) | 36¢–48¢ |
| Edensign Starter | $29 | 15 | $1.93 |
| Edensign Pro | $59 | 50 | $1.18 |
| Edensign Premium | $129 | 150 | $0.86 |
Edensign figures are their own advertised monthly prices and per-photo rates. Yearly billing lowers them — Starter $20, Pro $45, Premium $117 a month — but even at their best yearly rate the per-image cost stays well above ours.
Our credit model takes a minute to get your head around (plain-English version here), but it's built in your favour: you pay a little to look, and only pay in full for the images you actually want. And the common-sense move for multiple angles still applies today — pick the one angle that shows the room best, and stage that one well.
Appendix A: how we grade (the full rubric)
Every result is scored against the original photos and the approved staging by two independent automated judges. The second works adversarially: it derives the camera relationship itself from fixed landmarks (windows, doors, fireplaces) before reading the first judge's scores, and its verdict overrides. Judges are warned about both failure modes we've caught: passing results that copied the wrong viewpoint, and failing correct results by misreading mirrored layouts.
| Dimension | 2 | 1 | 0 |
|---|---|---|---|
| Furniture match | Same pieces, style, colours, count as the approved staging | Mostly same, one swap or missing piece | Different furniture set |
| Placement coherence | Every major piece in the geometrically correct real-world spot for the new camera | Most correct, one piece drifted | Layout re-invented |
| Architecture fidelity | Walls, windows, doors, floors exactly as the real photo | Minor texture drift | Structure altered or invented |
Works well (pass) = furniture 2, placement 2, architecture at least 1. OK (minor) = one small defect a casual viewer might miss. Fail = obviously inconsistent. Placement is judged in real-world terms: a sofa on the north wall stays on the north wall, whatever the camera does. Judges also verify each pair really shows the same room - two pairs in our fresh set turned out to be different rooms of the same house (matching paint and flooring, irreconcilable walls) and were disqualified.
Appendix B: the work, shown
Full per-model results on the 10-room benchmark (real listing photos, two angles each, camera moves from 30° to nearly 180°):
| Model | Idea | Works well | OK or better | Cost/room |
|---|---|---|---|---|
| SG 1.0 · Echo | Reference the approved photo (3 prompt strategies) | 0/10 | 1/10 | $0.27 |
| SG 1.2 · Twin | Both angles in one image (+ hardened variants) | 1/10 | 5/10 | $0.24–$1.93 |
| SG 2.0 · Compass | Floor-plane geometry hints | 0/10 | 4/10 | $0.15–$0.30 |
| SG 2.4 · Dolly | Video camera-move bridge (+ completeness repair) | 0/10 | 5/10 | ~$1.10 |
| SG 3.0 · Parallax | True 3D warp + AI refine (2 hardening levels) | 1/10 | 6/10 | $0.35–$0.45 |
| SG 3.4 · Keystone | 3D warp + hard architecture lock (fresh held-out set) | 1/10 valid | 2/10 valid | ~$0.26 |
| SG 4.0 · Stable Generation (current best) | Best model per room | 2/10 (20%) | 6/10 (60%) | $0.35–$1.25 |
Methodology, so you can hold us to it: benchmark rooms are real listing photos from our production system, not studio shots. A second fresh 12-pair set (never used during development) was held out to keep us honest; two of those twelve turned out to be different rooms of the same house, so we disqualified them and report Keystone on the 10 valid pairs. Numbers on fresh pairs run lower than on the development set - which is exactly why we publish both. Grading cost more than generating: roughly 6 million tokens of automated judging across 18 recipes, four model endpoints and 22 rooms.
Related: Can I stage multiple angles of the same room with the same furniture?