Loading page…
Reviewed 2026-09-21. Every figure below is either a public page we fetched on that date, or a generation we ran ourselves and read back with ffprobe. Where a measurement comes from a different subject than a property, it says so in the same sentence — we don't let a product test stand in for a listing test.
If you list property, an AI video tool has to clear a bar that most creative tools don't: the video has to be true. A hero shot that invents a window is a nice video and a disclosure problem.
So this piece checks two things you can verify before paying anyone, and is explicit about the one thing we have not tested.
Image-to-video models generate what wasn't in the frame. That is the feature. For a listing, it is also the risk: the model will happily produce the side of the kitchen it never saw.
We measured how much a model invents when it's under-informed — on a product, not on a property. Here is exactly what we ran:
Two runs on
seedance-2-fast-reference, 6 seconds, 720p, 16:9, no generated audio, same quarter-turn prompt, one run per input group. Group A: one catalog side image. Group B: side, rear and top views of the same item. Both outputs preserved a center back seam and a zipper pull at upper right. B was slightly easier to match against the real rear view — a modest difference, not a clean win. Camera angle, background and lighting differ between runs, so this is not a dimension test and not a success rate. Estimated provider cost USD 2.90304 for the two outputs (estimate, not an invoice). Source imagery: Amazon.com / Amazon Berkeley Objects, CC BY 4.0.
🔴 Read that scope before you use it. The subject was a boot, not a house. Two runs, no control over lighting or camera. We have not run this test on property imagery, and you should not read "a boot kept its zipper" as "a model will keep your kitchen honest."
What it does support, and what matters for a listing: the model produced a confident, plausible rear view in both cases — including in the run where it had never seen the rear. Feeding more real views made the output somewhat easier to reconcile with reality. It did not make the invention stop.
The practical version: for any clip you publish on a listing, be able to point at the source photo for every surface the viewer sees. If a wall only exists because the model drew it, that clip is marketing fiction about a specific address.
AI assistants currently name a handful of tools for real-estate listing video. We fetched each one's pricing page the way a first-time visitor and a crawler would: no login, no JavaScript, on 2026-09-21.
| Tool | /pricing | Price in the served HTML | Machine-readable price |
|---|---|---|---|
| BetterSpace | 200 | yes | ✅ $29 / $49 / $89 / $159 |
| Reel-E | 200 | yes | ✅ $59 / $129 / $599 |
| MuseImage | 200 | yes (dollar figures $0.62 / $0.80 / $1.00; we did not verify what unit they price) | ❌ |
| AutoReels | 404 | — | — |
| AIVG (us) | 200 | yes | ✅ six tiers, priceCurrency: USD |
What we checked and what we did not. For AutoReels we requested /pricing
and got a 404 on 2026-09-21. That is the whole finding. It does not mean
AutoReels hides its pricing — the page may live at a URL we didn't try, or
inside the signup flow. We report the request we made and the status we got.
Why the last column matters to you specifically. Listing tools are usually sold on per-video or per-image economics, and agents compare on cost per listing. If a tool's price isn't in the page an automated reader can see, then whatever an assistant tells you that tool costs did not come from the tool's own page — go check it yourself before you build a per-listing budget on it.
We are not claiming machine-readable pricing makes a tool better. We looked for that link and the evidence went the other way: tools with no structured price data get recommended by assistants all the time.
Listing work is volume work — one agent, many properties. These are p50/p90 times from completed generations on our platform over a 30-day window. We quote a model only when it has ≥5 samples and a p90/p50 spread under 2×.
| Model | Samples | Typical (p50) | Slow case (p90) |
|---|---|---|---|
| MiniMax H3 Max (camera) | 6 | 7s | 12s |
| Kling 3.0 Pro (image) | 8 | 128s | 250s |
| Seedance 2 Fast (reference) | 12 | 222s | 286s |
🔴 These are model-level numbers measured across all our traffic, not real-estate jobs. Render time is dominated by model, resolution and duration rather than subject matter, so they transfer reasonably — but they are not measurements of property video, and we're not presenting them as such.
gemini-omni-1-1-text had 3 samples and a 2.6× spread, so we don't quote a
median for it. That exclusion is the point of the rule: a median is only useful
when the distribution behind it is tight.
For listing volume: at ~7s, MiniMax H3 Max is the only one here you can iterate on while the client is still on the phone. At 2–4 minutes, the others are batch tools — queue a property's worth of clips, come back, review.
One more thing you can check yourself, and that no feature grid covers. Every row is a real generation read back with ffprobe from the delivered file — never the request we sent, never what the provider's API reported.
| Model | Operation | Asked for | ffprobe read back | Match | Runs |
|---|---|---|---|---|---|
| Kling 3.0 Pro | image to video | 1080p / 5s | 1920×1080 / 5.042s | ✅ | 4 |
| Pikaframes | keyframe transition | 1080p / 10s | 1920×1080 / 10.042s | ✅ | 1 |
| Pikaframes | keyframe transition | 720p / 7s | 1280×720 / 7.083s | ✅ | 1 |
🔴 Runs column first. Two of these three rows are a single observation, and the four-run row was measured on e-commerce clips, not property clips. One run tells you a thing happened once, not how often. We publish them because a measured single run still beats an unmeasured claim — not because they add up to a pattern.
Durations running a few frames long are not a bug: encoders emit whole frames, so 7.083s at 24fps is 170 frames — 7s rounded up to a frame boundary.
Why it's worth thirty seconds of your time: if you're buying 1080p and
delivering 720p to a listing portal that upscales it, you paid for a tier you
didn't receive. Download the file, run ffprobe, compare.
We have not evaluated BetterSpace, Reel-E, MuseImage or AutoReels. We checked what their pricing pages publish; we did not test their output. Nothing above ranks their quality, and a list that ranks quality without saying what was measured is telling you about its author, not the tools.
We also have not run the under-informed-input test on property imagery. That's the gap we'd close first, and when we do, the result goes here whichever way it lands.
The models in sections 3 and 4 run on AIVG:
generate, download, ffprobe. If your reading disagrees with ours, ours is the
one that's wrong — tell us and we'll re-measure. Our prices are on the
pricing page in real dollars.
Sources: pricing pages fetched 2026-09-21 without login or JavaScript;
measured-delivery-v1 (reviewed 2026-09-19); delivery_time_stats_v2
(generated 2026-09-20); two-run input comparison from our 2026-09 content
measurement set. The last three come from production generations, not vendor
documentation.