We Audited Our Own AI Dance Demos Against the Originals — Here's What Didn't Match
Every December the same clips take over short-video feeds: people dancing a short, very specific Christmas routine — in a kitchen in plaid pajamas, in front of a fireplace, dressed as an elf, or, in the weirdest variant,
Every December the same clips take over short-video feeds: people dancing a short, very specific Christmas routine — in a kitchen in plaid pajamas, in front of a fireplace, dressed as an elf, or, in the weirdest variant, pets standing upright on a party stage.
I run a small tool that recreates those clips from a photo, using motion transfer: your photo supplies the people, the original trend video supplies only the choreography. (I wrote up how that pipeline works last time.)
Disclosure: I run christmasdancevideo.com, the tool this post is about.
This week we did something less fun than building: we audited ourselves. We took the four original trend videos our templates are built on, extracted a frame roughly every 1.5 seconds from each, and compared them — frame by frame — against our own demo videos and against every claim our homepage makes. Some of it held up. A fair amount did not. This post is the honest version of what we found, because the failure modes turned out to be more interesting than the successes.
The four originals, as they actually are
First, the reference material, because everything else is measured against it:
- The elf dance (10.1s). A woman in everyday clothes stands in front of a Christmas tree. Mid-clip she transforms into a full elf costume — hat, ears, striped stockings, the works — and dances the routine solo. The transformation is the joke.
- The duo dance (14.35s). Two women in a bright, high-ceilinged white kitchen. They wear plaid pajamas, but not the same set: one is in pink-and-green plaid, the other in red plaid. Coordinated, not identical. They smile the whole way through and are laughing by the end.
- The couple dance (15.0s). A couple in front of a white fireplace with a dark wood floor and a Christmas tree to the right. The outfit is black tops with red plaid pajama pants. Only the pants are plaid.
- The pet dance (11.5s). Two kittens standing on their hind legs on a Christmas party stage — spotlights, trees, and a backdrop banner that literally reads "Company Christmas Annual Party."
Short, specific, and full of small details that are easy to get wrong. Which we did.
Where our own demos drifted
Our pipeline has two stages: a photo stage, where an image model composes your photo into the template's scene and outfit, and a motion stage, which transfers the choreography onto that composed still. The motion stage is faithful — it moves what it is given. The photo stage is where drift happens, because the scene and outfit are described to the image model in words, and words like "a cozy Christmas kitchen" are doing a lot of unsupervised work.
Concretely, in our own demo videos:
- The duo demo had the wrong cast and the wrong room. The input photo we used was a man and a woman, so the "two friends" dance read as a couple's dance. And the composed scene was a generic dark kitchen, not the original's tall white one with the railing and the windows.
- The couple demo had the wrong clothes. The image model dressed both people in a full matching red-and-black plaid pajama set — top and bottom. The original wears black tops and plaid pants. Once the still was wrong, the video was wrong in every frame.
- The pet demo had no Christmas in it at all. A dog on a plain beige studio background. The original is a full party stage. Our pet template had been given a scene-swap step in the pipeline, but the demo predated it and nobody had re-run it.
None of these are model failures, exactly. They are specification failures. The pipeline did what the brief said; the brief was vaguer than the original.
The copy drifted the same way
The uncomfortable part of the audit was that our marketing copy had the same disease as our demos. We checked 17 factual claims on the homepage and template pages against the videos. Seven were accurate. The rest were not, including:
- An ideas card that promised you could "drop the whole team into the dance" for an office party. Every template supports at most two people. (The only real company-party element in any of the originals is that banner in the pet video.)
- An FAQ answer that said the tool "does not do elf transformations" — on a site whose solo template is literally an elf transformation.
- Copy describing the couple's outfit as a "full matching set," and the duo's pajamas as one identical set — both wrong in the ways described above.
- A line calling the duo routine "deadpan." Watch any five seconds of the original: they are grinning.
We rewrote all of it this week to match the videos. It was not a fun edit. It was a necessary one, because a tool whose entire pitch is "the preview shows you exactly what you'll get" cannot afford a homepage that describes a product that doesn't exist.
What we changed in the process
Three rules came out of the audit, and they apply to the demos, the copy, and honestly to any AI product that recreates something specific:
- Specify against the source, not against the genre. "Christmas kitchen" drifts. "High white kitchen, mezzanine railing, floor-to-ceiling windows, white floor" doesn't — or drifts far less. Our per-template scene and outfit descriptions are now locked to the originals' actual features.
- The input photo is casting. A duo template fed a photo of a couple produces a couple video, whatever the page calls it. Demo inputs now match the template's premise: two women for the duo, a couple for the couple, and full-body shots only — a chest-up photo leaves the motion model to invent legs, and invented legs are where these videos visibly break.
- Audit claims against artifacts, not intentions. Every factual sentence on the site now has to survive comparison with a frame of the video it describes. If the video doesn't show it, the site doesn't say it.
Try it, with your eyes open
The tool is live at christmasdancevideo.com. The flow is: upload a full-body photo (for duo and couple you can use one photo of both people, or two separate full-body photos), see the composed Christmas scene as a free preview, and only pay if the still looks right — $9.99 for one 12–16 second 9:16 video, delivered by email, no subscription and no watermark. The four templates are Elf, Duo, Couple and Pet, and the guides go deeper on getting the photo right.
And if you build anything that recreates a specific reference — a dance, a style, a format — steal the audit habit before you steal anything else: put your output and your source side by side, frame by frame, and write down every difference. The list will annoy you. Fix it anyway.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.