The dumbest-sounding constraint is often the whole product
A few weeks ago I watched a friend lose an entire evening to a trend he could not join. The video everyone was posting was a specific thing: one face screams into the camera, another person dances behind them, the two swap places, and then they both sprint toward the lens together. The choreography is fixed. The music is fixed. The only real input is who the two people are.
He had two photos and no way in. The tools he tried wanted a prompt. They wanted a script. They wanted him to describe the dance, the lighting, the camera move. He didn't know how to describe a dance. He just wanted to put his face — and his cat's face — into a video everyone would recognize instantly.
That constraint — two photos in, no prompt, one very specific output — is a genuinely interesting engineering boundary. It's also the thing nobody writes about, because "pick two photos, get a 21-second vertical 720p video of a scream-and-swap dance" sounds too silly to take seriously. But the constraint is exactly why it works.
Why "no prompt" is a feature, not a shortcut
A general-purpose video model is a blank canvas. That's its strength and its tax. The moment you ask a general model for a dance video, you inherit a thousand hidden decisions: what choreography, what camera angle, what song, what length, what happens when two identities start to drift mid-frame. Most people solve zero of these and then blame the tool.
A template-based generator inverts the problem. It hardcodes the choreography, the timing, the music, and the camera — and leaves you exactly one free variable: the two faces. When the output space is small, the model can be tuned for that specific output. Face consistency across a 21-second swap, the moment the two people change places, the final sprint — those stop being "creative choices" and become a pipeline you can actually optimize.
This is the same logic behind why the best developer tools feel boring on the surface. A constrained interface is a product decision, not a limitation. It's the difference between a text field and a well-typed function signature.
The two axes that actually matter for selection
When I compare the "two photos → dance video" tools that have popped up around this trend, the ones that hold up tend to win on two axes and lose on a third.
Axis one: identity preservation under a swap. The hard part isn't rendering two people. It's rendering two people who stay themselves while they swap places. When the model swaps the roles, the cheap implementations drift — hairline changes, eye color shifts, the second face starts to look like a blend of both. This is why the underlying model matters. Tools built on a stronger base model (the current crop runs on things like Seedance 2.5) do measurably better at keeping each face locked while the bodies move past each other. Ask any of these tools what their base model is before you pay.
Axis two: pets, and "unexpected combinations." Here's a thing I did not expect to be a real feature: cats. And dogs. And cat-plus-dog. A dedicated cat dance is not a joke — it's a signal that the tool treats its subjects as more than human faces. A human face has a lot of high-level semantic structure the model can latch onto. A cat is mostly low-level texture. If a tool handles a cat without drifting into a weird chimera, it usually means the identity pipeline is genuinely robust, not just tuned for one face shape.
Axis three, where everyone loses: consent. This is the one no tool can reliably solve, and you should be suspicious of any that claims it can. None of these generators can verify that you have permission to use the second person's face. That check lives entirely with the person holding the photos. The honest tools say so. Before you put someone else's face — a friend, an ex, a colleague, a stranger — into a video, get their consent out loud. "It's just a joke" is not a rule. A technical pipeline will never be able to automate that for you, so don't let a product's marketing imply otherwise.
The pricing tells you the real model
One thing that reliably separates the serious tools from the trend-chasing copies is how they handle failure.
A one-off purchase with a credit system — say, ten credits for one standard video — plus a refund when a generation fails, is a cost model, and it's usually a sign the operator actually knows what their GPU bill looks like. Every failed render is real money for them, so a tool that refunds failed generations is admitting that failure happens, and paying for it rather than making you pay for it. A tool that charges you for garbage output is either unaware of its own failure rate or indifferent to it.
The same logic applies to the "no watermark" question. A watermark isn't just branding — it's often a licensing and cost signal. A tool that lets you download a clean file and post it straight to TikTok or Reels is treating the output as yours, which is the whole point of a one-time purchase. A tool that locks the clean export behind a subscription is selling you the file, not the video.
What I'd check before paying
If you're evaluating one of these tools, here's the short list:
- What's the base model? "Seedance 2.5" or a comparable current model is a good sign; a vague "AI-powered" is not.
- Does it handle pets cleanly? It's a surprisingly good proxy for identity robustness.
- Does it refund failed generations? This tells you the operator's real confidence, not their marketing copy.
- Is there a watermark, and can you download clean? One-time purchase tools should hand you a clean, shareable file.
- Does it ask you — in plain language — to get consent first? If a tool pretends the ethics problem doesn't exist, treat everything else it says with more skepticism.
The trend itself will fade; it always does. But the underlying pattern — shrink the output space to one recognizable thing, and optimize a pipeline until that one thing is bulletproof — is exactly how a lot of genuinely useful AI tooling gets built. The silly 21-second dance is a decent case study in a very serious design principle.
The tool I've been describing here is Smalltown Boy AI (https://smalltownboyai.app/), a two-photos-to-vertical-video generator for the scream-and-swap "Smalltown Boy" dance trend. I'm not affiliated with it — I just find its constraints more interesting than most of the blank-canvas generators people throw at this kind of video.
