We reshot our store's hero film with Seedance 2.5 in one evening. Here is what that says about the battle for AI cinema.

Last night we noticed a problem with the hero film on our merch store: in one shot, a rider’s hand had melted into the door panel. The kind of artifact AI video was famous for two years ago. We rewrote the shot brief, regenerated the scene on Seedance 2.5, cut it into a seamless loop, and published it to production. The whole fix, from spotting the bad hand to the new film being live on the site, took under an hour and cost less than a takeout dinner.

This is the shot that came back:

Every garment in that car is a real product from our store. The caps carry our actual embroidered logo. No models were hired, no convertible was rented, no camera operator hung off a chase car at golden hour. We fed the model our product photography as references and described the shot.

What Seedance 2.5 actually is

Seedance is ByteDance’s video generation model family, and 2.5 is the release that moved it from impressive to production-grade. The headline specs, per MindStudio’s breakdown, are generations up to 30 seconds long, native 4K output, and support for up to 50 simultaneous multimodal references in a single generation: images of characters and products, style frames, motion clips, even audio.

That last number is the one that matters for anyone selling something real. Reference-driven generation means the model is not inventing your product from a text description. It is copying it. Our store’s hoodies have an all-over house-circuit print, and the caps have a specific embroidered mark. Earlier model generations would garble prints and turn logos into alphabet soup. With product photos passed as references, Seedance reproduced the actual garments closely enough that we would put the footage next to the real product photography, and we did.

The honest caveats: platform access to the model often caps below its native ceiling (our runs render at 720p and we upscale), hands and fingers still occasionally go wrong, which is what triggered last night’s reshoot in the first place, and getting a seamless loop still takes old-fashioned editing work on the tail of the clip. The difference is that a bad take now costs a regeneration, not a reshoot day.

The battle: own the model, or own the workflow

The more interesting story is the industry forming around models like this, because two very different bets are being placed.

Runway is betting on owning the model and selling to Hollywood. Its flagship Gen-4.5 targets production teams rather than consumers, with frame-level control, motion tools, and integrations into Premiere and Resolve. It has production deals with Lionsgate and AMC Networks, raised $315 million in February at a $5.3 billion valuation, and is hiring cinematographers and colorists alongside ML engineers. Its long game is what it calls world models: simulations of physics, lighting, and camera behavior. Runway wants to be the studio-grade instrument, vertically integrated from research to render.

Higgsfield is betting that nobody should marry a model at all. It is an aggregator: one subscription in front of more than 15 models, including Seedance, Kling 3.0, Google’s Veo line, and its own in-house models, with a reasoning layer that translates creative intent into shot specifications. Fifteen months after launch it reported $500 million in annualized revenue and is reportedly raising at a $5 billion valuation. The pitch is simple: the best model changes every quarter, so own the workflow and swap the engine.

The cautionary tale sits between them. OpenAI’s consumer Sora app, per the same reporting, shut down this spring after reportedly burning roughly a million dollars a day against negligible revenue. Consumer novelty did not pay for the compute. The money in AI cinema is in production workflows, either the high end that Runway is chasing or the high-volume commercial work flowing through aggregators.

We run our creative through the aggregator side, and last night is why. The store’s previous hero was generated on a different engine a day earlier. When Seedance 2.5 became the better tool for the shot, we switched engines without switching platforms, accounts, or reference libraries. That flexibility is the same argument we make about AI in your back office: own the system and the assets, not the subscription.

Why a real estate operator should care

This is not a post about making movies. It is about what happens when broadcast-quality footage costs almost nothing and takes an evening.

A brokerage’s listing film, a builder’s lifestyle spot for a new community, a property manager’s brand film for owner acquisition: these were five-figure line items with multi-week lead times. The reference-driven workflow changes the input requirement too. You do not need a shoot to get footage. You need good photography of the thing you are selling, which most operators already have, and a clear brief.

Our whole store campaign runs this way now. Every distinct product has a saved reference asset built from its own product photo, so any future shot, still or moving, starts from the real merchandise instead of a description of it. Building that library took one evening and generated nothing; it is a filing exercise, not a production budget.

The same discipline applies here as everywhere else we put AI to work: the tool is the cheap part. Knowing what shot to ask for, catching the melted hand before your customers do, and wiring the output into a page that actually sells something is the work. That part still belongs to people.

The loop above is live right now on our store. If you want footage like this working for your listings, your communities, or your brand, that is a 30 minute conversation.

Sources: MindStudio on Seedance 2.5, Metaintro on Runway’s 2026 strategy, TechTimes on Higgsfield’s growth. Industry figures are as reported by these outlets.