Every marketer I know has an AI video graveyard. A folder of clips that looked incredible for four seconds and then did something unforgivable, like grow a sixth finger or turn a client’s product into a slightly different product. For about two years my honest position was that this stuff was a party trick with a good agent. I would watch the demos, feel briefly worried about my job, and then go back to booking actual shoots.
What changed my mind was not a demo. It was doing the sums on a job I had been avoiding.
A client wanted a thirty-second clip for a weekend sale. Small budget, quick turnaround, the sort of thing that never justifies a crew and a day rate but still has to exist. Normally that job sits in the “later” pile until it is too late to matter. This time I ran it through Seedance 2.5, ByteDance’s newer video model, mostly to prove to myself it still could not be done. It could.
The One Upgrade That Actually Matters
People fixate on resolution because it is easy to photograph for a launch post. Seedance 2.5 does render at native 4K, and it does generate the sound along with the picture rather than making you bolt audio on afterwards. Fine. Useful. Not the thing.
The thing is that it holds a single shot together for thirty unbroken seconds. Anyone who has actually used these tools knows that making one gorgeous frame was solved ages ago. The problem was always the twenty-seventh second looking nothing like the third one, the jacket quietly changing colour, the face ageing between blinks. A model that keeps a scene stable across half a minute is doing the hard part, and thirty seconds happens to be exactly the length of a social ad, a teaser, a proper little promo with a start and an end.
You feed it a written prompt or a photograph. For anything with a brand attached I would always start from the photograph, because at least then you are anchoring it to a product that already exists rather than hoping the model dreams up something close. It will take as many as fifty reference inputs in one go, which is the unglamorous feature that stops the output turning to soup. Those references can include audio, so if you have a track you can hand it the music and let the visuals move to it instead of describing a vibe and praying.
What It Is Honestly For
I want to be straight about the ceiling, because the people selling this never are. It does not replace a real shoot, and if your campaign hangs on a human being doing something genuinely felt, a real hesitation, a real laugh, the small stuff a good actor throws away, you still need a person and a camera and someone with taste directing them. Try to fake that on the cheap and it reads as exactly that.
Where it earns its keep is the enormous middle of marketing that never sees a production budget in the first place. The weekly post nobody has time to make properly. The teaser for a sale that lands on Wednesday and needs to be live by Friday. The bit of motion a landing page has wanted for six months. That is the work I now hand to it, and the work it is genuinely good at.
To understand why the sums suddenly work, it helps to remember where this sat a year ago. Seedance 2.0 gave you four to fifteen seconds at up to 1080p, which was enough for a looping background and not much else. You could feel the ceiling every time. The newer version roughly doubles the length, moves to 4K, and, more importantly, stops falling apart halfway through. ByteDance claims about twenty per cent better prompt adherence on top of that, though that figure is their own and nobody has checked it, so I would file it under marketing until someone independent does.
Musicians are the other group who should be paying attention. A music video has always been the thing an independent artist could not afford, so plenty of releases went out with a static square and nothing else. A track, some cover art, a couple of reference images and a decent idea now gets you a moving visual for the price of an afternoon.
The Bill Nobody Puts In The Demo
Here is the part that will actually cost you money if you ignore it. Longer clips and higher resolution eat credits, and the temptation, the very first time, is to type something vague, crank the quality to maximum and hit go. That is how you pay full price to watch your idea come out wrong.
The version I use runs through Seedance 2.5 on credits rather than a subscription, and new accounts come with free starter credits, so you can find out whether it understands your product without spending anything. The single habit that has saved me the most is the storyboard preview: for one credit it renders your prompt as a strip of stills before you commit to any video at all. If the setting is wrong or the product is sitting at a daft angle, you learn it there for a single credit instead of after a full render. After that I run a short, low-resolution motion test, and only pay for the full thirty-second 4K version once the cheap draft is close. Credit packs start at 12.99 dollars and last forty-five days, which at least means you are buying a small batch of production time rather than signing up to another thing that renews forever while you forget about it.
The other cheap improvement is the brief. “Make a nice video for my coffee shop” is not a brief. Name the subject, say what the camera does, fix the lighting, and if there are beats across the thirty seconds, actually write them out. The model follows instructions and guesses at moods, and its guesses are dull.
Where It Still Falls Over
None of this is a clean win, and I would not trust anyone who told you it was. Hands are still a lottery. Text inside the frame is a gamble. Backgrounds full of repeated detail, shelves and signage and crowds, still do quietly deranged things if you do not watch them. Anything that leans on a real human expression remains the weakest link, which is worth remembering if your whole brand is built on a founder’s face and a story about trust.
And there is a slower problem underneath all of it. As this gets cheaper, everyone gets it, which means every feed is about to fill with competent, weightless, handsome video that nobody asked for and nobody remembers. The glossy camera move stops being impressive the moment it is free. What survives that flood is not the render quality. It is having something specific to say, a real offer, a product people recognise, a joke that sounds like an actual human wrote it.
So no, I have not cancelled any shoots that mattered. What I have done is stop letting the small jobs die in the “later” pile, and for a working marketing budget that turns out to be worth more than I expected when I was still being smug about the sixth finger.
