Three shots from one request: a 15-second bakery ad generated with Seedance 2.5 at 480p for $1.30.
Every small business owner in a lake town knows the problem. The season is short, the marketing budget is shorter, and a video crew for a 15-second spot can cost more than a slow weekend brings in. Meanwhile, the feeds are full of slick clips that look as if a drone team filmed them at sunrise. More and more of them were not filmed at all. They were typed.
We wanted to know whether a winery, a bakery or a kayak rental could realistically do the same. Not a marketing agency with a production budget, but the owner, on a laptop, after closing time. So we invented three businesses of the kind you find along any of the Finger Lakes: a winery tasting room with a lake view, a bakery on a brick Main Street, and a kayak rental with a wooden dock. They are fictional. No real business, staff member or customer appears in this article, and every scene you see was generated from a text description.
The tool we used is Seedance 2.5, a video model from ByteDance that turns a written description into a short clip with sound. We ran seven requests in total: three cheap drafts on an older, smaller model, three five-second clips on Seedance 2.5, and one complete 15-second ad with a voice-over. The whole experiment cost $2.85 and took about three and a half minutes of waiting, because all the requests ran at the same time. Below we go through what came back, what looked real, what looked fake, and what a business owner should think about before putting an AI-generated scene in front of customers.
How we ran the test
Seedance 2.5 is available through several online services and developer platforms. We used it through an API, which simply means a small script sends the description and gets a video file back a few minutes later. If you have never written code, you do not need to: most platforms also offer a web page where you paste the same text. For the curious, this is all the script does:
import time, requests
BASE = “https://api.apimart.ai/v1”
H = {“Authorization”: “Bearer YOUR_KEY”}
def make_clip(prompt, model=”seedance-2.5″, seconds=5):
job = requests.post(f”{BASE}/videos/generations”, headers=H, json={
“model”: model, “prompt”: prompt,
“size”: “16:9”, “resolution”: “480p”, “duration”: seconds,
}).json()[“data”][0][“task_id”]
while True:
d = requests.get(f”{BASE}/tasks/{job}”, headers=H).json()[“data”]
if d[“status”] in (“completed”, “failed”):
return d # contains the video URL and the cost
time.sleep(6)
We kept every setting deliberately modest. All clips were rendered at 480p, the lowest resolution offered, which is still fine for a phone screen in a social feed. Sound was switched on, because Seedance 2.5 generates its own audio along with the picture, and we wanted to see whether that saved a step. We did not upload any photos. A real business would probably want to start from its own pictures, and we come back to that at the end, but for this test we wanted to see how far a plain description goes.
Test 1: writing the descriptions
The most useful lesson came before any video was made. A vague description like “a nice winery ad” gives the model nothing to work with. What works is a short paragraph that reads like a shot list: where we are, what time of day, what one person does, how the camera moves, and what we hear. These are the three descriptions we used, exactly as sent:
Winery:
Inside a small lakeside winery tasting room in upstate New York in late
afternoon, a server pours white wine into three tasting glasses lined up on
a wooden bar, autumn vineyard rows and a long blue lake visible through the
big window behind, slow push-in, soft natural light, quiet chatter and the
clink of glass.
Bakery:
Early morning in a small-town bakery, a baker in a flour-dusted apron pulls
a tray of golden bread loaves from a brick oven and sets it on the counter
next to a small chalkboard sign that reads FRESH SOURDOUGH $7, steam rising,
handheld camera, warm light, sound of the oven door and a quiet street
outside.
Kayak rental:
Sunrise on a calm narrow lake, two people paddle a red kayak and a yellow
kayak away from a small wooden dock, light mist on the water, gentle hills
with orange and red autumn trees, drone shot slowly rising and pulling back,
sound of paddles and birds.
Each description contains one deliberate trap. The winery asks for three glasses, which tests whether the model can count and keep track of objects. The bakery asks for a chalkboard with a price, because text and numbers are a classic weakness of image and video models, and a wrong price in an ad is worse than no price. The kayak scene asks for a specific camera move and two specific colours. Writing these took about ten minutes, which turned out to be the longest single step of the whole exercise.
Test 2: cheap drafts first
Before spending money on the better model, we ran all three descriptions through Seedance 2.0 mini, the smallest and cheapest model in the same family. Each five-second draft cost 8 cents and came back in 1 minute and 52 seconds. The idea is the same as a pencil sketch before a painting: you find out whether the description produces the scene you had in mind before paying five times as much for the final version.
Top row: 8-cent drafts. Bottom row: Seedance 2.5 at 44 cents. Same descriptions, frame taken at four seconds.
The drafts were better than we expected. All three were recognisably the right scene, with the right season, the right light and the right props. The kayak draft was arguably the prettiest clip in the whole test, with low sun, mist and a bright orange maple in the foreground. The weaknesses were in the details. In the winery draft, the glasses already contained wine before the pour began, so the action made no sense on a second viewing. In the bakery draft, the chalkboard said “FRESH SOURDOUGH $7”, but the lettering of “SOURDOUGH” was uneven, with letters of mixed size and case, the kind of thing a customer notices. The drafts are also slightly larger than true 480p (864 by 496 pixels), which does not matter in practice.
The verdict: drafts are worth it for checking the idea, the framing and the mood. For 25 cents we learned that all three descriptions worked and did not need rewriting, which is exactly what a draft is for.
Test 3: the winery tasting room
The Seedance 2.5 winery clip. The lake, the vineyard rows and the glassware look convincing. Only the middle glass gets wine.
The Seedance 2.5 version took 2 minutes and 31 seconds and cost 44 cents. The first impression is strong. The view through the window shows a long lake with terraced vineyards in autumn colour, which is exactly the look of a Finger Lakes tasting room in October. The light is soft and low, the glasses catch it properly, and the bottle looks like a real bottle of white wine with a believable label shape.
Two things give it away on a closer look. First, the server’s face is never shown; she is filmed from behind and then only as an arm and a hand. That is actually a sensible choice by the model, because faces are where viewers spot artificiality fastest, but it means you cannot use this clip to introduce your staff. Second, the description said the server pours wine into three glasses. She pours into one, the middle glass, and the clip ends. In a five-second shot that is fine, and many viewers would not notice. But it shows a general rule: the model is good at atmosphere and weaker at following a sequence of actions to the end.
The sound was the weakest part. We asked for quiet chatter and the clink of glass. What we got was very quiet room ambience. Measured with a standard loudness meter, the clip averaged −35.9 LUFS, which is far below the roughly −14 LUFS that social platforms use as a reference. In plain terms, on a phone it will sound almost silent next to the clips before and after it in a feed.
Test 4: the bakery and the price on the chalkboard
The price sign, cropped from both versions. Seedance 2.5 wrote it cleanly; the draft is legible but its second word is uneven.
This clip took 2 minutes and 51 seconds and cost 44 cents. It is the one that would fool most people. A baker in a canvas apron carries a tray of scored loaves away from a brick oven, steam drifts through the light from the window, and the tray lands on the counter next to a small chalkboard. The sign reads “FRESH SOURDOUGH $7” in clean chalk lettering, with the price correct, which we checked frame by frame. Two years ago, signs in AI video were usually gibberish. This one would pass.
The description said nothing about the baker, and the model chose a man in his forties. It makes these choices whenever you leave them open: if your bakery is run by two sisters, say so, or the ad will show someone who does not work there. That matters more than it sounds, as the section on disclosure explains.
The soundtrack had a little more going on, at −27.0 LUFS, with the scrape of the tray and room noise. When we ran it through a speech recognition tool as a check, the tool “heard” a subtitle-credit phrase in another language, a known quirk of that tool when there is no speech at all. The bakery clip, in other words, contained no voices, which is what we asked for.
Test 5: the kayak rental and the drone shot
The kayak clip: the camera rises and pulls back exactly as described, and the mist lifts off the water.
The kayak clip took 2 minutes and 38 seconds and cost 44 cents. This is the shot that most small businesses could never afford to film themselves, because it requires a drone, a licensed pilot, a calm morning and luck with the mist. Seedance 2.5 delivered the camera move precisely: the shot starts close above the dock and rises and pulls back over five seconds while two kayaks, one red and one yellow as requested, paddle away into the mist. The hills behind are covered in orange and red trees, and the mist sits on the water in a believable band.
It is also where you see why AI landscapes feel slightly too perfect. The water is still apart from the wakes, every tree is at peak colour, and nothing is out of place: no buoy, no moored boat, no cottage roof. People who know the lake will see that this is not it. A generated clip shows a mood, not your shoreline.
Across these three clips, the price was the same, 44 cents for five seconds, or about 8.7 cents per second at 480p. That price, and how it climbs with longer clips and higher resolution, is the number most owners will want before they start; the current Seedance 2.5 API pricing lists each resolution and the per-second rate, so you can estimate a whole campaign before generating anything.
Test 6: a complete 15-second ad in one request
Five-second clips are building blocks. What a business actually needs is a finished spot. So for the last request we asked Seedance 2.5 for an entire 15-second ad, with three shots and a voice-over, in a single description:
A 15-second local TV ad for a small-town bakery, three shots. Shot 1: the
brick storefront of a bakery on a quiet main street at dawn, lights glowing
inside, a striped awning. Shot 2: close-up of a baker’s hands scoring a
round sourdough loaf with a blade and sliding it into the oven. Shot 3: a
smiling customer receives a paper bag of warm bread across the counter.
A warm male voice-over says: “Fresh bread on Main Street, every morning
from six. Come say hello.” Soft acoustic guitar music underneath.
The request took 3 minutes and 17 seconds and cost $1.30, which works out to the same 8.7 cents per second as the short clips. What came back was a real, edited ad. The model cut between the shots on its own, at 5.7 and 10.9 seconds, which we measured with a scene-detection tool, so each shot lasts roughly five seconds. The voice-over is a calm, warm male voice, and when we transcribed it, every word of our script was there: “Fresh bread on Main Street, every morning from six. Come say hello.” The guitar sits underneath at a sensible level, and the whole ad measured −18.6 LUFS, much closer to broadcast and social levels than the ambience-only clips.
The model placed its own cuts at 5.7 and 10.9 seconds. Note that the round sourdough becomes a baguette in the last shot.
Now the problems. The storefront in shot one is a handsome brick building on an empty street at dawn, but the awning is plain grey, not striped, and there is no readable bakery name anywhere on the building. In shot two, the baker scores the loaf with a wide kitchen knife held flat, not the small razor blade bakers actually use, and the loaf already has score marks before the knife touches it. Anyone who bakes will see this. In shot three, the customer receives a paper bag with a baguette sticking out of it, although the ad had just shown a round sourdough. This kind of continuity slip is typical of generated ads: each shot is plausible on its own, but the story between them is not checked.
None of these flaws would stop the clip from working as an eye-catching post. All of them would bother a careful viewer, and some of them, like the baguette, quietly misrepresent what the bakery sells.
Test 7: what looked real and what looked fake
To make the results easier to compare, we went through every clip and sorted the details into two piles. This is our judgement from frame-by-frame viewing, not a survey, but the pattern was consistent.
Looked real: light and weather, especially low autumn sun, steam and morning mist; landscapes and lakes; textures such as bread crust, brick, wood and glass; camera moves including the drone rise and the handheld feel; the chalkboard text and price; the voice-over, which matched the script word for word.
Looked fake or wrong: actions with several steps, such as pouring three glasses; craft details that experts know, such as how bread is scored; continuity between shots, with the loaf turning into a baguette; ambient sound that was much too quiet; and a general “too perfect” quality to the landscapes, with no clutter, no other people and no imperfections.
The soundtracks, measured. Only the full ad with a voice-over came close to normal social media loudness.
The practical advice is simple. Use generated video for things that are expensive to film and do not claim to show your product: the lake at dawn, autumn colour, steam in morning light. Film your staff, your bread and your dock yourself; a phone is good enough. Then mix the two, and raise the volume of the ambience clips or add music, because left as they are they will sound broken in a feed.
Test 8: disclosure, labels and what the file says about itself
This is the part most small businesses skip, and it is the part that can cause real trouble. There are three separate questions: what the file itself reveals, what the platforms ask of you, and what the law in New York now expects.
What the file reveals. We inspected the downloaded video files. None of the seven clips carried a visible watermark. But every one of them contained an embedded “content credential”, a digital label using an industry standard called C2PA, which states that the video was created by a generative model and names the model. Some platforms read these credentials and add an “AI” tag automatically. However, the label is fragile. When we trimmed a clip, or simply re-saved it with a common free video tool without changing a single frame, the credential disappeared. So you cannot rely on the file to disclose for you, especially after you edit it into an ad.
What the platforms ask. The large social platforms have their own rules for realistic AI-generated content. YouTube asks creators to disclose realistic synthetic content when they upload, and TikTok and Meta both have labels for AI-generated material. The rules change often, so check each platform’s current help pages before you post, and use the built-in labelling option where there is one. It costs you nothing and avoids a takedown.
What the law says. New York signed a law in December 2025, in force since June 2026, that requires an advertisement to “conspicuously disclose” when it uses a “synthetic performer”, meaning digitally created media that appears to be a real person. Civil penalties start at $1,000 for a first violation. Our bakery ad contains exactly that: the smiling customer in the last shot does not exist, and neither does the baker. We are not lawyers, the law does not prescribe the exact wording, and the details are worth checking with someone who is, but the direction is clear. If your ad shows a generated person, say so. Beyond that, the general rule of advertising applies: an ad must not mislead customers about what you sell. A generated baguette in an ad for a bakery that does not sell baguettes is a small but real example.
A small corner label is enough. It is honest, and in our view it does not weaken the ad.
Our suggestion is a short label in a corner, such as “AI-generated scene”, or a line in the caption. In the mock-up above we added it to the final frame. It does not spoil the ad. If anything, a local audience that values honesty from the shops on its own Main Street is more likely to appreciate it than to be put off.
Test 9: the real cost of a do-it-yourself ad
Seven requests, $2.85 in total. All of them ran at the same time.
The whole test cost $2.85: 25 cents for three drafts, $1.31 for three Seedance 2.5 clips, and $1.30 for the finished 15-second ad. Because the requests can run in parallel, the waiting time was about three and a half minutes, not twenty. If you add the ten minutes of writing descriptions, ten minutes of checking the results frame by frame, and twenty minutes of editing in a free phone app, a complete ad takes well under an hour.
A realistic owner would not get a perfect ad on the first try. Expect two or three versions of the shot you care about most, which could bring the total to five or ten dollars, still a small fraction of a half-day shoot. What the money does not buy is accuracy about your own business: your sign, your staff and your bread are either missing or invented.
What we would do differently next time
Seedance 2.5 can also start from a photo instead of a description. For a real business that is the obvious next step: take a good photo of your actual storefront or your actual tasting room window, and ask the model to animate it with steam, mist or a slow camera move. That keeps the place real and uses the model only for motion and atmosphere. It is also a more honest use of the technology, because the location in the ad is your location.
We would also request each shot separately and assemble them in an editor. The single request was impressive, but the only way to fix the baguette is to regenerate the whole ad; separate shots cost the same per second and can be replaced one at a time.
Checklist for a small-business AI ad
- Write each shot like a shot list. Place, time of day, one action, camera move, sound.
- Draft on the cheapest model first. Our drafts cost 8 cents and confirmed all three ideas worked.
- Name everything you care about. Who works there, what the product looks like, what the sign says. Anything left open will be invented.
- Check text and prices frame by frame. Ours was clean on Seedance 2.5 and uneven on the draft model.
- Watch for continuity between shots. Our sourdough became a baguette.
- Fix the sound. Ambience-only clips were 13 to 22 dB quieter than normal feed level.
- Film your real product yourself and use AI for mood. Generated clips show a place like yours, not yours.
- Label it. Use the platform’s AI label, add a small on-screen note, and disclose any generated person.
The walkthrough video
The video below shows all seven clips, the draft comparison and the finished ad, with a short narration of what worked and what did not.
Walkthrough: a winery, a bakery and a kayak rental, generated with Seedance 2.5 (about 3 minutes).
The bottom line
For under three dollars and less than an hour of work, a small-town business can now produce a 15-second ad with a drone shot, golden light, a voice-over and music, and most viewers scrolling past on a phone will not know it was never filmed. The model was strongest exactly where small businesses are weakest: atmosphere, camera moves and landscape. It was weakest where small businesses are strongest: the specific, real details of what they sell and who sells it. The sensible path is to combine the two, keep the real product real, and label the rest. If you want to try the same experiment yourself, including the cheap draft model we used, the video models in this test are available through APIMart.
This content is brought to you by the FingerLakes1.com Team. Support our mission by visiting www.patreon.com/fl1 or learn how you send us your local content here.
