All posts
tutorials

How We Made a 90-Second E-Bike Ad With Wan 3.0 Prime

We made a 90-second ad for a fictional teal e-bike from six 15-second Wan 3.0 Prime shots on Flixly, three takes each. Here are the prompts, the takes we picked and why, and an honest look at where the rider and the bike stopped matching from shot to shot.

By Flixly TeamNovember 13, 2026
How We Made a 90-Second E-Bike Ad With Wan 3.0 Prime

TL;DR

We made a 90-second e-bike ad on Flixly from six 15-second Wan 3.0 Prime shots at 1080p, generating three takes of each; every take rendered in 136 to 261 seconds. Single shots look like commercial footage, but continuity broke: the rider's hair and clothes change at shot 4, a helmet appears in shot 5, and the closing bike has a different frame. Anchor every shot to the same images.

We made a 90-second e-bike ad with Wan 3.0 Prime by writing six 15-second shot prompts that share one style suffix, generating three takes of each shot at 1080p, picking one take per shot and joining the six end to end, with no other editing. Each shot on its own looks like commercial footage. The cut as a whole does not hold together: the rider's hair and clothes change halfway through, she suddenly has a helmet, and the bike in the closing shot has a different frame. Below is the finished cut, every prompt, the takes we picked and why, and how we would fix the continuity next time.

The product is fictional: a matte teal electric bicycle with no brand. Everything you see and hear came out of Alibaba's Wan 3.0 Prime on Flixly.

The final cut: six text-to-video shots, 1080p, 15 seconds each, shown here at 720p. Shots start at 0:00, 0:15, 0:30, 0:45, 1:00 and 1:15. The only edits are the joins and a 0.15-second fade out and in on the sound at each cut.

What is Wan 3.0 Prime?

Wan 3.0 is Alibaba's video generation model. Alibaba announced Wan 3.0 in August 2026 with clips of up to 30 seconds in a single generation, about twice the length of the previous version, and inputs that go beyond text and images to video, audio and documents. Alibaba also says Wan 3.0 can carry characters, props and styles over from reference material, which matters for the problem this case study runs into.

On Flixly, Wan 3.0 Prime is the premium tier of Wan 3.0, listed in the video generator. These are Flixly's settings for it:

Setting Wan 3.0 Prime on Flixly
Modes Text to video, image to video (first frame), reference to video
Length 2 to 30 seconds per generation
Resolution 480p, 720p and 1080p
Aspect ratios Auto, 16:9, 4:3, 1:1, 3:4, 9:16
Reference images Up to 10
Reference videos and audio Up to 5 of each
Sound Generated with the video
What we got back 1920 × 1080, 30 frames per second, 15.0 seconds, with a stereo soundtrack

How did we plan the ad?

Like a real commercial: a shot list first. The story is one morning-to-evening ride through a European city, told in six beats, each one a separate 15-second generation. Six times 15 seconds gives exactly 90 seconds, so nothing had to be trimmed.

Shot What happens in the take we used Take used Render time
1 Dawn by a misty canal. She walks up to the bike, standing next to a bike stand, and rides off Seed 1002 163 seconds
2 She rides along cobbled streets in warm, low morning sun, past a flower stall Seed 1001 166 seconds
3 Close-ups: the display lights up, the crank turns, a hand on the grip shifter, the rear gears Seed 1002 136 seconds
4 She rides toward a low camera on a country road, low sun behind her Seed 1002 260 seconds
5 At a viewpoint over a city and river she takes off her helmet and smiles Seed 1003 200 seconds
6 The bike alone against a stone wall at sunset as the camera pulls back Seed 1002 203 seconds

Every shot was text to video at 1080p and 15 seconds. We generated each one three times, with seeds 1001, 1002 and 1003, which is 18 clips in total.

What prompts did we use?

These are the six prompts, word for word:

  1. "Opening shot: dawn over a quiet European city, mist on the river, a woman unlocks a matte teal electric bicycle from a rack outside her apartment, cinematic commercial, 35mm film look"
  2. "She rides the matte teal electric bicycle through narrow cobblestone streets as shops open, tracking shot from the side, warm morning light, cinematic commercial, 35mm film look"
  3. "Close-up details of the matte teal electric bicycle: the handlebar display lights up, the chain turns smoothly, her hand shifts gears, macro commercial shots, 35mm film look"
  4. "She climbs a steep hill on the matte teal electric bicycle with ease, passing other cyclists, camera low and in front, sun flare, cinematic commercial, 35mm film look"
  5. "At the top of the hill she stops at a viewpoint over the city and river, takes off her helmet and smiles, slow push-in, golden light, cinematic commercial, 35mm film look"
  6. "Closing shot: the matte teal electric bicycle parked against a stone wall overlooking the city at sunset, the camera slowly pulls back leaving space for a logo, cinematic commercial, 35mm film look"

Two techniques hold them together:

  • A shared style suffix. Five prompts end with "cinematic commercial, 35mm film look", and the close-up shot swaps in "macro commercial shots, 35mm film look". Repeating the same look at the end of every prompt is the cheapest way to get six separate generations graded alike, and it worked: the six takes we used share the same filmic, commercial look.
  • The same product words every time. "Matte teal electric bicycle" appears in all six prompts, word for word. Every take that shows the bike shows a teal one, but as you will see, not always the same bike.

Each prompt also names one camera move (tracking from the side, low and in front, slow push-in, pull back) and one kind of light (dawn, warm morning, sun flare, golden, sunset). The light cues give the cut a clear arc from dawn to sunset.

How did we pick the best take of each shot?

We watched all 18 takes, then checked each one frame by frame at two frames a second. For every shot we asked three questions: does the rider look like the woman in the shots around it, is it the same bike, and does the light fit the time of day? Prompt accuracy came second.

The three takes of a shot can differ completely. Here is shot 1, the opening, from the same prompt:

Prompt: "Opening shot: dawn over a quiet European city, mist on the river, a woman unlocks a matte teal electric bicycle from a rack outside her apartment, cinematic commercial, 35mm film look" Text to video, 1080p, 15 seconds. Seed 1001 took 163 seconds, seed 1002 took 163 seconds, seed 1003 took 149 seconds.

Three takes, no sound. Three different cities, three different women, three slightly different bikes. Seed 1001 puts a woman with a dark ponytail on a canal-side street under a pink sky. Seed 1002 gives a woman with long, straight light-brown hair at a misty canal lined with gabled houses. Seed 1003 moves to a wide riverside boulevard with towers on the skyline, and ends with her coat filling the frame as she walks past the camera.

  • Shot 1: seed 1002. It has the mist on the water the prompt asked for, a clean step-through bike with a rear rack, and a beige trench coat with dark jeans that a following shot could match. None of the three takes shows a clear unlocking moment, so the "unlocks it from a rack" part of the prompt never really happened.
  • Shot 2: seed 1001. It has the warmest morning light and a rider in a beige trench coat with light-brown wavy hair, the closest of the three to shot 1. Seed 1002 followed the prompt more literally, with a shopkeeper raising a shutter as she passes, but gave her curly auburn hair and cooler light. Seed 1003 dropped the coat for a cream jumper and a shoulder bag.
  • Shot 3: seed 1002. Its bike is a deep teal close to shot 2, the display lights up in the first second, the crank turns, and the hand on the grip has a cream sleeve that could pass for the coat. Seed 1001 made the bike a lighter, brighter turquoise, and seed 1003 shot it against a plain grey studio backdrop with a rider in olive shorts.
  • Shot 4: seed 1002. All three takes took her out of the city onto a country road and swapped the coat for sportswear. Seed 1002 has a grey long-sleeved top, dark leggings and a low sun behind her, which leads into the golden light of shot 5. Seed 1001 is a dark tank top under a high, hard sun, and seed 1003 an olive tank top and shorts on a mountain road.
  • Shot 5: seed 1003. She wears a grey zip-up top close to the one in shot 4 and takes off a vented bicycle helmet. In the other two takes she wears a black leather jacket or an olive jacket and takes off a closed, motorcycle-style helmet.
  • Shot 6: seed 1002. It has the warmest sunset and empty sky above the bike, as the prompt asked. Its bike has a high, sloping top tube instead of the low step-through frame from shots 1 and 2, but seed 1001 has the same problem, and seed 1003 has a worse one.

Here is shot 6, the closing shot:

Prompt: "Closing shot: the matte teal electric bicycle parked against a stone wall overlooking the city at sunset, the camera slowly pulls back leaving space for a logo, cinematic commercial, 35mm film look" Text to video, 1080p, 15 seconds. Seed 1001 took 202 seconds, seed 1002 took 203 seconds, seed 1003 took 168 seconds.

Three takes, no sound. Seed 1003 is the only one with a step-through frame like the opening shots, but it took "leaving space for a logo" literally: about ten seconds in, an invented wordmark fades into the sky. The model filled the space it was asked to leave. For an ad that is a deal-breaker, so we used seed 1002.

Where does the continuity break?

This is the honest part. The six shots were six independent text-to-video generations, and a text prompt alone cannot pin down a face, a haircut or a frame design. Wan 3.0 Prime invented a new rider and a slightly different bike for every generation, and the best we could do was pick the closest matches.

What should stay the same Shots 1 to 3 Shots 4 to 6
The rider's hair Long, light brown, straight in shot 1 and wavy in shot 2 Long brown hair in shot 4, dark brown curls in shot 5
The rider's clothes Beige trench coat and jeans Grey long-sleeved top and dark leggings
Helmet None None in shot 4, then one to take off in shot 5
The bike Low step-through frame with a rear rack Shot 6 has a high, sloping top tube and no rack
The bike's colour Lighter teal in shot 1, deeper teal by shot 3 Deep teal
The place Old European city streets A country road in shot 4, then a city viewpoint
Lettering Invented shop signs in shot 2 Invented lettering on the shot 6 bike's frame

Other things a viewer may notice:

  • The "steep hill" is a country road. None of the three shot 4 takes showed a city hill, and the road does not look especially steep. The "other cyclists" ride away from the camera, so they are heading the other way rather than being passed.
  • The bike never appears in shot 5. All three takes frame her against the view, with no bike in sight.
  • The sound level jumps between shots. Each take came with its own soundtrack, and we kept them as they were. Measured integrated loudness ranges from -20.1 LUFS in shot 5 to -10.9 LUFS in shot 4, a spread of about 9 LU. A real edit would level the sound or lay one music track over the whole cut.

What did work: the light. The cut moves from dawn to warm morning to low sun, golden hour and sunset, because each prompt asked for it, and the shared style suffix kept the grade consistent from shot to shot. Single shots like the cobbled-street ride and the sunset pull-back are good enough to stand in a real spot.

How to keep the same rider and bike across shots

These are the fixes we would try next. We have not tested them in this case study, so treat them as the plan, not the result:

  1. Make one hero still first. Generate or photograph a single image of the rider with the bike, then use it as the first frame of each shot with image to video. Every shot then starts from the same face, clothes and frame.
  2. Use reference images. Wan 3.0 Prime's reference mode on Flixly takes up to 10 images, so you can give it the rider from a few angles and the bike side-on and in close-up. Alibaba says Wan 3.0 is built to carry characters and props over from references like these.
  3. Describe the rider and bike in full, the same way, in every prompt. Hair, clothes, helmet or no helmet, frame shape, rack: "a woman with shoulder-length straight light-brown hair in a beige trench coat, dark jeans and brown ankle boots, no helmet, on a matte teal step-through e-bike with a black rear rack". It narrows the range but will not lock a face on its own.
  4. Keep wardrobe changes out of the story. Our prompt for shot 5 asked her to take off a helmet she never wore before. Either give her the helmet in every shot or drop it.
  5. Never ask for a logo in the prompt. Ask for "empty sky in the top third, no text" and add the real logo afterwards in an editor.

How to make an ad like this on Flixly

  1. Open the video generator with Wan 3.0 Prime selected.
  2. Write a shot list, then one prompt per shot. Give every prompt the same product description and the same style suffix at the end.
  3. Pick 1080p and the shot length. Six 15-second shots make 90 seconds; a single Wan 3.0 Prime generation can run up to 30 seconds.
  4. Generate each shot two or three times. Flixly's generator has no seed setting for Wan 3.0 Prime, so a new generation of the same prompt gives you a new take. Start them together: Flixly runs up to 10 generations at once by default.
  5. Download the takes you like and join them in any video editor. Level the sound while you are there.
  6. Check the price shown on the generate button before you start: an ad like this is 6 to 18 generations at 1080p.

How long did it take to render?

Each 15-second 1080p take took between 136 and 261 seconds; the slowest, shot 5 with seed 1002, took 261 seconds. We did not wait for one take to finish before starting the next: several ran at the same time, and Flixly runs up to 10 generations at once by default. Added up one after another, the 18 renders come to 3,266 seconds, about 54 minutes, so running them side by side saves most of that wait.

Shot Seed 1001 Seed 1002 Seed 1003
1 163 seconds 163 seconds (used) 149 seconds
2 166 seconds (used) 172 seconds 193 seconds
3 151 seconds 136 seconds (used) 154 seconds
4 154 seconds 260 seconds (used) 150 seconds
5 221 seconds 261 seconds 200 seconds (used)
6 202 seconds 203 seconds (used) 168 seconds

The median was 167 seconds per 15-second 1080p clip. Times are wall-clock on Flixly from pressing generate to the finished file, so they include queueing.

Is Wan 3.0 Prime good enough for a real ad?

For single shots, yes. For a sequence with the same person and product in every shot, not from text prompts alone. Wan 3.0 Prime gave us 1080p shots with commercial lighting and camera moves in under five minutes each, and the shared style suffix kept the look consistent. It could not keep the rider or the bike the same across six separate generations, so our 90-second cut works as an animatic or a first draft, not a finished spot. For the real thing, anchor every shot to the same images.

Model page: Wan 3.0 Prime. To choose between the two tiers, see Wan 3.0 vs Wan 3.0 Prime.

Frequently Asked Questions

Can Wan 3.0 Prime make a 90-second video?▾

Not in one generation. On Flixly, Alibaba's Wan 3.0 Prime makes clips of 2 to 30 seconds, so a 90-second video is several shots joined together. We made six 15-second shots at 1080p and joined them end to end in a video editor.

How do you keep the same character across shots in Wan 3.0 Prime?▾

Anchor every shot to the same images. In Flixly's case study, six text-only Wan 3.0 Prime prompts produced a rider whose hair and clothes changed between shots. The fixes to try are image to video from one shared first frame, or reference to video with up to 10 images of the rider and product, which Alibaba says Wan 3.0 is built to follow.

How long does Wan 3.0 Prime take to render a 15-second 1080p clip?▾

In Flixly's test, a 15-second 1080p text-to-video clip from Alibaba's Wan 3.0 Prime took between 136 and 261 seconds, with a median of 167 seconds, measured from pressing generate to the finished file. Several of our takes ran at the same time, and Flixly runs up to 10 generations at once by default.

Does Wan 3.0 Prime generate sound?▾

Yes. Alibaba's Wan 3.0 Prime generates a soundtrack together with the picture, and on Flixly every one of our 18 clips came back with stereo audio. Each clip has its own sound, so when you join several shots, expect the level to jump: loudness varied by about 9 LU across the six shots we used.

Why generate several takes of the same shot with Wan 3.0 Prime?▾

Because each take can be very different. From one prompt, Alibaba's Wan 3.0 Prime gave Flixly three different cities, riders and bikes for the same opening shot. Two or three takes per shot let you pick the one that best matches the shots around it, not just the prettiest one.

Can Wan 3.0 Prime add a logo to the end of an ad?▾

Do not ask it to. When Flixly's prompt asked Alibaba's Wan 3.0 Prime to leave space for a logo, one of three takes drew an invented wordmark into the sky. Ask for empty sky and no text instead, then add your real logo afterwards in a video editor.

Tools mentioned in this post

Wan 3.0 PrimeWan 3.0AlibabaAI Video AdAI CommercialText to VideoAI Video GeneratorCharacter ConsistencyCase StudyFlixly

Related Articles

More on tutorials from the Flixly blog

Ready to create with tutorials?

Jump straight into Flixly's AI studio and try tutorials with 193+ models — free to start.