A holiday video of a holiday that never happened
A couple, a camper van and a week in the Austrian Alps — 2 minutes 44 seconds of it, in widescreen, with an original song and the words on screen. There was no trip, no camera and no cast. The hard part was not making pretty pictures; it was making the same two people appear in fifteen of them. This is how that is done, in the order we did it, with what it cost.
The four decisions that made it work
1 Spend the first two generations on nothing
Before any real shot, make a plate of the cast and a plate of the vehicle. Neither appears in the finished film. They exist to be attached to everything after, and they are the reason shot fifteen matches shot one.
2 Write attractive AND specific, never "beautiful"
Asking for a beautiful couple returns the generic AI face. Asking for heritage, age, bone structure, hair and one charming mark returns a person. Our first attempt stacked realism rules — chipped tooth, uneven skin, receding hairline — and produced a couple nobody wanted to watch for three minutes. Delete the decay list; keep one positive clause about real skin texture.
3 Describe the silhouette, never the brand
We asked for a specific make of camper van and explicitly banned its badge from four named places on the bodywork. The badge appeared anyway. Naming the marque is what summons the trademark — "a boxy 1980s European camper, flat upright front, two round headlights" gives you the same van and no logo.
4 Pick the shape before the first picture
Widescreen or vertical is the one decision you cannot cheaply undo once there is content in the project. This one is 16:9 because it was going to YouTube as a real video — vertical under three minutes is filed as a Short, and links in a Short description are not clickable.
Generating one good picture of a couple on a mountain road is easy. Generating fifteen of them, where it is recognisably the same couple and the same van each time, is the entire job — and it is where most AI films fall apart. By shot six the man has a different nose, the woman has aged five years, and the vehicle has changed shape, so what you have is fifteen postcards rather than one story. The fix is not a better prompt. It is that your first two pictures stop being pictures and become references: you attach them to every generation afterwards, so each new shot is drawn as a variation on people who already exist rather than invented from the words again. We made a film to prove it — "A Brown Cow Ate The Map", about a couple who get pleasantly lost in Austria — and everything below is reproducible in the editor with the same buttons.
Perfect for
Some moments this fits — there are more.
A trip you did take
Your own photos, cut to a song written about the actual week.
A trip you are planning
Make the film first and see whether you want the holiday.
An anniversary
The road trip you keep telling people about, finally with pictures.
A travel channel intro
Widescreen, free of an OpenMuse watermark, with commercially licensed included music.
A song about a place
Bring the lyrics; the track gets written and sung here.
A moving card
Two minutes about two people beats a paragraph in a message.
Learning the editor
One real project teaches the whole tool faster than a tour.
Tips for the best result
A few small things make a noticeable difference.
- 1
Read the prompt back as one physical pose
A shot asking for a couple with their backs to the camera AND their bare legs hanging towards it cannot exist. The model will not refuse — it renders backs of heads with fronts of legs, confidently, at full quality, and charges you for it. Anything describing more than one body part deserves one read-through as a real position before you spend.
- 2
Name what a strange object is NOT
A knitted marmot came back as a teddy bear twice. What fixed it was naming the thing it must not be, next to the shapes that tell them apart: "a plump ground squirrel, definitely not a teddy bear — wide flat head, small round ears set low and far apart, a blunt muzzle." For any prop outside the everyday, budget one clause for the wrong answer and one for the silhouette.
- 3
Give the film one small running thing
Ours is a toy marmot on the dashboard that gains a pair of sunglasses, then a plaster, then a crumb. It costs nothing, it survives every generation because it is attached to the vehicle plate, and it is what turns fifteen pretty shots into something a viewer follows to the end.
- 4
Check the model and the ratio on every generation
Reloading the page resets the AI model to the default and leaves the aspect ratio where it was. A square picture on a widescreen canvas is not an error message — it is a shot with blurred bars that looks deliberate, and you find out at render.
- 5
Keep every rejected picture until the edit is locked
Reference plates, near-misses and superseded shots are worth holding on to while you work. Delete them at the end if you need the space — a finished film is a couple of hundred megabytes, and they add up faster than the pictures do.
How it works
- 1
Choose the shape, then write the song
Start a project and pick Widescreen if it is going to YouTube, vertical if it is going to Reels or TikTok. Then write the song — paste your own lyrics or have them written, describe how it should sound, and generate the track for 30 tokens. Do this before the pictures: the song's length decides how many shots you need. Ours came back at 2:44, which is about fifteen.
- 2
Make two plates that never appear in the film
Generate one picture of your cast and one of the vehicle or the place. Judge them hard — everything downstream inherits them, so a face you are lukewarm about now is a face you will dislike fifteen times. 32 tokens each at 2K. When you like them, they stop being shots and become your references.
- 3
Write each shot, and keep the references attached
A prompt is a shot description: where they are, what the light is doing, and what must not be in the frame. Attach both plates to every generation. "Create again" keeps the whole form — prompt, both references, model, aspect ratio, quality — so each new shot is edit-the-prompt and submit. Budget 90 to 150 seconds per image; the on-screen estimate is optimistic.
- 4
Animate the few that earn it
Send four or five stills to the video model for about five seconds of real movement each, at 15 tokens a second. Spend it on motion the eye expects — water falling, a vehicle travelling, wind — not evenly across the film. Twenty stills with four moving among them reads as a film; twenty moving clips reads as a showreel and costs six times more.
- 5
Add the clips, then fit them to the song
Add the shots you want to the timeline and use Auto-arrange to fit them to the song's length. Fifteen shots became twenty-five cuts at about six seconds each, ending exactly when the music does. Two things worth knowing: turn collages off unless you want photos grouped four-to-a-frame, and check the arrangement drew from the clips you added rather than from everything the project has ever held.
- 6
Fix the first shot and the last shot
Nothing in the arrangement knows which of your pictures is the punchline. Ours landed at position eleven and the film ended on a dinner table; two drags in the Organize sheet fixed it. Opening and closing frames are the only two positions worth setting by hand — the middle takes care of itself.
- 7
Put the words on screen, and render
Turn lyrics on so the words light up in time with the music. In widescreen, raise the size — the default is set for a phone-shaped frame and looks small on a television. Then render at 1 token a second. The whole film, from empty project to finished file, came to 1,167 tokens.
Frequently asked questions
Other ways to make it
Make an AI music video
The full method behind one real 2:19 film — including what it actually cost.
Read moreTravel slideshow
Pick 3–12 photos. AI cuts a cinematic trip reel.
Read morePhoto to time-lapse
One still, turned into a sweeping time-lapse.
Read moreNeon ring video
Your photo in a neon ring that pulses with your song.
Read more