AI video · Directed production
The problem was never AI.It was that nobody was directing it.
Twenty-four AI-generated shots, directed one by one. No shoot, no set, no actors. And none of the look everyone associates with AI.
Everything above is synthetic.
No camera, no set, no actors. Not a single frame filmed.
- 24 pieces
- 6 sectors
- 4 directing methods
What people mean by “you can tell it's AI”
It is not a vague feeling. It is four specific failures, and all four are failures of direction, not of the model. Each one links to the piece on this page that resolves it.
-
Contact physics
A hand closing on an object without deforming it. A foot that lands and does not press the grass down. Slop fails here because the model draws the look of touching without resolving the touch.
See the piece that resolves it -
The character coming apart
The face drifts from shot to shot: the nose shortens, the mole moves, three years fall off the age. It is the failure anyone spots fastest, even without being able to name it.
See the piece that resolves it -
Light from nowhere
Highlights with no source, shadows that ignore the window, faces lit from the front inside a backlight. When something feels fake and you cannot say why, it is almost always this.
See the piece that resolves it -
The floating camera
A smooth continuous move no operator would make, with no weight, no decision and no narrative reason. You recognise it because it is not looking at anything in particular.
See the piece that resolves it
Fifteen finished shots
No labels, no explanation, no order of importance. Watch first, read afterwards.
How a shot gets directed
This is the difference between generating and directing. Nine pieces with their input on show: what went in on one side and what came out on the other. On four of them you can drag the slider and watch both layers move at once.
Visual reference
The input is not text, it is images. A sketch, a product, a shot already filmed.
From park to vineyard
The same two people, the same stride, the same framing, moved from a gravel path with takeaway coffee to a vineyard with wine. The world changes, the performance does not.
Input: Couple walking through a park with takeaway coffee Result: Same couple, same stride, in a vineyard at golden hour
Industrial design diptych
The same vehicle, pencil sketch on paper on one side and the car running through a desert on the other, alternating panel by panel. It is not a before and after: it is one idea held in two languages at once. Notice that the sketch lines, the fall of the roof and the arch over the rear wheel, are still there once the car has bodywork, sky reflections and road dust. That is what a design studio asks for: a render that does not betray the drawing.
Eight references, one commercial
The input panel is not a written prompt: it is eight stills, among them a fabric soft toy, a real child's bedroom, a family drawn by a child and the product on white. Below, the resulting commercial honours all eight: the toy keeps its stitching and its folded ear, the bedroom keeps the bed and the rug where they were, and the drawing appears taped to the wall exactly as it was. This is how you work with a brand that already has a product and a manual: you invent nothing, you respect everything.
Input: Eight reference images: soft toy, bedroom, drawing, product Result: Commercial with the toy animated inside the bedroom
3D blockout
Volume and camera path go in first, in grey, with no materials. Dressing comes after.
Coloured boxes, a whole world
The shot was composed first with boxes floating over a grey sea. Each box marks where a cloud, the plane or the creature that turns up later will sit.
Input: Colour blocks marking volume and camera path Result: Flight through clouds, plane and islands fully dressed
From CAD to finished interior
A clay-grey chassis, no materials and no lights, becomes a cabin with leather, stitching and trim. The geometry does not move a millimetre between the two states: the mounting points, the gap between seats and the curve of the dashboard stay where they were. This is the natural workflow for any manufacturer that already has the 3D model and needs to show it before the first physical unit exists.
Input: Chassis and structure in clay render Result: Finished cabin with materials and trim
Grey previsualisation
Above, the scene in black and white with no textures: volumes only, the camera path and where the bike enters. Below, the same shot with its time of day, its hanging dust and the heat of the desert. Previs is the cheap part, the one you can redo ten times in an afternoon. When the move works in grey it works dressed, and that order is what saves money.
Input: Untextured previs, volume and movement only Result: Desert at dusk, with dust and heat
Green screen and compositing
You shoot what has to be real and build the rest around it, not behind it.
Green screen to forest
An actress walks down a staircase against a green screen with tracking marks. On the other side of the slider, the same walk inside a forest full of fireflies.
Input: Green screen, practical staircase and tracking marks Result: Night forest, stone arch and floating lights
Camera direction
Every shot carries its movement order in writing before it exists.
Every shot carries its camera order
The labels over the image are not decoration: they are the movement instruction each shot was generated with, from the crane rise to the frozen orbit. You see the whole stage here, crane, track and crew, because the piece does not show a result, it shows a method. Directing with AI is not asking for a nice shot, it is deciding where the camera starts, where it goes and how fast, exactly as you would sign it off on a shooting schedule.
Trimmed excerpt: labels naming real people have been left out.
Nine panels, one sequence
Above, the full storyboard as a nine-panel grid. Below, the assembled sequence, honouring the order, the camera axis and the character of each panel. The firefighter is the same across all nine, same helmet, same smear of soot, and so is the woman he carries out. That is exactly what a storyboard is for: so the client signs off the film before it exists, and so what was signed off is what gets delivered.
Input: Nine-panel storyboard grid Result: The sequence resolved, shot by shot
Four shots, up close
What was decided before generating, and the specific detail that shows the craft. These four play in full, not as an excerpt.
Full piece, with sound if you turn it on.
The last haircut
This piece exists to answer the most common objection: that AI cannot hold a character beyond a single shot. There are six here, with two actors entering and leaving frame, meeting each other's eyes, shaking hands and saying goodbye. The old man's face is the same in the first shot and the last, same age spots, same way of pursing his mouth when he smiles.
What was directed was not the scene, it was the consistency. Each shot was generated anchored to the previous one, with face and wardrobe locked before any movement was requested. The room is anchored too: the diplomas on the wall, the framed photographs and the row of chairs hold their positions even when the camera changes axis.
The lighting is where the craft shows. The whole scene is lit by the mirror bulbs, which is where light would come from in a real barbershop, and that forces shadow under the cheekbone and a highlight on the forehead. When generated footage looks fake it is usually because the light comes from nowhere in particular. Here it does, and that is why the final shot, the empty shop seen from the street at night, works without a line of dialogue.
The audio is generated as well: the hum of the clippers, the conversation in the background and the silence of that last shot.
Full piece, with sound if you turn it on.
A tray on the lake
A cosmetics packshot lives or dies on two things: the product has to read, and the world around it must not contradict it. Here the shot begins tight on hydrangeas beaded with water, pulls back to frame the full tray with the bottle and the lipstick, and keeps pulling back until it reveals that the tray is floating on a lake ringed by misted mountains. The brand is fictional and we say so: LUNYAO does not exist.
Watch the water. As the camera retreats, the tray's reflection stretches across the surface with the right distortion, and the mountains behind double downward with less contrast, which is what a real reflection does. A model with no direction resolves water as a shiny texture and the mistake shows in the first second.
Then watch the scale. The same object goes from filling the screen to being a dot in a landscape without its proportions or its finish drifting. That continuous macro to wide move does not exist on a real shoot without changing lenses, and that is precisely the argument: this is not imitating a camera, it is directing one that could not exist.
The closing logo over the water was generated in the same pass, not composited afterwards.
Full piece, with sound if you turn it on.
One house, three generations
This is the piece that sells renovation work most directly, and not because it shows attractive furniture. It shows time. It opens in a whitewashed village house with a wooden table and a rush chair, nobody inside. The camera pushes in, enters the grain of the wood, and comes out the other side as a family toasting around a dinner table in the eighties. It pushes again, passes through the screen of a cathode television, and the two-thousands living room appears. It ends in a present-day flat with a marble kitchen.
The device is old and it is cinema: the transition hidden inside an object that exists in both scenes. The idea is not the hard part. The hard part is that the space stays recognisable while everything else changes. The window sits in the same place across all four eras. The proportions of the room never drift. That forces you to direct the floor plan before generating anything.
For a renovation studio this is worth more than a render: a render shows how it will look, this shows what it will feel like to live in. The same technique applied to a real project lets you travel through a home from its current state to the finished design, with no building work and no scale model.
The palette does the rest. Olive green and dark wood for the eighties, browns and leather for the two-thousands, white and marble for today. Nobody has to tell you which decade you are in.
Full piece, with sound if you turn it on.
Green screen to forest
This is the shot that best explains the whole page, and up in the process section you can drag the slider and see it for yourself. The starting material is a conventional shoot: a prop staircase, a green backdrop and tracking crosses taped to the fabric. The result is that same footage, frame for frame, turned into a night forest with a stone arch and lights drifting through the undergrowth.
This is not a background pasted behind. In the source the blue dress picks up green bounce from the screen; in the result it picks up amber from the forest lights, which means the lighting on the actress was rebuilt, not merely replaced. The hair edges, where a bad matte always falls apart, resolve against branches rather than against a cut-out silhouette.
The staircase deserves its own note. In the original it is a smooth practical; in the result it has moss, wear and uneven treads, and the feet still land exactly where they landed. The geometry was respected before it was dressed.
For a client the practical translation is this: any location you can build from a staircase and a green cloth can end up anywhere, and the actor is still the actor. No travel, no forest to build.
Drag it yourself in the comparatorThe questions everybody asks
Can you really not tell they were made with AI?
It depends who is watching and for how long. A normal viewer does not tell these apart from conventional footage. A cinematographer will find details if they freeze frames. The useful question is not whether it fools an expert, it is whether it holds attention and says what it needs to say.
Can the generated material be used commercially?
Yes. Material is delivered cleared for commercial use, documented with the tool and licence behind every shot. No real people's faces and no third-party brands are used without permission.
Who owns the rights to the finished video?
The client. Footage and working files are handed over. We keep the right to show the piece as a reference unless agreed otherwise.
How long does a directed shot take?
A single shot can be ready the same day. A sequence with character continuity across several framings takes between three days and two weeks, depending on how much iteration the art direction asks for.
What resolution do you deliver in?
We work to deliver 1080p or 4K depending on the destination. The pieces on this page are shown at their test resolution, which is lower: they are directing trials made for this page, not client deliveries.
Does the same character hold up across different shots?
Yes, and it is the part that takes the most work. Face and wardrobe references are locked before any movement is generated, and each shot is anchored to the previous one. The barbershop piece runs six shots with two characters.
Is the audio generated too?
Yes. Atmosphere, effects and voice come out of the same process. On this page you can unmute any piece and hear it for yourself.
From trial to commission
What you have just watched can be applied to your brand, your product or your space. It starts with a twenty-minute conversation: what has to be told, to whom and on what budget.