A Century of Filming Places the Camera Never Went

A brand that wants its product on a glacier, or its presenter on a Lisbon rooftop, can now generate the place with AI video. Before approving that shot, know that film crews have faked locations for more than a century, and that every method broke in the same spot: where the real subject meets the invented world.

Generated video still struggles in exactly that spot. Its history is a practical guide to what goes wrong in a generated set, and what a director can do about it.

Paint on glass, 1907

In 1907, filming California’s old missions, Norman O. Dawn painted the missing parts of crumbling buildings on glass in front of the camera and shot the real scene through it. American Cinematographer calls him the first known cinematographer to execute glass shots; his collection at the Harry Ransom Center covers more than 75 films made through 1950.

The rule: the painting must share the real scene’s horizon, vanishing lines and sunlight, for one camera position. Move the camera and the painted building slides against the real hill.

The same rule governs a generated background behind a filmed product. A product shot from a low camera needs a room with a low horizon and lines converging on the same vanishing points. Given the product plate as a reference image, a model will often invent a pleasant eye-level room instead. Dawn also locked his camera. A model will often add a slow push or drift unless the prompt holds the frame still, and every move is another chance for the two lines to part.

Painted backdrops and projected streets

The studio era brought the world indoors. The Mount Rushmore backdrop painted for Hitchcock’s North by Northwest (1959) measured 91 by 30 feet, in a fast, loose style made for the lens. Up close, it looks nothing like what the camera saw.

Generated environments share that property. A model builds a place that reads well from the requested framing, with no floor plan behind it. When the camera moves, a doorway changes width, a row of windows gains a floor, a street rearranges itself behind a turning presenter. A director can keep moves short, or generate the environment as a still and animate from that frame, so the model has less to invent.

For anything that moved, Hollywood used rear projection: footage of a real place played on a screen behind the actors. Film historian Julie Turnock calls it the main composite technology of the studio system from about 1935 to about 1970. On Vertigo, second units shot the plates Farciot Edouart ran on his rear projection stage.

The familiar tell is the car interior with a street sliding past the back window. Turnock names the problem as “the discrepancy of the image quality between the foreground and background.” Studio lamps lit the actors; the sun had lit the street, on another day.

Putting a filmed product into a generated world is rear projection in reverse, with the same discrepancy. The product plate carries real lens softness and grain; the generated world may be cleaner, sharper after upscaling, at another frame rate. Before the two are combined, they have to be conformed: one frame rate, matched grain and sharpness, and light on the product that agrees with the generated sun.

Blue screen and front projection

Putting a performer inside a fantasy world takes a matte that follows the actor through the frame. The Thief of Bagdad (1940) did it with blue screen, under effects supervisor Lawrence W. Butler. Roger Ebert credited the method with keeping real-world subjects in the shot. It also lit actors with a coloured wall instead of the world they would join, a spill compositors have painted out ever since.

A presenter filmed in a neutral studio and dropped into a generated sunset carries the same studio light. Relighting, by hand or with generative tools, can shift the direction and colour of light on the subject, but it has to guess the shape of a face or a bottle. The sturdier fix comes first: settle the generated world, then light the real shoot to match.

For 2001: A Space Odyssey (1968), Stanley Kubrick projected stills of Southwest Africa along the lens axis onto a screen 110 feet wide covered in tiny glass beads. According to American Cinematographer, it returned an image 100 times brighter than the light bouncing off the actors, so the picture falling on them vanished. The tell came from the light anyway: a leopard’s eyes caught the beam and glowed. Kubrick called it “a happy accident.”

Generated scenes slip on reflective surfaces too. A model paints a plausible reflection in eyes, sunglasses or a car’s paint, and plausible is not always the place around the subject. On product shots, check chrome and glass frame by frame.

The LED volume, 2019

The Mandalorian (2019) was shot largely inside ILM’s StageCraft: an LED wall 20 feet high curving 270 degrees around a 75-foot performance space, under an LED ceiling. ILM says more than 50 percent of the first season was filmed this way, with environments rendered in real time from the camera’s point of view so the background moved with correct parallax. The screens also lit actors and sets with the environment itself, reflected in the title character’s armor. Visual effects supervisor Richard Bluff said wraparound green screens “cause confusion for both actors and crew.”

That is the standard for a generated set: the light on the subject comes from the world. A model inventing subject and place together tends to get the broad direction right. The risk grows when a real person or product is inserted, or when the camera moves, because a model has no 3D scene to compute parallax from; it predicts what should appear. Location reference images, a fixed first and last frame, or a motion reference from a real camera move all narrow what it must guess.

How crews have filmed where they couldn’t go

Virtual set techniques, each with a documented example

Virtual set techniques and documented examples

Year	Technique	Example
1907	Glass shot	Norman O. Dawn, California missions
c. 1935	Rear projection	Studio standard until about 1970
1940	Blue screen	The Thief of Bagdad
1959	Painted backdrop	North by Northwest, Mount Rushmore
1968	Front projection	2001: A Space Odyssey, Dawn of Man
2019	LED volume	The Mandalorian, ILM StageCraft
Now	Generated set	AI video models

Generated environments, now

Studios already put generated environments on screen. In July 2025, Netflix co-CEO Ted Sarandos said a building collapse in the Argentine series El Eternauta was the company’s first generative AI final footage, completed 10 times faster than with traditional visual effects tools. Amazon’s House of David used around 70 AI images in its first season and between 350 and 400 in its second, for battle sequences, locations and environments, TheWrap reported. Showrunner Jon Erwin describes a real camera on a real actor as “the hand inside a puppet,” with the digital world as the puppet.

The tell lives where it always has, between the subject and the world, with one difference. Dawn, Edouart and Kubrick set light and perspective by hand, once. A model decides them frame by frame, and they can drift within a single shot: a shadow that swings, a skyline that shifts as a presenter walks. The same place can also change between separately generated shots, so a sequence needs shared reference images and a check that the sun stays put from cut to cut.

What to check in any virtual set

Viewers catch mismatched light and perspective before anything else. Before approving a generated or composited shot, ask for these checks.

One light. Do shadows on the subject and in the world fall the same way, with the same softness, from the first frame to the last?

One horizon. Does the horizon sit where the camera height says it should, and do the set’s lines converge with the subject’s perspective?

Contact. Where the subject touches the world, feet on ground or product on table, is there a shadow?

Camera moves. If the camera moves, does the architecture hold its shape, and does the place match the shots before and after?

Reflections. Do eyes, glasses, chrome and windows reflect the place they are in? Kubrick’s leopard shows that one glowing detail is enough.