Why Horror Is the Genre Where Small Budgets Win

Every October, brands want something scary. Generated video makes the monster cheap: describe it and a model renders it, in focus, in the middle of the frame, often lit like a product shot. The history of horror suggests that is the wrong instinct. The genre’s most profitable films are remembered for what they did not show.

A small budget used to force that restraint on a director. With AI video, showing everything costs almost nothing, so restraint has to be chosen.

Three of horror’s best returns cost almost nothing

John Carpenter’s Halloween was produced for $325,000 in 1978, according to Time, and grossed $47 million in the United States and $70 million worldwide, about $169 million and $251 million in 2013 dollars. The Blair Witch Project, released in 1999, carries a reported production budget of $60,000 on Box Office Mojo, against $248.6 million worldwide. Paranormal Activity, shot for $15,000 in a house in suburban San Diego according to NPR, reached $193.4 million worldwide after its 2009 release.

None of the three could afford to show much. Their threats live in moonlight, off camera or in the sound mix. Ebert saw The Blair Witch Project as proof that scaring an audience takes no expensive machinery. Restraint was a constraint before it became a style.

What a small horror budget returned

Worldwide gross as a multiple of the reported production budget

Worldwide gross as a multiple of reported production budget

Film	Reported budget	Worldwide gross	Multiple
Halloween (1978)	$325,000	$70 million	about 215×
The Blair Witch Project (1999)	$60,000	$248.6 million	about 4,100×
Paranormal Activity (2009 release)	$15,000	$193.4 million	about 12,900×

Generated video removes that constraint. A creature that once took a workshop months can appear in seconds, and a brief that says “show the monster” gets the monster. The cheaper the reveal, the easier it is to reveal too much, too early.

The shark that kept breaking

Jaws (1975) was no small production, and it learned the same lesson by accident. The Conversation recounts that its malfunctioning mechanical shark limited how often it could appear. Steven Spielberg shot many scenes in which the shark is only hinted at; for much of the hunt, floating yellow barrels mark where it is, and many shots show only the dorsal fin. “The shark not working was a godsend,” he said years later.

Carpenter built that restraint by design. Reviewing Halloween on October 31, 1979, Roger Ebert called him “uncannily skilled” at using foregrounds: the camera sets up a situation, pans to one side, and something looms up. He admired how the film set ordinary people against drab daylight and impenetrable night. Time’s 2013 appreciation singles out flecks of blue moonlight, a jack-o’-lantern’s grin and the breathing of a Michael Myers we cannot see. The piece quotes Carpenter calling his killer human, “only part supernatural,” and the ending is built to leave you sure he is standing right behind you.

An AI video model never breaks down like that shark; it renders the creature whenever asked. So the barrel goes into the shot list, with prompts that describe the evidence: water parting near a swimmer, a blurred shape in the foreground. Image-to-video helps, because the first frame can be composed with the threat half hidden before any motion is generated.

Darkness is a decision

The Blair Witch Project was filmed by its own characters. Ebert noted that all the footage came from two cameras, a color video camcorder and a 16 mm black-and-white camera, run by actors Heather Donahue and Joshua Leonard, whose characters carry their own names. They get lost, hear noises in the night and find disturbing stick figures. The worst of it happens at night, in woods that become a hiding place, mostly as noise. Writing when digital effects could already show almost anything, Ebert concluded that “what really scares us is the stuff we can’t see.”

Generated images have tended the other way. In 2023, researchers Shanchuan Lin and colleagues showed that a common flaw in how diffusion models are trained and sampled kept a widely used open image model at medium brightness, unable to produce very dark or very bright images. They proposed fixes, and models differ: test night shots before a campaign depends on them.

Darkness also has to survive delivery. Social platforms compress video heavily, and smooth dark gradients are where compression shows first, as blocks and bands. Good practice on an AI project: generate from a first frame graded dark on purpose, motivate the light with one source such as a flashlight, a window or a phone screen, and approve every shot on a phone at low brightness.

A locked camera and a thud in the hallway

Paranormal Activity rests on a camera on a tripod at the end of a couple’s bed. Ebert noted that no shot breaks the premise that the couple filmed everything themselves, and that long stretches where nothing happens never bore. NPR’s review lays out the routine: a timer runs in the frame, then a low rumble, on some nights faint walking, sometimes a thud that sends Micah into the hallway with the camera. In one stretch, Katie stands motionless by the bed while the timer runs in fast motion. For that critic, even a recognizable shadow, a door swinging shut or covers rising up “dilute the dread.”

That format suits generated video. A locked-off wide shot with one small change asks little of a model: no camera move, few moving parts, less room for a hand or a face to deform. The fear comes from timing and sound, which people design and edit.

Why horror is where AI shorts go first

At a 2024 film festival run by an AI video company, which drew 3,000 submissions according to TheWrap, the grand prix went to Daniel Antebi’s Get Me Out, a disturbing short in which a second self lays bare muscle, blood and bone. Artist and musician Claire Evans, on the festival’s industry panel, told TheWrap: “AI has given me an ear for the weird.”

Horror forgives what other genres punish. A generated hand with one joint too many sinks a product ad and becomes a scare in a haunted hallway. The condition is that the director places the strangeness: an artifact on the logo, the product or the host’s face is still a defect.

The seasonal stakes are real. The US National Retail Federation expects Halloween spending to reach $13.5 billion in 2026, and watching scary movies ranks among the top ways people plan to celebrate, at 42%. A lot of brands will publish something spooky this month, much of it generated.

What a brand should ask of a Halloween AI video

Ask what stays off-screen, and where the threat is only suggested; a storyboard that shows the monster in every shot is a warning sign. Ask for night shots to be tested early, and review them on a phone at low brightness, after platform compression. Listen with your eyes closed: the sound should carry the fear before the picture does. Keep your product and logo out of the uncanny zone, lit and exact, with the strangeness placed around them. Then check that every unsettling detail in the generated shots was put there on purpose.