🔍 Read the full analysis: Hollow House: How We Made An AI Horror-Comedy Overnight, And Why Version One Fell Flat on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer says he made Hollow House, a 97-second AI-generated horror-comedy, overnight on October 27–28, 2026, after rejecting two earlier versions. The account describes how a tighter edit and using one video model with the full cast as references shaped the final cut; the production cost about $95 in Higgsfield credits, according to Meyer.
Filmmaker Thorsten Meyer says he created Hollow House, a 97-second AI-generated horror-comedy, overnight after making and rejecting two earlier versions. In an account published October 28, Meyer said the final version worked because he changed the editing rhythm and gave a single video model references for the full cast—not because the earlier images looked poor.
The short follows a ghost joining the staff of an extreme haunted attraction, where employees treat the supernatural threat as a workplace problem. When a guest says his safe word and the scares continue, the crew realise the Tall Man is not one of their actors. Meyer says the film’s characters, voices and sets were AI-generated. The finished runtime is 1 minute 37 seconds.
Meyer describes three attempts made between early evening on October 27 and about 1:30 a.m. the next day. The first cut used 20 shots and ran 2 minutes 15 seconds; the second contained 66 fast cuts and ran 2 minutes 21 seconds. The third version used three segments of about 30 seconds each, generated with one model given references for the whole cast. Meyer says only that final version worked.
The production account puts the total at 2,220 Higgsfield credits across all three versions, worth roughly $95 at Meyer’s plan price, plus about 2,100 ElevenLabs characters for music and sound effects. He says character reference sheets were made first with Google’s Nano Banana 2 image model through Higgsfield. The account also describes using Veo 3.1 for video in the first version and Whisper speech recognition to check generated dialogue against the script.
Behind the scenes · AI short film · Oct 27–28, 2026
Hollow House: How We Made an AI Horror-Comedy Overnight
Thorsten Meyer says the images were never the problem. After two rejected cuts, a sharper edit and a cast-aware video workflow shaped a 97-second short about a ghost who becomes a workplace problem.
“The pictures were never the problem. The rhythm was.”Thorsten Meyer · on the main creative shift
From long takes to a tighter cut
The overnight process ran from early evening on October 27 to about 1:30 a.m. the next day. Each pass changed the way the story moved.
Good-looking, slow-moving
The first cut used generated starting images animated with Veo 3.1 fast mode. Meyer described it as visually good but boring and amateurish.
More cuts, still too long
The second attempt accelerated the cutting, but the complete version still ran longer than the finished film.
Three cast-aware segments
The story was grouped into three roughly 30-second segments. One video model received references for the whole cast; the supplied account does not name it.
Comedy depends on timing
Meyer compared average shot duration in the first cut with the short that inspired the project. This is his count, not an independently verified measurement.
A ghost, a safe word, a waiver
The premise puts a supernatural threat inside the rules and routines of an extreme haunted attraction. The absurdity comes from the crew treating the haunting like an operational issue.
“Did you even sign the waiver?”
A guest named Derek says “cotton candy,” the attraction’s safe word, but the scares continue. Shift lead Ray asks whether the apparent ghost signed a waiver.
Build the cast
Eight character reference images were made to help limit drift in hair and costume details.
Generate the scenes
The first version used 20 starting images, animated with Veo 3.1 fast mode.
Review the rhythm
The team counted shots, compared pacing, and rejected two versions that did not work.
Restructure and grade
Three segments, full-cast references, human editing decisions, and a final color grade shaped the release cut.
What the models got wrong
The account describes practical snags that affected performance, revision time, and the shots Meyer could use.
The video model spelled out capital letters in dialogue. Whisper speech recognition was used to check generated lines against the script.
A requested full head turn failed; a planned backward bend became a forward crawl instead.
A lift scene was blocked when interpreted as hanging. Rephrasing the action as floating was accepted, but the effect was weaker.
“Terribly boring and amateurish.”Thorsten Meyer · describing the first version
What this account can—and cannot—show
The production details come from Meyer’s own behind-the-scenes account, published October 28, 2026.
Hollow House is a single production case. Meyer says its characters, voices, and sets were AI-generated, while editing, pacing decisions, restructuring, and final color grading were part of the human workflow. The premise’s structure was inspired by an AI horror-comedy short shared on X; Meyer describes the script, characters, and jokes as his own.
Why the First Cut Fell Flat
Meyer’s account points to a practical limit of AI-assisted filmmaking: generating visually consistent shots did not automatically produce a coherent comedy. He says the first version looked good, but its average shot lasted 5.6 seconds, compared with about 1.8 seconds in the short that inspired the project. That comparison is Meyer’s own count, not an independently verified measure.
The creator identifies editing rhythm and human decisions as the difference between the attempts. He says he had to recognise that the first cut was boring, restructure the film, and apply the final colour grade. The account is a single production case, not evidence that one workflow will work for every AI film, but it illustrates how people still shape pacing, performance and continuity in a generative-video project.
It also records concrete production problems. Meyer says the video model spelled out capital letters in dialogue, failed to perform a requested full head turn, and changed a planned backward bend into a forward crawl. One lift scene was blocked after the prompt was interpreted as depicting hanging; rephrasing it as floating was accepted, but produced a weaker effect. These examples show that model outputs and content filters can affect both creative choices and revision time.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Bose S1 Pro+ Portable PA Speaker
As an affiliate, we earn on qualifying purchases.
From Character Sheets to Three Cuts
Meyer says the project began with a premise built around the rules of extreme haunted attractions: visitors sign waivers and can use a safe word to stop the scares. In Hollow House, a guest named Derek says “cotton candy,” but the threat does not stop. The shift lead, Ray, responds by asking whether the apparent ghost signed a waiver. The account says the premise’s structure was inspired by an AI horror-comedy short shared on X, while Meyer describes the script, characters and jokes as his own.
Before generating video, the team created eight character reference images, one for each character. Meyer says those references were intended to limit visual drift, such as hair or costume details changing between shots. In the first version, the production then generated 20 starting images and animated them with Veo 3.1 fast mode, which Meyer says produced dialogue and sound alongside the video.
After reviewing that cut, the team counted its shots and compared their duration with the inspiration short. Meyer says the second attempt raised the cut count to 66, but its 2:21 runtime still exceeded the final film. For version three, he says the team grouped the story into three roughly 30-second segments and supplied one model with references for the entire cast. The supplied account does not name that model.
““Terribly boring and amateurish.””
— Thorsten Meyer, describing his first version
What the Account Leaves Open
The production details come from Meyer’s own behind-the-scenes account, and the source material does not include an independent review of the film, the cost estimate or the workflow. The roughly $95 credit cost is based on Meyer’s plan price; it does not establish the full cost of labour, subscriptions or other production inputs.
It is also unclear which model generated the final three segments, how much footage was discarded within each version, or how the second version differed beyond its cut count and runtime. Meyer gives no audience figures or reception data, so the account does not show how viewers responded to the finished short. The exact publication time and the dates of any subsequent changes are not provided.
The Finished Short and Its Workflow
Meyer’s account points readers to the completed Hollow House short and behind-the-scenes material, including character sheets, starting frames, failed attempts and a contact sheet of the final cut. The supplied source does not announce a follow-up film, release schedule or additional production details.
For now, the next step for readers is to compare the finished 97-second film with Meyer’s description of the three versions. Any broader assessment of the workflow would depend on further examples, independently documented costs and information about how audiences respond.
Key Questions
What is Hollow House?
Hollow House is a 97-second horror-comedy short about a ghost working at an extreme haunted attraction. Its creator, Thorsten Meyer, says the characters, voices and sets were AI-generated.
How many versions did Meyer make?
Meyer says there were three versions. The first had 20 shots, the second had 66 cuts, and the third used three roughly 30-second segments. He says only the third version worked.
Why did Meyer reject the first version?
He said it looked good but felt boring and amateurish. Meyer counted an average of 5.6 seconds per cut in the first version, compared with about 1.8 seconds per cut in the short that inspired the project.
How much did the AI generation cost?
Meyer estimates that the three versions used 2,220 Higgsfield credits, worth about $95 at his plan price, plus roughly 2,100 ElevenLabs characters for music and sound effects. The estimate does not detail all production costs.
What remains unknown about the production?
The account does not identify the model used for the final version or provide independent verification of its cost and shot counts. It also gives no audience or reception figures and does not announce a follow-up project.
Source: ThorstenMeyerAI.com
