🔍 Read the full analysis: How AI Helped Us Make Hollow House—and Why The First Version Fell Flat on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Hollow House, a 97-second AI-generated horror-comedy, was made in three versions over one night, according to creator Thorsten Meyer. The first cut looked polished but felt slow; Meyer says the successful version came from faster editing, shared character references and human decisions about timing and color.
Creator Thorsten Meyer says his 97-second AI-generated horror-comedy Hollow House took three versions to get right, after the first cut looked polished but struck him as boring and amateurish. In a production account published October 28, Meyer described how the project used generative image, video and audio tools—and why he believes editing rhythm, not image quality, was the main problem.
Meyer says the film was made between the evening of October 27 and about 1:30 a.m. the next day. Its characters, voices and sets were AI-generated. The story puts a real ghost among the workers at an extreme haunted attraction, where staff treat the supernatural intruder as a workplace problem. Meyer says the script, characters and jokes were his own, while the structure was inspired by a horror-comedy short shared on X.
The first version used 20 shots and ran 2 minutes 15 seconds. Meyer generated character reference images, then start frames for each shot, before animating them with Google’s Veo 3.1 fast mode. He says the visuals and costumes held together, but the edit averaged 5.6 seconds per cut. By his count, the short that inspired the project averaged about 1.8 seconds per cut, a pace he felt better suited to the comedy.
Meyer reports that the second version used 66 fast cuts and ran 2 minutes 21 seconds. The final version was assembled from three roughly 30-second segments generated with one model using the entire cast as references; it runs 1:37. Across all three versions, he says the production used 2,220 Higgsfield credits, worth roughly $95 at his plan price, plus about 2,100 ElevenLabs characters for music and sound effects. These are the creator’s reported figures, not independently verified production costs.
Why the First Cut Felt Slow
The account offers a specific example of how AI-generated footage can change the work without eliminating human judgment. Meyer says the images in the first version were not the main weakness: shot length, pacing and sound left the jokes without the timing he wanted. The revisions shifted attention from generating attractive frames to deciding how the scenes should move and land.
His experience also shows that consistency and control remain practical concerns. Meyer created eight character reference sheets before generating video, using them to help maintain faces, costumes and other visual details. He also describes model errors that required retakes or workarounds, including misspoken dialogue and actions that did not match the script. The production account is a single creator’s case study, not evidence that the same workflow will work for other films or tools.
Popular right now
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Dyson V8 Cyclone Cordless Vacuum
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
As an affiliate, we earn on qualifying purchases.
From Character Sheets to Final Cut
Before generating video, Meyer says he made one full-length reference image for each of the film’s eight characters using Nano Banana 2, Google’s image model, through the Higgsfield platform. He says each image cost two credits at 2K resolution and all eight were created in under two minutes. He used these references in later image and video generation to reduce changes in character appearance across scenes.
For the first cut, Meyer paired generated start frames with Veo 3.1, which he says produced dialogue and sound alongside the images. He checked the resulting clips with Whisper, OpenAI’s speech-recognition model, and compared transcripts with the script. That review caught an instance where a shouted “HEY!” was spoken as the letters “H-E-Y,” as well as garbled syllables in a line that needed another take.
Meyer also describes attempts that did not produce the intended action. Veo repeatedly turned a character’s whole body instead of turning only his head; a scene intended to show a guest lifted off the ground was blocked in one attempt and barely lifted him after the prompt was changed. Meyer says a planned backward bend became a crawl. These examples are his account of specific generations and do not establish how the tools would behave in other prompts or settings.
“The first version looked great, and when I watched it I called it what it was: terribly boring and amateurish.”
— Thorsten Meyer, creator of Hollow House
What the Account Does Not Verify
The production details and judgments in the account come from Meyer; there is no independent verification of the reported credit use, costs, generation times or comparison with the short that inspired the project. The source does not identify the original short or provide a link to it. It also does not detail the precise tools, settings or prompts used to make the final version, beyond saying it drew on three segments and references for the full cast.
The account describes a successful final cut but does not provide audience testing or other evidence that viewers found it funnier or more effective than the earlier versions. It is also unclear how much of the finished film’s sound and music was generated, beyond the reported ElevenLabs use for music and effects. Those limits make the account useful as a maker’s description of one workflow, rather than a measured evaluation of AI filmmaking more broadly.
The Finished Short and Workflow
Meyer’s account links to the finished Hollow House short, which runs 1 minute 37 seconds. The source material does not announce another version, a release schedule beyond the posted film, or plans for a sequel. Any next steps for the project remain unconfirmed.
For readers assessing the production claims, the next useful point of reference is the finished short itself alongside Meyer’s described versions and behind-the-scenes images. The published account presents the final edit as the result of three attempts, but does not say whether the creator plans to publish further production data or revise the film again.
Key Questions
What is Hollow House?
Hollow House is a 97-second horror-comedy short about a real ghost joining the staff of an extreme haunted attraction. Meyer says its characters, voices and sets were AI-generated.
Why did the first version fall flat?
Meyer says the first cut looked good but averaged 5.6 seconds per cut, which he felt made the comedy too slow. He says the short that inspired the project averaged about 1.8 seconds per cut.
How many versions did Meyer make?
Meyer reports making three versions overnight: a 20-shot, 2:15 cut; a 66-cut, 2:21 version; and the final 1:37 film made from three segments using references for the full cast.
How much did the AI production cost?
Meyer reports spending 2,220 Higgsfield credits, which he valued at roughly $95 under his plan, plus about 2,100 ElevenLabs characters for music and sound effects. The figures come from his account and have not been independently verified.
What remains uncertain about the production?
The account does not independently verify the costs or timings, identify the short that inspired the project, or provide audience data comparing the versions. It also does not give a complete breakdown of the final cut’s prompts and settings.
Source: ThorstenMeyerAI.com
