Read the full analysis: Making Hollow House With AI: A One-Night Horror-Comedy Experiment on ThorstenMeyerAI.com
TL;DR
Hollow House, a 97-second horror-comedy made with AI-generated characters, voices and sets, was produced overnight on October 27-28, 2026, according to its creator, Thorsten Meyer. Meyer says the first two versions missed the mark, and that tighter editing and using one video model with the full cast as references made the final cut work.
Filmmaker and AI creator Thorsten Meyer says he made Hollow House, a 97-second horror-comedy short featuring AI-generated characters, voices and sets, during one night of work from October 27 into October 28. His account describes three versions of the film and argues that the decisive improvement came not from better-looking images, but from changing the cut’s rhythm and production method.
In the short, a ghost joins the staff of an extreme haunted attraction, where employees respond to the supernatural disruption as if it were a workplace issue. A guest’s safe word fails to stop the scare, and the shift lead asks whether the ghost signed a waiver. Meyer’s account says the script, characters and jokes were his own, while the structure was inspired by an AI horror-comedy short shared on X. The source does not identify that short.
Meyer reports that the first version used 20 shots and ran 2 minutes 15 seconds. He judged it boring despite its consistent faces, costumes and lighting. The second version used 66 fast cuts and ran 2 minutes 21 seconds. The final version, which he says worked, was assembled from three roughly 30-second segments generated by one model with the cast supplied as references; it runs 1 minute 37 seconds.
The production account lists 2,220 Higgsfield credits used across the three versions, valued by Meyer at about $95 under his plan, and about 2,100 ElevenLabs characters for music and sound effects. He says he first generated reference images for eight characters using Google’s Nano Banana 2 through Higgsfield, then used generated frames and video tools for the earlier cut. Those costs and workflow details are the creator’s report, not an independently audited production record.
AI filmmaking · Behind the cut · October 2026
Making Hollow House With AI
A one-night horror-comedy experiment. Thorsten Meyer reports making a 97-second short with AI-generated characters, voices and sets. Three versions later, his stated breakthrough was the edit’s rhythm.
The production lesson
“The pictures were never the problem. The rhythm was.”
— Thorsten Meyer, summarizing his process
01 / Iteration
Three cuts, one turning point
Meyer says attractive images did not solve the storytelling problem. He first changed the pace, then rebuilt the workflow around a single model with cast references.
Version 01 · Rejected
Consistent, but slow
2:15
20 shots. Meyer described this cut as boring despite consistent faces, costumes and lighting.
Version 02 · Recut
Faster, still not right
2:21
66 fast cuts. The shorter beats shifted the rhythm, but Meyer still did not consider this the working version.
Version 03 · Final
Built for the rhythm
1:37
Three roughly 30-second segments, generated with one model and the cast supplied as references.
Reading the bars: The first bar scales average shot length against 5.6 seconds; the second shows the 66-cut count; the third scales final runtime against 2:15. These metrics describe different measures.
02 / Reported workflow
1
Cast references
Generated reference images for eight characters to guide appearance and costume continuity.
2
Generate scenes
Early versions used generated stills and video clips; the first workflow paired Nano Banana 2 with Veo 3.1 fast mode.
3
Revise the cut
After rejecting the first pace, Meyer says he supplied the full cast as references to one model for the final segments.
4
Finish by hand
He credits human choices in selection, timing and final colour grade; speech recognition helped flag unwanted dialogue.
03 / What changed
Direction still mattered
Hollow House follows a ghost joining the staff of an extreme haunted attraction. A guest’s safe word fails; the shift lead asks whether the ghost signed a waiver.
Timing over polish
Meyer says the first cut looked consistent but felt slow. His own count put its average shot at 5.6 seconds, compared with about 1.8 seconds for the unnamed short that inspired the structure.
References for continuity
Eight character images served as visual guides for hair, clothing and appearance across generated frames and clips.
Human decisions at every pass
Choosing what to reject, how to recut, which references to provide and how to grade the final image remained part of Meyer’s process.
04 / Evidence & limits
A reported experiment
The production details are drawn from Meyer’s own behind-the-scenes account, published October 28, 2026.
Creator-reported inputs
Credits and sound
Meyer reports 2,220 Higgsfield credits, valued at about $95 under his plan, and roughly 2,100 ElevenLabs characters for music and sound effects. The figures have not been independently audited.
What remains unclear
Tools, comparison and response
The account does not name the model used for the final segments, identify the X short used for the pacing comparison, or provide an independent review, audience response or detailed post-production record.
Source: ThorstenMeyerAI.com. The short is titled Hollow House | A Halloween Horror-Comedy Short and is described as available through the YouTube link in Meyer’s account.
Editing Shaped the Final Short
The account offers a small-scale example of a broader practical issue in generative filmmaking: producing attractive shots does not by itself produce a coherent film. Meyer says the first cut’s average shot length was 5.6 seconds, while the short that inspired the project averaged about 1.8 seconds, based on his own count. That comparison led him to treat pacing, rather than image quality, as the main problem.
The case also shows how much creative judgment remained with the person making the film. Meyer identifies three important decisions: rejecting the first cut as boring, giving one model references for the whole cast, and applying the final colour grade. Models generated the material, but the selection, revision and timing described in the account were human choices. For readers tracking AI video, the production is a reported experiment, not evidence that a particular tool can reliably make films without substantial direction and editing.
From Character Sheets to Three Cuts
Before generating video, Meyer made one reference image for each of eight characters. He says these images helped keep details such as hair, clothing and appearance consistent across later frames and clips. In his account, this served as a stand-in for casting and costume references in a conventional shoot.
The first approach paired generated stills with animated clips. Meyer says he used Nano Banana 2 for start frames and Google’s Veo 3.1 fast mode for video, which generated speech and sound along with the picture. He describes problems with particular instructions: a shouted “HEY!” was spoken as the letters “H-E-Y,” a requested full head turn changed the character’s whole body, and some physical actions did not generate as intended. He says he used speech recognition to compare clips against the script and catch unwanted dialogue.
These are observations from one creator’s workflow. The source gives no independent evaluation of the models, and its description of the final version does not name the model used for the three segments. It does say that using a single model with all cast members as references was part of the final approach.
““Terribly boring and amateurish.””
— Thorsten Meyer, describing his reaction to the first version
What the Account Does Not Establish
The production details and costs come from Meyer’s own account; the source does not include an independent review of the workflow, credit use or final film. It is also unclear exactly which model generated the final three segments, or how much editing and other post-production work went into them beyond the stated colour grade and sound work.
Meyer compares his shot count with a short shared on X, but the source does not name the original film or provide a link to it, so readers cannot check that comparison from the account alone. There is no information on audience response, distribution beyond the supplied YouTube link, or whether the production will be repeated using the same process.
The Published Short and Its Process
The short is available at the YouTube link included with Meyer’s account, under the title Hollow House | A Halloween Horror-Comedy Short. Meyer’s October 28 article presents the making-of material, including character sheets, start frames, failed attempts and a contact sheet of the final cut. The source does not announce a follow-up film or a further production milestone.
Any broader assessment will depend on viewers being able to judge the finished short and on additional reporting about the tools and production choices. For now, the documented development is one creator’s overnight experiment: three iterations, a reported cost estimate and a final edit shaped by a change in pacing.
Key Questions
What is Hollow House?
Hollow House is a 97-second horror-comedy short about a ghost joining the staff of an extreme haunted attraction. Its creator says the characters, voices and sets were AI-generated.
How long did it take to make?
Thorsten Meyer says work began in the early evening on October 27, 2026, and continued until about 1:30 a.m. on October 28. This timing comes from his account.
How many versions were made?
Meyer describes three versions: a 20-shot cut running 2:15, a 66-cut version running 2:21, and the final 1:37 short made from three approximately 30-second segments.
What did the production cost?
Meyer reports using 2,220 Higgsfield credits, which he values at about $95 at his plan price, plus roughly 2,100 ElevenLabs characters for music and sound effects. The figures have not been independently verified.
What changed between the first and final versions?
Meyer says the first cut looked good but felt too slow. He increased the cutting pace and used one model with the whole cast as references for the final version; he also credits the final colour grade as part of the process.
Source: ThorstenMeyerAI.com