🔍 Read the full analysis: Hollow House: Why Our AI Horror-Comedy Needed More Than Version One 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
Hollow House, a 97-second AI-generated horror-comedy, was made in three versions between the evening of October 27 and about 1:30 a.m. on October 28, 2026. Its creator says the first cut looked good but felt boring, prompting a faster edit and a production change that gave one video model references for the full cast.
Thorsten Meyer has published a behind-the-scenes account of Hollow House, a 97-second AI-generated horror-comedy made in three versions over one night. Meyer says the first cut looked polished but was boring, and that the final version improved when the team changed its editing rhythm and production approach—a reminder that generated images alone do not determine whether a short film works.
The film follows a real ghost joining the staff of an extreme haunted attraction, where employees treat the supernatural newcomer as a workplace problem. In one scene, a guest says the safe word, “cotton candy,” but the frightening encounter continues. The source account says the cast, voices and sets were all AI-generated. Meyer describes the script, characters and jokes as original, while saying the premise’s structure was inspired by an AI horror-comedy short shared on X.
Production ran from early evening on October 27 until about 1:30 a.m. the next day. The first version used 20 shots and ran 2 minutes 15 seconds; the second used 66 fast cuts and ran 2 minutes 21 seconds. The final cut was assembled from three roughly 30-second segments generated with one model using references for the full cast, and runs 1 minute 37 seconds. Meyer says only the third version worked as a film.
The project used 2,220 Higgsfield credits across the three versions, which Meyer estimates at about $95 under the team’s plan, plus approximately 2,100 ElevenLabs characters for music and sound effects. The account says character reference images were created first with Google’s Nano Banana 2 model, then used to guide later frames and clips. These figures and production details come from Meyer’s account; the article does not provide an independent audit of the costs or tools.
Hollow House: Why Our AI Horror-Comedy Needed More Than Version One
TL;DR A 97-second AI-generated horror-comedy took three versions and one overnight production. The images looked polished early; the film only clicked after a faster edit and a production change that gave one video model references for the full cast.
A ghost joins the shift
A real ghost becomes staff at an extreme haunted attraction. The employees treat the supernatural newcomer as a workplace problem, while one guest’s safe word—“cotton candy”—fails to end the frightening encounter.
Supernatural HR issue
The short turns a real ghost’s arrival into an absurd staff-management problem at a haunted attraction.
Cast, voices & sets
Meyer says the characters, voices and sets were AI-generated. He describes the script, characters and jokes as original.
A structure, not a script
The premise’s structure was inspired by an AI horror-comedy short shared on X. The supplied account does not identify it.
Iteration changed the film
The creator says only the third version worked as a film. The sequence shows a shift in cut length and generation approach, not a controlled comparison of models.
Polished, but slow
20 shots
2:15 runtime · average cut: 5.6 seconds
Static cameras and no music left the jokes without enough rhythm, Meyer says.
More cuts, still long
66 fast cuts
2:21 runtime
The team compared pacing with the inspiration short, which Meyer counted at roughly 74 cuts.
A film that clicked
3 segments of roughly 30 seconds
1:37 runtime · one model with full-cast references
Faster rhythm, cast guidance and a final colour grade shaped the version Meyer says worked.
From character sheets to final cut
Generated assets were only part of the workflow. The account describes scripting, prompting, dialogue checks, selection, editing and grading alongside them.
Build the cast
Eight character reference images, made first with Google’s Nano Banana 2. Meyer reports two credits per 2K image and under two minutes for all eight.
Generate shots
The first version used character-guided starting frames for 20 planned shots, animated with Veo 3.1 fast mode.
Check the dialogue
Whisper speech recognition checked generated lines against the script and helped flag “H-E-Y” and a garbled final line.
Edit & grade
The team revised the rhythm, generated three longer final segments with full-cast references, then applied a colour grade.
“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
Human decisions shaped the result
The account’s turning points were editorial and practical choices. Its examples describe this production, rather than a systematic test of the tools.
Call the cut boring
Rejecting a visually polished first cut made room to rethink pace, camera stillness and music.
Reference the whole cast
Character sheets aimed to reduce drift in hairstyles and costumes across generated shots.
Grade the final image
A final colour grade was one of the three choices Meyer identifies as consequential.
Tools, spend & constraints
Figures below come from Meyer’s account. The article provides no independent audit of costs, tools or performance.
| Item | Reported detail | What it tells us |
|---|---|---|
| Production window | Evening of October 27 to about 1:30 a.m. October 28, 2026 | One overnight experiment |
| Video generation | 2,220 Higgsfield credits across three versions | Creator estimates about $95 under the team’s plan |
| Music & sound | About 2,100 ElevenLabs characters | Separate usage figure; no cost stated |
| Generation limits | “HEY!” became “H-E-Y”; a head turn missed; a stunt changed after rewording | Examples reported by the creator, not a model benchmark |
| Record completeness | Supplied source ends partway through a revised starting-frame description | Full production record is not available here |
What this account does not establish
The story documents one creator’s workflow. Broader claims about tools, costs or outcomes need more projects and independent checks.
No independent assessment
The supplied account does not independently verify the final film, credit estimate, or comparison with the inspiration short. How cut counts were measured is not explained.
Some details remain open
The final model is not named in the supplied material. Audience response, distribution beyond the linked YouTube video, future versions and follow-up projects are not reported.
One overnight experiment
The account links the finished 1:37 short with production images and descriptions of the iterations. For now, it is a production record—not a controlled comparison.
Script & characters
Original script, characters and jokes, according to Meyer; premise structure inspired by a short shared on X.
Generate, inspect, revise
Reference images and generated clips met human review, dialogue checks and repeated editing decisions.
Watch the finished short
Meyer’s account points to the 1:37 film on YouTube. Further versions and follow-up work have no stated schedule.
Quick answers
What is Hollow House?
A 97-second horror-comedy about a real ghost who joins an extreme haunted attraction’s staff. The creator says its characters, voices and sets were AI-generated.
Why three versions?
The first looked good but felt boring. Meyer says two revisions changed the pacing and production approach before the final 97-second cut.
How much did it cost?
Meyer reports 2,220 Higgsfield credits, estimated at about $95, plus roughly 2,100 ElevenLabs characters for music and sound effects. These figures are not audited.
Were the story and jokes copied?
Meyer says the structure was inspired by an AI horror-comedy short shared on X, while describing Hollow House’s script, characters and jokes as original.
Why the First Cut Fell Flat
Meyer’s account focuses on a familiar production problem in a new workflow: visual consistency did not guarantee effective pacing. The first version averaged 5.6 seconds per cut, while Meyer counted roughly 74 cuts in the short that inspired the project, or about one new image every 1.8 seconds. He says the slower edit, static cameras and lack of music left the jokes without enough rhythm.
That distinction matters to readers following generative video because the account describes a human editorial decision, rather than a single model improvement, as the turning point. Meyer identifies three consequential choices: rejecting the first cut as boring, supplying one model with references for the whole cast, and applying a final colour grade. The story does not establish that the same approach will produce comparable results elsewhere, but it shows how much revision and judgment this particular short required.
The production also records limitations that shaped the final work. Meyer says a shouted “HEY!” was spoken as the letters “H-E-Y,” a requested head turn did not work as intended, and an attempted stunt was blocked or rendered differently after rewording. These examples are the creator’s report of specific generations, not a systematic test of the models.
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
DJI Osmo Pocket 3 Vlogging Camera with 1-Inch CMOS Sensor and 4K Video
As an affiliate, we earn on qualifying purchases.
From Character Sheets to Final Cut
Before generating video, Meyer created one reference image for each of the film’s eight characters. The account says each image cost two credits at 2K resolution and all eight were made in under two minutes. Meyer describes these sheets as a way to reduce character drift, such as a hairstyle or costume changing between shots.
For the first version, the team planned 20 shots, made a starting frame for each with the character references, and animated them using Veo 3.1 fast mode. The generated clips included speech and sound. After reviewing the result, the team compared its cut count with the short that inspired it and revised the pacing. The final approach grouped the film into three longer segments generated using a single model with the complete cast as references.
The account also describes using OpenAI’s Whisper speech-recognition model to check generated dialogue against the script. Meyer says that process caught the spelled-out “H-E-Y” and helped prompt a retake when the ghost’s final line contained extra, garbled syllables. The reported workflow combines generated images, video and sound with human scripting, selection, editing and grading.
“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 Establish
The production details are presented by Meyer, and the source account does not include an independent assessment of the final film, the credit estimate or the claimed comparison with the inspiration short. The short’s stated cut counts and average shot lengths are Meyer’s measurements; how those cuts were counted is not explained in the supplied material.
It is also unclear which model produced the final three segments, beyond the account’s description of using one model with references for the cast. The supplied source ends partway through a description of a revised starting frame, so it does not provide the complete behind-the-scenes record promised in its introduction. No details are given about audience response, distribution beyond the linked YouTube video, or whether the team plans further versions.
The Short and Its Production Record
Hollow House is presented as a completed short, and Meyer’s account links to the 1:37 film on YouTube alongside production images and descriptions of the iterations. The next point readers can verify is the finished video itself and any further material Meyer publishes about the work. The source gives no release schedule for additional versions or a follow-up project.
For now, the production account documents one overnight experiment rather than a controlled comparison of AI filmmaking tools. Any broader conclusions about model performance, cost or the value of a particular workflow would require more projects and independently checked results.
Key Questions
What is Hollow House?
Hollow House is a 97-second horror-comedy about a real ghost who joins the staff of an extreme haunted attraction. Its creator says the characters, voices and sets were AI-generated.
Why were there three versions?
Creator Thorsten Meyer says the first version looked good but felt boring, with 20 shots averaging 5.6 seconds per cut. The team revised the pacing and workflow twice before settling on a 97-second final cut.
How much did making the short cost?
Meyer reports using 2,220 Higgsfield credits, estimated at about $95 at the team’s plan price, plus roughly 2,100 ElevenLabs characters for music and sound effects. These are figures from the creator’s account, not independently audited totals.
Were the film’s story and jokes copied?
Meyer says the structure was inspired by an AI horror-comedy short shared on X, but describes Hollow House’s script, characters and jokes as original. The supplied account does not identify the other short.
Source: ThorstenMeyerAI.com
