🔍 Read the full analysis: The Boring First Version Behind Our AI Horror-Comedy, Hollow House on ThorstenMeyerAI.com
Get audio and creator gear 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 the 97-second AI horror-comedy Hollow House in three versions between the evening of Oct. 27 and about 1:30 a.m. on Oct. 28. The first cut looked polished but felt boring, prompting a shift from 20 longer shots to a faster, three-segment edit. The account documents the workflow and its reported costs, but does not independently verify them.
Thorsten Meyer says he made Hollow House, a 97-second AI-generated horror-comedy, in three versions over one night, changing its structure after the first cut looked polished but felt “terribly boring and amateurish.” In a behind-the-scenes account published Oct. 28, Meyer describes how faster cuts and a single video-generation model supplied with the full cast’s references shaped the version he released.
Meyer’s account says work began in the early evening of Oct. 27 and continued until about 1:30 a.m. the next day. The first edit used 20 shots and ran 2 minutes 15 seconds. The second used 66 fast cuts and ran 2 minutes 21 seconds. The final short runs 1 minute 37 seconds, arranged in three roughly 30-second segments generated with one model using the cast as references. Meyer says only the third version worked.
The film’s premise puts a ghost among the workers at an extreme haunted attraction. A guest says the safe word, “cotton candy,” but the scare does not stop; the staff then treat the unexplained figure as a workplace problem. Meyer says the characters, voices and setting were AI-generated. Before making video, he created reference images for eight characters with Google’s Nano Banana 2 image model through Higgsfield, then used those images to help maintain consistent appearances across later frames and clips.
For the first version, Meyer generated 20 starting images and animated them with Google’s Veo 3.1 fast mode, which he says produced speech and sound alongside the visuals. He compared the pacing with a short that inspired the project: his count put that film at about 74 cuts over 2 minutes 14 seconds, or roughly 1.8 seconds per image, against 5.6 seconds per cut in his first edit. He says the slower rhythm undermined the comedy. He also reports using Whisper speech recognition to check generated dialogue against the script and identify lines that needed retakes.
The Boring First Version Behind Our AI Horror-Comedy, Hollow House
Three versions in one night turned a polished but sluggish cut into a 97-second horror-comedy. Thorsten Meyer’s account traces the shift from long shots to faster timing, character references and retakes.
“The pictures were never the problem. The rhythm was.”Thorsten Meyer · Production lesson
Three versions, one overnight rethink
Meyer says the first cut looked polished but felt “terribly boring and amateurish.” The changes reshaped both pace and structure.
Polished, too slow
Twenty longer shots, static cameras and no musical score left the jokes without the pace Meyer wanted.
2 min 15 sec · 5.6 sec per cut
More cuts, still long
The edit jumped to 66 fast cuts. It moved quicker, but the final structure was still to come.
2 min 21 sec · 66 cuts
Three compact segments
Roughly 30-second segments used one video model with the cast supplied as references.
1 min 37 sec · The version that worked
Spot the mismatch
Strong images did not make the first edit feel funny or engaging.
Change the rhythm
Shorter beats brought the timing closer to the reference short’s reported pace.
Rebuild the structure
Three segments and shared cast references shaped the released version.
Comedy lives in the timing
Meyer compared his first edit with a short that inspired the project. His count showed a much quicker image rhythm in the reference.
“Terribly boring and amateurish.”
“Comedy lives on timing, and we had none.”
A practical AI production chain
The account describes a workflow where visual generation, editing choices and speech checks each played a part.
Eight character reference images were made with Google’s Nano Banana 2 through Higgsfield.
Reference images helped keep hair, clothing and appearances consistent across later shots.
The first version used Google’s Veo 3.1 fast mode for visuals, speech and sound.
Whisper speech recognition was used to compare generated dialogue with the script.
Generate
Create character sheets and starting images.
Review
Inspect clips for visual and spoken mismatches.
Adjust
Retake, reframe or solve a moment in the edit.
When the model missed the beat
Meyer describes several imperfect generations and a few practical fixes during production.
A shouted “HEY” became spelled-out letters; a 180-degree head turn moved the body too. Meyer used a still with the head already turned, then added a bone-crack sound.
Reported workaroundA ghost joins the crew at an extreme haunted attraction.
A guest says “cotton candy,” but the scare does not stop.
What the account says it used
These are figures reported by Meyer for this production, not independently checked costs or a general estimate for AI shorts.
2,220 credits
Reported Higgsfield credit use across the three versions.
About $95 at his plan price
~2,100
ElevenLabs characters reported for music and sound effects.
Approximate figure
8 characters
Created at 2K resolution, with a stated cost of two credits each.
Under two minutes, per account
What remains unverified
The production details below come from Meyer’s published account.
The account describes the process, cut counts, model behavior and resource use, but those details have not been independently verified here. The figures should be read as the creator’s report of one production.
- No full shot count for the final three-segment version.
- The exact final video model is not identified in the supplied account.
- The inspiring short shared on X is not named.
- The reported $95 may not include subscriptions, other software or labor.
Why the Edit Changed the Film
Meyer’s account points to a practical distinction in AI filmmaking: producing attractive individual images did not, in his telling, produce an effective short. He says the first version’s visuals and character consistency were strong, but its longer shots, static cameras and lack of a musical score left the jokes without the pace he wanted. The final result depended on decisions about timing and structure as much as on image generation.
That matters to viewers and creators because AI-generated video is often discussed through the tools used to make it. This production account describes a less automated part of the process: deciding that a technically polished cut was not working, changing the editing approach and checking whether generated speech matched the intended lines. Those are Meyer’s observations about this project, not proof that the same method will work for other films.
Meyer reports spending 2,220 Higgsfield credits across the three versions, which he values at about $95 at his plan price, plus approximately 2,100 ElevenLabs characters for music and sound effects. The figures give readers an estimate of this production’s reported resource use, though the account does not provide an independently checked invoice or a broader cost comparison.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
From Reference Sheets to Retakes
The production began with character reference sheets rather than animated footage. Meyer says all eight were made in under two minutes at 2K resolution, at a stated cost of two credits each. He used the images to keep details such as hair, clothing and character appearance consistent as he generated the later shots. The process reflects a recurring challenge in AI video production: a model may change a character’s features between clips unless the creator supplies visual references.
Meyer also describes limitations in the generated footage. In his account, the model spelled out the letters in a shouted “HEY,” failed to make a character turn his head 180 degrees without turning his body, and produced a less dramatic version of a planned floating scene after an earlier attempt was blocked. He says a backward-bending stunt also came out differently than intended. For the head-turn moment, he used a still image with the head already turned and added a bone-crack sound effect.
The film’s comic setup draws on extreme haunted attractions, where guests may sign waivers and use a safe word to stop scares. Meyer says the story structure was inspired by an AI horror-comedy short shared on X, while describing the script, characters and jokes as his own. The source account does not identify that earlier short by title.
““Terribly boring and amateurish.””
— Thorsten Meyer, describing his first cut
What the Account Does Not Verify
The production details, costs, cut counts and account of the model’s behavior come from Meyer’s own published report. They have not been independently verified here. The source also does not provide a full breakdown of the three-segment version’s shot count, the exact model used for that final generation, or how much time was spent on editing beyond the overnight schedule.
Meyer says the short that inspired the project was shared on X, but does not name it in the material provided. It is also unclear whether the reported $95 credit value includes every production expense, such as subscriptions, additional software or Meyer’s labor. The account describes the specific tools and figures he used, not a standardized estimate of the cost of making comparable AI shorts.
The Published Short and Its Process
Hollow House is available as a 1:37 short on YouTube, according to Meyer’s post. His account presents the production through character sheets, starting frames, failed attempts and a contact sheet of the final cut. No further release, follow-up project or additional production milestone is announced in the source material.
For now, the stated next step for readers is to view the short alongside the behind-the-scenes materials and compare the finished pacing with Meyer’s description of the earlier edits. Any further details about the final model, the costs or future versions would depend on additional information from the creator.
Key Questions
What is Hollow House?
Hollow House is a 97-second horror-comedy short in which a ghost joins the staff of an extreme haunted attraction. Meyer says its characters, voices and setting were AI-generated.
How many versions did Meyer make?
Meyer describes three versions made during one overnight production. The first ran 2:15, the second 2:21, and the final version runs 1:37.
Why did the first version change?
Meyer says the initial cut looked good but felt boring. He attributed that to pacing: it averaged 5.6 seconds per cut, compared with roughly 1.8 seconds per image in the short he used as a reference.
How much did the AI generation cost?
Meyer reports using 2,220 Higgsfield credits, valued at about $95 at his plan price, plus around 2,100 ElevenLabs characters for music and sound effects. Those figures are from his account and are not independently verified.
What remains unknown about the production?
The source does not name the short shared on X that inspired the structure, give a complete expense breakdown, or specify the exact model used for the final three-segment generation. The production figures are reported by Meyer rather than independently audited.
Source: ThorstenMeyerAI.com
