The Way Home: an AI film pipeline
The Way Home is a short film I made on my own with AI in 2026. I built it as a chain of small apps in Google Opal: each step has one job and its own output, a review scores the plan before anything is made, and repair steps fix only what the review flags.
A short film, made through ten small AI apps
A four minute sci-fi short, made by one person. Watch it first. Below is the system that kept the same faces and places from the first shot to the last, and the review that caught three errors before any shot was made.
One person, one film, and an AI that forgets
Image tools don’t remember the last shot. A prisoner escapes, gets home to his family and grows old with them, until his world starts to glitch. Every shot was generated, so Jack had to look like Jack in his 40s and in his 70s, and the facility had to be the same white room at the start and the end.
The brief
A cinematic short film made by one person, with the same characters and places from the first shot to the last.
My role
I designed and built the pipeline, ran every step, and made the final edit.
Tools
Small AI apps I built in Google Opal, Higgsfield for the images and video, and Suno for the music.
Year
2026, about four minutes long.
Ten steps, each with one job
Instead of one long prompt, the film runs through a chain of small apps. Each one takes the previous step’s output, does one job and hands over a document the next step can read. Every document is something I can check before the work moves on.
Plan
- 01Script refinerScreenplay and production report
- 02Character castingCasting sheet and angle studies
- 03World buildingLocation guide
Check and repair
- 04Greenlight reviewAudits all three together and scores the plan
- 05RepairScript, characters and locations, only what was flagged
Make
- 06Asset validatorEach image against its profile
- 07Shot listShot by shot, with a pacing review
- 08ShotsGenerated in Higgsfield
- 09ScoreMusic cues made in Suno
- 10EditThe final cut
Check the whole plan before a single shot
The greenlight step reads the script, the casting sheet and the location guide together, and scores the plan before anything is made. It came back Yellow: “Minor logic/asset issues identified; requires asset updates before filming.” Each problem would otherwise have shown up as a broken shot, when it costs the most to fix.
Scene 1 · Visual logic
The script has red alarms flashing, but the location guide sets the facility as monochromatic and shadowless.
Scene 7 · Character logic
The story spans decades, but the casting sheet only describes Jack in his mid 40s.
Missing asset
Scene 3 sends the family into drainage tunnels that have no entry in the location guide.
Fix only what the review flagged
Three repair editors take the report and change only what it points to. The script editor works to one rule: fix the flagged errors and keep every other line as it was. It ends by confirming “Fixed 2 errors. 100% of original dialogue and action preserved.”
Script
Two fixes. Every other line unchanged.
Characters
An elderly Jack added to the casting sheet, marked “Updated per Greenlight Report”.
Locations
The facility now allows red emergency light, and the drainage tunnels have their own entry.
Check every image, then shoot
Before an image becomes a shot, the asset validator compares it with its profile: age, scars, clothing texture, lighting and camera angle. Then the shot list breaks the script down, the shots are made in Higgsfield, the music in Suno, and the edit finishes the film.
What the AI does
Writes, casts, describes, audits, repairs and generates, one step at a time.
What I decide
What each step does, the rules it follows and what it hands to the next one, and every cut in the edit.
Result
A finished short film, about four minutes long.
Fix it on paper, not in the shot
The plan took an evening. The shots took a week. These numbers come from the project files, and they changed how I build every AI system since: check the plan before the expensive part starts.
- 4 hoursto plan, check and repairThe script, casting, locations, review and all three repairs ran in one evening. Making the shots took the next six days. The cheap part is where the errors were caught.
- 35 clipsfor 22 planned shotsOne generated clip rarely held a whole moment, so many shots became two or three short clips. Plan for more clips than shots.
- 3 cutsbefore the final editTwo full edits were set aside before this one. The masked figures were also redesigned after the first pass.
- Next timecheck the video, not just the stillsThe validator compared still images with their profiles. What still drifted happened in motion, so the next version checks the clips too.