Three films for a Hajj exhibition, built in AI.
Three looping short films for a physical exhibition in Saudi Arabia, generated shot by shot and held to a hard religious and cultural rulebook. On one of the most scrutinized subjects on earth, for a government-facing audience, looks good enough was never the bar.
Zero room for error
A cultural exhibition in Saudi Arabia needed three storytelling films to run on loop in a physical space. One narrated piece following the pilgrim's journey. One silent, numbers-only film built as a four-quadrant split screen around the services that carry the pilgrimage. One narrated hospitality story, moving from scarcity to mercy.
The subject matter allowed no mistakes, and the audience was government-facing. Every frame had to be right before it was allowed to be beautiful.
Identity locked before motion
Every character, environment, and prop started as a multi-panel reference sheet locking the same face, dress, geometry, and palette across every angle. Those sheets anchored the video generation, so a pilgrim in scene fourteen is the same person who set out in scene one.
Around the generated footage sat traditional post: editing and grade, a Najdi Arabic voiceover, and real archival inserts and real Arabic calligraphy composited by hand. Over two hundred generated clips were delivered across the three films, built from a bank of approved reference sheets, with three full client feedback rounds and a hard approval gate before any animation began.
The identity locks
Four of the locked reference sheets behind the films. One pilgrim per region, two dress variants each, the same face across every angle and every scene she appears in.
Correctness in the prompt, not the crop
The rulebook was literal. No sacred figures. Ritual dress had to match a canonical seven-stage timeline. One holy site has dense tent grids, its neighbor has none. Arabic was always real calligraphy, placed in post by an Arabic-literate designer.
The exterior plate of the Grand Mosque took more than ten regeneration passes to match the real, current-day architecture against client references. We do not ship until it matches reality.
The engine's job ended where cultural judgment began.
Fighting the filter
The strangest production problem: the video model's own safety filter kept rejecting legitimate religious imagery. Pilgrims in traditional Ihram dress read to the moderation system as something to censor, and place names alone could trip it. Roughly a third of first-pass generations were wrongly flagged.
We reverse-engineered a playbook: strip religion-coded nouns even from negative prompts, refresh flagged references, abstract sensitive place descriptions into neutral language. And when two group portraits would not clear the filter no matter what, we did not fake it or give up. We composited the group in the edit from individually generated portraits. A mature AI production knows when to fight the model and when to solve it in the edit suite.
Three films, signed off, on the wall
All three films were delivered, approved by the client, and run in the physical exhibition. The final cuts will land on this page once cleared for web.
The takeaway is not speed. It is discipline: a rulebook every prompt had to satisfy, real archival and real calligraphy done by hand, and the judgment to know where the model's job ended and ours began. Taste and correctness cannot be automated. Everything else around them can.
Working on something this sensitive?
If your brief has cultural, religious, or institutional stakes where good enough is not good enough, this is the discipline it takes. Fifteen minutes.
Book a 15 minute call