Student Project Showcase logoStudent Project Showcase
← Back to all projects
Generative MediaAug 2026

Hong Kong Reimagined: An AI-Generated Travel Film

by Lau Shing Long

Google VeoGeminiSunoLyria 3Google VidsModel EvaluationPrompt EngineeringCantoneseAI Music

A travel film is one of the hardest things to make well, because the genre is so worn — every city has its thousand drone shots and its swelling soundtrack. "Hong Kong Reimagined" sidesteps that by making one decision early and holding it: the film is scored in **Cantonese**.

That choice does more work than it appears. Generative music tools are English-first — training data, prompt conventions and best results all lean that way — so an original Cantonese song is genuinely harder than taking the default. It also changes what the film *is*. A Hong Kong travel film scored in English is made for people looking at the city from outside. One that sings in Cantonese is made from inside it, in the language actually spoken on the streets being shown.

**The most professional thing here is what happened before production started.** Rather than using the first tool to hand, Lau ran a bake-off: HeyGen, Seedance, Higgsfield and Kling were each trialled against the same brief before Google Veo was chosen for the final generation. Those tools are built for different jobs — avatar and presenter video, cinematic camera motion, general text-to-video — and testing them against your own material rather than against a demo reel is how studios actually pick a pipeline. Most people skip this step and let their first tool quietly define what their film can be. Choosing a tool *after* you know what you need is the difference between using AI and directing it.

**The finished pipeline gives each tool one job.** Gemini acts as the writers' room — generating prompts, drafting the lyrics, breaking the piece into scenes, and producing still images to art-direct the look before anything moves. Google Veo turns those scenes into footage. Suno and Lyria 3 handle the music, and running two music models rather than one means each cue can be auditioned twice and the better take kept — the same reason a composer writes alternates. Google Vids assembles the final edit.

Settling the still image first and the motion second is a quietly important habit: a frame you can iterate on cheaply is far easier to art-direct than a clip, and locking the look before generating movement is how a film stays visually consistent across shots.

The word in the title carries the rest. This isn't documentary footage of Hong Kong — it's Hong Kong **re-imagined**, generated rather than filmed. That's the honest framing for AI travel media, and it opens the question the whole piece rests on: what does a city look like rendered from a description of itself rather than photographed? Treating that as the point rather than as a limitation is what separates AI standing in for a camera from AI doing something a camera can't.

More like this

Ready to start building?

Book a free trial class with Harvey.

Enroll