AI hero-video pipeline

A good hero video makes a hotel homepage feel alive, but commissioning one means a videographer, a shoot day, and a serious bill. Most independent properties never get one. I wanted a hero video built from the photography we already own, on demand.

How it works

Image-to-video with Google's Veo model, driven from Python. The key call is image-to-video from a property's real photos, not text-to-video, which cheerfully invents a fake building. A locking prompt holds the architecture and sign text steady while only clouds, foliage, and light move. Assembly runs on a static ffmpeg (via imageio-ffmpeg, since there is no Homebrew on the machine): xfade crossfades between shots, a reprised opening for a near-seamless loop, exported as a small VP9 webm with an H.264 mp4 fallback and a poster frame.

• Image-to-video from real photos, so the geography and motion are authentic
• A lock prompt, so signage and buildings do not morph or distort
• webm-first with an mp4 fallback, served autoplay-muted-loop as the hero
• A trivial per-clip cost: no shoot, no crew, no location day

The first hero is delivered and live on one property's site. The generation scripts are ready to productionize into a repo, so the same treatment rolls across the rest of the portfolio, each of which already has its own professional photo library to draw from.

This is a small video-production house replaced by an API and an ffmpeg pipeline. It will not out-shoot a real crew, but it ships a genuine, on-brand hero for next to nothing, and once it is a repo one owner covers the whole portfolio's video seat. The exception shrinks from "who do we hire" to "which photos."