Finsbury

The problem. Turning a catalogue of still product photos into short, consistent videos, shoe after shoe, without the leather warping, the frame drifting or the motion feeling random.

Prompts that describe what you want to see. Naming a defect in a prompt tends to invoke it, so every prompt was rewritten in positive form: no lists of things to avoid, only the state the frame should reach. "The leather and laces stay firm, smooth and structured." "The air is perfectly still and completely empty." The camera works the same way: a vague adjective barely moves it, a visible destination does. "Until the shoes fill the lower half of the frame, and the camera holds that distance."

A pipeline, not a batch of prompts. Models were tested shot by shot and chosen on results and cost. Faces are handled before generation, by reframing the reference image, never by cropping the video afterwards. Every fix that works becomes a permanent step of the chain, with the condition that triggers it. Colour grading is tuned per material, and upscaling only happens on validated videos.

The eye decides. Automated detectors flag issues and a review tool lines the clips up side by side, but no threshold decides on its own: every clip is validated by eye. A shared registry tracks each product's status, so a validated video is never regenerated by accident.