Summary
Dutch filmmaker Youri van Hofwegen presents a complete professional workflow for producing coherent AI-generated short films — addressing the most common failure mode where AI videos look like disconnected clips rather than a single unified piece.
The process begins before any generation happens. Van Hofwegen uses Claude to develop a story with intentional design constraints: keeping technically difficult elements (ski tricks, crowd scenes) in distant wide shots where models don’t need to resolve fine detail, and restricting dialogue to audio-only to avoid animating additional on-screen characters. Once the story is locked, he builds reference assets in OpenArt using GPT Image 2 — a character sheet showing the subject from multiple angles, and empty location images for each scene. The deliberate absence of people in location references preserves control over who appears in each shot during video generation.
For video generation, the tutorial uses Cedance 2 (integrated into OpenArt), with prompts written by Claude combining character and location reference images with detailed cinematic direction. The video walks through shot sequencing, camera language, and how transitions are planned at the asset-building stage rather than fixed in post. A key principle throughout is that every decision made before generation reduces the number of variables the model must handle, and that pre-checking all reference assets before building shots on top of them prevents compounding errors. Viewers leave with a repeatable system applicable to any short-form AI film project.
📺 Source: Youri van Hofwegen · Published August 09, 2026
🏷️ Format: Tutorial Demo
