Pocobello: a short film made entirely with Wan 3
Every moving image in our short film Pocobello comes from Wan 3. A production report on what the model does well, where it struggles and what stays human work.
Miraki MindAI-assisted editorial · 5 November 2026 · 3 min read
Watch the videoPocobello is a short film we made ourselves, for the TopView Wan 3 Challenge. Every moving image in it comes from Wan 3. Above, Miraki Mind tells in two and a half minutes how working with the model actually went, presented by an AI-generated character. Here is the same in writing, with images from the production.
The story is simple. An empty promenade out of season, one last box of popcorn, and two people who are far too polite to take it. Above them, a narrator who simply cannot keep out of it.
Images first
Characters, places and props were designed as stills first, long before anything moved. By the time everything was in place, there were 199 images and 20 character sheets. Oswald, the narrator, went through seven versions before he looked right. The popcorn van went through six. And one image had to be made on a local machine, because of the blood.
The van shows how much happens before the first moving image. A cream delivery van became a striped popcorn van, in the colours of the apron worn by Beatrix, who sells from it.
Omni-reference or start frame
For most shots we worked with Wan’s omni-reference: characters and places as reference images, and Wan was left to find the framing itself. Most of the time, that did not work cleanly. So we often went back to a start frame, where framing, lens and light were already set.
One lesson came early. Short generations work far better than long ones. If you come from Seedance 2.5, that is a different way of thinking.
Where it struggles
Hands and objects morphed into one another. Characters got mixed up. And people rarely stayed where we had placed them. Blocking is not where Wan 3 is strongest, and if you are used to Seedance 2.5, you notice quickly.
Longer prompts did not fix that. What helped was a new keyframe, or a cut around the difficult moment. Faces, costumes and colours were kept consistent by people, shot by shot.
The model moves. People decide.
Where Wan 3 shines
In character. Wan 3 painted these people with real care, and the mood of the film comes through. Even speech and dialogue worked surprisingly well. Afterwards we made the voices consistent in ElevenLabs, so that each character sounds the same throughout the film. All characters and voices are invented, no real person, no cloned voice.
What stays human work
The images, the motion and the voices are generated. The writing, the direction, the edit, the colour and the sound are human work.
At the TopView Wan 3 Challenge, among thousands of entries, Pocobello received the Outstanding Visual Achievement Award. We were rather pleased.