Dan Osterman— AI Product Engineer & Creative Technologist

Case 01 · An AI video platform for filmmakers

Control in the shot

The problemI can't direct a prompt. I need to block the scene and move the camera, then generate.

The source keyframe used to build the 3D blocking scene
01.1The source image for the scene build
View full size
01.2Choose the framing with the 50 mm camera and precise movement controls
Play full video
A frame from the previz clip
01.3The previz the stage exported, sent with the shot as a video reference
View full size
A frame from the Seedance result generated with the previz reference
01.4What Seedance made from it. Framing and lighting drift from the previz
View full size

Notes · Case 01

In short

A 3D stage that opens from any storyboard shot, with proxy characters and a camera the director can move, and a previz clip that the video model is conditioned on. Blender users get the same workflow inside Blender.

Role
Designed and built the 3D stage and the Blender integration
Client
An AI video platform for filmmakers
Year
2026
Stack
Three.js · Blender extension · image-to-video and reference-to-video models on fal

What came out

  • A shot image becomes an editable proxy scene with a camera the director can move
  • The camera path renders to a previz clip that can be attached to video generation
  • Blender users keep their own tools and send the clip back without leaving Blender

Where it started

Every director who tried generative video said a version of the same thing. The output was often beautiful and almost never the shot they had in their head. They could describe a camera move in the prompt and get a different one. They could ask for two characters on the left and get three in the middle. Rerolling is not directing.

What I built

The 3D stage opens from any shot on the storyboard. It drops the shot’s characters in as proxies, gives the director a camera, and lets them block the scene the way they would on set: who stands where, where the camera is, how it moves. When they’re happy, the stage renders a short previz clip.

That clip goes into generation as a reference, alongside the shot’s images and prompt. The director picks any model that takes a video reference; here it’s Seedance 2.5 in Reference mode. The previz steers the model rather than dictating to it, and the side-by-side below shows what carried over and what drifted.

For people who already live in Blender, I built the same loop as a Blender add-on. Open the shot from the platform, build or animate the scene with the assistant or by hand, send the clip back, and generate with it attached. Nothing to export or re-upload.

What was hard

The previz has to be good enough to steer the model and cheap enough to make in a minute. Proxy characters, not finished ones. A real camera, not a guess. And the whole thing had to feel like a step inside the shot, not a separate app, or directors would skip it.