Why Reroll
Traditional Director Tools
3D scene → rendered previs video
Traditional director tools are harder for beginners. They may require building a scene in 3D software and rendering a previs video. Not everyone knows 3D, and the process takes time.
From Image to Editable Scene
YOLOv8-seg separates subjects, while DPT estimates relative depth. Together they turn one image into editable scene data for the iPhone.
Update — SAM 3
Tests showed that SAM 3 produces cleaner segmentation than YOLOv8-seg. Its masks preserve each object's shape, so the AR camera displays object silhouettes instead of boxes.
Direct the Scene in AR
Place the detected scene on a surface, then edit each object. The iPhone becomes the camera: move through the space to test angle, framing, and motion.
A Unity test validated the JSON pipeline. The prototype then moved to ARKit and RealityKit for a phone-first experience.
Speak Your Direction
Apple's Speech framework uses SFSpeechRecognizer to transcribe each direction. Voice adds action, mood, light, and timing, while every note stays linked to the selected object.
Final Refined Prompt
An AI agent combines scene objects, camera movement, and the user's edit tips into one refined prompt. That prompt guides the AI video model.
Scene Objects
Masks, position, scale, and depth
Camera Direction
Angle, framing, and movement path
Edit Tips
Voice notes for action, mood, and light
