CinemaTraj: LLM Agent Plans Cinematic Camera Trajectories in 3D Scenes

joemeno · x · 2026-08-02

Researchers from TU Munich and Huawei Dresden propose CinemaTraj, a framework that reframes camera trajectory planning as language-grounded spatial reasoning. Given RGB-D images and a user prompt, an LLM agent uses a structured 3D scene graph to decompose the request into atomic cinematographic movements (dolly, orbit, crane, pan, tilt, zoom, arc), and plans collision-free trajectories. It generates cinematic videos with synchronized voiceover and subtitles, with applications in real-estate ads and virtual tours. Accepted at ACM Multimedia 2026.

Original post →

More from Multimodal

Multimodal channel →