Research map / Video / Video generation
Unveiling the Value of Motion for Cinematic Camera Trajectories
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DirSpeed, a motion-centric camera trajectory representation (normalized direction plus log-speed), shows it improves text-trajectory alignment and generation, and introduces the CineGEN masked autoregressive generator and the CineScript dataset.
Problem
Text-conditioned camera trajectory generation and trajectory-text alignment rely on absolute per-frame pose representations, which entangle direction, speed and location, making language alignment hard; prior alignment evaluation (CLaTr) also mixes contrastive and reconstruction objectives.
Approach
Replace absolute per-frame poses with frame-to-frame translational and rotational velocities decomposed into unit directions and log-speeds (DirSpeed), paired with a continuous masked autoregressive generator (CineGEN) using variance-guided unmasking, and a purely contrastive evaluator.
Key results
- DirSpeed raises CLaTr-style-protocol R@1 from 17.8 (Pose9D) to 25.2 on motion-caption retrieval with the contrastive evaluator.
- CineGEN reaches R@1 3.35 vs 0.93 for retrained GenDoP and reduces FCD from 22.10 to 6.77 on CineScript.
- In a blinded human study, CineGEN is selected in 70.1% of judgements versus 28.9% for the strongest baseline GenDoP.
Surprising findings
- Retrofitting baselines (CCD, E.T., GenDoP) with DirSpeed output improves motion F1 and alignment but can worsen FCD and Coverage for GenDoP.
- A caption-only CineGEN variant achieves higher alignment (R@1 5.82) than the full model with logline conditioning (R@1 3.35).
Limitations
- Attribute probes achieve only modest absolute F1, so high-level style signal is weak.
- Many generated metrics depend on self-trained evaluator spaces; AlignScore is not comparable across evaluators.
- Loglines are underdetermined and only partly aid generation; movie-attribute subset is small (3.2K clips).
Experimental setup
Datasets: CineScript ShotBench CineTechBench MovieShots CMD VADB CameraBench
Benchmarks: CineScript
Metrics: R@1 R@5 R@10 MedR AlignScore motion-primitive F1 FCD Coverage macro-F1 attribute probes
Baselines: CCD E.T. GenDoP CLaTr
Builds on: masked autoregressive models (MAR) DDPM CLaTr E.T. GenDoP
Closest papers by meaning
- CineTechBench: A Benchmark for Cinematographic Technique Understanding and Generation (NeurIPS 2025 · similarity 0.89)
- Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation (NeurIPS 2026 · similarity 0.89)
- CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation (NeurIPS 2026 · similarity 0.87)
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.87)
- Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text (NeurIPS 2024 · similarity 0.87)
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.