Source-linked AI summary

ShotVerse: Advancing Cinematic Camera Control for Text-Driven Multi-Shot Video Creation

Songlin Yang, Zhe Wang, Xuyi Yang, Songchun Zhang, Xianghao Kong, Taiyi Wu, Xiaotong Zhao, Ran Zhang, Alan Zhao, Anyi Rao

arXiv:2603.11421v2cs.CV

TL;DR

Cinematic multi-shot video generation lacks precise, practical camera control because textual instructions are unreliable and explicit trajectories are laborious and difficult to execute. ShotVerse addresses this with aligned data, a VLM Planner, and a camera-conditioned Controller, achieving camera-accurate, cross-shot-consistent videos with superior aesthetics while remaining limited to single-scene multi-shot generation.

  • Problem

    Cinematic multi-shot camera control remains difficult because textual conditions are unreliable, while manual explicit trajectory plotting is laborious and complex trajectories can fail during execution.

  • Method

    ShotVerse uses aligned (Caption, Trajectory, Video) data to decouple VLM-based cinematic trajectory planning from camera-conditioned holistic video rendering.

  • Results

    ShotVerse achieves the lowest FVD and highest Aesthetic Quality (5.465), while its Controller achieves the lowest translation and rotation errors and highest CAS.

  • Takeaways & Limitations

    ShotVerse bridges unreliable textual control and labor-intensive manual plotting, producing multi-shot videos that are camera-accurate and cross-shot consistent.

  • Takeaways & Limitations

    The current method is limited to single-scene, multi-shot video, constraining duration and flexibility; longer multi-scene generation remains future work.

Abstract

from arXiv · show

Text-driven video generation has democratized film creation, but camera control in cinematic multi-shot scenarios remains a significant block. Implicit textual prompts lack precision, while explicit trajectory conditioning imposes prohibitive manual overhead and often triggers execution failures in current models. To overcome this bottleneck, we propose a data-centric paradigm shift, positing that aligned (Caption, Trajectory, Video) triplets form an inherent joint distribution that can connect automated plotting and precise execution. Guided by this insight, we present ShotVerse, a ``Plan-then-Control'' framework that decouples generation into two collaborative agents: a VLM (Vision-Language Model)-based Planner that leverages spatial priors to obtain cinematic, globally aligned trajectories from text, and a Controller that renders these trajectories into multi-shot video content via a camera adapter. Central to our approach is the construction of a data foundation: we design an automated multi-shot camera calibration pipeline aligns disjoint single-shot trajectories into a unified global coordinate system. This facilitates the curation of ShotVerse-Bench, a high-fidelity cinematic dataset with a three-track evaluation protocol that serves as the bedrock for our framework. Extensive experiments demonstrate that ShotVerse effectively bridges the gap between unreliable textual control and labor-intensive manual plotting, achieving superior cinematic aesthetics and generating multi-shot videos that are both camera-accurate and cross-shot consistent.

1 Introduction

ShotVerse addresses the difficulty of precise, cinematic multi-shot camera control by separating trajectory planning from video rendering and grounding both in aligned data.

  • Existing methods struggle with textual camera control, unified multi-shot coordinates, and trajectories that follow cinematic patterns.
  • Manual cinematic trajectory plotting imposes a prohibitive design burden because poses must align with narrative flow in a global coordinate system.
  • Complex cinematic trajectories can fall outside camera-controlled video models’ training distribution, causing generation failures.
  • Aligned (Caption, Trajectory, Video) triplets support separate optimization of trajectory planning and video generation.This data-centric view models P(Trajectory | Caption) and P(Video | Caption, Trajectory), avoiding unstable joint training while preserving compatibility.
  • ShotVerse uses a VLM Planner for explicit cinematic trajectories and a Controller with a camera adapter for high-fidelity multi-shot rendering.The two agents share the aligned data distribution and the Controller uses a holistic multi-shot backbone.
  • ShotVerse-Bench aligns disjoint trajectories through automated camera calibration and evaluates planning, execution fidelity, and multi-shot consistency.

2 Related Work

Related work spans multi-shot generation and camera-control datasets, while ShotVerse emphasizes holistic generation and automated cinematic plotting from textual descriptions.

  • Multi-Shot Video Creation: Cascaded multi-shot approaches struggle with consistency across temporal gaps, while memory-based methods face error accumulation.
  • Camera Trajectory Datasets: ShotVerse-Bench is presented as a unified trajectory dataset supporting cinematic camera-control research.
  • Multi-Shot Video Creation: Holistic generation models entire video sequences jointly instead of iteratively synthesizing separate shots.
  • Camera Control for Text-Driven Video Generation: Earlier camera-control methods encode camera extrinsics or use 3D priors, but they lack multishot storytelling capability.

3 Methodology: ShotVerse

ShotVerse’s methodology combines hierarchical, VLM-based trajectory planning with camera-conditioned holistic video generation, using unified global coordinates and cross-shot temporal modeling.

  • Overview: ShotVerse decouples multi-shot camera control into a Planner that generates trajectories and a Controller that renders them into video.
  • Planner: Shot-Aware Cinematic Trajectory Plotting: The Planner represents each shot as a trajectory of poses in a unified global coordinate system.Camera poses are represented with translation and rotation parameters and converted through reversible tokenization and de-tokenization.
  • Planner: Shot-Aware Cinematic Trajectory Plotting: Hierarchical prompts combine global and per-shot text with learnable query placeholders that provide slots for shot-specific camera plans.
  • Planner: Shot-Aware Cinematic Trajectory Plotting: VLM query-token states aggregate global context, previous shots, and current-shot instructions to encode context-aware camera codes.
  • Planner: Shot-Aware Cinematic Trajectory Plotting: A lightweight autoregressive Transformer decodes concatenated shot codes into variable-length trajectory-token sequences with temporal modeling across shots.
  • Controller: Cinematic Camera Control: The Controller injects camera features into the video model and uses 4D Rotary Positional Embedding to encode shot-aware structure.
  • Controller: Cinematic Camera Control: The controller objective conditions flow matching on textual embeddings and explicit camera-condition embeddings.

4 Dataset and Benchmark: ShotVerse-Bench

ShotVerse-Bench provides aligned, globally unified multi-shot camera trajectories with hierarchical captions for training and evaluating cinematic video generation. Its three-track protocol separately assesses planning, execution, and end-to-end quality.

  • Dataset Curation: ShotVerse-Bench contains 20,500 high-production clips with hierarchical captions and unified multi-shot trajectories.The dataset spans 27+ fine-grained trajectory categories, multiple framing scales, explicit transitions, and four content domains.
  • Dataset Curation: The dataset includes diverse camera motions, compositions, shot transitions, durations, and content domains.Examples include 1,639 zoom/scale samples, 1,720 explicit cuts or transitions, and clips averaging 5.50 seconds.
  • Evaluation Benchmark: The three-track protocol evaluates text-to-trajectory planning, trajectory-to-video execution fidelity, and end-to-end text-to-video generation quality.Track A uses hierarchical prompts and globally aligned camera sequences, while Track B evaluates execution from trajectories.
  • Dataset Curation: ShotVerse-Bench is presented as the first large-scale dataset combining multi-shot cinematic camera trajectories with rich, multi-level caption annotations.This alignment addresses the lack of semantic descriptions paired with globally unified camera trajectory annotations.
  • Evaluation Benchmark: Track A compares representative trajectory planners, while Track C includes open-source and closed-source multi-shot video models.The benchmark adapts trajectory-generation baselines for multi-shot evaluation using the proposed calibration pipeline.

5 Experiments

Experiments evaluate ShotVerse across alignment, camera-control fidelity, cinematic quality, and component ablations. The method achieves strong benchmark performance while error analysis identifies limits in long-context persistence, dense scenes, and single-scene scope.

  • Benchmark Results: ShotVerse’s three-track evaluation measures alignment, control, and cinematic quality using rendered outputs, VLM scoring, and human studies.The evaluation emphasizes rendered video because numerical trajectory metrics may not fully reflect cinematographic experience.
  • Benchmark Results: The VLM-driven Planner achieves the best results on both evaluated benchmarks, indicating stronger cross-domain generalization than the compared planners.GenDoP performs competitively on its native benchmark but shows a domain gap after retraining on ShotVerse-Bench.
  • Benchmark Results: The Controller achieves the lowest translation and rotation errors and the highest CAS among compared methods under ground-truth trajectory control.These results suggest stronger cross-shot consistency under the CAS proxy.
  • Benchmark Results: ShotVerse achieves the lowest FVD and an Aesthetic Quality score of 5.465 in end-to-end text-to-video evaluation.Its 4D RoPE raises Shot Transition Accuracy to 0.933, compared with 0.927 for MultiShotMaster and 0.645 for HoloCine.
  • Ablation Studies: Ablations show that removing the VLM encoder degrades F1-Score and CLaTr-CLIP, supporting the role of semantic-spatial priors in trajectory planning.The reported degradation compromises narrative alignment and trajectory-text consistency.
  • Generalizability: Long shot-reverse-shot sequences and high-density crowds remain difficult, producing occasional off-target trajectories, background drift, or subject flickering.These failures arise in settings involving long-context persistence, many independently moving subjects, and complex occlusions.
  • Generalizability: The method is limited to single-scene, multi-shot video, constraining duration and flexibility despite compatibility with autoregressive generation.Extending the approach to longer, multi-scene generation remains future work.

6 Conclusions

ShotVerse decouples cinematic multi-shot generation into VLM-driven trajectory planning and geometry-aware rendering, supported by the ShotVerse-Bench dataset. Its evaluation reports technical camera accuracy alongside cinematic pacing and visual salience.

  • 6 Conclusions: ShotVerse uses a Plan-then-Control framework that separates VLM-driven cognitive plotting from geometry-aware video rendering.This decoupling targets the spatial logic required for multi-shot storytelling.
  • 6 Conclusions: ShotVerse-Bench unifies disjoint shot trajectories into a global coordinate system through a novel calibration pipeline.The dataset is presented as a high-fidelity foundation for teaching cinematic structure.
  • 6 Conclusions: A three-track evaluation protocol assesses ShotVerse's technical accuracy and its implicit understanding of cinematic pacing and visual salience.The conclusion frames these evaluations as confirming both geometric and cinematic capabilities.
Loading 2603.11421v2…