OmniCamera: A Unified Framework for Multi-task Video Generation with Arbitrary Camera Control

Yukun Wang; Ruihuang Li; Jiale Tao; Shiyuan Yang; Liyi Chen; Zhantao Yang; Handz; Yulan Guo; Shuai Shao; Qinglin Lu

doi:10.48550/arXiv.2604.06010

One model · Nine tasks · Arbitrary camera control

OmniCamera

A Unified Framework for Multi-task Video Generation with Arbitrary Camera Control

Yukun Wang^1,2,* Ruihuang Li^2,† Jiale Tao² Shiyuan Yang^2,3 Liyi Chen^2,4 Zhantao Yang² Handz² Yulan Guo^1,† Shuai Shao² Qinglin Lu²

¹ Sun Yat-sen University · ² Hunyuan, Tencent · ³ CityU · ⁴ PolyU

^* Work done during an internship at Tencent Hunyuan. ^† Corresponding authors.

Explore video results Read the paper Code

01 Camera control

Camera condition

Motion Text01 / 01

Click or scroll to switch

Content condition

Source Video01 / 01

Click or scroll to switch

All results 01 / 01

Motion Text×Source Video

Dolly in

Re-camera an existing video with natural-language camera commands while preserving the source content.

02 Content input

The idea

Decouple what happens
from how it is seen.

Video intertwines two axes: the dynamic content of a scene and the camera motion through which it is observed. OmniCamera explicitly disentangles and commands both.

3 × 3control combinations

830K500K synthetic + 330K real videos

50predefined motion types

5Bparameter backbone

Method

Why one model handles all nine tasks.

Three ideas make the video control stable, precise, and photorealistic.

Unified architecture

Decoupled injection & Condition RoPE

Text prompts and visual conditions are fused through joint attention, while explicit camera trajectories are injected through a dedicated MLP pathway. 3D Condition RoPE assigns modality-specific spatial-temporal offsets to reduce interference.

OmniCAM dataset

Precision meets realism

Accurate UE5 trajectories provide geometric supervision; curated real videos restore photorealism and broaden scene diversity.

Training strategy

Dual-level curriculum

Training progresses from text to reference video to trajectory control, then from precise synthetic supervision to photorealistic real-world adaptation. Dual-condition classifier-free guidance provides separate scales for semantic text and camera-motion conditions.

Current scope. Finer-grained controls such as multiple reference images and localized motion guidance remain future work.

Interactive results

Nine tasks, explored in one place.

Each case exposes its camera condition, content condition, and generated result together. Videos load as they enter the viewport.

01 / 09

Text-controlled V2V

0 results

Citation

Build on OmniCamera.

If this work supports your research, please cite the paper.

View on arXiv

@misc{wang2026omnicamera,
  title         = {OmniCamera: A Unified Framework for Multi-task Video Generation with Arbitrary Camera Control},
  author        = {Wang, Yukun and Li, Ruihuang and Tao, Jiale and Yang, Shiyuan and Chen, Liyi and Yang, Zhantao and Handz and Guo, Yulan and Shao, Shuai and Lu, Qinglin},
  year          = {2026},
  eprint        = {2604.06010},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2604.06010}
}