OmniCamera

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

OmniCamera is a unified camera-controllable video generation framework built on Wan2.2-TI2V-5B. It separates camera control from content conditioning and supports arbitrary pairings of three camera conditions and three content conditions.

• Paper: https://arxiv.org/abs/2604.06010
• Project page: https://yukun66.github.io/omnicamera-webdemo/
• Code: coming soon

Supported modes

Camera condition	Text content	Image content	Video content
Camera-motion text	T2V	I2V	V2V
Explicit 3D trajectory	T2V	I2V	V2V
Reference motion video	T2V	I2V	V2V

Checkpoint

joint_params_step61000.pth is a partial full-finetuning checkpoint, not a LoRA adapter and not a standalone copy of Wan2.2-TI2V-5B.

It contains BF16 parameters for:

• self-attention and cross-attention in all 30 DiT blocks;
• camera encoders in DiT blocks 0, 5, 10, 15, 20, and 25;
• camera projectors in the same six blocks.

The unchanged text encoder, VAE, tokenizer, and remaining DiT parameters must be loaded from Wan-AI/Wan2.2-TI2V-5B.

Loading contract

1. Load the Wan2.2-TI2V-5B base model.
2. For trajectory-conditioned modes, create cam_encoder: Linear(12, 3072) and projector: Linear(3072, 3072) in blocks 0, 5, 10, 15, 20, and 25.
3. Remove the leading dit. prefix from checkpoint keys.
4. Load the resulting state dictionary into the base DiT.

The original training/inference configuration uses:

• architecture version: 3.0;
• output size: 704 x 1248;
• number of frames: 41;
• output frame rate: 16 FPS;
• inference steps: 40;
• CFG scale: 5.0.

An end-to-end inference entry point and Hugging Face Space are being prepared.

License

The base model is distributed under the Apache 2.0 license. The license for this OmniCamera checkpoint will be finalized before the public release. This private repository is currently intended for deployment validation.

Citation

@article{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},
  journal = {arXiv preprint arXiv:2604.06010},
  year    = {2026}
}