WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control

       

Lei Cao^1,2* · Yu Liu^2* · Wenhua Liao² · Peng Zhang^2‡ · Fei Wang² · Hao Zhang¹ · Jiajun Le¹ · Daiguo Zhou^2† · Jian Luan² · Jiayi Ma^1,3†

¹Electronic Information School, Wuhan University   ²MiLM Plus, Xiaomi Inc.   ³School of Robotics, Wuhan University

^*Equal Contribution   ^‡Project Lead   ^†Co-corresponding Authors

Given one reference image, a target camera trajectory, and 1–4 viewpoint anchor images of the subject (from arbitrary viewpoints), WorldCast generates a video that follows the target camera motion while preserving the subject's identity across viewpoints.

• [2026.10] Initial release: inference code, bundled demo case, and project page.

• Release inference code and demo
• Release paper (arXiv)
• Release model checkpoints
• Release CLACT dataset
• Release training code

git clone https://github.com/PixCtrol/WorldCast
cd WorldCast

Python >= 3.10, a CUDA GPU (>= 48 GB recommended for the 14B model at 480x832; memory scales with resolution and frame count), bf16.

pip install -r requirements.txt

Note: WorldCast trained checkpoints (identity bank + anchor adapter) are coming soon (see TODO). The demos below additionally require them and cannot run until they are available.

Bundled five-trajectory demo (preprocessed case included; requires the WorldCast checkpoints, see the note above):

python demo.py \
  --manifest demo_case_traj5.json \
  --model_root pretrained/Wan2.1-I2V-14B-720P \
  --ckpt checkpoints/worldcast_bank.safetensors \
  --init_dit_from checkpoints/base_dit.safetensors

--traj presets (built by scripts/data/trajectories.py):

The orbit family is degree-parameterized: orbit_<N>[_cw|_ccw] builds a horizontal arc of N degrees (1–360), clockwise or counterclockwise viewed from above; without a suffix the arc continues the base trajectory's own direction. orbit_360 is the closed-ring preset itself.

For batch inference over custom cases, see scripts/infer/infer_worldcast.py and launch/infer/run_anchor_inference_example.sh; the manifest format is shown in examples/anchor_case_manifest.example.json, and the bundled demo_case/ illustrates the full case layout.

If WorldCast is useful for your research, please cite our paper:

@misc{worldcast,
  title         = {WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control},
  author        = {Cao, Lei and Liu, Yu and Liao, Wenhua and Zhang, Peng and Wang, Fei and Zhang, Hao and Le, Jiajun and Zhou, Daiguo and Luan, Jian and Ma, Jiayi},
  year          = {2026},
  eprint        = {2610.33135},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.33135}
}

Apache-2.0 (inherited from DiffSynth-Studio). Wan base weights follow their own license from Wan-AI.

WorldCast is built on DiffSynth-Studio.

This work uses SpatialVID under CC BY-NC-SA 4.0 and its Terms of Use. The authors confirm that all uses of this dataset are strictly for academic research purposes and not for any commercial applications.