ViBT: Vision Bridge Transformer at Scale
Zhenxiong Tan, Zeqing Wang, Xingyi Yang, Songhua Liu, Xinchao Wang
xML Lab, National University of Singapore; The Hong Kong Polytechnic University; Shanghai Jiao Tong University

• Bridge formulation: Data-to-data trajectories between inputs and outputs instead of noise-to-data diffusion.
• Scaled transformers: 20B and 1.3B parameter ViBT variants for image/video translation.
• Stabilized training: Variance-stabilized velocity-matching objective for robust large-model optimization.
• Fast inference: Removal of conditional tokens yields up to 4× faster runs versus token-heavy baselines.

1. Environment

conda create -n ViBT python=3.12
conda activate ViBT

2. Install requirements

pip install -e .

• Image instruction-based editing and stylization: examples/image_stylization.ipynb
• Video stylization: examples/video_stylization.ipynb
• Video colorization: examples/video_colorization.ipynb
• Video frame interpolation: examples/video_frame_interpolation.ipynb

We keep different models for image and video tasks.

• Image tasks (stylization, editing) are trained on Qwen-Image-Editing.
• Video tasks (stylization, depth-to-video, colorization, frame interpolation) are trained on Wan2.1 1.3B.

Training code is under development; we will add full instructions once released.

@article{tan2025vision,
  title={Vision Bridge Transformer at Scale},
  author={Tan, Zhenxiong and Wang, Zeqing and Yang, Xingyi and Liu, Songhua and Wang, Xinchao},
  journal={arXiv preprint arXiv:2511.23199},
  year={2025}
}