[CVPR 2026 重点] 使用 Alpha 通道实现高质量的文字到视频生成
Qualitative results of video generation using Wan-Alpha-v2.0. Our model successfully generates various scenes with accurate and clearly rendered transparency. Notably, it can synthesize diverse semi-transparent objects, glowing effects, and fine-grained details such as hair.
| Prompt | Preview Video | Alpha Video |
|---|---|---|
| "The background of this video is transparent. It features a beige, woven rattan hanging chair with soft seat and back cushions. Realistic style. Medium shot." |
# Clone the project repository
git clone https://github.com/WeChatCV/Wan-Alpha.git
cd Wan-Alpha
# Create and activate Conda environment
conda create -n Wan-Alpha python=3.11 -y
conda activate Wan-Alpha
# Install dependencies
pip install -r requirements.txtDownload Wan2.1-T2V-14B
Download Lightx2v-T2V-14B
Download Wan-Alpha-v1.0, Wan-Alpha-v2.0
Download Wan-Alpha-v1.0-VAE, Wan-Alpha-v2.0-VAE
You can test our model through:
…You can specify the weights of Wan2.1-T2V-14B with --ckpt_dir, LightX2V-T2V-14B with --lightx2v_path, Wan-Alpha-VAE with --vae_lora_checkpoint, and Wan-Alpha-T2V with --lora_path. Finally, you can find the rendered RGBA videos with a checkerboard background and PNG frames at --output_dir.
We provide an example of Gaussian mask. You can also use gen_gaussian_mask.py to generate a Gaussian mask from an existing alpha video. Alternatively, you can directly create a Gaussian ellipse video, which can be either static or dynamic (e.g., moving from left to right). Note that alpha_shift_mean is a fixed parameter.
Prompt Writing Tip: You need to specify that the background of the video is transparent, the visual style, the shot type (such as close-up, medium shot, wide shot, or extreme close-up), and a description of the main subject. Prompts support both Chinese and English input.
# An example of prompt.
This video has a transparent background. Close-up shot. A colorful parrot flying. Realistic style.VAE
For the VAE training dataset, please refer to Section 3.3 of our paper for preparation details.
You can train our VAE model through
…Before VAE training, you need to cache empty text ("") for VAE training.
T2V
You can download the training dataset from Google Drive or Hugging Face.
You can train our T2V model through
…We recommend caching the processed data before starting the training process to improve efficiency.
Coming soon...
This project is built upon the following excellent open-source projects:
We sincerely thank the authors and contributors of these projects.
If you find our work helpful for your research, please consider citing our paper:
@misc{dong2025wanalpha,
title={Video Generation with Stable Transparency via Shiftable RGB-A Distribution Learner},
author={Haotian Dong and Wenjing Wang and Chen Li and Jing Lyu and Di Lin},
year={2025},
eprint={2509.24979},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.24979},
}If you have any questions or suggestions, feel free to reach out via GitHub Issues . We look forward to your feedback!
暂无开放 Issues，或尚未同步最近议题。