Luozhou Wang*, Yijun Li**, Zhifei Chen, Jui-Hsien Wang, Zhifei Zhang, He Zhang, Zhe Lin, Yingcong Chen†
HKUST(GZ), HKUST, Adobe Research.
* Intership Project. ** Project Leader. † Corresponding Author.
Text-to-video generative models have made significant strides, enabling diverse applications in entertainment, advertising, and education. However, generating RGBA video, which includes alpha channels for transparency, remains a challenge due to limited datasets and the difficulty of adapting existing models. Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixar, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. TransPixar leverages a diffusion transformer (DiT) architecture, incorporating alpha-specific tokens and using LoRA-based fine-tuning to jointly generate RGB and alpha channels with high consistency. By optimizing attention mechanisms, TransPixar preserves the strengths of the original RGB model and achieves strong alignment between RGB and alpha channels despite limited training data. Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation.
- [2024.01.07] We have released the project page, arXiv paper, inference code and huggingface demo for TransPixar + CogVideoX.
- Release code, paper and demo.
- Release checkpoints of joint generation (RGB + Alpha).
conda create -n TransPixar python=3.10
conda activate TransPixar
pip install -r requirements.txt
Our pipeline is designed to support various video tasks, including Text-to-RGBA Video, Image-to-RGBA Video.
We provide the following pre-trained LoRA weights for different tasks:
Task | Base Model | Frames | LoRA weights |
---|---|---|---|
T2V + RGBA | genmo/mochi-1-preview | 37 | Coming soon |
T2V + RGBA | THUDM/CogVideoX-5B | 49 | link |
I2V + RGBA | THUDM/CogVideoX-5b-I2V | 49 | Coming soon |
We have open-sourced the training code for Mochi on RGBA joint generation. Please refer to the Mochi README for details.
In addition to the Hugging Face online demo, users can also launch a local inference demo based on CogVideoX-5B by running the following command:
python app.py
To generate RGBA videos, navigate to the corresponding directory for the video model and execute the following command:
python cli.py \
--lora_path /path/to/lora \
--prompt "..." \
- finetrainers: We followed their implementation of Mochi training and inference.
- CogVideoX: We followed their implementation of CogVideoX training and inference.
We are grateful for their exceptional work and generous contribution to the open-source community.
@misc{wang2025transpixar,
title={TransPixar: Advancing Text-to-Video Generation with Transparency},
author={Luozhou Wang and Yijun Li and Zhifei Chen and Jui-Hsien Wang and Zhifei Zhang and He Zhang and Zhe Lin and Yingcong Chen},
year={2025},
eprint={2501.03006},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2501.03006},
}