WeChatCV/Wan-Alpha
[CVPR 2026 Highlight] High-Quality Text-to-Video Generation with Alpha Channel
What it solves
Wan-Alpha is designed to solve the difficulty of generating high-quality videos with stable transparency (alpha channels). It allows for the creation of videos where the background is transparent, making it easier to integrate these videos into other scenes without manual rotoscoping.
How it works
The project implements a "Shiftable RGB-A Distribution Learner" to enable text-to-RGBA video generation. It builds upon the Wan2.1-T2V-14B base model and utilizes a specialized VAE (Variational Autoencoder) and LoRA weights to handle the alpha channel. The system uses Gaussian masks to guide the generation process and supports both English and Chinese prompts.
Who it’s for
This tool is for AI researchers, video editors, and digital content creators who need high-fidelity video assets with built-in transparency for professional compositing.
Highlights
- Stable Transparency: Generates videos with accurate alpha channels for semi-transparent objects, glowing effects, and fine details like hair.
- HuggingFace Integration: Provides open-source checkpoints for v1.0 and v2.0 models and VAEs.
- Training Pipeline: Includes full training scripts for both the VAE and the text-to-RGBA video generation model.
- Base Model Adaptation: Adapted from the Wan2.1-14B-T2V model for enhanced video synthesis capabilities.