Wan-Alpha

Overview

Wan-Alpha is a novel framework designed for generating high-quality RGBA (Red, Green, Blue, Alpha) videos from text prompts, where the alpha channel encodes transparency. Unlike traditional text-to-video models that generate only RGB frames, Wan-Alpha jointly learns both RGB and alpha channels to produce videos with transparent backgrounds and fine-grained transparency effects such as semi-transparent objects, glowing effects, and detailed elements like hair strands. The framework employs a variational autoencoder (VAE) that encodes the alpha channel into the RGB latent space, enabling efficient joint representation. A diffusion transformer is then trained on a newly constructed high-quality and diverse RGBA video dataset to generate realistic motion and transparency. This approach significantly improves visual quality, motion realism, and transparency rendering compared to prior methods.