Wan-Alpha is a text-to-video generation model developed by WeChat AI’s Computer Vision (WeChatCV) team. Released on September 30, 2025, this open-source model is notable for producing high-quality videos with alpha channels — meaning it supports transparent backgrounds, ideal for compositing and visual effects.
Built on Wan 2.1 (14B) architecture, Wan-Alpha can generate realistic 832×480 videos with semi-transparent elements like smoke, glass, and hair — details most T2V models struggle with. It uses an enhanced VAE and LightX2V distillation for efficiency.
There seems to be no VRAM requirement published, but judging by the weights and lora sizes alone that'd be way above 24GB.
Some folks on Reddit say if you're using the Wan2.1 T2V model you can try running the first frame with no training using VACE. Might need a few code tricks though.
Another user said they just drop in a PNG with alpha, set the first frame flag, and a see-through video comes out in one go.
If you'd like to access this model, you can explore the following possibilities:
Weights