Adobe Research and the Hong Kong University of Science and Technology (HKUST) have teamed up to create TransPixar, a cool new AI model that makes Premultiplied visual effects stock footages like smoke, reflections, and magical portals—way easier and better.
“Alpha channels are super important in VFX. They let transparent elements like smoke and reflections fit perfectly into scenes,” explained Yijun Li, the project lead at Adobe Research. “But making RGBA videos (the ones that include alpha channels) has been tricky because of limited data and outdated models.”
That’s where TransPixar comes in. It uses a fancy tech called Diffusion Transformers (DiT) to generate RGBA videos, including those tricky alpha channels, with amazing quality—even with very little training data. Instead of building something completely new, the team upgraded existing AI models, making it both smart and efficient.
What’s really cool? TransPixar can create all kinds of effects, like swirling storm clouds, shattering glass, and glowing magical portals, just from text prompts. It can even bring still images to life with animated transparency effects.
Best of all, it’s open-source! The team has shared the code on GitHub and set up a demo on Hugging Face so anyone—artists, developers, or just curious tech fans—can try it out.
TransPixar is shaping up to be a game-changer for VFX and compositing artists, helping them create jaw-dropping effects with less hassle.
Check out the code here : https://github.com/wileewang/TransPixar
Paper Link : https://wileewang.github.io/TransPixar/
