Adobe brings transparent background video to Firefly

Anyone who edits video knows that keying out green screens and cleaning up edges is tedious work. Adobe wants generative AI to skip that step entirely.

Adobe just added transparent background video generation to the web version of its Firefly Video model. The tool supports both text-to-image and image-to-video workflows, allowing users to export assets with built-in alpha channels.

484 clips and a LoRA

The feature stems from an Adobe Research project originally dubbed TransPixar, later renamed Transpixeler.

The wild part is how it was built: the team trained the capability using just 484 video clips paired with a LoRA adapter. Instead of burning massive compute, a tiny curated dataset was enough to teach the model how to deliver isolated motion graphics, quick visual effects, and stock elements without a solid background.

Why it matters

For motion designers, social video creators, and VFX artists, transparent assets are pure gold. Being able to prompt a floating fireball, an animated sticker, or a UI element directly onto a transparent canvas cuts out the annoying step of rotoscoping or keying out green.

We think the smart move here is the training efficiency. Showing that 484 clips and a LoRA can crack alpha channel generation proves that targeted, lightweight tuning can solve real production bottlenecks without retraining massive foundation models from scratch. The real test now will be edge fidelity when dealing with tough elements like hair, smoke, and semi-transparent glass.