OmnimatteZero cleanly erases objects and shadows from video, open code released

OmnimatteZero is a new video editing model that can remove selected objects from footage. This includes shadows and reflections, making for cleaner backgrounds.

In examples, it deletes cats while also erasing their reflections in puddles, removes people and their shadows. There’s also an example where it takes out a swan while cleaning up its reflection in water.

The system supports multiple object removals in the same scene, such as four puppies (with their shadows) disappearing from a lawn. In challenging tests, it can remove only a background person while preserving a foreground pole, something rival tools struggle with.

Benchmark comparisons show Omnimatte Zero consistently outperforming other methods on quality metrics. Beyond removal, the model can extract objects into transparent foreground layers for compositing on new backgrounds.

However, its current examples show blending limitations. Granted, its major plus is that the team has already open-sourced the code.

Its GitHub repo includes instructions for running the model locally, allowing editors to integrate it into desktop workflows without round-tripping to web apps.