Netflix Explores Controllable AI Video Editing with Vera and VOID

Netflix's research team introduced two new AI models, Vera and VOID, aimed at closing gaps in current generative video editing tools used for promotional content like trailers and social videos. Existing methods often regenerate entire frames to execute a single edit, risking unintended changes to character identity, performance, or background, while object-removal techniques frequently ignore scene physics, producing unnatural results.

Vera addresses this through a layered video diffusion architecture that generates only an edit layer and alpha matte for the targeted change, then composites it back onto the untouched source footage. Lacking a suitable public dataset, the team built a custom 486k-frame layered video dataset at 832x480 resolution across three complexity tiers. Architecturally, Vera employs a Mixture-of-Transformers design with separate diffusion transformer branches for the edit, alpha, and composite outputs, joined via shared self-attention, and was trained in 1.3B and 14B parameter variants for object addition and background replacement tasks.

VOID complements this as a video inpainting model specialized in physically plausible object and interaction removal, reconstructing scenes as though the removed object was never present rather than merely erasing pixels. Netflix has publicly released research papers detailing both models, offering the broader research community a concrete example of how generative AI can be engineered to preserve, rather than override, an artist's creative control.