Netflix and Caltech Introduce Vera, a Layered Diffusion Model for Content-Preserving Video Editing

Researchers from Netflix and Caltech have published Vera, a layered diffusion model that jointly generates an edit layer, alpha matte, and composite video from a source clip and a text instruction. By separating "what to generate" from "what to preserve," Vera enables precise, content-preserving edits such as object addition and background replacement while keeping unaffected regions intact.

The model uses a Mixture-of-Transformers (MoT) architecture in which separate diffusion transformers handle the edit layer, alpha matte, and composite video independently, coordinating through joint self-attention. Quantitative benchmarks show Vera-14B outperforming baselines including VACE, ReCo, and VideoPainter on content-preservation metrics (PSNR, SSIM, LPIPS) for both object-addition and background-change tasks. The team also released a 486K-frame layered video dataset combining synthetic composites and real stock footage to support training. The work is described as a research prototype and not an official Netflix product.