Vera: A Layered Diffusion Model for Content-Preserving Video Editing

Vera, developed by California Institute of Technology and Netflix, Inc., is a layered diffusion framework for content-preserving video editing that avoids regenerating the full frame. Instead, it jointly predicts an edit layer, an alpha matte, and a natural composite video, then uses a Mixture-of-Transformers design with three separate DiTs and joint self-attention so the edit branch can stay creative while the preserved content is retained by construction. The method is trained with flow matching on a curated layered dataset of 486K frames at 832 × 480, built from synthetic composites and realistic single- and multi-object videos with accurate alpha mattes and effects such as shadows and reflections. On 72 object-addition tests and 69 background-change tests, Vera-1.3B reaches 25.3 dB PSNR for object addition and 35.2 dB for background change, while Vera-14B reaches 26.1 dB and 36.2 dB, respectively; it also tops instruction compliance on object addition with 4.13 IS and remains competitive on edit quality. The paper’s ablations show that the layered formulation, the composite branch, and MoT joint attention are all necessary, and a human 2AFC study over 513 valid trials preferred Vera on content preservation and instruction compliance.