Netflix researchers present two early-stage AI video editing models: Vera and VOID. Vera is a layered video diffusion model that generates only the edited regions (plus an alpha matte) rather than regenerating the entire clip, preserving unedited content pixel-perfectly. It uses a Mixture-of-Transformers (MoT) architecture with three specialized DiTs and was trained on a custom 486k-frame layered dataset. VOID is a video object removal model that handles physically plausible inpainting — not just erasing objects but reconstructing how the scene would realistically evolve without them, using a VLM-based reasoning pipeline and quadmask conditioning. Both models outperform existing baselines in human preference studies. The research papers for both models are publicly released.
