{"slug": "netflix-vera-layered-video-dataset", "title": "Netflix/Vera-Layered-Video-Dataset", "summary": "Netflix researchers released Vera, a layered diffusion model for content-preserving video editing, along with a dataset of over 18,000 video samples for training and evaluation. The model jointly generates an edit layer, alpha matte, and composite video to separate generated content from preserved elements. The Vera dataset includes background replacement and object addition edits across 49-frame and 81-frame sequences.", "body_md": "#### Vera: A Layered Diffusion Model for Content-Preserving Video Editing\n\nPaper • 2606.23610 • Published • 11\n\n[Hongkai Zheng](https://devzhk.github.io/)¹²* ·\n[Ta-Ying Cheng](https://ttchengab.github.io/)² ·\n[Benjamin Klein](https://scholar.google.com/citations?user=xkX9W9QAAAAJ&hl=en)² ·\n[Yisong Yue](https://yisongyue.com/)² ·\n[Zhuoning Yuan](https://zhuoning.cc/)²†\n\n¹California Institute of Technology ²Netflix, Inc.\n\n*Work done during an internship at Netflix †Project Lead\n\nTL;DR: A layered diffusion framework for video editing. Vera jointly generates an edit layer, an alpha matte, and a composite video, separating what to generate from what to preserve.\n\nDisclaimer:This is a research prototype, not an official product.\n\n**Note:** The current Vera models are trained on 49-frame sequences.\n\n| Split | Edit Type | # Samples |\n|---|---|---|\n| train / 49-frames / realistic-set1-bg-change | background_replace | 914 |\n| train / 49-frames / realistic-set1-obj-add | obj_add | 470 |\n| train / 49-frames / realistic-set2-obj-add | obj_add | 770 |\n| train / 49-frames / synthetic-bg-change | background_replace | 4,994 |\n| train / 49-frames / synthetic-obj-add | obj_add | 4,848 |\n49-Frame Train Total |\n11,996 |\n\n| Split | Edit Type | # Samples |\n|---|---|---|\n| train / 81-frames / realistic-set1-bg-change | background_replace | 457 |\n| train / 81-frames / realistic-set1-obj-add | obj_add | 235 |\n| train / 81-frames / realistic-set2-obj-add | obj_add | 385 |\n| train / 81-frames / synthetic-bg-change | background_replace | 2,497 |\n| train / 81-frames / synthetic-obj-add | obj_add | 2,431 |\n81-Frame Train Total |\n6,005 |\n\n| Split | Edit Type | # Samples |\n|---|---|---|\n| test / bg-change | background_replace | 69 |\n| test / obj-add | obj_add | 72 |\nTest Total |\n141 |\n\n| Source | License |\n|---|---|\n|\n\nThe test set is sourced from the training sources above, plus:\n\n| Source | License |\n|---|---|\n|\n\n```\n@article{zheng2026vera,\n    title     = {Vera: A Layered Diffusion Model for Content-Preserving Video Editing},\n    author    = {Zheng, Hongkai and Cheng, Ta-Ying and Klein, Benjamin and Yue, Yisong and Yuan, Zhuoning},\n    journal   = {arXiv preprint arXiv:2606.23610},\n    year      = {2026}\n}\n```\n\n", "url": "https://wpnews.pro/news/netflix-vera-layered-video-dataset", "canonical_source": "https://huggingface.co/datasets/netflix/Vera-Layered-Video-Dataset", "published_at": "2026-07-09 19:52:30+00:00", "updated_at": "2026-07-09 20:07:21.686091+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "computer-vision", "ai-research", "ai-products"], "entities": ["Netflix", "California Institute of Technology", "Hongkai Zheng", "Ta-Ying Cheng", "Benjamin Klein", "Yisong Yue", "Zhuoning Yuan", "Vera"], "alternates": {"html": "https://wpnews.pro/news/netflix-vera-layered-video-dataset", "markdown": "https://wpnews.pro/news/netflix-vera-layered-video-dataset.md", "text": "https://wpnews.pro/news/netflix-vera-layered-video-dataset.txt", "jsonld": "https://wpnews.pro/news/netflix-vera-layered-video-dataset.jsonld"}}