Vera: A Layered Diffusion Model for Content-Preserving Video Editing
Paper • 2606.23610 • Published • 11
Hongkai Zheng¹²* · Ta-Ying Cheng² · Benjamin Klein² · Yisong Yue² · Zhuoning Yuan²†
¹California Institute of Technology ²Netflix, Inc.
*Work done during an internship at Netflix †Project Lead
TL;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.
Disclaimer:This is a research prototype, not an official product.
Note: The current Vera models are trained on 49-frame sequences.
| Split | Edit Type | # Samples |
|---|---|---|
| train / 49-frames / realistic-set1-bg-change | background_replace | 914 |
| train / 49-frames / realistic-set1-obj-add | obj_add | 470 |
| train / 49-frames / realistic-set2-obj-add | obj_add | 770 |
| train / 49-frames / synthetic-bg-change | background_replace | 4,994 |
| train / 49-frames / synthetic-obj-add | obj_add | 4,848 |
| 49-Frame Train Total | ||
| 11,996 |
| Split | Edit Type | # Samples |
|---|---|---|
| train / 81-frames / realistic-set1-bg-change | background_replace | 457 |
| train / 81-frames / realistic-set1-obj-add | obj_add | 235 |
| train / 81-frames / realistic-set2-obj-add | obj_add | 385 |
| train / 81-frames / synthetic-bg-change | background_replace | 2,497 |
| train / 81-frames / synthetic-obj-add | obj_add | 2,431 |
| 81-Frame Train Total | ||
| 6,005 |
| Split | Edit Type | # Samples |
|---|---|---|
| test / bg-change | background_replace | 69 |
| test / obj-add | obj_add | 72 |
| Test Total | ||
| 141 |
| Source | License |
|---|---|
The test set is sourced from the training sources above, plus:
| Source | License |
|---|---|
@article{zheng2026vera,
title = {Vera: A Layered Diffusion Model for Content-Preserving Video Editing},
author = {Zheng, Hongkai and Cheng, Ta-Ying and Klein, Benjamin and Yue, Yisong and Yuan, Zhuoning},
journal = {arXiv preprint arXiv:2606.23610},
year = {2026}
}
source & further reading
huggingface.co — original article
Agentic Coding Harness - July - Daily Driver < 120GB VRAM
For All AI Researchers, Developers
Gradio Spaces now require a paid plan; big impact for our education use case