Netflix/Vera-Layered-Video-Dataset

petethomas1 pts0 comments

netflix/Vera-Layered-Video-Dataset · Datasets at Hugging Face

Log In<br>Sign Up

Dataset Viewer

Auto-converted to Parquet API Embed Duplicate Data Studio<br>Subset (1) default · 90.1k rows<br>default (90.1k rows) Split (2)<br>train · 90k rows<br>train (90k rows)test (86 rows)

Search is not available for this dataset

SQL<br>Console

video<br>video

End of preview. Expand in Data Studio

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

Hongkai Zheng¹²* ·<br>Ta-Ying Cheng² ·<br>Benjamin Klein² ·<br>Yisong Yue² ·<br>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.

📋 Dataset Description

Curated by: Hongkai Zheng, Ta-Ying Cheng, Benjamin Klein, Yisong Yue, Zhuoning Yuan

License: Apache License 2.0

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

📦 Dataset Structure

🎞️ Splits — 49 Frames (3 sec)

Note: The current Vera models are trained on 49-frame sequences.

Split<br>Edit Type<br># Samples

train / 49-frames / realistic-set1-bg-change<br>background_replace<br>914

train / 49-frames / realistic-set1-obj-add<br>obj_add<br>470

train / 49-frames / realistic-set2-obj-add<br>obj_add<br>770

train / 49-frames / synthetic-bg-change<br>background_replace<br>4,994

train / 49-frames / synthetic-obj-add<br>obj_add<br>4,848

49-Frame Train Total

11,996

🎞️ Splits — 81 Frames (5 sec)

Split<br>Edit Type<br># Samples

train / 81-frames / realistic-set1-bg-change<br>background_replace<br>457

train / 81-frames / realistic-set1-obj-add<br>obj_add<br>235

train / 81-frames / realistic-set2-obj-add<br>obj_add<br>385

train / 81-frames / synthetic-bg-change<br>background_replace<br>2,497

train / 81-frames / synthetic-obj-add<br>obj_add<br>2,431

81-Frame Train Total

6,005

🧪 Test Splits

Split<br>Edit Type<br># Samples

test / bg-change<br>background_replace<br>69

test / obj-add<br>obj_add<br>72

Test Total

141

🗂️ Data Sources

🏋️ Training Set

Source<br>License

Pexels<br>Pexels License

Mixkit<br>Mixkit License

VideoMatte240K<br>MIT License

🧪 Test Set

The test set is sourced from the training sources above, plus:

Source<br>License

DAVIS<br>CC BY-NC 4.0

VACEBench<br>Apache License 2.0

📝 Citation

@article{zheng2026vera,<br>title = {Vera: A Layered Diffusion Model for Content-Preserving Video Editing},<br>author = {Zheng, Hongkai and Cheng, Ta-Ying and Klein, Benjamin and Yue, Yisong and Yuan, Zhuoning},<br>journal = {arXiv preprint arXiv:2606.23610},<br>year = {2026}

Copy to bucket new

Downloads last month<br>11,661<br>Number of rows:<br>90,091<br>Total file size:<br>320 GB

Paper for netflix/Vera-Layered-Video-Dataset<br>Paper • 2606.23610 • Published 18 days ago • 11

train frames video vera dataset license

Related Articles