Why this matters
Content-preserving video editing needs training data that explicitly separates what should be regenerated from what should be preserved. This dataset supplies layered supervision (edit layer + alpha matte + composite target) across short clips, letting diffusion-based editing models learn to synthesize edits while retaining unchanged content.
What Sets It Apart
- Layered targets: each sample includes an edit layer, an alpha matte, and the final composite, which is uncommon in public video corpora and directly supports architectures that predict a separate edit and blending mask.
- Two edit types and synthetic/real mixes: organizes samples into background-replace and object-add edits and provides both realistic and synthetic subsets to help researchers evaluate robustness to domain gap and controllability.
- Framed for model training: models in the associated work are trained on 49-frame sequences (3s) — the dataset also offers 81-frame (5s) variants to test longer temporal consistency.
Key Details
- Typical sequence lengths: 49-frame (3 s) and 81-frame (5 s) splits. The 49-frame train total is 11,996 samples; the 81-frame train total is 6,005 samples; test set totals 141 samples (69 bg-change, 72 obj-add).
- Source licenses and provenance: training sources include Pexels, Mixkit and VideoMatte240K; test set also draws from DAVIS and VACEBench. The dataset is released under Apache-2.0, but downstream use should respect original source licenses where required.
- Intended labels: per-frame edit layers and alpha mattes suitable for supervised training of layered diffusion or compositing-aware video editors.
- Associated paper and resources: accompanies the Vera layered-diffusion paper (arXiv) and project page with demos and model descriptions.
Who It's For and Trade-offs
Great fit if you are developing or evaluating video editing models that must preserve scene content while applying localized edits—especially diffusion-based or layered architectures that predict both appearance changes and blending masks. It helps benchmark temporal consistency on short clips and compare synthetic vs realistic edit training.
Look elsewhere if you need long-form video (>5 s), dense semantic annotations beyond alpha mattes (e.g., dense pose/keypoints for every object), or very large-scale video corpora for pretraining; the dataset focuses on edit-localized supervision and relatively short sequences. Also verify original-source licenses for commercial redistribution of derived assets.
