HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Fri, 26 Dec 2025 19:51:59 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"694ee75f-5fd8" expires: Tue, 30 Dec 2025 08:30:26 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: E0F6:2D8B9D:9D34E1:B08FDF:69538B4A accept-ranges: bytes age: 0 date: Tue, 30 Dec 2025 08:20:26 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210052-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1767082827.747958,VS0,VE208 vary: Accept-Encoding x-fastly-request-id: 6811e6459603669c1f26b0df6523fcdc83a0b8da content-length: 5864

Over++: Generative Video Compositing for Layer Interaction Effects

Luchao Qi¹   Jiaye Wu²   Jun Myeong Choi¹   Cary Phillips³   Roni Sengupta¹   Dan B Goldman³  

¹University of North Carolina at Chapel Hill    ²University of Maryland    ³Industrial Light & Magic

   arXiv    Video    Code    BibTeX

TL;DR: Generate environmental effects between any foreground and background layers.

Over++ enables effect generation and editing, with or without mask or prompt guidance. Explore our applications below.

Effect generation Effect editing Keyframe masking Background swapping

I. Effect Generation

II. Effect Editing

III. Keyframe masking

IV. Background Swapping

---

We introduce Over++, a framework for generating environmental effects and enabling effect editing through mask- or prompt-guided control. Our method trains on both paired and unpaired data, providing a unified framework that remains effective even with limited training data (≈50 paired real-world videos, 600 paired synthetic videos, and 500 unpaired synthetic videos—≈1.1K total), which is significantly less than the ≈15K paired videos used in Force Prompting for force-guided video generation. Explore the sections below for more details:

Baseline Comparisons Our framework Training data Robustness Failure cases

Baseline Comparisons

Our Framework

Naively compositing the foreground over the background layer (copy-paste: $\mathcal{I}_{\text{over}} = \mathcal{I}_{\text{fg}} \oplus \mathcal{I}_{\text{bg}}$) produces a video that lacks environmental effects such as shadows or wakes. Given such an input composite and an optional binary mask ($\mathcal{M}_{\text{effect}}$) indicating the target effect regions, our model generates the desired effects within those regions.

Our method is trained on both paired and unpaired data. For unpaired data, we zero out the latent codes of $\mathcal{I}_{\text{over}}$ and $\mathcal{M}_{\text{effect}}$. (Text prompts $\mathcal{T}$ are not shown here for simplicity.)

Back to top

Training Data

Robustness

Failure Cases

References

Baseline comparisons

• Ku et al. AnyV2V: A Tuning-Free Framework For Any Video-to-Video Editing Tasks. TMLR, 2024.
• Gao et al. LoRA-Edit: Controllable First-Frame-Guided Video Editing via Mask-Aware LoRA Fine-Tuning. ArXiv, 2025.
• Jiang et al. VACE: All-in-One Video Creation and Editing. ICCV, 2025.
• Runway. Runway Aleph. 2025.

Data collection

• Gillman et al. Force Prompting: Video Generation Models Can Learn and Generalize Physics-based Control Signals. NeurIPS, 2025.
• Ruiz et al. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. CVPR, 2023.
• Lu et al. Omnimatte: Associating Objects and Their Effects in Video. CVPR, 2021.
• Lee et al. Generative Omnimatte: Learning to Decompose Video into Layers. CVPR, 2025.
• Lin et al. OmnimatteRF: Robust Omnimatte with 3D Background Modeling. ICCV, 2023.
• Greff et al. Kubric: A scalable dataset generator. CVPR, 2022.

Failure cases

• Sadat et al. Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models ICLR, 2025.

Back to baseline comparisons     Back to top

Societal Impact

We acknowledge that powerful video editing tools, including ours, may raise ethical considerations depending on their context of use. While our work is intended to augment video compositing and professional workflows, such capabilities could potentially be misused. We therefore encourage responsible use aligned with community guidelines and emphasize transparency regarding any applied edits.

Acknowledgements

Thank you to all ILM staff who assisted in preparing this work, especially Miguel Perez Senent for the 3D boat and ocean elements used in Figure 3 (row 2) and Figure 6 (row 3), and ILM leaders Rob Bredow, Francois Chardavoine, and Greg Grusby for their assistance in clearing this work for publication.

BibTeX

@misc{qi2025overgenerativevideocompositing,
  title={Over++: Generative Video Compositing for Layer Interaction Effects}, 
  author={Luchao Qi and Jiaye Wu and Jun Myeong Choi and Cary Phillips and Roni Sengupta and Dan B Goldman},
  year={2025},
  eprint={2512.19661},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.19661}, 
}