Paper 2604.02296
VOID: Video Object and Interaction Deletion
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 4
- GitHub
- 2K stars
01 In brief
Summary
VOID is a video object removal framework that generates physically plausible counterfactual videos when an object is removed, addressing limitations of existing methods that only handle photometric effects like shadows.
It uses a two-pass approach: first, a video diffusion model (CogVideoX) synthesizes a counterfactual trajectory guided by a quadmask, which extends prior trimasks to resolve ambiguities between the removed object and affected regions.
Second, a flow-warped noise stabilization pass corrects structural deformations in objects undergoing new motion.
Training data comes from Kubric simulations (~1900 pairs) and HUMOTO motion capture (~4500 pairs), providing counterfactual supervision.
At inference, a VLM (e.g., Gemini 3 Pro) generates the quadmask by identifying affected objects and their counterfactual positions.
Evaluations on 75 real-world and 40 synthetic videos show VOID outperforms baselines (ProPainter, DiffuEraser, ROSE, Gen-Omnimatte, Runway, MiniMax-Remover) in human preference (64.8% wins) and VLM-judge scores, with best results in interaction physics.
Ablations confirm the benefits of mixed datasets and quadmask conditioning.
VOID generalizes to unseen effects like balloons floating and blenders stopping, demonstrating high-level causal reasoning.
02 From the paper
Abstract
Existing video object removal methods excel at inpainting content "behind" the object and correcting appearance-level artifacts such as shadows and reflections. However, when the removed object has more significant interactions, such as collisions with other objects, current models fail to correct them and produce implausible results. We present VOID, a video object removal framework designed to perform physically-plausible inpainting in these complex scenarios. To train the model, we generate a new paired dataset of counterfactual object removals using Kubric and HUMOTO, where removing an object requires altering downstream physical interactions. During inference, a vision-language model identifies regions of the scene affected by the removed object. These regions are then used to guide a video diffusion model that generates physically consistent counterfactual outcomes. Experiments on both synthetic and real data show that our approach better preserves consistent scene dynamics after object removal compared to prior video object removal methods. We hope this framework sheds light on how to make video editing models better simulators of the world through high-level causal reasoning.