VOID: Video Object and Interaction Deletion

Saman Motamed; William Harvey; Benjamin Klein; Luc Van Gool; Zhuoning Yuan; Ta-Ying Cheng

cs.CVApr 2, 2026

VOID: Video Object and Interaction Deletion

Authors: Saman Motamed, William Harvey, Benjamin Klein, Luc Van Gool, Zhuoning Yuan, Ta-Ying Cheng

Organizations: Netflix · INSAIT, Sofia University “St. Kliment Ohridski”

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.

Figures & tables

Figure 2: Counterfactual supervision examples. Top: videos $\mathbf{V}$ where $O$ is outlined in red. Bottom: re-simulated counterfactuals $\hat{\mathbf{V}}$ generated without $O$ . In Kubric scenes, downstream motion changes when the initiating object is removed. In HUMOTO scenes, supported objects transition naturally under gravity.

Figure 3: Void : Interaction-Aware Counterfactual Video Generation. A user provides an input video and clicks on an object to mask it for removal. A VLM-based pipeline expands the mask to identify other areas that will be affected. VOID’s first pass then predicts a counterfactual trajectory. The optional second pass stabilizes object deformation using flow-warped noise derived from the initially predicted motion.

Figure 4: Frames from generated videos featuring a guitar entering free-fall and a thrown ball following a new trajectory. Pass 1 (left) produces correct counterfactual trajectories but exhibits structural deformation. Pass 2 (right) better preserves object rigidity by using motion-aligned warped noise.

Figure 5: Qualitative comparisons on real-world videos. VOID maintains object structure and produces plausible motion over time, while the baselines exhibit deformation (the kettlebell on the pillow, and the floaty deforming), incomplete removal (two cars crashing), or implausible outputs (paint appearing after the roller is removed).

Figure 6: Generalization on various object interactions. VOID removes the target object and resolves downstream physical consequences (e.g., released objects fall; prevented collisions do not occur; shadows/reflections disappear).

Model	Win %
Void (ours)	64.8
Runway	18.4
Gen-Omni.	11.2
DiffuEraser	4.0
ROSE	1.6
MiniMax-Rem.	0.0

Table 1: Human preferences on real-world edits. 25 participants each evaluated 5 scenarios.

Judge	Model	Int.Phys $\uparrow$	Obj.Rem $\uparrow$	Bg.Art $\uparrow$	Temp $\uparrow$	Pres $\uparrow$	Sharp $\uparrow$	Total $\uparrow$
Gemini-3 Pro	ProPainter	$0.82$	$3.81$	$2.64$	$3.64$	$4.81$	$3.35$	$19.05$
	DiffuEraser	$1.48$	$4.43$	$3.53$	$4.12$	$4.84$	$4.13$	$22.53$
	MiniMax-Remover	$1.94$	$4.47$	$3.70$	$4.30$	$4.80$	4.24	$23.46$
	ROSE	$2.25$	4.77	$3.86$	$4.26$	4.92	$4.17$	$24.22$
	Gen-Omnimatte	$2.30$	$4.75$	$3.81$	$4.32$	4.92	$4.23$	$24.34$
	Runway	2.61	$4.62$	4.16	4.49	$4.82$	4.35	25.05

Table 2: VLM-as-a-judge evaluation on real-world videos. Each criterion is scored in

[0,5]

. Best per judge and column is green ; second best is orange .

Model	PSNR $\uparrow$	LPIPS $\downarrow$	DreamSim $\downarrow$	DINOv2 $\uparrow$	FVD $\downarrow$	VLM-Judge $\uparrow$
MiniMax-Remover	$29.96$	0.1091	0.0879	0.91	$448.43$	22.83
ProPainter	30.48	0.0981	$0.0987$	$0.8881$	$471.13$	$21.38$
DiffuEraser	$30.11$	$0.1213$	$0.0992$	$0.8920$	$496.61$	$21.30$
ROSE	$29.21$	$0.1296$	$0.1104$	$0.8905$	$480.18$	$21.62$
Gen-Omnimatte	$29.44$	$0.1186$	$0.1219$	$0.8688$	437.88	$20.40$
Runway	$26.68$	0.11	$0.1464$	$0.8522$	$442.76$	$21.67$

Table 3: Synthetic benchmark evaluation on 10 classic shadow/reflection removal cases and 30 dynamic interaction cases (Kubric + HUMOTO) capturing a wide range of effects. All metrics measure fidelity to the ground-truth counterfactual targets.

Model (Dataset Size)	Int.Phys $\uparrow$	Obj.Rem $\uparrow$	Bg.Art $\uparrow$	Temp $\uparrow$	Pres $\uparrow$	Sharp $\uparrow$	Total $\uparrow$
Kubric-Only (1200)	2.63	4.06	2.33	3.46	4.34	3.54	20.36
HUMOTO-Only (1200)	2.50	4.22	2.36	3.33	4.30	3.41	20.12
Both Datasets (1200)	3.04	4.30	2.31	3.87	4.45	3.96	21.93
Gen-Omni. Mask (Full)	3.30	4.73	3.04	4.04	4.22	4.06	23.39
VOID (Full)	3.66	4.82	4.10	4.44	4.88	4.22	26.12

Table 4: Ablation study evaluated by VLM judge on 75 real-world test cases.

Mask Reasoner	Int.Phys	Obj.Rem	Bg.Art	Temp	Pres	Sharp	Total
Qwen3-32B	3.75	4.23	3.54	4.11	4.86	3.42	23.91
GPT 5.2	3.49	4.17	3.60	4.03	4.94	4.11	24.34
Gemini 3-Pro	3.66	4.82	4.10	4.44	4.88	4.22	26.12

Table 5: Gemini 3 VLM-judge evaluation on 75 real-world videos when Void uses different VLMs for mask generation during inference.

Pass	Int.Phys	Obj.Rem	Bg.Art	Temp	Pres	Sharp	Total
Pass 1	2.90	4.20	3.70	3.80	4.90	4.00	23.5
Pass 2	3.90	4.90	4.00	4.20	4.80	4.20	26.0

Table 6: Per-category average scores (out of 5) across the 10 samples that are selected by the VLM for pass 2 refinement.

Figure 7: User interface for mask generation. The user selects sparse points on the target object, and the VLM generates the removal mask.

Figure 8: Interface used in the human evaluation study.

Figure 9: Examples where standard video similarity metrics favor visually implausible results.

Model	LPIPS $\downarrow$	PSNR $\uparrow$	DreamSim $\downarrow$	DINOv2 $\uparrow$
ProPainter	0.0879	27.98	0.0459	0.9824
DiffuEraser	0.0823	28.14	0.0384	0.9836
Ours	0.1081	28.96	0.0486	0.9786

Table 7: Metrics calculated for the example shown in Figure 9 .

Object Removal

See all 27 papers

BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal
Aug 20, 2026Yigit Ekin, Enes Sanli, Aykut Erdem +2Object Removal Video Editing
Bilkent University, Department of Computer Engineering, Ankara, Türkiye · Koç University, Department of Computer Engineering, Istanbul, Türkiye · KUIS AI Center, Istanbul, Türkiye +1
EffectLearner: World-Aware Object-Effect Reasoning for Real-World Video Object Removal
Aug 6, 2026Feier Wu, Wanke Xia, Xu He +8Video Diffusion Models Video Editing
Tsinghua University · China Agricultural University · Jiangnan University +1
TripleFlow: Training-Free Video Object Removal by Bridging Residual Editing and Native Generation
Sep 30, 2026Songhe Wang, Lifu Wei, Shuolin Xu +2Video Editing Image Inpainting
CSE Department, Penn State University · Department of Computer Science, The University of British Columbia, Canada · Department of Computing, Bournemouth University, United Kingdom +2

Aug 20, 2026J/K move · Enter open · S save

BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal

Yigit Ekin, Enes Sanli, Aykut Erdem, Erkut Erdem+1

Bilkent University, Department of Computer Engineering, Ankara, Türkiye · Koç University, Department of Computer Engineering, Istanbul, Türkiye · KUIS AI Center, Istanbul, Türkiye+1

Object Removal Video Editing Physical Reasoning in Video Causal Reasoning in Videos Video Generation Evaluation

Open paper AskPDF arXiv

Recent advances in generative video models have significantly improved visual realism in video object removal, yet evaluation protocols still focus on masked region fidelity, treating removal as local inpainting. In real scenes, object removal is a causal intervention: eliminating an object also requires removing its induced physical effects, such as shadows, reflections, illumination changes, translucency, and dynamic traces. Existing benchmarks lack aligned clean references or remain limited to simplified synthetic settings, preventing systematic evaluation of causal consistency. We introduce BeyondMasks, a paired benchmark for causally consistent video object removal, consisting of temporally aligned synthetic and real world video pairs with clean background references. The dataset spans diverse photometric, geometric, volumetric, and dynamic interactions, and supports both mask based and instruction driven editing. We further propose CORE, a structured vision language model based evaluation protocol that jointly measures object disappearance and after effect consistency, aligning more closely with human judgments than existing metrics. Benchmarking state of the art methods reveals systematic failures in removing secondary physical effects despite high masked region fidelity, exposing a gap between visual plausibility and causal correctness. BeyondMasks reframes video object removal as causal scene consistency rather than local reconstruction and provides a unified framework for its evaluation.

Abstract

Object Removal

Explore similar work

BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal

EffectLearner: World-Aware Object-Effect Reasoning for Real-World Video Object Removal

TripleFlow: Training-Free Video Object Removal by Bridging Residual Editing and Native Generation