VidEdit3D: Flexible 3D Gaussian Editing with Video Diffusion

Yanming Yang; Chenxi Song; Tong Zhao; Juntao Xu; Wenhao Rao; Yiwen Chen; Yiwei Wang; Chi Zhang

cs.CVOct 26, 2026

VidEdit3D: Flexible 3D Gaussian Editing with Video Diffusion

Authors: Yanming Yang, Chenxi Song, Tong Zhao, Juntao Xu, Wenhao Rao, Yiwen Chen, Yiwei Wang, Chi Zhang

Organizations: Westlake University · University of Leeds · Nanyang Technological University · University of California, Merced

Abstract

Recent advancements in text-guided 3D editing, particularly with 3D Gaussian Splatting (3DGS), have shown promise but still lag behind the flexibility of 2D editing tools. Existing methods face a critical trade-off between preserving multi-view consistency and allowing structural modifications. Image-based editors can also struggle to apply the intended edit to objects viewed from unusual angles. To address these limitations, we introduce VidEdit3D, a framework that reformulates 3D scene editing as a trajectory-based video editing task. By drawing a parallel between camera movement in static 3D scenes and temporal dynamics in video, we leverage the robust spatiotemporal priors of pre-trained Video Diffusion Models (VDMs) to encourage semantic and structural coherence across views. Specifically, we propose Trajectory-Guided Rendering (TGR) to select informative keyframes and construct smooth video clips aligned with VDM input requirements. To propagate edits from an anchor view to these clips, we develop Flow-Aligned Video Editing (FAVE), a training-free framework incorporating an Editing Consistency Guidance (ECG) mechanism. ECG contrasts source and edited image conditions at a shared latent state, using scheduled source-image amplification to strengthen adherence to the edited reference while maintaining original scene context. Finally, we introduce Region-Selective 3DGS Refinement (RS-GS) to strictly confine geometric and appearance updates to target primitives, reducing background attribute leakage. Experiments demonstrate stronger instruction alignment and better multi-view consistency than the baselines.

Figures & tables

Fig. 1 : Flexible and consistent 3D editing with VidEdit3D. The baseline (left) shows inconsistent appearance, incomplete headset insertion and vase removal, and misplaced glasses. VidEdit3D (right) propagates the anchor edit across views, producing coherent appearance changes and localized object insertion and removal while preserving nearby content.

Fig. 2 : Overview of VidEdit3D. We reformulate 3D scene editing as a trajectory-based video editing task. The pipeline consists of three key stages: (1) TGR : We organize selected views into smooth video clips. (2) FAVE : Given a user-edited anchor frame, we propagate the edit across the clip using a pretrained VDM. This module incorporates ECG for semantic alignment and a self-attention-based mask for background preservation. (3) RS-GS : We lift the 2D edits back to 3D via inverse rendering to confine updates to the target Gaussians.

Method	CLIP-D $\uparrow$	CLIP-S $\uparrow$	DINO $\uparrow$	MEt3R $\downarrow$
GaussianEditor	0.0607	0.3001	0.8309	0.1195
DGE	0.0747	0.3053	0.7674	0.1259
VCEdit	0.0214	0.3015	0.8004	0.1203
GaussCtrl	0.0450	0.3023	0.8302	0.1137
EditSplat	0.0813	0.3061	0.7943	0.1535
Omni3DEdit	0.0588	0.3010	0.7559	0.1471

Table 1: Quantitative comparison of edit fidelity and multi-view consistency. VidEdit3D achieves the highest CLIP-D, CLIP-S, and DINO and the lowest MEt3R. Bold denotes the best value in each column.

Fig. 3 : Qualitative comparison of reference-guided scene editing. VidEdit3D inserts accessories, removes a vase, and localizes appearance changes to the requested objects. Source and edited results are shown at matched viewpoints, with insets providing a second view.

Method	CLIP-D $\uparrow$	CLIP-S $\uparrow$	DINO $\uparrow$	MEt3R $\downarrow$
w/o ECG amplification	0.1064	0.3116	0.8213	0.1164
w/o FAVE (vanilla Wan I2V)	-0.0022	0.2480	0.6693	0.1838
w/o RS-GS region selection	0.1246	0.3116	0.8276	0.1148
w/o RS-GS depth culling	0.1260	0.3116	0.8309	0.1137
w/o FAVE source regularization	0.1270	0.3116	0.8149	0.1187
TGR w/ unordered views	0.0405	0.3001	0.8085	0.1200

Table 2: Ablation of FAVE, ECG, RS-GS, and TGR. VidEdit3D achieves the best scores across all four metrics. Bold denotes the best value in each column.

Fig. 4 : Visual effects of ECG amplification and TGR view ordering. ECG amplification is disabled by setting $\lambda=1$ , while the unordered TGR variant shuffles the source sequence. Without amplification, the red-chair edit weakens in the oblique view; VidEdit3D retains the red appearance across both views. Ordered views support hat insertion.

Fig. 5 : Effect of RS-GS region selection. Without restricting Gaussian updates, recoloring the armchair also alters the adjacent sofa. VidEdit3D preserves the sofa appearance while completing the red-chair edit. Colored boxes mark corresponding regions enlarged below.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	CLIP-D $\uparrow$	CLIP-S $\uparrow$	DINO $\uparrow$	MEt3R $\downarrow$
ViP3DE-Wan	0.0472	0.3017	0.7946	0.1193
FlowAlign-Wan	0.0970	0.3083	0.8079	0.1186
VidEdit3D	0.1275	0.3129	0.8327	0.1126

Appendix

Table 3: Comparison using the same Wan2.1-I2V-14B-480P backbone. Bold denotes the best value in each column.

Fig. 6 : Additional qualitative comparisons of text transfer, accessory insertion, and localized recoloring. Each case shows two views of VidEdit3D and the compared baselines.

Panel of Figure 10 — Fig. 7 : More results. Each task shows a source image followed by two edited views.

Panel of Figure 11 — Fig. 8 : More results (continued). Each task shows a source image followed by two edited views.

Fig. 9 : Edits with user-provided accessory references. User-provided images specify the hat and mask, sunglasses, and AI pendant used in the edited anchor. Each row shows the source, the reference image, and three views rendered from the edited 3D scene.

Method	Multi-view consistency	Prompt fidelity	Unedited-region preservation
GaussianEditor	$1.98\pm 0.83$	$1.63\pm 0.76$	$1.84\pm 0.80$
DGE	$2.05\pm 0.83$	$1.68\pm 0.85$	$1.83\pm 0.81$
VCEdit	$1.95\pm 0.86$	$1.70\pm 0.83$	$1.80\pm 0.78$
GaussCtrl	$1.76\pm 0.83$	$1.60\pm 0.78$	$1.82\pm 0.81$
EditSplat	$2.04\pm 0.83$	$1.60\pm 0.74$	$1.95\pm 0.76$
Omni3DEdit	$1.74\pm 0.82$	$1.38\pm 0.76$	$1.75\pm 0.79$

Appendix

Table 4: User ratings (mean

\pm

standard deviation). Bold denotes the highest mean score in each column.

Fig. 10 : Failure-mode and scope analysis: occlusion, imperfect anchors, thin structures, and segmentation.

Fig. 11 : Field of view affects the balance between unchanged context and target resolution. Wider views retain surrounding context and better sustain the jersey and gold-material edits, at the cost of target resolution. The localized 3D edit (bottom left) remains stable. Video editing (bottom middle and right) is unaffected by this appearance fading: temporal continuity provides stronger frame-to-frame correspondence and better matches the video diffusion model’s learned prior, allowing the target appearance to persist. Rows show source and edited frames.

Text-Guided 3D Scene Editing

See all 19 papers

3D-Consistent Multi-View Editing by Correspondence Guidance
Nov 27, 2025Josef Bengtson, David Nilsson, Dong In Lee +2Multi-View Consistency Diffusion Model Guidance
Chalmers University of Technology · Korea University
DisCo3D: Distilling Multi-View Consistency for 3D Scene Editing
Aug 3, 2025Yufeng Chi, Huimin Ma, Kafeng Wang +1Multi-View Consistency 3D Scene Editing
Institute for Artificial Intelligence, Beijing National Research Center for Information Science and Technology, Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China · College of Computer, National University of Defense Technology, Changsha 410073, China
VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction
May 14, 2026Kaixin Zhu, Yiwen Tang, Yifan Yang +93D Shape Editing 3D Scene Editing
1Peking University · 2Tencent · 3The Chinese University of Hong Kong +4

Nov 27, 2025J/K move · Enter open · S save

3D-Consistent Multi-View Editing by Correspondence Guidance

Josef Bengtson, David Nilsson, Dong In Lee, Yaroslava Lochman+1

Chalmers University of Technology · Korea University

Multi-View Consistency Diffusion Model Guidance 3D Scene Editing Text-Guided Image Editing Diffusion-Based Image Editing Image Editing

Open paper AskPDF arXiv

Recent advancements in diffusion and flow models have greatly improved text-based image editing, yet methods that edit images independently often produce geometrically and photometrically inconsistent results across different views of the same scene. Such inconsistencies are particularly problematic for editing of 3D representations such as NeRFs or Gaussian splat models. We propose a training-free guidance framework that enforces multi-view consistency during the image editing process. The key idea is that corresponding points should look similar after editing. To achieve this, we introduce a consistency loss that guides the denoising process toward coherent edits. The framework is flexible and can be combined with widely varying image editing methods, supporting both dense and sparse multi-view editing setups. Experimental results show that our approach significantly improves 3D consistency compared to existing multi-view editing methods. We also show that this increased consistency enables high-quality Gaussian splat editing with sharp details and strong fidelity to user-specified text prompts. Please refer to our project page for video results: https://3d-consistent-editing.github.io/

Abstract

Text-Guided 3D Scene Editing

Explore similar work

3D-Consistent Multi-View Editing by Correspondence Guidance

DisCo3D: Distilling Multi-View Consistency for 3D Scene Editing

VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction