EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting �Dong In Lee, et al�2025

Siyu Li (siyul@seas.upenn.edu)

Oct 2^nd, 2025

Contents

• Problem
• Method
• Experiments

​

2

Problem

3

Problem

4

• Text-driven 3D editing
• Given text prompt, source 3D represent (3D-GS in the setting of this paper)
• Output target 3D represent aligned with text prompt

​

​

Recap: Classifier-free Guidance

5

For 2D image editing, it is widely used to generate image aligned with prompt

Naïve pipeline for 3D editing

6

1. Forward 3D-GS from source multi-view images
2. Edit images given text prompt by Diffusion Model
3. Retrain 3D-GS on edited multi-view images

Challenge

7

• Multi-View Inconsistency
• Images share consistent prior! Style transfer/ Inpainting may fail!
• Inefficient Optimization
• Finetuning GS may not handle details well occasionally.

Method

8

Method: Pipeline

9

Method: Preprocess

10

• A source 3D representation by pretrained 3D-GS model
• A series of edited source images by pretrained text-image Diffusion Model

Multi-View Fusion

11

• Consistency Filtering
• Filter based on ImageReward
• Adjacent Selection
• 5-top source views near target
• Depth Blending
• Project source views & blend
• Background Refinement
• via SAM

Multi-View Fusion

12

MVF Guidance

13

Attention Guided Trimming

14

• Attention Weighting
• where K=text vector, Q=image vector from diffusion
• Trimming
• Exclude gaussian with low attention score
• Pruning
• Improving GS training efficiency with more reasonable gradient

​

Attention Guided Trimming

15

• Optimization of Gaussian Splatting
• Finetune Gaussians selected with edited images
• LPIPS loss + L1 regularization

​

Experiments

Experiments: Quantitative

17

Experiments: AGT Ablation

18

Experiments: Comparison

19