Contents

1. Introduction
2. Dataset Construction
   • Dataset Pipeline
   • SAVEBench Examples
3. SAVE Removal Performance
4. Generalizability
   • Single-Object Removal
   • Multi-Object Removal
5. Interactive Demo
6. More Examples

---

Introduction

Schrödinger Audio-Visual Editor (SAVE) presents a novel approach to object-level audio-visual removal. Our method enables precise removal of both visual objects and their corresponding sounds simultaneously, addressing a critical challenge in video editing.

Traditional video editing methods often struggle with joint audio-visual manipulation, either removing visual elements while leaving their sounds intact or vice versa. SAVE addresses this by:

• Joint Audio-Visual Processing: Simultaneous removal of visual objects and their corresponding audio
• Object-Level Control: Precise selection and removal of specific objects
• High-Quality Results: Maintaining synchronization after removal
• Generalization: Works across diverse real-world scenarios

Dataset Construction

In this section, we present **SAVEBench**, the first audio-visual paired dataset for object-level editing task

I. Dataset Pipeline

We show our synthetic pair-generation pipelines. For audio, we use Qwen-VL to enumerate sounding objects, then synthesize an object-centric track for each with MMAudio, and keep only tracks validated as clean by Qwen-Audio. We form (N!-!1) by mixing all retained tracks except one, and (N) by reintroducing the held-out track. For visual, we obtain object boxes with GroundingDINO and segmentation masks with SAM2. With the segmentation masks, we remove the corresponding objects with Inpaint-Anything on each frame of the video.

---

II. SAVEBench Examples

Example pairs from our SAVEBench dataset demonstrating various object removal scenarios.

Source	Target (Object Removed)
Sample 1: guitar	Sample 1: guitar removed
Sample 2: speaker	Sample 2: speaker removed
Sample 3: man	Sample 3: man removed

SAVE Removal Performance

In this section, we present some editing results with SAVE editor

Source	Target (Object Removed)
Sample 1: violin	Sample 1: violin removed
Sample 2: machine	Sample 2: machine removed
Sample 3: cat	Sample 3: cat removed

Generalizability

Examples showing SAVE's ability to generalize to diverse real-world scenarios

Single-Object Removal

Source	Output (SAVE)
Target Object: tree	tree removed
Target Object: orange	orange removed
Target Object: table	table removed

Multi-Object Removal

Source	Output (SAVE)
Target Objects: cat and dog	cat and dog removed

Interactive Demo

Coming Soon!

More Examples

Additional examples demonstrating the versatility and robustness of our audio-visual editor.

Source	SAVE (Ours)
Target Object: alarm clock	alarm clock removed

Source	SAVE (Ours)
Target Object: plow	plow removed

Source	SAVE (Ours)
Target Object: fire truck	fire truck removed

Source	SAVE (Ours)
Target Object: man	man removed

Source	SAVE (Ours)
Target Object: fireworks	fireworks removed

Source	SAVE (Ours)
Target Object: motorcycle	motorcycle removed

Source	SAVE (Ours)
Target Object: airplane	airplane removed

Source	SAVE (Ours)
Target Object: bulldozer	bulldozer removed