Zero-shot · Text-guided video editing

Edit-Your-Interest

Efficient Video Editing via
Feature Most-Similar Propagation

Yi Zuo*Zitao WangLingLing Li*Xu LiuFang LiuLicheng Jiao

School of Artificial Intelligence, Xidian University

Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education

* Corresponding authors: Yi Zuo and LingLing Li

arxiv Pattern Recognition Watch the demo Code

Edit the objects you care about.
Preserve the motion and the world around them.

See it in motion

01 / DEMONSTRATION

Text-driven edits. Consistent across frames.

Download video

The idea

Less memory.
More consistency.

Video editing with text-to-image diffusion models often comes with high computational and memory costs, while feature smoothing can introduce blurring and temporal artifacts. Edit-Your-Interest is a lightweight, text-driven, zero-shot method designed to balance efficiency and visual fidelity.

A Spatio-Temporal Feature Memory bank (SFM) caches representative spatial-attention features under a fixed memory budget. Feature Most-Similar Propagation (FMP) retrieves relevant cached tokens for the current frame, while a feature-diversity-based update keeps the memory representative. Automatically extracted cross-attention masks support precise object edits and preserve the surrounding background.

02 / Method

A memory that follows the video.

Cache representative features. Retrieve what matters.
Keep the edit focused.

Edit-Your-Interest pipeline: DDIM inversion and sampling, feature memory with most-similar propagation, and automatic mask extraction and injection for instance-level editing.
Framework overviewView full resolution
01

Remember efficiently

SFM stores a representative pool of feature tokens under a fixed memory budget. Diversity-based updates keep the pool useful as the video progresses.

02

Propagate by similarity

FMP retrieves the most similar cached features for the current frame, improving temporal consistency without full-sequence temporal attention.

03

Edit your instance

Cross-attention maps automatically provide object masks. Injecting these masks during denoising helps preserve the background and suppress boundary artifacts.

03 / Results

Same scene. New possibilities.

Explore the original and edited videos side by side.
Use “Play pair” to compare them together.

04 / Citation

BibTeX

@article{ZUO2027115113,
title = {Edit-Your-Interest: Efficient video editing via Feature Most-Similar Propagation},
journal = {Pattern Recognition},
volume = {183},
pages = {115113},
year = {2027},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2026.115113},
url = {https://www.sciencedirect.com/science/article/pii/S0031320326020777},
author = {Yi Zuo and Zitao Wang and Lingling Li and Xu Liu and Fang Liu and Licheng Jiao},
}