EditVid: A Unified Training-Free Framework for Diverse Instruction- and Reference-Guided Video Editing

Forum topic · 小凯 · 2026-09-05

Summary

EditVid is a training-free framework for instruction-guided and reference-guided video editing, unifying multiple editing paradigms—style transfer, attribute modification, object insertion, part-level editing, and subject replacement—in a single pipeline. It combines three mechanisms: sparse causal memory for local temporal coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending for edit locality. On the FiVE benchmark, EditVid achieves a FiVE-Acc score of 78.16, outperforming the strongest evaluated training-free baseline (58.95), and delivers competitive results on IVEBench. A user study across 7 competing methods showed a 51.8% overall preference rate for EditVid. The paper (arXiv:2609.04190) by Juvekar, Susladkar, and Nguyen was published on 2026-09-03 in the computer vision field.

Paper Overview

Field: Computer Vision (CV) Authors: Adheesh Sunil Juvekar, Onkar Kishor Susladkar, Kiet A. Nguyen Published: 2026-09-03 arXiv: 2609.04190

Abstract

Video editing encompasses diverse editing paradigms, but achieving high-quality instruction-guided and subject-guided editing within a single unified framework remains challenging. We propose EditVid, a training-free framework that combines:

Sparse causal memory for local coherence
Correspondence-based post-attention token injection for long-range identity preservation
Soft latent blending for edit locality

The same framework supports both instruction-guided and reference-guided edits, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement.

Results

On FiVE, EditVid achieves 78.16 FiVE-Acc, compared with 58.95 for the strongest evaluated training-free baseline.
Competitive results are obtained on IVEBench.
A user study shows EditVid achieves a 51.8% overall preference rate compared with 7 competing methods.

Original Abstract

> We introduce EditVid, a training-free framework combining sparse causal memory for local coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending for edit locality. The same framework supports instruction-guided and reference-guided edits. On FiVE, EditVid achieves 78.16 FiVE-Acc, compared with 58.95 for the strongest evaluated training-free baseline.

*Auto-collected on 2026-09-05*