EditVid: A Training-Free Unified Framework for Diverse Video Editing

Forum topic · 小凯 · 2026-09-05

Summary

EditVid is a training-free framework for instruction-guided and reference-guided video editing, presented in arXiv paper 2609.04190 by researchers including Adheesh Sunil Juvekar, Onkar Kishor Susladkar, and Kiet A. Nguyen. The method combines three techniques: sparse causal memory for local temporal coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending to localize edits. A single unified framework supports multiple editing paradigms, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement. On the FiVE benchmark, EditVid achieves a FiVE-Acc of 78.16, substantially outperforming the strongest evaluated training-free baseline at 58.95, while also delivering competitive results on IVEBench. A user study across seven competing methods reported a 51.8% overall preference rate for EditVid. The paper falls under computer vision and was automatically collected on 2026-09-05.

Paper Overview

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

Summary (translated from the original Chinese post)

Video editing spans diverse editing paradigms, but achieving high-quality instruction-guided and subject-guided editing within a single unified framework remains challenging. The authors 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 editing, covering style transfer, attribute modification, object insertion, part-level editing, and subject replacement.

Key Results

On FiVE, EditVid achieves a FiVE-Acc of 78.16, compared with 58.95 for the strongest evaluated training-free baseline.
EditVid also achieves competitive results on IVEBench.
In a user study against 7 competing methods, EditVid received a 51.8% overall preference rate.

Original Abstract (excerpt)

> 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.

---

*Automatically collected on 2026-09-05.*