Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MLV-Edit: Towards Consistent and Highly Efficient Editing for Minute-Level Videos

Yangyi Cao, Yuanhang Li, Lan Chen, Qi Mao

TL;DR

MLV-Edit tackles the challenge of editing minute-to-minute videos with a training-free, scalable approach. It segments long videos and uses two innovations—Velocity Blend to smooth cross-segment velocity transitions and Attention Sink to maintain a global semantic anchor—thereby mitigating boundary flicker and editing drift. The method builds on Wan-Edit's latent-space transport and is validated on the newly introduced MLV-EVAL minute-level benchmark, where it achieves state-of-the-art temporal stability and semantic fidelity. The work offers a practical solution for long-form video manipulation with significant gains in consistency and quality, useful for developers, search engines, and AI systems requiring reliable long-video context.

Abstract

We propose MLV-Edit, a training-free, flow-based framework that address the unique challenges of minute-level video editing. While existing techniques excel in short-form video manipulation, scaling them to long-duration videos remains challenging due to prohibitive computational overhead and the difficulty of maintaining global temporal consistency across thousands of frames. To address this, MLV-Edit employs a divide-and-conquer strategy for segment-wise editing, facilitated by two core modules: Velocity Blend rectifies motion inconsistencies at segment boundaries by aligning the flow fields of adjacent chunks, eliminating flickering and boundary artifacts commonly observed in fragmented video processing; and Attention Sink anchors local segment features to global reference frames, effectively suppressing cumulative structural drift. Extensive quantitative and qualitative experiments demonstrate that MLV-Edit consistently outperforms state-of-the-art methods in terms of temporal stability and semantic fidelity.

MLV-Edit: Towards Consistent and Highly Efficient Editing for Minute-Level Videos

TL;DR

MLV-Edit tackles the challenge of editing minute-to-minute videos with a training-free, scalable approach. It segments long videos and uses two innovations—Velocity Blend to smooth cross-segment velocity transitions and Attention Sink to maintain a global semantic anchor—thereby mitigating boundary flicker and editing drift. The method builds on Wan-Edit's latent-space transport and is validated on the newly introduced MLV-EVAL minute-level benchmark, where it achieves state-of-the-art temporal stability and semantic fidelity. The work offers a practical solution for long-form video manipulation with significant gains in consistency and quality, useful for developers, search engines, and AI systems requiring reliable long-video context.

Abstract

We propose MLV-Edit, a training-free, flow-based framework that address the unique challenges of minute-level video editing. While existing techniques excel in short-form video manipulation, scaling them to long-duration videos remains challenging due to prohibitive computational overhead and the difficulty of maintaining global temporal consistency across thousands of frames. To address this, MLV-Edit employs a divide-and-conquer strategy for segment-wise editing, facilitated by two core modules: Velocity Blend rectifies motion inconsistencies at segment boundaries by aligning the flow fields of adjacent chunks, eliminating flickering and boundary artifacts commonly observed in fragmented video processing; and Attention Sink anchors local segment features to global reference frames, effectively suppressing cumulative structural drift. Extensive quantitative and qualitative experiments demonstrate that MLV-Edit consistently outperforms state-of-the-art methods in terms of temporal stability and semantic fidelity.
Paper Structure (12 sections, 7 equations, 6 figures, 2 tables, 1 algorithm)

This paper contains 12 sections, 7 equations, 6 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Preliminary
  4. Overall Framework
  5. Velocity Blend
  6. Attention Sink
  7. Experiments
  8. Experimental Settings
  9. Comparisons with Baselines
  10. User Study
  11. Ablation Studies
  12. Conclusion

Figures (6)

  • Figure 1: Temporal inconsistencies in long video editing. Temporal slices are visualized by extracting pixels along a fixed vertical line across frames. Existing segmented editing methods fail to preserve consistent editing effects between segments. By employing Velocity Blend and Attention Sink, MLV-Edit effectively maintains coherent temporal evolution over long video sequences.
  • Figure 2: The framework of the proposed MLV-Edit. (a) The overall pipeline of MLV-Edit for long video editing. MLV-Edit first encodes the source video into the latent space and partitions it into multiple overlapping segments. (b) Velocity Blend fuses the velocity fields in overlapping regions between adjacent segments. (c) Attention Sink caches the key and value pairs from the first frame and injects them into subsequent segments.
  • Figure 3: Comparisons of MLV-Edit with video editing baselines. The MLV-Edit not only effectively suppresses the flickering and artifacts at the segment boundaries, but also preserves consistent editing effects throughout the long video sequences.
  • Figure 4: User study result. The values presented reflect the proportion of users who favor our proposed method over comparative approaches.
  • Figure 5: Ablation study on the influence of Velocity Blend.
  • ...and 1 more figures