[论文解读] Versatile Editing of Video Content, Actions, and Dynamics without Training

Vladimir Kulikov, Roni Paiss+4 more|arXiv (Cornell University)|Mar 18, 2026

Generative Adversarial Networks and Image Synthesis被引用 0

一句话总结

DynaEdit 是一种无需训练的方法，通过以反演-free 路径引导预训练的文本到视频流模型来编辑复杂的视频动力学与交互，结合相似性引导聚合与退火噪声相关性，在表达力与源视频保真度之间取得平衡。

ABSTRACT

Controlled video generation has seen drastic improvements in recent years. However, editing actions and dynamic events, or inserting contents that should affect the behaviors of other objects in real-world videos, remains a major challenge. Existing trained models struggle with complex edits, likely due to the difficulty of collecting relevant training data. Similarly, existing training-free methods are inherently restricted to structure- and motion-preserving edits and do not support modification of motion or interactions. Here, we introduce DynaEdit, a training-free editing method that unlocks versatile video editing capabilities with pretrained text-to-video flow models. Our method relies on the recently introduced inversion-free approach, which does not intervene in the model internals, and is thus model-agnostic. We show that naively attempting to adapt this approach to general unconstrained editing results in severe low-frequency misalignment and high-frequency jitter. We explain the sources for these phenomena and introduce novel mechanisms for overcoming them. Through extensive experiments, we show that DynaEdit achieves state-of-the-art results on complex text-based video editing tasks, including modifying actions, inserting objects that interact with the scene, and introducing global effects.

研究动机与目标

• Aim to enable unconstrained, text-driven editing of actions, dynamics, and interactions in real-world videos without training data.
• Overcome limitations of prior inversion-free methods that struggle with non-structural and dynamic edits.
• Maintain fidelity to original content while applying rich edits described by text prompts.
• Address the challenge of inserting interacting objects and global stylistic changes without degrading motion or identity.

提出的方法

• Adopt an inversion-free flow-based editing framework that transforms the source video along a noise-free path to the edited video.
• Introduce Similarity Guided Aggregation (SGA) to soft-select edit velocities based on their similarity to the source video.
• Introduce Annealed Noise Correlation (ANC) to gradually increase temporal noise correlation, reducing high-frequency jitter while preserving alignment.
• Leverage an image-to-video (I2V) flow model trained on triplets (text, first frame, video) to condition edits on content.
• Formulate the editing as an ODE where dZedit/dt equals the averaged velocity differences between target- and source-conditioned flows.
• Provide a practical pseudocode implementation (FlowEdit baseline and DynaEdit with SGA and ANC) for reference.
• Demonstrate model-agnostic applicability by showcasing results with WAN2.1 and Hunyuan I2V models.

实验结果

研究问题

• RQ1 Can training-free, text-based editing extend to unconstrained modifications of motion and object interactions in real videos?
• RQ2 How can inversion-free editing be adapted to avoid low-frequency misalignment and high-frequency jitter when making dynamic edits?
• RQ3 Do mechanisms like similarity-guided aggregation and annealed noise correlation improve quality and fidelity compared to prior methods?
• RQ4 How does DynaEdit perform versus trained models and other training-free baselines across diverse editing tasks?

主要发现

• DynaEdit achieves state-of-the-art results among training-free methods for complex edits (actions, dynamics, interactions) in real videos.
• The method attains competitive performance with a trained Aleph model in terms of text adherence and visual quality.
• SGA improves alignment to the source video over simple velocity averaging in FlowEdit.
• ANC reduces high-frequency jitter without sacrificing low-frequency alignment.
• Qualitative and user studies show DynaEdit preferred over leading baselines in content preservation, text adherence, and visual quality.