BROWSE
Detail
Conference Paper
MoSA-Edit: Motion- and Source-Aware Inversion-Free Video Editing
Cited - time in 
Share 
Share
- Authors
- Soonchul Jung, Jae Woo Kim, Jin-Seo Kim, Seungjoon Kwon
- Issue Date
- 2026-10
- Citation
- International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-4
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Inversion-free, training-free video editing follows a direct trajectory from a source video to a prompt-conditioned result. This freedom enables large changes, but it can also alter appearance or motion that the prompt does not ask to change. Choosing candidate updates only by whole-video similarity makes the problem worse because the intended edit is also counted as an error, which can favor weak edits. We present MoSA-Edit, which separates candidate selection from source preservation. Motion- and Edit-aware Similarity Aggregation (MESA) estimates where change is intended and compares candidate appearance and motion only in the remaining region. Checkpoint Source Re-anchoring (CSR) then restores temporally reliable source-like latent components at intermediate trajectory checkpoints. The method requires neither fine-tuning nor inversion. On 20 in-house and 20 Pexels videos, MoSA-Edit achieves the best appearance preservation, motion preservation, prompt adherence, and four-metric composite among the evaluated methods, while VACE attains the highest visual quality.
- Keyword
- video editing, training-free editing, inversion-free editing, optical flow, source preservation
- KSP Keywords
- Candidate selection, Optical Flow, Similarity aggregation, fine-tuning, video editing, visual quality