첫 프레임 편집을 전체 영상에 자연스럽게 전파하는 방법

In-Context Sync-LoRA for Portrait Video Editing Sagi Polaczek1 Or Patashnik1 Ali Mahdavi-Amiri2 Daniel Cohen-Or1 1Tel Aviv University 2Simon Fraser University “a person wearing decorative forehead jewelry, speaking” ”a person speaking with burning trees in the background“ Source Video Edited First Frame “a person wearing a fedora hat, speaking and moving his hand” Figure 1. Given a source video and an edited first frame, our method propagates the visual edit. The resulting video precisely retains the subject’s identity and frame-accurate source motion, while faithfully adopting the new appearance from the edited frame. Abstract Editing portrait videos is a challenging task that requires flexible yet precise control over a wide range of modi- fications, such as appearance changes, expression edits, or the addition of objects. The key difficulty lies in pre- serving the subject’s original temporal behavior, demand- ing that every edited frame remains precisely synchronized with the corresponding source frame. We present Sync- LoRA, a method for editing portrait videos that achieves high-quality visual modifications while maintaining frame- accurate synchronization and identity consistency. Our ap- proach uses an image-to-video diffusion model, where the edit is defined by modifying the first frame and then propa- gated to the entire sequence. To enable accurate synchro- nization, we train an in-context LoRA using paired videos that depict identical motion trajectories but differ in ap- pearance. These pairs are automatically generated and curated through a synchronization-based filtering process that selects only the most temporally aligned examples for training. This training setup teaches the model to combine motion cues from the source video with the visual changes introduced in the edited first frame. Trained on a com- pact, highly curated set of synchronized human portraits, Sync-LoRA generalizes to unseen identities and diverse ed- its (e.g., modifying appearance, adding objects, or chang- ing backgrounds), robustly handling variations in pose and expression. Our results demonstrate high visual fidelity and strong temporal coherence, achieving a robust balance be- tween edit fidelity and precise motion preservation. Project page: https://sagipolaczek.github.io/Sync-LoRA/ . 1. Introduction With the rapid advancement of diffusion-based video edit- ing techniques, text- and image-driven methods have gained prominence for achieving high-quality, temporally consis- tent edits [19, 27, 57]. These techniques hold strong promise for applications in advertising, film production, game development, and interactive media, where fine- grained control over visual content is essential. A particular interest in video editing lies in the editing of portrait videos, as a large portion of visual media fea- tures humans [17, 20, 35]. Editing portrait videos presents a uniquely demanding challenge. On the one hand, the vi- sual appearance of the subject must be modified according arXiv:2512.03013v1 [cs.CV] 2 Dec 2025 to a user-defined instruction. On the other hand, the result- ing video is expected to preserve the exact motion patterns of the source video. This requirement goes beyond visual consistency across frames. The edited video should follow the source on a frame-by-frame basis so that every blink, gaze shift, or articulation occurs at the same moment as in the source. This level of temporal synchronization is critical in talking-head scenarios, where even small misalignment may render the output unfaithful to the source performance. At the same time, the desired edit is often localized and spe- cific, such as adding a mustache, changing an expression, or inserting a visual accessory. The goal is to alter only what is strictly required by the instruction, leaving all other visual content and motion untouched. Achieving this combination of precise synchronization, a minimal editing footprint, and temporal identity preservation remains an open challenge for most existing video editing techniques [17, 27, 29, 37]. Inspired by the In-Context LoRA (IC-LoRA) paradigm [1, 9, 11, 26, 46, 65], we introduce Sync-LoRA, a method designed to generate precisely synchronized edited portrait videos. The edit is guided by modifying only the first frame of the source video, which, together with a text prompt, de- fines the visual transformation to be applied. Our key idea is to fine-tune an image-to-video diffusion model using cu- rated pairs of videos, one showing the original sequence and the other its edited counterpart. These paired videos are constructed to be frame-accurately synchronized, meaning that corresponding frames exhibit identical motion trajecto- ries while differing only in the aspects altered by the edit. To obtain such pairs, we develop a two-stage pipeline: generation followed by synchronization-based filtering. In the generation stage, a vision-language model produces diverse subject prompts and edit

…(Full text truncated)…

This content is AI-processed based on ArXiv data.