AvatarEdit: Optical Flow-Guided Consistent Talking Head Avatar Editing

View at publisher Summarise with VIRAT
Authors: Haijie Yang, Zhenyu Zhang, Jianjun Qian, Jian Jun Yang
Cited by: 0
References: 42

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Pre-trained conditional diffusion models have shown impressive potential in image editing. While these methods can produce satisfactory editing results, they often suffer from inconsistencies in appearance, expression, and temporal coherence, especially in facial images. These issues arise from the independent editing of individual images and the inherent loss of temporal continuity during the editing process. In this paper, we introduce AvatarEdit, a novel framework designed for generating and editing consistent virtual avatars. Unlike traditional image editing methods, AvatarEdit leverages the intuitive and efficient optical f low technique to establish correspondences across different views and frames, thereby enhancing editing consistency. Specifically, given a set of avatar images rendered by a pre-trained 3D Gaussian Splatting model, we propose a multi-view consistency module that selects reference views based on optical flow variations. These selected frames are then used to provide complementary view consistency for the target image, ensuring a unified edited appearance. Additionally, we introduce a temporal consistency module that incorporates optical flow and depth changes from source images into the diffusion process, refining the final edited virtual avatar images to improve temporal stability. Extensive experiments demonstrate that AvatarEdit outperforms state-of the-art methods in terms of appearance, expression, and temporal consistency in the edited avatars.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis 3D Shape Modeling and Analysis Face recognition and analysis

Publication details

Journal / source
IEEE Transactions on Multimedia
Volume
28
Pages
8473–8480
Publication date
1 January 2026
Language
EN