MeInTime: Bridging Age Gap in Identity-Preserving Face Restoration
Teer Song, Yue Zhang, Yu Tian, Ziyang Wang, Xianlin Zhang, Guixuan Zhang, Xuan Liu, Xueming Li, Yasen Zhang
TL;DR
MeInTime is a diffusion-based face restoration method that effectively preserves identity and maintains age consistency in degraded images using cross-age references and a novel age-aware guidance strategy.
Contribution
It introduces a new approach extending reference-based face restoration to cross-age scenarios with a novel age-aware guidance and attention mechanisms.
Findings
Outperforms existing methods in identity preservation.
Achieves superior age consistency in restored images.
Effective in real-world cross-age face restoration tasks.
Abstract
To better preserve an individual's identity, face restoration has evolved from reference-free to reference-based approaches, which leverage high-quality reference images of the same identity to enhance identity fidelity in the restored outputs. However, most existing methods implicitly assume that the reference and degraded input are age-aligned, limiting their effectiveness in real-world scenarios where only cross-age references are available, such as historical photo restoration. This paper proposes MeInTime, a diffusion-based face restoration method that extends reference-based restoration from same-age to cross-age settings. Given one or few reference images along with an age prompt corresponding to the degraded input, MeInTime achieves faithful restoration with both identity fidelity and age consistency. Specifically, we decouple the modeling of identity and age conditions. During…
Peer Reviews
Decision·Submitted to ICLR 2026
+Pioneering Cross-Age Reference-Based Framework: This work introduces the first reference-based face restoration framework specifically designed for cross-age scenarios, effectively extending the capability of existing methods from same-age to cross-age restoration by incorporating target age prompts. + Novel Disentangled Training-Inference Strategy: The proposed method employs a decoupled approach that separately handles identity preservation during training through dedicated attention mechanis
- According to Table 1, the performances of the proposed method are not always the best. The authors should explain the reasons in detail. - The authors do not compare the speed and the number of parameters of the proposed method, compared to existing methods. - The authors do not present the failure cases of the proposed method. I think it is better to analyze the limitations. - The fonts in the Figures are too small. - The effectiveness of the Age-Aware Gradient Guidance is not verified in th
1.The paper targets a practically relevant and previously underexplored problem: high-fidelity, identity-preserving face restoration when only cross-age references are available. 2. The use of a decoupled training-inference strategy—training for identity preservation and introducing age controllability at inference—is a thoughtful response to the lack of large-scale cross-age paired datasets, as discussed with supporting data in Appendix B/Figure 12 (Page 15–16).
1. Limited Theoretical Analysis of Attribute Decoupling: The paper claims that identity and age are decoupled by design (identity during training, age only via gradient guidance at inference), yet it lacks a more principled investigation or proof of whether and to what extent this decoupling is reliably achieved. For example, no formal analysis or visualization is provided to demonstrate that the injected identity embeddings or the age gradients are indeed orthogonal in the learned space. Withou
Solid engineering design, the combination of GRF for stable identity fusion and age-aware gradient guidance is well implemented and empirically effective. Clear structure and writing,the paper is clearly written, visually engaging, and the methodology section is easy to follow. Training-free age control: The gradient-based age guidance is conceptually clean and avoids extra finetuning.
1. The cross-age reference setting is a rare and contrived use case. It’s unclear how many real restoration tasks actually require explicit age matching. The work does not convincingly show that this setting matters beyond a few illustrative examples. The paper starts from the observation that current reference-based face restoration methods assume the reference and degraded faces are of similar age. While this is technically true, the practical importance of bridging “cross-age” gaps in face re
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsFace recognition and analysis · Generative Adversarial Networks and Image Synthesis · Advanced Image Processing Techniques