AgeBooth: Controllable Facial Aging and Rejuvenation via Diffusion Models

TL;DR

AgeBooth tackles controllable facial aging with identity preservation in diffusion-based text-to-image generation. It introduces a two-pronged approach: few-shot age-specific finetuning with LoRA adapters to embed aging concepts, and a training-free LoRA and prompt fusion framework using SVD-based fusion (SVDMix) plus prompt interpolation to achieve smooth, continuous age transitions while maintaining identity via an ID adapter. The method yields accurate age control and high-quality, identity-consistent portraits across ages from a single reference image, outperforming prior editing-based approaches in both age accuracy (via MiVOLO) and aesthetics (via LAION predictor). Extensive ablations and cross-backbone evaluations demonstrate the robustness of the aging interpolation and the practical, plug-in nature of AgeBooth for personalized diffusion-based generation.

Abstract

Recent diffusion model research focuses on generating identity-consistent images from a reference photo, but they struggle to accurately control age while preserving identity, and fine-tuning such models often requires costly paired images across ages. In this paper, we propose AgeBooth, a novel age-specific finetuning approach that can effectively enhance the age control capability of adapterbased identity personalization models without the need for expensive age-varied datasets. To reduce dependence on a large amount of age-labeled data, we exploit the linear nature of aging by introducing age-conditioned prompt blending and an age-specific LoRA fusion strategy that leverages SVDMix, a matrix fusion technique. These techniques enable high-quality generation of intermediate-age portraits. Our AgeBooth produces realistic and identity-consistent face images across different ages from a single reference image. Experiments show that AgeBooth achieves superior age control and visual quality compared to previous state-of-the-art editing-based methods.

Figures (14)

• Figure 1: We propose an age transformation method, AgeBooth, which can be applied to various Adapter-based ID customization models, enabling them to generate portraits of a person across different age stages: (a) shows a series of portraits generated by InfiniteYou with AgeBooth; (b) shows results generated by PuLID with AgeBooth using the same input.
• Figure 2: Overview of the AgeBooth framework: AgeBooth first performs LoRA fine-tuning on both young and old age groups, enabling the model to learn the concepts of youth and aging. The resulting LoRA modules for young and old are then fused using the proposed SVD-based method and applied to the base model. To preserve identity and further control the generated age via prompts, a pre-trained ID Adapter is integrated into the model, and the input prompts are interpolated accordingly.
• Figure 3: Comparisons of identity-personalized results with and without our proposed AgeBooth, alongside existing age editing methods.
• Figure 4: Ablation study. (a) Impact of different identity modulation factor $\gamma$ of the pretrained ID adapter. (b) Evaluations of different LoRA interpolation methods introduced in Sec. \ref{['subsec:fusions']}.
• Figure 5: Sensitivity experiment. We vary training size $N \in \{1, 5, 10, 20, 50\}$ to assess data dependence. Identity drift is evident when $N \leq 5$, but stabilizes for $N \geq 10$.