AgeBooth: Diffusion Framework for Facial Aging

• AgeBooth is a diffusion-based framework for controllable facial aging and rejuvenation that parametrizes age as a visual factor while preserving identity.
• It employs few-shot age-specific finetuning paired with LoRA adapter fusion via SVDMix to achieve smooth interpolation between age states.
• Empirical results show reduced age error and improved visual realism compared to conventional GAN and editing methods.

AgeBooth is a diffusion-based framework for controllable facial aging and rejuvenation that addresses the challenges of precise age control and identity preservation in personalized face image synthesis. The method enables the generation of visually realistic, identity-consistent facial images at arbitrary points along an age spectrum—using limited age-labeled data—by introducing an age-specific finetuning procedure coupled with advanced LoRA module fusion and prompt blending. AgeBooth represents a major methodological advance over previous GAN-based or editing pipelines by formalizing aging as a parametrizable visual factor, efficiently decoupled from identity information, and enabling high-fidelity aging and rejuvenation for practical applications.

1. Motivation and Problem Setting

Facial aging synthesis plays a critical role in domains such as entertainment, digital forensics, and identity management, yet consistently suffers from a trade-off between accurate age control and preservation of subject identity. Traditional GAN-based and editing methods typically require large paired datasets spanning multiple age stages or risk identity drift and visual artifacts, especially when attempting to interpolate intermediate ages. Earlier adapter-based diffusion personalization models are limited by weak age conditioning and tend to “copy” reference image details, making arbitrary age manipulation unreliable. AgeBooth’s aim is to provide a solution that allows for continuous and finely controlled facial aging/rejuvenation, robustly and efficiently, in a few-shot setting (Zhu et al., 7 Oct 2025).

2. Methodological Framework

AgeBooth comprises two sequential methodological advances: (i) few-shot age-specific finetuning, and (ii) age-specific LoRA and prompt fusion enabled by SVDMix.

Few-Shot Age-Specific Finetuning

The process begins by collecting a small (~10) set of age-labeled images centered around the “young” and “old” age groups. The base diffusion model (e.g., InstantID, PuLID, or InfiniteYou) is then finetuned with age-specific cues. Prompts such as “a person in sbu age” are used to guide the model to associate these age-related features—e.g., wrinkles, skin tone changes, facial geometry—with particular prompt embeddings.

Age-Specific LoRA Fusion via SVDMix

To avoid retraining across the full age spectrum and to facilitate continuous interpolation, AgeBooth uses dedicated Low-Rank Adaptation (LoRA) adapters for each discrete age group (e.g., “young” and “old”). To synthesize intermediate ages, the model interpolates between the learned LoRA modules and associated textual prompt embeddings. SVDMix—an SVD-based matrix fusion scheme—decomposes the LoRA weight matrices of both adapters, linearly interpolates each factor (U, Σ, V, and analogously for A, Λ, Q in the paper’s notation), and recombines them:

$$\begin{align*} B_* & = U_* \Sigma_* V_*^{T} \ A_* & = P_* \Lambda_* Q_*^{T} \quad (* \in \{\text{young, old}\}) \ X_\alpha & = \alpha \cdot X_\text{young} + (1-\alpha)\cdot X_\text{old} \ B_\text{fused}(\alpha) & = U_\alpha \Sigma_\alpha V_\alpha^T \ A_\text{fused}(\alpha) & = P_\alpha \Lambda_\alpha Q_\alpha^T \ \Delta\theta_\text{fused}(\alpha) & = B_\text{fused}(\alpha)\cdot A_\text{fused}(\alpha) \end{align*}$$

This ensures a smooth, continuous mapping in weight space between age extremes, parametrized by an interpolation factor $\alpha\in[0,1]$. Simultaneously, prompt embeddings are blended at the same interpolation rate to effect semantic alignment in the text encoder.

Identity Preservation

An integrated pre-trained ID adapter is injected in the cross-attention layers of the diffusion backbone, stabilizing identity features during inference, independent of the interpolated age target.

3. Technical Innovation and Properties

AgeBooth introduces several technical mechanisms advancing the state of the art:

• SVDMix Matrix Fuse Operator: By interpolating SVD components of LoRA weights, AgeBooth achieves continuity in the visual transition that crude direct linear mixing cannot.
• Prompt Fusion: Ensures that age-relevant semantic guidance from prompt encoders is smoothly matched to the corresponding visual transformation, rather than switching abruptly between discrete age conditions.
• Adapter Modularity: The architecture allows for separate plug-and-play adaptation modules for any number of visual factors—not limited to age—extending the framework's flexibility.
• Data Efficiency: The approach reduces dependence on large age-labeled datasets by using few-shot finetuning, thereby making it feasible in practical data-scarce settings.
• Compatibility with Multiple Diffusion Bases: The method has been applied on various base T2I diffusion models, demonstrating robust transferability and adaptability.

4. Experimental Results and Evaluation

Empirical validation was carried out on multiple public T2I diffusion personalization foundations using datasets spanning two age groups (e.g., 10–20, 70–80 years):

• Qualitative Results: AgeBooth produces visually realistic, identity-preserving outputs across the interpolation spectrum, as evidenced by sharper details, natural skin transformations, and absence of artifacts common in GAN/editing pipelines.
• Quantitative Metrics: Reduction in mean age error (MAE) and mean absolute error compared to prior art is observed. Aesthetic predictors (e.g., LAION aesthetic scores) confirm improved global image quality.
• Ablation Studies: SVDMix interpolation is shown to outperform basic linear blending of LoRA weights, yielding smoother transitions and better correspondence of facial features to target age.
• Sample Efficiency: Performance stabilizes when at least 10 age-specific samples are used for adapter training, indicating high data efficiency.

A summary table of the evaluation protocol is shown below.

Component	Evaluation Metric	AgeBooth Performance
Age Accuracy	MAE	Lower than prior methods
Visual Realism	LAION Aest. Score	Higher
Identity Drift	ID Similarity	Lower drift, more stable

5. Comparative Analysis and Impact

AgeBooth achieves superiority in age control and visual quality relative to state-of-the-art editing-based approaches (e.g., SAM, FADING). Unlike methods requiring paired cross-age data or relying on single-stage prompt embedding, AgeBooth’s LoRA and prompt fusion framework ensures:

• Identity Consistency: Avoids identity drift across large age shifts.
• Continuous Transformation: Enables arbitrary intermediate ages (not just preset bins).
• Absence of Artifacts: Visual output is notably free from abrupt transitions and common generative artifacts.

Its decoupling of age and identity factors provides more robust and scalable deployment prospects, minimizing the risk of overfitting or hallucinated features. The approach opens the domain of plug-and-play age conditioning, supporting broader applications across age progression/regression, digital entertainment, forensics, and avatar generation.

6. Applications and Implications

AgeBooth’s controllable face aging underpins a spectrum of real-world use cases, including:

• Entertainment and Visual Effects: Enables character spanning and realistic age editing for film or game production.
• Security and Digital Forensics: Permits synthetic but identity-consistent “current” faces to be extrapolated for missing-persons searches and age-invariant face recognition.
• Social Media and Digital Avatars: Facilitates customizable virtual representations with realistic and smooth aging.
• Data-Efficient Personalization: Reduces annotation requirements, making age-conditional modules deployable in privacy-sensitive or resource-constrained environments.

A plausible implication is that the methodology could be extended to other attributes (e.g., facial expression, hairstyle) by training discrete LoRA modules, broadening the modular personalized editing landscape.

7. Future Directions

Potential advancements include:

• Learnable Alpha Regression: Directly regressing the interpolation parameter $\alpha$ to a specified age, thereby tightly coupling control signal and output for increased accuracy.
• Broader Attribute Fusion: Incorporating multi-dimensional attribute adapters to simultaneously control factors such as emotion, lighting, or artistic style.
• Further Optimization of Matrix Fusion: Exploring more advanced fusion strategies—potentially non-linear or learned fusions—to improve visual quality and flexibility.
• Extension to Lifespan Modeling: Embedding temporal consistency constraints for life-long aging synthesis, potentially leveraging datasets with more granular temporal sampling.

Additionally, integrating AgeBooth’s LoRA-prompt fusion into future base diffusion models and deploying it in privacy-preserving or federated settings may further improve scalability and security for deployment in sensitive domains.

---

AgeBooth introduces a modular, mathematically robust approach to fine-grained, identity-consistent facial aging via diffusion models, setting a new standard for controllable visual factor editing under limited data conditions (Zhu et al., 7 Oct 2025).