Abstract
We introduce the Aging Multiverse, a framework for generating multiple plausible facial aging trajectories from a single image, each conditioned on external factors such as environment, health, and lifestyle. Unlike prior methods that model aging as a single deterministic path, our approach creates an aging tree that visualizes diverse futures. To enable this, we propose a training-free diffusion-based method that balances identity preservation, age accuracy, and condition control. Our key contributions include attention mixing to modulate editing strength and a Simulated Aging Regularization strategy to stabilize edits. Extensive experiments and user studies demonstrate state-of-the-art performance across identity preservation, aging realism, and conditional alignment, outperforming existing editing and age-progression models. By transforming aging into a multi-dimensional, controllable, and interpretable process, our approach opens up new creative and practical avenues in digital storytelling, health education, and personalized visualization.
Introduction: Beyond the Linear Aging Path
Traditional facial aging models typically conceptualize the process as a single, deterministic "aging line." This approach assumes that every individual follows a fixed trajectory from youth to old age. However, real-world aging is inherently stochastic and heavily influenced by a multitude of external variables, including environment, health choices, and lifestyle habits. A person who lives in an arid climate with high sun exposure will age differently than someone with a rigorous skincare routine in a temperate zone.
We introduce the "Aging Multiverse" framework, which moves beyond linear progression to generate an "aging tree." This tree visualizes multiple plausible futures for a single identity, each conditioned on specific external factors such as skincare, lifestyle, and environment.
The Three-Way Trade-off in Facial Editing
The primary challenge in conditional facial aging is balancing three competing objectives:
- Identity Preservation: Maintaining the unique facial features that make a person recognizable.
- Age Accuracy: Ensuring the generated image accurately reflects the target chronological age.
- Condition Control: Faithfully representing external prompts such as "alcoholism," "sun damage," or "weight gain."
Existing image-editing techniques often fail this "three-way tug-of-war." For instance, some models preserve identity perfectly but produce "under-aged" results, while others achieve realistic aging at the cost of "identity drift," where the subject no longer looks like themselves.
Methodology: Training-Free Diffusion and Attention Mixing
Our framework utilizes a training-free approach built upon Rectified Flow (RF) models, specifically the Flux.1-dev architecture. This allows the system to leverage the vast prior knowledge of large-scale diffusion models without the need for expensive fine-tuning.
Prompt Refinement via LLMs
High-level conditions like "lifestyle stress" or "alcoholism" are difficult for diffusion models to interpret directly at the pixel level. We employ GPT-4o to expand these abstract concepts into detailed visual descriptions (e.g., "sunken eyes, sallow skin, and increased forehead furrowing"). These refined prompts provide more precise guidance for the generation process.
Attention Mixing
To navigate the identity-vs-editability trade-off, we introduce Attention Mixing. By modulating the Key (K) and Value (V) tensors within the model's attention blocks, we project inversion features (identity-preserving) onto editing features. This mechanism amplifies signals where the edit aligns with the subject's structure and suppresses conflicting signals, allowing for significant aging transformations that remain anchored to the original identity.
Simulated Aging Regularization (SAR)
To ensure the results remain within the realm of realistic human aging, we implement SAR. By pre-calculating "age directions" in the attention space using clusters of young and old faces, we provide a semantic guide that prevents the model from generating distorted or out-of-distribution facial structures.
Experimental Results and Evaluation
We evaluated the Aging Multiverse framework using both celebrity datasets (providing real-world longitudinal ground truth) and a new "in-the-wild" non-celebrity dataset.
Quantitative and Qualitative Performance
Our method achieved state-of-the-art performance across several metrics. In celebrity comparisons, the framework ranked highest in CLIP-T (prompt alignment) and AgeMAE (age accuracy). Qualitatively, the model produces distinct, realistic textures—such as skin leathering or specific hair loss patterns—that surpass the generic wrinkling seen in baseline models like FADING or FlowEdit.
User Study
A pairwise preference study involving human participants confirmed these findings. Participants preferred the Aging Multiverse results in 85% of cases, noting a superior balance between the realism of the aging effect and the preservation of the subject's original identity.
Conclusion and Ethical Considerations
The Aging Multiverse transforms facial aging into a multi-dimensional, controllable process. While the results are highly realistic, it is important to note that these are algorithmic simulations, not deterministic predictions. These tools are intended for creative storytelling, health education, and personalized visualization, and should not be used for formal facial identification. By treating aging as a branching path, we offer a more nuanced and representative view of how our lives and environments shape our appearance over time.
Summary updated