cs.CVFeb 26, 2026
SaveFace Time Traveller : Travel Through Ages Without Losing Identity
Organizations: Sony Research India · IIIT Hyderabad
Abstract
Face aging, an ill-posed problem shaped by environmental and genetic factors, is vital in entertainment, forensics, and digital archiving, where realistic age transformations must preserve both identity and visual realism. However, existing works relying on numerical age representations overlook the interplay of biological and contextual cues. Despite progress in recent face aging models, they struggle with identity preservation in wide age transformations, also static attention and optimization-heavy inversion in diffusion limit adaptability, fine-grained control and background consistency. To address these challenges, we propose Face Time Traveller (FaceTT), a diffusion-based framework that achieves high-fidelity, identity-consistent age transformation. Here, we introduce a Face-Attribute-Aware Prompt Refinement strategy that encodes intrinsic (biological) and extrinsic (environmental) aging cues for context-aware conditioning. A tuning-free Angular Inversion method is proposed that efficiently maps real faces into the diffusion latent space for fast and accurate reconstruction. Moreover, an Adaptive Attention Control mechanism is introduced that dynamically balances cross-attention for semantic aging cues and self-attention for structural and identity preservation. Extensive experiments on benchmark datasets and in-the-wild testset demonstrate that FaceTT achieves superior identity retention, background preservation and aging realism over state-of-the-art (SOTA) methods.
Figures & tables
Figure 1: Illustration of real-world use cases of face aging. Top: The Benjamin Button example from The Curious Case of Benjamin Button (2008) , where Brad Pitt was digitally aged using a complex hybrid VFX pipeline [ 8 ] . Modern face aging models can achieve similar visual realism at significantly lower time and cost without prosthetics or manual VFX—while preserving the actor’s identity across different lifespans. Middle (left): Application in personalized character and gaming environments. Middle (right): Digital historical reconstruction. Bottom: Additional use cases.
Figure 2: Illustration of cyclic identity similarity protocol. Input image (age 40) is first re-aged to the target age (60) and then reverted to its original age (40). Top row depicts the reference-based identity similarity between the re-aged and reference image, while bottom row shows the cyclic identity similarity between the input and reconstructed faces.
Figure 3: Overview of the proposed FaceTT framework. Given an input face, the Face-Attribute-Aware Prompt Refinement strategy produces a refined, attribute-rich facial prompts. These prompts are encoded using the embedding network, where the source age embeddings guide the Angular Inversion process to obtain high-fidelity latent reconstructions. Both source and target age embeddings are then injected into the diffusion model through Adaptive Attention Control , enabling precise and identity-consistent age editing. Finally, the age-transformed face is generated.
Figure 4: Illustration of the Angular Inversion technique with prior inversion techniques.
Figure 5: Illustration of Adaptive Attention Control Mechanism.
Figure 6: Visual comparison with SOTA face aging methods.
Figure 7: Visual comparison with Aging Multiverse [ 12 ] . Our method generates better attribute-aware aging while preserving background details (see rows 1 & 2), skin tone (see rows 2 & 3) and accessories such as eyeglasses (see rows 1 & 3).
Figure 8: Visual comparison with MyTM [ 30 ] . MyTM fails to preserve the background (missing dark shadow in the lower-left region) and misses the accessory (earrings). FaceTT maintains consistent background and accessory details.
Figure 9: Short-Range Aging analysis: FaceTT maintains strong identity consistency across closely spaced age targets.
Figure 10: Ablation analysis on hyper-parameters (a) ; (b) ; (c) , . (More examples are provided in supplementary-S3 ).
Figure 11: User study Analysis: our method consistently scores high in both aging and identity preservation. HRFAE’s high score for identity preservation stems from its limited aging ability, often producing results nearly identical to the input.
Figure 12: Ablation analysis on CelebA-HQ for young-to-60 task.
Figure S2 : Visualization of Cross- & Self-attention maps for age progression to 60-year-old appearance. Cross-attention maps highlight regions in the source image (e.g., eyes, nose, mouth) that align with target prompt <Photo of a 60 years old woman/man> . While self-attention visualization showcases top-7 components derived from singular value decomposition (SVD) [ 39 ] , which is important to maintain identity and consistency.
Figure S3 : Comparison of facial re-aging across different age targets using HRFAE, CUSP, FADING, and our FaceTT method (Part 1). Across both examples, FaceTT produces the most consistent identity preservation and the most realistic aging trajectories from early childhood (Age: 2) to older adulthood (Age: 60). Competing methods exhibit notable artifacts: HRFAE often fails to introduce meaningful age changes and tends to output visually similar faces across age ranges; CUSP frequently produces structural distortions and texture inconsistencies (blue boxes) , especially in the eyes and mouth regions; and FADING, while capable of stronger aging effects, suffers from identity drift and unnatural facial modifications (blue boxes) . In contrast, FaceTT generates smooth, coherent age transitions with preserved facial geometry and identity cues, while accurately reflecting age-specific features such as fuller cheeks in youth, subtle mid-life changes, and realistic wrinkle patterns at older ages.
Figure S4 : Additional visual comparison of facial re-aging methods. (Part 2, continued from Figure S3 )
Figure S5 : Additional visual comparison of facial re-aging methods. (Part 3, continued from Figure S3 )
Figure S6 : Cyclic Identity Similarity protocol for evaluating identity preservation across age transitions. For each input celebrity image, we perform age progression and regression cycles (e.g., 40→20→40 (top left), 40→60→40 (top right), 20→40→20 (bottom left), 60→25→60 (bottom right)) and compute both Reference Identity Similarity (between the re-aged output and a real image at the target age) and Cyclic Identity Similarity (between the input and its cyclic reconstruction). Across all age cycles, our method consistently achieves higher identity similarity compared to HRFAE, CUSP, and FADING, demonstrating more faithful identity retention during both forward and reverse age transformations.
Figure S7 : Short-range aging analysis for identity preservation. Each row shows aging results within a narrow age window (±2 years) around the input age. Our method preserves identity exceptionally well across all nearby target ages, producing nearly indistinguishable facial structure, expression, and texture from the original input. This demonstrates that even under fine-grained age perturbations—where inconsistencies are easiest to detect—FaceTT maintains stable and coherent identity features without drifting or introducing artifacts.
Figure S8 : Ablation analysis on different values being used in the proposed Angular Inversion technique. The results demonstrate how different values influence the balance between preserving identity and achieving accurate transformations to match the target prompts (e.g., age 2 or 60 years). Among the tested values, = 1.2 provides the best results, striking an optimal trade-off between transformation fidelity and identity preservation.
Figure S9 : Ablation analysis on different values in the proposed Adaptive Attention Control technique. The results illustrate how varying impacts the trade-off between maintaining structural consistency and achieving accurate aging transformations. Lower values (e.g., = 0.03) retain excessive source attributes, leading to incomplete aging effects, while higher values (e.g., = 0.06) introduce artifacts and distortions in fine details (highlighted in black boxes). The optimal value, = 0.05, provides the best balance, ensuring realistic aging transformations while preserving important identity and contextual details.
Figure S10 : Ablation analysis on different ( , ) values in the Adaptive Attention Control technique, investigating their effect on balancing structural preservation and non-rigid transformations. Lower values of (e.g., 25) retain more source characteristics but limit the effectiveness of the transformation, while higher values (e.g., 40) can lead to excessive modifications, causing distortions in fine details (highlighted in red boxes). Similarly, an improper choice of can result in either overly rigid transformations or undesirable warping effects. The best balance is achieved with ( , ) = (35,15), ensuring natural age transformation while preserving key identity features.
Figure S11 : Statistical analysis. Targeted predicted age: 65.14 4.86
Figure S12 : Comparison of different inversion methods for facial re-aging across multiple target ages (2 → 60 years). Proximal, Null-Text, Direct, and DDPM inversion each introduce distortions or identity drift during the re-aging process. Proximal Inversion often produces inconsistent geometry and blurry structural changes, leading to unnatural aging transitions. Null-Text Inversion exhibits significant shape warping and texture artifacts (blue boxes) , particularly in early-age predictions. Direct Inversion suffers from identity inconsistency and incorrect facial proportions across ages, while DDPM Inversion frequently alters key identity features (blue boxes) , causing the re-aged outputs to deviate from the input subject. In contrast, Angular Inversion yields stable, high-fidelity reconstructions with smooth and realistic age progression, maintaining consistent identity cues from childhood to old age. The results highlight that Angular Inversion achieves the most reliable and artifact-free initialization for age editing, enabling accurate transformations across the entire age spectrum.
Figure S13 : Comparison of age editing using P2P, PnP, MasaCtrl, FPE, and our Adaptive Attention Control ( AAC ) across target ages 2→60. Existing attention-based editing methods exhibit noticeable artifacts and identity inconsistencies: P2P often produces over-smoothed faces and distorted geometry (blue boxes) , especially for younger ages; PnP struggles with preserving identity and frequently introduces incorrect mouth and eye structures; MasaCtrl shows unstable attention modulation, leading to drifted identity and inconsistent aging patterns; and FPE tends to hallucinate textures or distort facial regions (blue boxes) , particularly in older targets. In contrast, AAC achieves smooth and coherent age progression while consistently preserving identity, facial structure, and attribute integrity across the entire age spectrum. The results demonstrate that Adaptive Attention Control provides more precise, stable, and semantically aligned guidance than prior attention-editing approaches.
More in this topic
Identity-Preserving Image Generation
Identity-preserving image generation creates or edits images while retaining a person's recognizable features across changes in pose, style, or scene.
See all 48 papers