From Cradle to Cane: A Two-Pass Framework for High-Fidelity Lifespan Face Aging

We sincerely appreciate your constructive feedback. We are grateful for your recognition that the paper is well-written and the experiments are solid; that our proposed framework is reasonable and well-validated. Below, we provide detailed responses to each of your specific points.

W1: The novelty of the experimental components is relatively limited. Techniques like SVD decomposition and CLIP-space interpolation have been extensively explored during the StyleGAN era. The first stage is essentially an image-to-image translation (SD-edit).

We agree and acknowledge that the individual components we employ, such as SVD reweighting, and CLIP/ArcFace embeddings, have established precedents in the broader fields of generative modeling and diffusion-based editing.

However, we respectfully argue that our core novelty lies not in the invention of the base components, but in the creation of a novel framework to solve a challenging and previously under-explored problem: the persistent Age-Identity (Age-ID) tradeoff in facial age editing.

Our main contributions and points of novelty are:

(1) Problem Formulation and Analysis: To the best of our knowledge, we are the first to conduct an in-depth, systematic investigation of the Age-ID tradeoff. While implicitly known, this problem has not been formally analyzed or addressed with a targeted solution. Our work formally identifies this critical challenge, substantiating it with empirical evidence (Fig. 1 and Fig. 2), which itself is a valuable contribution to the field.

(2) Novel Two-Pass Architectural Design: Our central innovation is Cradle2Cane, a novel two-pass framework architected specifically to resolve the Age-ID tradeoff. Prior methods attempt to modify age and preserve identity simultaneously in a single, entangled step, leading to suboptimal results. Our approach's novelty lies in decoupling these competing objectives. By separating age accuracy from identity preservation into two distinct passes, we introduce a new and effective strategy that fundamentally differs from previous single-pass editing paradigms.

(3) Fine-Grained Control via Adaptive Noise Injection: While the concept of noise controlling editing strength is known, our approach introduces an adapative mechanism. We propose an adaptive noise injection strategy that calibrates noise intensity based on the magnitude of the desired age transformation. Specifically, our method dynamically adjusts the degree of injected noise for each edit—a small age change receives a subtle noise injection, while a large transformation receives a proportionally stronger one. This marks a crucial departure from prior methods, which typically use a static, one-size-fits-all approach (i.e., the same network and fixed parameters) for all age edits, regardless of the transformation's scale. This dynamic calibration is the key to achieving fine-grained control and successfully navigating the delicate balance between age transformation and identity preservation.

In summary, our novelty stems from (1) the formal analysis of a key problem, (2) a novel system architecture (Cradle2Cane) to solve it, and (3) a refined control mechanism (AdaNI) enabling superior performance. We believe this multifaceted approach presents a significant and robust contribution to the specialized domain of facial age editing.

W2: The results indicate that some age-irrelevant features may still be altered. For instance, in Fig. 9, the male subject's clothing changes, which is not expected to correlate with aging.

Thank you for pointing out this issue. We acknowledge this limitation and have discussed it in Appendix A. The issue stems from our adaptive noise injection design, which increases editability for large age gaps but can potentially alter finer image semantics beyond facial identity.

A potential solution is to employ a mask-guided editing strategy. This would involve using a segmentation model to generate a mask that isolates the facial region, thereby protecting accessories and the background. During each denoising step, the latent features from the original image could then be used to restore the non-facial regions. This approach would ensure edits are confined to the facial region, better preserving accessories and background details, especially during large age transformations.

We have updated our limitations section to include this discussion and plan to explore this solution in future work.

We sincerely appreciate your feedback. We are grateful for your recognition that our proposed framework is novel and elegant; that the techniques are well-motivated. Below, we provide detailed responses to each of your specific points.

W1 & Q1 & W3 & Q3: Limited benchmark scope. All quantitative results are on CelebA-HQ (aligned and “in-the-wild” variants). Generalisation to standard ageing sets such as AgeDB-30, CACD, or FG-NET is not shown. Comparative breadth. No comparisons against recent generic face-editing systems (e.g., StyleCLIP[1] or FaceDNeRF[2]) that could also perform age edits.

We sincerely thank the reviewer for you valuable feedback. Following the suggestions, we have expanded our experiments to include the standard aging datasets AgeDB, CACD and FG-NET, and broadened our comparisons to include the suggested baselines, StyleCLIP and FaceDNeRF. We also include the top-performing methods from our original comparison for context.

For each dataset, we generated approximately 200 images and computed metrics for age difference (age dif), identity similarity (id sim), image quality (quality), and harmonic consistency score (hcs) using the Qwen-VL evaluation. The comprehensive results are summarized in the table below.

AgeDB

CACD

FGNET

Our method, cradle2cane, demonstrates exceptional performance across all three benchmarks. On AgeDB-30, it achieves state-of-the-art results, securing the best age accuracy (age dif) and the top harmonic score (hcs), which underscores its superior overall balance. This strong performance continues on the CACD dataset, where our method again delivers the best age accuracy while maintaining highly competitive scores on other metrics. Even on the highly challenging FG-NET dataset, cradle2cane proves its robustness by delivering the second-best age accuracy and demonstrating strong, consistent performance against all competitors.

W2 & Q2: Lower identity scores vs. GAN methods. Methods like HRFAE retain higher ArcFace similarity. The paper lacks an ablation study on how the loss weights λ₁, λ₂ (ID) versus λ₃, λ₄ (age) affect the age-ID trade-off.

Regarding the ablation study on loss weights, our approach was twofold. First, we initialized the weights by adopting established values from pervious works [1,2,3]. This standard practice ensures a balanced starting point where neither the age nor identity objective initially overpowers the other.

More importantly, a key architectural advantage of our two-pass framework is its reduced sensitivity to these initial λ values. Unlike single-pass methods that must perfectly balance competing losses simultaneously, our first pass aggressively optimizes for age accuracy, and the second pass focuses almost exclusively on restoring identity. This decoupling makes the final output more robust to minor variations in the initial loss weights.

While a full grid search on all λ parameters is computationally prohibitive within the short rebuttal period, we did conduct a dedicated ablation on the most critical new hyperparameter in our design: the AdaNI threshold, which directly controls the trade-off. The results confirm our choice provides an optimal balance.

We believe this, combined with our framework's inherent robustness, validates our design. However, we agree a comprehensive analysis is valuable and commit to including a full ablation study on the λ weights in the appendix of the final manuscript.

[1]: Alaluf, Yuval, Or Patashnik, and Daniel Cohen-Or. "Only a matter of style: Age transformation using a style-based regression model." ACM Transactions on Graphics (TOG) 40.4 (2021): 1-12.
[2]: Li, Senmao, et al. "Interlcm: Low-quality images as intermediate states of latent consistency models for effective blind face restoration." The Thirteenth International Conference on Learning Representations (2025).
[3]: Liu, Tao, et al. "One-prompt-one-story: Free-lunch consistent text-to-image generation using a single prompt." The Thirteenth International Conference on Learning Representations (2025).

We sincerely appreciate your constructive and insightful feedback. We are grateful for your recognition that our proposed framework is novel, intuitive and elegant; that the techniques are well-motivated and achieve comprehensive evaluations. Below, we provide detailed responses to each of your specific points.

W1: While the overall framework is novel, the core ideas behind the individual components build heavily on existing work. The concept that noise level controls editing strength in diffusion models is well-established , and the use of SVD to disentangle features has also been explored previously.

We agree and acknowledge that the individual components we employ, such as SVD reweighting, and CLIP/ArcFace embeddings, have established precedents in the broader fields of generative modeling and diffusion-based editing.

However, we respectfully argue that our core novelty lies not in the invention of the base components, but in the creation of a novel framework to solve a challenging and previously under-explored problem: the persistent Age-Identity (Age-ID) tradeoff in facial age editing.

Our main contributions and points of novelty are:

(1) Problem Formulation and Analysis: To the best of our knowledge, we are the first to conduct an in-depth, systematic investigation of the Age-ID tradeoff. While implicitly known, this problem has not been formally analyzed or addressed with a targeted solution. Our work formally identifies this critical challenge, substantiating it with empirical evidence (Fig. 1 and Fig. 2), which itself is a valuable contribution to the field.

(2) Novel Two-Pass Architectural Design: Our central innovation is Cradle2Cane, a novel two-pass framework architected specifically to resolve the Age-ID tradeoff. Prior methods attempt to modify age and preserve identity simultaneously in a single, entangled step, leading to suboptimal results. Our approach's novelty lies in decoupling these competing objectives. By separating age accuracy from identity preservation into two distinct passes, we introduce a new and effective strategy that fundamentally differs from previous single-pass editing paradigms.

(3) Fine-Grained Control via Adaptive Noise Injection: While the concept of noise controlling editing strength is known, our approach introduces an adapative mechanism. We propose an adaptive noise injection strategy that calibrates noise intensity based on the magnitude of the desired age transformation. Specifically, our method dynamically adjusts the degree of injected noise for each edit—a small age change receives a subtle noise injection, while a large transformation receives a proportionally stronger one. This marks a crucial departure from prior methods, which typically use a static, one-size-fits-all approach (i.e., the same network and fixed parameters) for all age edits, regardless of the transformation's scale. This dynamic calibration is the key to achieving fine-grained control and successfully navigating the delicate balance between age transformation and identity preservation.

In summary, our novelty stems from (1) the formal analysis of a key problem, (2) a novel system architecture (Cradle2Cane) to solve it, and (3) a refined control mechanism (AdaNI) enabling superior performance. We believe this multifaceted approach presents a significant and robust contribution to the specialized domain of facial age editing.

W2 & Q1: The method introduces a significant number of hyper parameters, including noise thresholds for AdaNI, parameters for SVR-ArcFace and different loss weight. The paper provides the final values used but does not include a sensitivity analysis or justification for how these values were chosen.

We thank the reviewer for this important point. Our strategy for hyperparameters was twofold. First, for established components, we initialized values by adopting standards from pervious works [1, 2, 3]. This practice provides a robust and balanced starting point.

More importantly, a key architectural advantage of our two-pass framework is its reduced sensitivity to some of these hyperparameters, particularly the loss weights. Unlike single-pass methods that must perfectly balance competing losses simultaneously, our first pass aggressively optimizes for age accuracy, and the second pass focuses almost exclusively on restoring identity. This decoupling makes the final output more robust to minor variations in the initial loss weights.

While a full sensitivity analysis on every parameter is computationally prohibitive within the rebuttal period, we conducted a dedicated ablation study on the most critical new hyperparameter introduced in our design: the AdaNI threshold, which directly controls the age-identity trade-off. The results confirm our chosen 5_20 threshold provides the optimal balance of age accuracy and identity preservation.

We believe this principled approach, combined with our framework's inherent robustness, validates our design. However, we agree a comprehensive analysis is valuable and commit to including it in the appendix of the final manuscript.

[1]: Alaluf, Yuval, Or Patashnik, and Daniel Cohen-Or. "Only a matter of style: Age transformation using a style-based regression model." ACM Transactions on Graphics (TOG) 40.4 (2021): 1-12.
[2]: Li, Senmao, et al. "Interlcm: Low-quality images as intermediate states of latent consistency models for effective blind face restoration." The Thirteenth International Conference on Learning Representations (2025).
[3]: Liu, Tao, et al. "One-prompt-one-story: Free-lunch consistent text-to-image generation using a single prompt." The Thirteenth International Conference on Learning Representations (2025).

Q2 & Q3: The paper claims superior robustness on in-the-wild human face images, is this majorly come from the standard pre-processing of background removal using Carvekit? If the robustness stems from the model architecture itself, could the authors provide a more detailed analysis of which components (e.g., the SDXL-Turbo backbone, the IDEmb conditioning) are most responsible for this generalization?

Thank you for this insightful question. While background removal with Carvekit is a helpful standardization step, our model's core robustness comes from its architecture, not the pre-processing step.

To confirm this, we ran an ablation study comparing our method with and without Carvekit. We trained both versions for 10 epochs on 1,000 images and tested them on the CelebA-in-the-wild dataset. The results below show that the performance difference is minimal, proving Carvekit is not the primary source of robustness.

The model's robustness is primarily due to two architectural components:

• SDXL-Turbo Backbone: This foundation model was pre-trained on diverse, "in-the-wild" data, giving it inherent generalization against variations in pose, lighting, and expression.
• Our IDEmb Module: This module is key for identity preservation.
  • SVR-ArcFace extracts a stable identity feature by isolating it from variable ones like age or expression.
  • Rotate-CLIP performs a precise, minimal edit in CLIP's robust feature space, adjusting age while preserving the image's original characteristics.

We will add this analysis and ablation study to the revised paper to clarify these contributions.

We sincerely appreciate your feedback. We are grateful for your recognition that our proposed method achieves high-fidelity results with efficient inference and effectively addresses the long-standing trade-off in facial aging. Below, we provide detailed responses to each of your specific points.

W1: The experimental descriptions are somewhat concise.

Thank you for your feedback. We would like to clarify that a detailed description of the experiments was included in Appendix C of our original submission. This section covers the implementation details of our method, the calculation of evaluation metrics, prompt settings, and the implementation details of the baselines. We hope this clarification helps in understanding our experimental setup.

W2: The idea of adaptive noise injection and controlling the degree of editing based on noise intensity has precedents in other diffusion editing works; identity preservation largely relies on existing technologies such as ArcFace/CLIP embeddings, SVD reweighting, and slerp interpolation. The novelty mainly lies in the combination approach rather than entirely new principles.

We agree and acknowledge that the individual components we employ, such as SVD reweighting, and CLIP/ArcFace embeddings, have established precedents in the broader fields of generative modeling and diffusion-based editing.

However, we respectfully argue that our core novelty lies not in the invention of the base components, but in the creation of a novel framework to solve a challenging and previously under-explored problem: the persistent Age-Identity (Age-ID) tradeoff in facial age editing.

Our main contributions and points of novelty are:

(1) Problem Formulation and Analysis: To the best of our knowledge, we are the first to conduct an in-depth, systematic investigation of the Age-ID tradeoff. While implicitly known, this problem has not been formally analyzed or addressed with a targeted solution. Our work formally identifies this critical challenge, substantiating it with empirical evidence (Fig. 1 and Fig. 2), which itself is a valuable contribution to the field.

(2) Novel Two-Pass Architectural Design: Our central innovation is Cradle2Cane, a novel two-pass framework architected specifically to resolve the Age-ID tradeoff. Prior methods attempt to modify age and preserve identity simultaneously in a single, entangled step, leading to suboptimal results. Our approach's novelty lies in decoupling these competing objectives. By separating age accuracy from identity preservation into two distinct passes, we introduce a new and effective strategy that fundamentally differs from previous single-pass editing paradigms.

(3) Fine-Grained Control via Adaptive Noise Injection: While the concept of noise controlling editing strength is known, our approach introduces an adapative mechanism. We propose an adaptive noise injection strategy that calibrates noise intensity based on the magnitude of the desired age transformation. Specifically, our method dynamically adjusts the degree of injected noise for each edit—a small age change receives a subtle noise injection, while a large transformation receives a proportionally stronger one. This marks a crucial departure from prior methods, which typically use a static, one-size-fits-all approach (i.e., the same network and fixed parameters) for all age edits, regardless of the transformation's scale. This dynamic calibration is the key to achieving fine-grained control and successfully navigating the delicate balance between age transformation and identity preservation.

In summary, our novelty stems from (1) the formal analysis of a key problem, (2) a novel system architecture (Cradle2Cane) to solve it, and (3) a refined control mechanism (AdaNI) enabling superior performance. We believe this multifaceted approach presents a significant and robust contribution to the specialized domain of facial age editing.

Q1: The rationale for selecting the thresholds (5 years old and 20 years old) is only briefly mentioned in the text. It is recommended to include ablation experiments to quantitatively analyze the validity of such threshold division.

We thank the reviewer for this valuable suggestion. We have conducted the requested ablation study to analyze the impact of different threshold settings. The results are presented below:

This study reveals a clear trade-off between age accuracy and identity preservation. While larger threshold gaps like 10_30 achieve a higher harmonic mean (hcs), this is driven by improved id_sim at the cost of significantly lower age accuracy (e.g., a higher age_dif of 5.46 vs. 4.92).

For our facial aging task, we prioritize age accuracy as the primary objective. We chose 5_20 because it represents the optimal balance; it delivers state-of-the-art age accuracy while maintaining a high, acceptable level of identity preservation.

We will add this ablation table and a discussion of this trade-off to the revised manuscript to make our parameter choice transparent and well-justified.

Q2: I suggest to the author spell out in pseudocode or a decision table: (a) when to apply each module, (b) how you combine their outputs.

Thank you for your suggestion. We have added pseudocode for our method to the appendix to provide a clearer description.

L1: I suggest that in the Limitation section, the authors could put forward future solutions or directions to address the shortcomings of the proposed method.

Thank you for your suggestion. A potential solution to this limitation is to employ a mask-guided editing strategy. First, a segmentation model would generate a mask to isolate the facial region for editing, thereby protecting accessories and the background. During each denoising step, the latent features from the original image can be used to replace the predicted output in the non-facial regions. This approach would ensure edits are confined to the facial region, better preserving accessories and background details, especially during large age transformations.

We have added this discussion to our limitations section and plan to explore this solution in future research.

Paper Formatting Concerns: It is recommended that Figure 1 specify in the caption what the horizontal and vertical axes represent, and remove the "Age shift value" written alone in the fourth subplot.

Thank you for the valuable suggestion. We have revised the figures as requested: we updated the caption of Fig.1 to specify the axes, removed the isolated text in the subplot, and applied corresponding formatting improvements to Fig.7 in the appendix for consistency. These changes will be reflected in the final manuscript.