ReAge3D: Re-Aging 3D Faces with View Consistency

Abstract: We present a novel framework for realistic and controllable 3D face re-aging which produces highly detailed, identity-preserving results. Existing 3D editing methods, while effective for coarse semantic changes, are not well suited for re-aging, as even small inconsistencies across re-aged 2D views can lead to over-smoothing of subtle but perceptually important age-related details. To address this challenge, we first introduce a 2D diffusion-based re-aging model, DiffReaging, trained on synthetically generated image pairs. We further propose a center-out editing propagation strategy that leverages this re-aging model to reconstruct multi-view-consistent re-aged images. Specifically, starting from a re-aged frontal pivot view, we reconstruct the remaining views through warping and our proposed Masked-DiffReaging process. By injecting existing content at every step of the diffusion process, Masked-DiffReaging ensures that the reconstructed regions remain coherent with existing pixels. The resulting consistent set of re-aged views supervises the optimization of the re-aged 3D representation. Our method outperforms existing 3D editing techniques both visually and quantitatively, enabling smooth, fine-grained control over age transformations in 3D face models.

• The paper introduces a diffusion-based 2D re-aging model integrated with multi-view propagation to maintain pixel-level consistency in 3D face editing.
• It employs a pivot frontal view and warping strategy with Masked-DiffReaging to accurately propagate detailed age cues across neighboring views.
• Quantitative results and ablation studies demonstrate that ReAge3D outperforms existing methods by reducing age prediction errors and preserving identity.

Authoritative Summary of "ReAge3D: Re-Aging 3D Faces with View Consistency" (2606.18156)

Introduction and Context

"ReAge3D" introduces a novel, practical framework for 3D face re-aging that targets high-fidelity, multi-view-consistent appearance modifications on explicit 3D representations (primarily 3D Gaussian Splatting, 3DGS). The methodology addresses core deficiencies of prior 3D face editing methods, which typically fail to preserve fine age cues (such as wrinkles) due to inconsistent supervision across re-aged 2D views. This work leverages a custom-trained diffusion-based model for 2D re-aging, achieves robust details and identity preservation, and extends these edits to 3D while ensuring pixel-level multi-view consistency.

Diffusion-Based Re-Aging Model (DiffReaging)

The pipeline centers around DiffReaging, a latent diffusion model fine-tuned for age transformation tasks using synthetically generated, identity-consistent image pairs from the FFHQ dataset manipulated by SAM [APCO21]. DiffReaging is conditioned on both the input image and a textual prompt specifying the target age, notably "Photo of a {target age} years old person," and is trained with standard diffusion loss in the LDM space [RBL*22].

The model exhibits strong control over age attributes, supporting continuous and precise transitions from 10 to 80 years. DiffReaging systematically outperforms prior diffusion-based re-aging methods (e.g., FADING [CL23a]) and encoder-decoder approaches (e.g., FRAN [ZCS*22]), both qualitatively and quantitatively, as measured by age estimation and identity similarity scores. Numerical results highlight average age prediction errors ranging 2.3–6.6 years and identity scores of 0.684±0.184, surpassing benchmarks.

Multi-View Consistency: Masked-DiffReaging and Propagation

To extend DiffReaging to 3D, ReAge3D introduces a robust multi-view consistent editing strategy. Rather than independently editing each render, the method begins with a "pivot" frontal view, applies its DiffReaging transformation, and warps this result to neighboring views using optical flow. Occluded regions in these warped images are reconstructed via Masked-DiffReaging, a novel variant that injects known content at every diffusion timestep, controlled by a per-pixel confidence mask. This ensures coherence with the existing content and fine local propagation of age cues.

The center-out strategy is employed, iteratively expanding edits from the pivot to concentric layers of neighboring views, avoiding redundant inpainting of overlapping regions and further enhancing consistency and detail retention across the full viewing hemisphere.

Optimization of 3D Representation

The set of multi-view-consistent, re-aged 2D images supervises optimization of the 3DGS model via a combined $L_1$ and SSIM loss between rendered and target images. Optimization updates are performed periodically, with re-aged target supervision updated every 400 steps, yielding improvements in both flow accuracy and final details.

The framework is agnostic to representation type and renderer, though 3DGS is adopted for demonstration due to computational efficiency. Fine-grained geometric changes (as seen in normal maps) validate that the model modifies physical facial structures in a manner consistent with actual age-related transformations.

Results and Comparisons

ReAge3D is evaluated against multiple baselines, including state-of-the-art 3D editing frameworks (InstructGS2GS [VH24], GaussianEditor [CCZ*24], DGE [CLV24]), hybrid methods integrating DiffReaging, and AgeTrans3D [LGLG24]. The method consistently produces more realistic, higher-resolution, identity-preserving, and age-accurate results, especially at advanced ages. Previous baselines show significant artifacting, over-smoothing, or underestimation of target ages (e.g., baselines regress to ~60 years for 80-year targets). ReAge3D preserves critical age cues (e.g., deep wrinkles, skin sagging) at the image and geometric level.

Quantitative analyses across a dataset of ten subjects and multiple facial expressions demonstrate robust control over age transformation (age error 3.0–9.6, 13.9% average), with substantially higher identity scores than alternatives (0.671±0.019 vs. 0.149–0.218).

Ablation studies further validate the technical contributions, showing that each modular addition (pivot reference, warping, center-out propagation) incrementally reduces artifacting, improves detail fidelity, and enforces consistency.

Limitations and Future Directions

The framework is presently limited by the treatment of occlusions in portrait images and by its focus on facial regions — other age-related features (hair, accessories, hands) are not addressed. Severe occlusions (e.g., sunglasses) are particularly challenging. The approach does not explicitly model biophysical skin appearance changes (albedo, redness, translucency [IGAJG15]), unlike AgeTrans3D.

Refinements could integrate biophysical models for improved realism and extend editing to non-facial regions. Combining the propagation framework with single-image or sparse-view 3D-aware diffusion generation would reduce reliance on dense multi-view captures.

Practical and Theoretical Implications

ReAge3D establishes a new standard for controllable, identity-preserving, and pixel-level consistent age editing in 3D assets. Practically, it supports productions in entertainment/VFX, forensics, and biometrics that require accurate and localized face aging or de-aging. Theoretically, it demonstrates an effective mechanism for propagating fine semantic edits across views in explicit 3D representations, bridging advances in generative diffusion modeling and 3D scene optimization. Masked-DiffReaging offers a generalizable paradigm for localized, view-consistent inpainting in multi-view editing.

Potential future developments in AI may leverage the core center-out propagation, iterative refinement, and masked content injection strategies for other 3D semantic editing tasks requiring spatial and identity consistency, beyond aging.

Conclusion

"ReAge3D" (2606.18156) advances fine-grained, multi-view-consistent 3D face re-aging by integrating a diffusion-based 2D re-aging model with pixel-level content propagation and warping strategies. The framework achieves high-resolution, identity-preserving edits with precise age control, surpassing state-of-the-art baselines both qualitatively and quantitatively. Its modular, representation-agnostic design and robust results across a range of expressions and subjects position it as a foundational solution for photorealistic 3D face age transformation, with promising extensibility to broader semantic editing scenarios and integration with biophysical modeling.