ReAge3D: Re-Aging 3D Faces with View Consistency

Eurographics Symposium on Rendering (2026) I. Gkioulekas and A. Jarabo (Editors)

ReAge3D: Re-Aging 3D Faces with View Consistency Libing Zeng1 , Li Ma2 , Mingming He2 , Ning Yu2 , Paul Debevec2 , Nima Khademi Kalantari1

arXiv:2606.18156v1 [cs.CV] 16 Jun 2026

1 Texas A&M University, 2 Netflix Eyeline Studios

Figure 1: Given a 3D face model, represented here using 3D Gaussian splatting, our proposed method enables precise age manipulation with fine-grained detail while maintaining multiview consistency and preserving identity. It exhibits smooth and continuous transitions across age ranges from 20 to 80 years, effectively capturing age-related changes such as the gradual formation and deepening of wrinkles. Additionally, the method is robust across a range of facial expressions (e.g. neutral, pucker, and smiles) and generalizes across different identities. 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. CCS Concepts • Computing methodologies → Image manipulation; Computational photography; Rendering;

1. Introduction Face re-aging – the task of synthesizing aged or de-aged versions of an individual’s face – is a complex and compelling problem © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

in computer graphics and vision, with diverse applications in entertainment, forensics, and biometrics. In recent years, there has been significant progress in 2D face re-aging, driven by advances in

2 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

deep generative models such as generative adversarial networks (GANs) [GPAM∗ 14] and diffusion models [HJA20]. These approaches typically either invert the input image into a pretrained GAN or diffusion model and perform re-aging via latent manipulation [LJA21, YPN∗ 21, DLQ∗ 19, HHLP20, SGTZ20, APCO21, CL23a,GTLMC22], or train a dedicated network for direct age transformation [ZCS∗ 22]. The latter approach, with FRAN [ZCS∗ 22] being the prominent example, produces high-quality re-aged images by avoiding the identity distortion commonly introduced by latent space projection. Despite these advances, 3D face re-aging, which aims to modify the apparent age of a 3D face model represented using techniques such as 3D Gaussian Splatting (3DGS) [KKLD23] or neural radiance fields (NeRFs) [MST∗ 20], remains relatively unexplored. A key challenge in extending 2D re-aging techniques to 3D is ensuring multi-view consistency. While not specifically designed for re-aging, several recent methods [HTE∗ 23, VH24, CCZ∗ 24, CLV24, LCZ∗ 24, WBL∗ 24, LDY∗ 24, WYW∗ 24] perform general 3D scene editing using pretrained 2D diffusion models. Specifically, these approaches operate by editing rendered or input 2D images using the diffusion model and then use the edited images to optimize a 3D model. To encourage multi-view consistency, they typically rely on either feature-space constraints from the diffusion model [CLV24, WBL∗ 24, LDY∗ 24, WYW∗ 24], controlled noise prediction [LCZ∗ 24], or specific strategies and/or losses during optimization [HTE∗ 23, VH24, CCZ∗ 24]. Although these techniques achieve impressive results for general edits (e.g., “turn him into a clown”), they struggle to produce detailed and age-consistent re-aged models. This is primarily because age-related cues, such as wrinkles and skin texture, are subtle yet perceptually critical, and even small inconsistencies in these features across views can lead to an over-smooth 3D appearance. Furthermore, many of these approaches rely on editing models like InstructPix2Pix [BHE23], which are not designed for fine-grained age control or identity preservation, both critical for realistic 3D face re-aging. To overcome these challenges, we propose a novel framework that utilizes the expressiveness of diffusion models for 2D re-aging and achieves multi-view consistent re-aging directly in the image domain, as in Fig. 1. At the core of our method is a 2D diffusion-based re-aging model, DiffReaging, trained to produce identity-preserving and age-consistent re-aged face images. Specifically, we propose to adapt a general-purpose editing diffusion model to the task of re-aging by fine-tuning it on synthetically re-aged images. Once trained, DiffReaging enables high-quality re-aging of 2D images and serves as a key component in our multi-view pipeline. Rather than re-aging each view independently, which often results in inconsistent age-related details across views, we propose to reconstruct the re-aged images by propagating the content of already re-aged images to other views to ensure multi-view consistency. Starting from a frontal pivot view, we apply DiffReaging to generate a high-quality, identity-preserving re-aged image at the desired target age. We then warp this re-aged pivot to neighboring views using optical flow and reconstruct the missing regions by our proposed Masked-DiffReaging process. This strategy leverages the iterative denoising nature of DiffReaging, injecting the warped

content at each timestep using a confidence mask to ensure that the reconstructed missing regions are coherent with the already existing ones. To avoid independently reconstructing overlapping content in different views, we perform the warping and Masked-DiffReaging in a center-out manner, beginning with the pivot and progressively re-aging neighboring views in concentric layers. The resulting set of multi-view-consistent 2D re-aged images is then used to supervise the optimization of the 3D representation, with the center-out re-aging repeated periodically to refine both the supervision and the final 3D model. In summary, we make the following contributions: • We introduce a novel and practical framework for realistic 3D face re-aging, tackling a previously underexplored and challenging problem. • We present a diffusion-based 2D face re-aging model capable of producing detailed re-aged images with strong identity preservation and age accuracy. • We propose a content propagation strategy to produce multi-view consistent and realistic re-aging across views. Extensive experiments show that our approach outperforms existing 3D editing methods both visually and quantitatively across a wide range of identities and expressions. 2. Related Work 2.1. Face Re-Aging Face re-aging modifies a person’s apparent age in images or videos while preserving identity, realism, and consistency. For in-depth reviews, we refer readers to surveys on face age estimation [RCB09, ATA18] and aging models [GPP18, SXLT16], and highlight key works below. 2D Face Re-Aging. Early methods attempt to capture age progression from limited or unpaired data, spurring the use of adversarial training [GPAM∗ 14] and cycle consistency [ZPIE17]. Notable examples include partitioning data into discrete age groups [KSSS14] and designing conditional GANs [ABD17b, ABD17a, WTLG18]. Later methods refine these ideas by enforcing identity preservation [DLQ∗ 19], disentangling identity from age [HXCL22], or providing continuous re-aging [LJA21, YPN∗ 21]. StyleGANbased approaches introduce latent-space editing [HHLP20,SGTZ20, AZMW21], enabling realistic transformations but often at the cost of identity or fine details [TAN∗ 21, RMBCO22, TMG∗ 22]. More recent works incorporate text-driven diffusion [CL23a, LWH∗ 23] or attribute disentanglement [GTLMC22] for refined results. A complementary approach uses a supervised encoder-decoder with synthetic training data, as in FRAN [ZCS∗ 22], which encodes input faces into a 5-channel representation (including target-age masks). Attentionbased strategies [ZHSea20, CN23] further localize aging to specific face regions, while PADA [LWH∗ 23] integrates CLIP [RKH∗ 21] to manage style-content disentanglement. Despite recent progress, these 2D methods often struggle to generate high-resolution details and exhibit inconsistencies when applied to multiple views. Applying 2D re-aging frame by frame can lead to flickering or identity drift in videos. Early work [DLQ∗ 19] addresses sequential aging via reinforcement learning, but high-quality, temporally consistent solutions remain elusive. Some methods rely on per-frame latent-space editing in StyleGAN [YNGH21], but without further © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

constraints, results exhibit temporal artifacts. Recent improvements suggest hybrid encoders [TMG∗ 22, TAN∗ 21] or specialized modifications [YJLL23] handle cropping and viewpoint changes, while diffusion-based editing [KSK∗ 23, PCWS22, ZLX∗ 25] disentangles time-dependent features from shared identity. However, maintaining stable aging effects across diverse videos remains challenging. More recently, several methods explore personalized or diverse face aging. TimeBooth [SLZ∗ 25] disentangles identity and age representations via bidirectional adversarial learning, and MyTimeMachine [QWG∗ 25] combines a global aging prior with a personal photo collection (∼50 images) through a StyleGAN2-based adapter network. Both methods require cross-age personal photo collections, which are unavailable in typical 3D face re-aging scenarios where the subject is captured at a single point in time. Moreover, their reliance on GAN-based architectures makes it difficult to extend them to masked inpainting, which is essential for propagating reaged content across views. Concurrently, Cradle2Cane [LZL∗ 25] proposes a two-pass diffusion framework that decouples age accuracy and identity preservation, and the Aging Multiverse [GQW∗ 25] generates condition-aware aging trees with diverse trajectories using training-free diffusion. While these methods advance 2D re-aging quality, they all operate without multi-view consistency guarantees. Independently editing multiple views of a 3D face leads to inconsistent age-related details across views, causing over-smoothing when optimizing the 3D representation. Our DiffReaging, by contrast, is designed as a diffusion-based model that naturally extends to Masked-DiffReaging, enabling pixel-level propagation of re-aged content across views with precise numerical age control. 3D-Aware Face Re-Aging. Recent trends aim to incorporate 3D structures into 2D models for more robust view synthesis. 3D-aware GANs [CLC∗ 22, GLWT22, XLSL22] achieve high-quality viewpoint synthesis, but GAN-based 3D inversions [KLA∗ 20, YZL∗ 23, BNS23] can miss fine details. Diffusion models better preserve subtle facial features [CL23b] but struggle with view consistency, motivating integration with 3D face structures [WZRT24, OT24]. DiffAge3D [WZRT24] uses 3D-aware GAN priors and diffusionbased refinements for faithful multi-view re-aging but depends on a large-scale 3D-aware dataset. AgeTrans3D [LGLG24] performs 3D-aware re-aging by uplifting the result of a 2D re-aging diffusion model, together with the simulation of biophysical skin properties [IGAJG15]. However, its 2D re-aging model is trained from scratch as a low-resolution pixel-space diffusion model, which limits its performance. Another group of 3D generative models [HZG∗ 23,XCG∗ 24,SWY∗ 23,LXJ∗ 23,LSC∗ 24] can reconstruct 3D models from single or sparse views by leveraging multi-view consistency in diffusion models. While they achieve consistent generation across views, their fidelity and novel-view synthesis quality remain limited. 2.2. 3D Representation and Editing Recent advances in 3D representations, such as neural radiance fields (NeRFs) [MST∗ 20] and 3D Gaussian Splatting (3DGS) [KKLD23], have significantly transformed the field of modeling complex 3D scenes, enabling free-viewpoint photorealistic rendering. A range of subsequent works further improve the performance of highquality 3D reconstruction [JTA21, YYTK21, BMT∗ 21, CFHT22, © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

3 of 14

MESK22, CXG∗ 22, BMV∗ 23, ZLW∗ 22, CLTS24, LRS∗ 24, NSW24, YLCL24, YCH∗ 24, CXZ∗ 24] and inspire applications like 3D asset creation, often combining with diffusion models [PJBM23, TRZ∗ 24, YFW∗ 24]. Editing 3D representations such as NeRFs and 3DGS remains challenging. The implicit nature of NeRF-based representations often necessitates global editing through text prompts [WJC∗ 24, BHE23, DW23] or reference images [HHY∗ 22, BZY∗ 23]. As an explicit representation, 3DGS supports direct edits by modifying Gaussian parameters. Prior methods [YJM∗ 24, HSY∗ 24] enable manipulation of 3DGS scene elements and extend this capability to motion control in dynamic scenes. However, fine-grained editing, such as re-aging, cannot be achieved by explicit editing and often benefits from the integration of text- or image-guided methods and diffusion models. Text-guided 3DGS editing methods typically incorporate 3DGS with text embeddings from large vision-language models, such as CLIP [RKH∗ 21], or utilize pretrained diffusion models [HJA20, ALF22, BHE23, KZL∗ 23, CWWL24] to edit the training images based on text prompts. Some approaches [PSBR24,CWWL24] leverage Score Distillation Sampling (SDS) [PJBM23] to guide image editing, while others [ZCJ∗ 24, CCZ∗ 24, WFZ∗ 24, XCH∗ 24, VH24, ZKC∗ 24] utilize image-conditioned diffusion models to improve spatial control. To enable localized editing, some works [ZCJ∗ 24, CCZ∗ 24, WFZ∗ 24, XCH∗ 24] propose first segmenting or tracking regions of interest and then applying text-guided modifications. Another line of research on global editing, such as InstructGS2GS [VH24] and TIP-Editor [ZKC∗ 24], iteratively edits input images using image-conditioned diffusion models like InstructPix2Pix [BHE23] while optimizing the underlying 3D scene. Although these methods can produce edits faithful to text prompts, they often suffer from multi-view inconsistency and blurry reconstruction due to discrepancies across guidance images. To address view inconsistency, recent approaches integrate 3D geometry cues, such as epipolar constraints [CLV24], depth supervision [LCZ∗ 24, WBL∗ 24, LPS∗ 24], or multi-view cross-attention maps [WYW∗ 24, LPS∗ 24, WQP∗ 25] into the diffusion models, or align attention-based latent codes using reference views [WBL∗ 24]. However, these constraints are typically applied in latent space and can not ensure pixel-level consistency. We propose an optimization framework that combines an image-conditioned re-aging model with 3DGS reconstruction to achieve fine-grained view-consistent results.

3. Preliminaries 3D Gaussian Splatting (3DGS) Although our technique is agonistic to the choice of 3D representation, we implement our system using 3DGS [KKLD23], which explicitly represents a 3D scene as a collection of N anisotropic Gaussians G = {Gi }N i=1 . Each Gaussian primitive in 3DGS is defined with a set of optimizable parameters, Gi = {pi , Σi , ci , αi }, consisting of 3D position pi , anisotropic covariance matrix Σi , RGB color ci , and opacity αi . Given the representation and a camera pose v, 3DGS uses a fast differentiable renderer R to produce an image I = R(G, v). We demonstrate the results of our approach on 3D faces, represented with 3DGS.

4 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

Figure 2: Overview of Our 3D Face Re-Aging Framework. We first train a diffusion-based re-aging model, DiffReaging, on a synthetic dataset. To extend its re-aging capability to 3D, we introduce a multi-view consistent editing framework. Starting with the pivot view, we re-age it using DiffReaging and warp the results to other views. The Masked-DiffReaging model, adapted from DiffReaging with additional mask constraints, is then used to reconstruct occluded regions. Using a center-out re-aging strategy, we progressively re-age surrounding views. Finally, all re-aged images are used to update the 3D model. Diffusion Models The diffusion model generates an image x0 by progressively denoising pure noise through a series of intermediate steps and consists of a forward and reverse process. In the forward process, noise ϵ with progressively increasing strength is added to clean data x0 to produce noisy images at different time √ steps √ t ∈ {0, . . . , T }, following the formulation xt = ᾱt x0 + 1 − ᾱt ϵ, where ᾱt is the cumulative product of noise scaling factors across time steps. The reverse process aims to invert the forward process by denoising xT back to x0 using the learned denoising function ϵθ , where θ denotes the model parameters. This denoising diffusion model takes as input the current image xt and typically a text prompt c and estimates the noise, i.e., ϵ̂ = ϵθ (xt , c). The model is trained by optimizing the following objective: h i Ldiff = E ∥ϵθ (xt , c) − ϵ∥2 . (1) In our method, we build our 2D re-aging model upon a variant of diffusion models, known as latent diffusion models (LDMs) [RBL∗ 22]. LDMs follow the same overall procedure but operate in the latent space of a variational autoencoder (VAE) to reduce computational and memory costs. 4. Method Given a 3D face representation Go , our objective is to produce a re-aged 3D face Gr that preserves the identity of the original face and accurately reflects the target age, as illustrated in Fig. 1. Note that our approach primarily focuses on re-aging the face, as other age-related features, such as hair, clothing, and accessories, are typically handled through practical effects in production [ZCS∗ 22]. A straightforward way to tackle this problem is to first set up a set of N predefined camera poses vi and render images using the original 3D model {Ii = R(Go , vi )}N i=1 . An existing 2D re-aging approach can then be used to produce the corresponding re-aged images Li . These re-aged images can, in turn, be used to optimize a new re-aged 3D representation. This simple approach, however, leads to unsatisfactory results, as the 2D re-aging model produces images that lack 3D consistency, i.e., it adds age-related details, such as wrinkles, inconsistently across different views. Using such images to optimize the 3D model thus leads to overblurring of the details. We address this challenge by proposing a novel framework that

produces re-aged, highly detailed 3D models with the age-consistent appearance. The key idea of our approach is to ensure multi-view consistency of age-related details by propagating the content of a re-aged image to other views, rather than re-aging each view independently. To this end, we introduce DiffReaging (Sec. 4.1), a 2D diffusion-based re-aging model capable of generating age-consistent, identity-preserving images. Starting from a pivot view (frontal in our implementation), we apply DiffReaging and propagate the reaged content to the neighboring views through warping (Sec. 4.2.1). We then reconstruct the missing regions in these views through our proposed Masked-DiffReaging strategy (Sec. 4.2.2). Crucially, Masked-DiffReaging leverages the iterative denoising process of diffusion models to inject known content at every step, allowing the model to reconstruct missing regions coherently with the existing ones. To avoid redundant or inconsistent reconstructions of the same 3D region across multiple views, we perform warping and Masked-DiffReaging in a center-out manner, starting from the pivot and progressively expanding to surrounding views (Sec.4.2.3). The resulting set of 2D re-aged images is then used to supervise the optimization of a 3D face model (Sec. 4.2.4), with periodic updates to the supervision using the center-out re-aging strategy. An overview of our framework is shown in Fig. 2, and the following sections detail each component. Note that our approach first generates view-consistent re-aged targets, which are then used to supervise a 3D face model. As a result, the framework is agnostic to the underlying 3D representation and supports optimization with any differentiable renderer (e.g., NeRF [MST∗ 20] or 3DGS [KKLD23]). In this work, we primarily demonstrate results using 3DGS due to its computational efficiency. 4.1. DiffReaging Given an input image I and a target age, our goal is to generate a re-aged image L of the same person at the specified age. Inspired by the existing diffusion-based image-to-image translation methods [BHE23], we propose to model the process using a conditional diffusion model. Specifically, our model, dubbed DiffReaging, takes the input image I and the target age in the form of a text prompt, “Photo of a {target age} years old person.” as the condition and perform the iterative denoising to produce the re-aged image x0 from pure noise xT . Since we use latent diffusion models, the denoising process occurs in the latent space and the final re-aged image L is obtained by decoding the clean latent x0 . © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

5 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

fashion. These re-aged images are then used to optimize a detailed and age-consistent 3D model. We begin by describing our warping module and proposed Masked-DiffReaging technique, which both are key to propagating details across views. We then explain how these components are used together to reconstruct the re-aged views in a center-out manner. Finally, we detail our optimization strategy. 4.2.1. Flow and Warping Figure 3: Masked-DiffReaging. Given a warped re-aged image with missing regions (represented as black holes), Masked-DiffReaging inject the incomplete result into the diffusion process of DiffReaging at every step t with a confidence mask. While all operations are conducted in latent space, we show the original image here for illustration purposes. To train our model, we require triplets consisting of pixelaligned source (input) images, corresponding re-aged (target) images, and the associated target age. We construct this dataset using SAM [APCO21], a GAN-based age-conditioned face image generator. Although SAM often fails to preserve the identity of the input image in its re-aged outputs, prior work [ZCS∗ 22, LGLG24] has shown that the identities across these re-aged outputs are internally consistent. Based on this observation, we use the re-aged images generated by SAM to form our training triplets. Specifically, we select 2,000 identities from the FFHQ dataset [KLA19] and use SAM to re-age each image to eight target ages ranging from 10 to 80 years. During training, we randomly sample two of the eight re-aged images for each identity. One image is used as the source I, while the other L′ serves as the basis for the target. To ensure that the model focuses on learning age-related changes primarily in the facial region, we compute a face mask M using BiSeNetV2 [YGW∗ 21] and form the final target as L = (1 − M) ⊙ I + M ⊙ L′ (see Fig. 2 left). This blending retains the background and non-face regions of the source while applying the age transformation only to the face. Note that the original images from the FFHQ are only used to produce the re-aged images and are not part of our training data. We base our model on pre-trained InstructPix2Pix [BHE23] and fine-tune it on our re-aging dataset using standard diffusion loss in Eq. 1. Note that the diffusion model in this case takes the input image I in addition to the text prompt c as the condition, i.e., ϵθ (xt , c, I). Once trained, DiffReaging is able to produce detailed, age-consistent, and identity-preserving re-aged results that are better than state-ofthe-art 2D re-aging methods (Fig. 5 and Table 1). 4.2. Multi-View Consistent Re-Aging Our goal is to use DiffReaging to produce a detailed re-aged 3D model Gr from the original representation Go . To this end, we render a set of N images Ii = R(Go , vi ) by placing virtual cameras in a grid pattern over a viewing hemisphere centered on the face. Rather than generating the re-aged images Li independently, our key idea is to enforce multi-view consistency by first re-aging the frontal pivot image I p and then reconstructing the remaining views by propagating the re-aged content outward from the pivot in a center-out © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Given the pivot image I p , rendered using the frontal camera pose v p and the original 3D model Go , we first apply DiffReaging to produce a re-aged pivot image L p . Our goal is then to warp the content of the re-aged pivot to the neighboring views. To do so, we compute optical flow from each neighboring view n to the pivot view, denoted as Fn
p , and then align the pivot image to the neighboring view through backward warping: L p
n = W(L p , Fn
p ),

(2)

where W denotes a backward warping function implemented via bilinear interpolation. The main challenge here is that since we aim to warp the reaged pivot, the optical flow must be computed between the re-aged neighboring and pivot views. However, the re-aged neighboring view is precisely what we are trying to reconstruct and is therefore unavailable. To overcome this, we propose computing the optical flow between the rendered neighboring and pivot images using the re-aged 3D model Gr : Fn
p = O (R(Gr , vn ), R(Gr , v p )) ,

(3)

where O is the optical flow operator, and we use SEARAFT [WLD24] in our implementation. We initialize Gr with Go at the start of optimization (Sec. 4.2.4). As optimization progresses, Gr moves towards a re-aged representation, leading to improved optical flow, which in turn enhances the 3D model (see Fig. 12). While optical flow provides dense correspondences, not all flow vectors are reliable due to occlusions and estimation errors. To address this, we compute a binary confidence mask C using the forward-backward consistency check, retaining only the warped pixels associated with reliable flow estimates. An example of the warped image L p
n and its corresponding confidence mask C is shown in Fig. 3. In the next section, we describe our strategy for reconstructing the missing regions in L p
n to obtain the complete re-aged neighboring view Ln . 4.2.2. Masked-DiffReaging Our goal is to reconstruct the missing regions with details that are consistent with the existing content in the warped image. While directly applying DiffReaging to the neighboring image In yields a high-quality re-aged result, the age-related details may not align with those in the warped image L p
n . Inspired by Lugmayr et al. [LDR∗ 22], we address this issue by leveraging the iterative denoising nature of the DiffReaging’s diffusion process. As illustrated in Fig. 3, in our method, dubbed

6 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

Figure 4: Center-Out Re-Aging. Progressive reconstruction of multiview images starting from a central pivot view (labeled as 0). Neighboring views are reconstructed outward in concentric layers. At each iteration, we apply warping and Masked-DiffReaging to generate the next set of views. The labels indicate the corresponding iteration steps. Masked-DiffReaging, we inject the warped image into the diffusion output at each time step t using the confidence mask as follows: xt = C↓ ⊙ x̂t + (1 −C↓) ⊙ E(L p
n )t

(4)

where C↓ is the confidence mask downsampled to the latent resolution, x̂t is the denoised latent at step t, and E is the VAE encoder that maps an image to its latent representation. Moreover, E(Lp
n)t denotes the latent of the warped image after adding noise at timestep t. Finally, xt is the blended latent at step t, used as the input to the diffusion model at the next time step. Starting from pure noise xT , we perform the iterative denoising and blending until we obtain the clean re-aged latent x0 . This latent is then passed through the VAE decoder to generate the final re-aged image at the neighboring view, Ln . 4.2.3. Center-Out Re-Aging While the proposed warping and Masked-DiffReaging can be used to reconstruct all non-pivot views directly from the pivot image, doing so results in independently filling the unmasked areas in each non-pivot view. This, in turn, leads to inconsistencies in the generated age-related details across views that correspond to the same 3D region and negatively impacts the quality of the results, as shown in Fig. 11 (“+Pivot+W”). To address this, we propose a center-out re-aging strategy, where we begin with the pivot and iteratively propagate the content outward by using already reconstructed neighbors to guide the reconstruction of adjacent views. In this strategy, illustrated in Fig. 4, we start from the pivot view at the center (marked as 0) and progressively reconstruct neighboring views in concentric layers. At each iteration, we perform warping and Masked-DiffReaging to reconstruct the next set of views, moving outward from the pivot. Specifically, in the first iteration, we reconstruct the four immediate neighbors (labeled 1) by warping the pivot view and applying Masked-DiffReaging. In the second iteration, we reconstruct the next ring of neighbors (labeled 2), and so on. For each target view, we combine information from multiple previously reconstructed views. As indicated by the arrows in Fig. 4, we warp the pivot view (not shown to reduce clutter) along with the horizontal and vertical neighboring views from previous iterations. At each step, we only add pixels that were not already filled by earlier warps. Once the combined warped image is constructed, Masked-DiffReaging is applied to inpainting the remaining unfilled regions. This strategy reduces the possibility of reconstructing overlapping regions independently

Figure 5: Here, we show comparisons against FRAN. FRAN exhibits limitations in de-aging (top row), as it is not able to effectively remove deep wrinkles. Additionally, its editing capability diminishes when the target age is close to the source age, as shown in the bottom row. In contrast, our method achieves precise age control, enabling effective modification of the age attribute in input images. and subsequently avoids blurring the details, as demonstrated in Fig. 11 (“+Pivot+W+COR”). 4.2.4. Optimization Here, we use the re-aged 2D images to supervise the optimization of the re-aged 3D representation Gr . Specifically, we initialize Gr with the original representation Go , and optimize it by minimizing the following objective:  

L3D = L R(G , v ), L +L R(G , v ), L rec r n n rec r p p , ∑ n∈{1,...,N}\{p}

(5) where the first term is the reconstruction loss between the rendered and re-aged non-pivot view, and the second term is the same loss computed for the pivot view. Since we primarily use 3DGS as our representation, we adopt the same reconstruction loss used in 3DGS, which combines L1 and SSIM. We perform the optimization for 2,000 iterations and update the re-aged non-pivot images by performing the center-out re-aging every 400 steps. As shown in Fig. 12, as our 3D model improves, the flow quality improves which in turn improves the quality of the 3D re-aged representation. 5. Results We implement our approach in PyTorch and use the ADAM optimizer [KB15] with default parameters for both the training of DiffReaging and the optimization of the 3D model. DiffReaging is fine-tuned on images with a resolution of 1024 × 1024 and a batch size of 4 for approximately three days using four NVIDIA A100 GPUs with 40GB of memory. For 3D model optimization, we use a single NVIDIA A5000 GPU with 24GB of memory. We first present results for 2D re-aging, followed by the evaluation of our 3D face re-aging method. 5.1. 2D Face Re-Aging We compare the 2D re-aging results of our DiffReaging with those of the diffusion-based face re-aging approach FADING [CL23a]. As shown in Fig. 6, FADING fails to accurately manipulate agerelated attributes and struggles to preserve the subject’s identity. © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

7 of 14

Figure 6: Compared to FADING, our re-aging diffusion model consistently produces realistic age transformations while preserving the subject’s identity. Furthermore, the age progression shows smooth transitions, exemplified by the gradual appearance and deepening of wrinkles as age increases. Table 1: Evaluating Age Accuracy. We measure age accuracy by calculating the difference between the estimated age of the re-aged results and the target ages. Lower values indicate better performance, with the best results highlighted in bold. Additional results in 5-year increments provided in the supplementary document. Age Error ↓ 10 20 30 40 50 FADING 10.2 17.6 13.1 8.2 8.4 IPix2Pix 9.4 12.3 11.3 12.4 9.3 FRAN 4.2 5.1 3.9 4.9 6.3 Ours 2.3 3.1 3.2 4.0 4.7

ID Score ↑ 60 70 80 7.0 12.3 14.8 0.555±0.232 13.1 17.8 13.6 0.604±0.248 5.1 11.2 13.3 0.675±0.199 3.6 5.0 6.6 0.684±0.184

In contrast, our method demonstrates precise control over the age attribute, successfully re-aging input images to target ages spanning from 10 to 80 years in 10-year increments. The results indicate that our re-aging diffusion model produces realistic outputs while maintaining identity across a wide range of target ages. We further compare our method against FRAN [ZCS∗ 22] in Fig. 5. As shown, FRAN exhibits clear shortcomings, particularly when attempting to de-age the subject. For instance, in the top row, the 20-year-old output still retains deep wrinkles, indicating incomplete removal of age-related features. Additionally, FRAN struggles with age transitions when the target age is close to the source age. This is evident in the bottom row, where the model generates a 42-year-old appearance despite the target age being 50. In contrast, our method produces age-consistent results, demonstrating its effectiveness in modifying age while preserving identity. Moreover, our method, DiffReaging, is explicitly designed as the foundation for Masked-DiffReaging, enabling multiview-consistent and localized re-aging, which further expands the applicability of our method in high-fidelity, view-consistent face editing scenarios. In addition, we conduct a quantitative evaluation of age manipu© 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

lation accuracy and identity preservation against FADING [CL23a], InstructPix2Pix [BHE23] and FRAN [ZCS∗ 22]. Specifically, we randomly sample 1,000 unseen images from FFHQ [KLA∗ 20] and re-age each to target ages from 10 to 80 in 10-year increments. We assess the re-aged outputs using an age estimation model [RTVG18] and a face recognition model [DGXZ19] to compute age prediction errors and identity similarity scores, respectively. The results are summarized in Table 1. Our method consistently outperforms all baselines in terms of age consistency and identity similarity. 5.2. 3D Face Re-Aging We evaluate 3D face re-aging results ten subjects from our internal dataset. Each subject in our internal dataset was captured using a light stage setup [DHT∗ 00] with 76 cameras. These images are used to optimize a 3DGS model using the standard process with densification and pruning, which serves as the input representation Go for our method. As shown in Fig. 1, our approach enables finegrained age manipulation with strong identity preservation and multiview consistency. It supports smooth transitions from ages 20 to 80, capturing gradual changes such as wrinkle formation. Additional results in Fig. 9 show both aging and de-aging across a diverse set of subjects, demonstrating the robustness of our method. Qualitative Comparisons We compare against five baselines: InstructGS2GS (IGS2GS) [VH24], GaussianEditor (GE) [CCZ∗ 24], DGE [CLV24], and two hybrid methods, IGS2GS+DiffReaging (IGS2GS+) and GE+DiffReaging (GE+), in which we integrate our DiffReaging model into the original algorithms. Notably, neither IGS2GS/IGS2GS+ nor our approach performs densification or pruning during the editing process. In contrast, both GE/GE+ and DGE apply explicit densification, following the configurations recommended in their official implementations: every 500 iterations for GE/GE+ and every 100 iterations for DGE.

8 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

Figure 7: We show comparisons against several methods. The baselines IGS2GS [VH24], GE [CCZ∗ 24], and DGE [CLV24], all based on InstructPix2Pix [BHE23], struggle to reage faces to 80 years, often introducing severe artifacts or over-smoothing due to reliance on feature-space constraints. Enhanced variants (IGS2GS+ and GE+) reduce artifacts using our re-aging diffusion model but still fail to reach the target age, often smoothing out critical features like wrinkles. For instance, they make the senior lady (about 60 years with deep wrinkles), appear younger than the input. In contrast, our method achieves accurate aging to 80 years, preserving fine age-related details and maintaining multi-view consistency. accurately reaches the target age of 80, preserving age-related features such as wrinkles. We emphasize that the final renderings produced by all methods are inherently view-consistent, as they are generated from optimized 3D representations. The key difference is that existing methods produce re-aged 2D target images that are not consistent across views. Optimizing a 3D model under such conflicting supervision inevitably averages out high-frequency age cues, resulting in over-smoothed and blurry renderings.

Figure 8: We present a comparison with 2D re-aging of AgeTrans3D [LGLG24]. Our method generates higher-resolution results and more realistic age-related details. As shown in Fig. 7, IGS2GS, GE, and DGE, each based on InstructPix2Pix, struggle to re-age subjects to 80 years, often introducing severe artifacts. In particular, DGE produces over-smoothed results, likely due to its reliance on feature-space consistency, which does not always translate to consistency in image space. While the enhanced variants IGS2GS+ and GE+ incorporate our diffusionbased re-aging model and thus exhibit fewer artifacts than their original versions, they tend to overblur age-related details and fail to achieve the desired target age. Notably, in the case of the senior woman shown at the top (60 years old with deep wrinkles), these methods smooth out key age cues in the re-aged image, making the output appear even younger than the input. In contrast, our method

We also provide comparisons with AgeTrans3D [LGLG24]. AgeTrans3D performs single-image 2D re-aging using a diffusion model while additionally considering biophysical skin properties [IGAJG15], and lifts the re-aged image to 3D. However, its 2D re-aging model is trained from scratch as a low-resolution pixelspace diffusion model, whereas our DiffReaging fine-tunes a pretrained image-editing latent diffusion model with strong priors for photorealistic image generation. This enables our method to produce higher-resolution re-aging results with finer facial details. Since the AgeTrans3D code is not publicly available, we compare against two representative examples from their paper that showcase their best reported results. As shown in Fig. 8, our method produces higher-resolution results with finer age-related details, including natural wrinkles around the eyes. However, unlike AgeTrans3D, we do not model the biophysical skin appearance changes associated with aging, such as the natural lightening of skin tone. To further demonstrate the geometric editing capability of our © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

9 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

Table 2: Age Accuracy Comparison. We measure age errors relative to target ages. Our method achieves low errors across all targets, while the enhanced baselines show significant underestimation for ages 70 and 80, often regressing to about 60, confirming their lack of age consistency. Age Error ↓ 10 IGS2GS+ 4.8 GE+ 4.4 Ours 3.0

20 5.4 4.8 3.8

30 4.6 6.3 3.1

40 5.4 6.0 4.6

50 6.9 7.7 5.6

60 4.8 5.1 2.9

70 11.0 12.1 8.4

80 Avg. % 19.5 20.7% 19.4 21.2% 9.6 13.9%

Table 3: Identity Preservation (ID Score) Comparison. Our method yields higher average ID scores and lower variance, reflecting stronger identity preservation. ID Score ↑

Figure 9: We show de-aging results in the top row and aging results in the bottom row. As observed, age-related features gradually fade during de-aging and progressively emerge during aging. The reaging process maintains strong multi-view consistency.

IGS2GS GE DGE Ours 0.218±0.042 0.149±0.047 0.206±0.026 0.671±0.019

Each subject is re-aged to the target ages of 10, 20, 30, 40, 50, 60, 70, and 80 years. From the final re-aged models, we render 30 views per subject, covering yaw angles from -60° to 60° with a fixed pitch angle of 0°. We compare against IGS2GS, GE, and DGE using ArcFace [DGXZ19], a face recognition method, to compute identity similarity scores (ID Scores) between original and re-aged images. As shown in Table 3, our method achieves higher average ID scores with lower variance, indicating better identity preservation. For enhanced baselines (IGS2GS+ and GE+), we report age errors relative to target ages in Table 2. As seen, the enhanced baselines show larger errors than ours, particularly for age 80, demonstrating their inability to produce age-consistent results. Additional comparisons on ten randomly selected subjects from NeRSemble dataset [KQG∗ 23], as well as further results with 3Daware diffusion models such as LRM [HZG∗ 23, XCG∗ 24] and One-2-3-45 [LXJ∗ 23, LSC∗ 24], are provided in the supplementary document.

Figure 10: Geometry editing effect. Our method enables geometric changes, as demonstrated through normal maps extracted from 3DGS [KKLD23] models. method, we extract normal maps from 3DGS [KKLD23] models, as illustrated in Fig. 10. The visualizations show that as the subjects de-aged or aged, age-related geometric details—such as wrinkles, and skin sagging—emerge clearly in the normal maps, indicating that our method introduces meaningful geometry changes consistent with natural aging. Quantitative Comparisons. To numerically assess the age control capability of our approach, we select ten subjects, each exhibiting four distinct facial expressions: neutral, pucker face, closed-mouth smile (smile without teeth), and open-mouth smile (smile with teeth). © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Ablation Study. To assess the contribution of each component in our proposed method, we conduct an ablation study by progressively incorporating key modules. We begin with the “No Pivot” variant, which is equivalent to the IGS2GS+ baseline. We then introduce the pivot editing reference along with the pivot consistency loss (full loss in Eq. 5), resulting in the “+Pivot” variant. Next, we incorporate the warping strategy to propagate the pivot edit across novel views, forming the “+Pivot+W” variant. Finally, we introduce the Center-Out Re-aging (COR) strategy to obtain the full model, denoted as “+Pivot+W+COR”. Comparative results are presented in Fig. 11. The “+Pivot” model ensures accurate editing of the pivot view, aligning it closely with the target aging attributes. The “+Pivot+W” variant enables effective propagation of this edit to other viewpoints, as illustrated by the blue inset, which shows improved details compared to the red inset. Lastly, the full “+Pivot+W+COR” model further enhances detail preservation by reducing the independent reconstruction of overlapping regions in different views. This is particularly evident in the orange inset, which exhibits sharper and more refined details compared to the corresponding green inset. The corresponding numerical ablation study is provided in the supplementary document.

10 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces Re-aged

Figure 11: Ablation Study. The "+Pivot" variation ensures that the pivot view of the 3D model is effectively edited to match the target attributes. The "+Pivot+W" variation facilitates the propagation of information from the pivot view to other views, as evidenced by the blue box, which demonstrates superior results compared to the red box region. Finally, the "+Pivot+W+COR" variation further reduces the randomness in the inpainted pixels, as highlighted by the orange box, which exhibits sharper and more refined details of the ear, compared to the green box region.

Figure 12: We present intermediate results from the iterative re-aging process. As the iterations progress, the accuracy of flow warping improves because the rendered images used to compute the flow increasingly resemble the target age. This leads to more precise reconstruction of the re-aged image, as highlighted by the green boxes. We further evaluate the necessity of iterative re-aging rather than a single-step transformation. As shown in Fig. 12, flow warping becomes more accurate with each iteration, as the rendered images used to compute flow progressively align with the target age. This leads to a more precise reconstruction of re-aged features, as highlighted in the green boxes. 6. Conclusion, Limitations, and Future Work We have presented a novel framework for realistic 3D face reaging that produces identity-preserving results with fine-grained age-related details. At the core of our approach is a diffusion-based 2D re-aging model, DiffReaging, trained on synthetically generated re-aged image pairs. To produce multi-view consistent re-aged images, we introduce a propagation strategy that re-ages a single pivot view and reconstructs the remaining views through warping and our proposed Masked-DiffReaging mechanism. Masked-DiffReaging

Input

20

40

60

80

Figure 13: Here, we highlight the limitations of DiffReaging on portrait images with occlusions. While age-related features gradually emerge during the process of re-aging, such as wrinkles forming on the lower face of the fair-skinned subject and around the eyes of the darker-skinned subject, visible artifacts also appear. leverages the iterative nature of DiffReaging to ensure coherence between the reconstructed regions and existing warped content. Moreover, we propose a center-out re-aging method to avoid redundant reconstruction of overlapping regions by progressively expanding from the pivot in concentric layers. The resulting re-aged images supervise the optimization of the re-aged 3D representation, enabling age-consistent and identity-preserving 3D re-aged results. Extensive experiments demonstrate that our method outperforms existing 3D editing approaches both visually and numerically. While DiffReaging achieves state-of-the-art results, it has limitations when re-aging faces with occlusions, as shown in Fig. 13, which may also affect downstream 3D face re-aging. Additionally, it currently does not account for aging in other body parts like hair or hands. Our method focuses on the facial region, which is often the most challenging and identity-critical component of re-aging in film production. We therefore restrict editing to masked facial regions while leaving non-facial regions unchanged, naturally preserving aging-irrelevant accessories such as earrings or necklaces. Severe facial occlusions, such as large sunglasses, remain challenging cases for future work. Moreover, our method does not explicitly model age-related biophysical albedo changes [IGAJG15], such as skin lightening, reduced redness, or changes in translucency. While the approach of AgeTrans3D [LGLG24] incorporates these effects, it trains a pixel-space diffusion model from scratch and consequently produces results with fewer age-related details (see Fig. 8). We believe incorporating a biophysical appearance model into our system is a promising direction and leave it for future investigation. Finally, our method targets high-fidelity re-aging of captured multi-view 3D faces, where multiple views provide reliable identity and geometry cues. Removing this requirement is an interesting direction for future work. For example, one could combine our re-aging propagation framework with single-image or sparse-view 3D-aware diffusion methods, such as Xiang et al. [XYHT23], to hallucinate missing viewpoints before optimization. 7. Acknowledgements We sincerely thank the anonymous reviewers for their valuable feedback and constructive suggestions. Additionally, portions of this research were conducted with the advanced computing resources provided by Texas A&M High Performance Research Computing. © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

References [ABD17a] A NTIPOV G., BACCOUCHE M., D UGELAY J.-L.: Boosting cross-age face verification via generative age normalization. In 2017 IEEE International Joint Conference on Biometrics (IJCB) (2017), IEEE, pp. 191–199. 2 [ABD17b] A NTIPOV G., BACCOUCHE M., D UGELAY J.-L.: Face aging with conditional generative adversarial networks. In 2017 IEEE international conference on image processing (ICIP) (2017), IEEE, pp. 2089– 2093. 2 [ALF22] AVRAHAMI O., L ISCHINSKI D., F RIED O.: Blended diffusion for text-driven editing of natural images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 (2022), IEEE, pp. 18187– 18197. URL: https://doi.org/10.1109/CVPR52688.2022. 01767, doi:10.1109/CVPR52688.2022.01767. 3 [APCO21] A LALUF Y., PATASHNIK O., C OHEN -O R D.: Only a matter of style: Age transformation using a style-based regression model. ACM Trans. Graph. 40, 4 (2021). URL: https://doi.org/10.1145/ 3450626.3459805. 2, 5 [ATA18] A NGULU R., TAPAMO J. R., A DEWUMI A. O.: Age estimation via face images: a survey. EURASIP Journal on Image and Video Processing 2018, 1 (2018), 1–35. 2 [AZMW21] A BDAL R., Z HU P., M ITRA N. J., W ONKA P.: Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (2021). 2 [BHE23] B ROOKS T., H OLYNSKI A., E FROS A. A.: Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 18392–18402. 2, 3, 4, 5, 7, 8 [BMT∗ 21] BARRON J. T., M ILDENHALL B., TANCIK M., H EDMAN P., M ARTIN -B RUALLA R., S RINIVASAN P. P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV (2021). 3 [BMV∗ 23] BARRON J. T., M ILDENHALL B., V ERBIN D., S RINIVASAN P. P., H EDMAN P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV (2023). 3 [BNS23] B HATTARAI A. R., N IESSNER M., S EVASTOPOLSKY A.: TriplaneNet: An encoder for eg3d inversion. arXiv preprint arXiv:2303.13497 (2023). 3 [BZY∗ 23] BAO C., Z HANG Y., YANG B., FAN T., YANG Z., BAO H., Z HANG G., C UI Z.: SINE: semantic-driven image-based nerf editing with prior-guided editing field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 (2023), IEEE, pp. 20919– 20929. URL: https://doi.org/10.1109/CVPR52729.2023. 02004, doi:10.1109/CVPR52729.2023.02004. 3 [CCZ∗ 24] C HEN Y., C HEN Z., Z HANG C., WANG F., YANG X., WANG Y., C AI Z., YANG L., L IU H., L IN G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 21476–21485. 2, 3, 7, 8 [CFHT22] C HEN Z., F UNKHOUSER T., H EDMAN P., TAGLIASACCHI A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022). 3 [CL23a] C HEN X., L ATHUILIÈRE S.: Face aging via diffusion-based editing. arXiv preprint arXiv:2309.11321 (2023). 2, 6, 7 [CL23b] C HEN X., L ATHUILIERE S.: Face aging via diffusion-based editing. In Proceedings of the 34th British Machine Vision Conference (BMVC) (2023). 3 [CLC∗ 22] C HAN E. R., L IN C. Z., C HAN M. A., NAGANO K., PAN B., D E M ELLO S., G ALLO O., G UIBAS L., T REMBLAY J., K HAMIS S., K ARRAS T., W ETZSTEIN G.: Efficient geometry-aware 3d generative © 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

11 of 14

adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022). 3 [CLTS24]

C HARATAN D., L I S. L., TAGLIASACCHI A., S ITZ V.: Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 19457– 19467. URL: https://doi.org/10.1109/CVPR52733.2024. 01840, doi:10.1109/CVPR52733.2024.01840. 3 MANN

[CLV24] C HEN M., L AINA I., V EDALDI A.: Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929 (2024). 2, 3, 7, 8 [CN23] C HANDALIYA P. K., NAIN N.: Aw-gan: Face aging and rejuvenation using attention with wavelet gan. Neural Computing and Applications 35, 3 (2023), 2811–2825. 2 [CWWL24] C HEN Z., WANG F., WANG Y., L IU H.: Text-to-3d using gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 21401–21412. URL: https://doi.org/10.1109/ CVPR52733.2024.02022, doi:10.1109/CVPR52733.2024. 02022. 3 [CXG∗ 22] C HEN A., X U Z., G EIGER A., Y U J., S U H.: Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517 (2022). 3 [CXZ∗ 24] C HEN Y., X U H., Z HENG C., Z HUANG B., P OLLEFEYS M., G EIGER A., C HAM T., C AI J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Computer Vision ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXI (2024), Leonardis A., Ricci E., Roth S., Russakovsky O., Sattler T., Varol G., (Eds.), vol. 15079 of Lecture Notes in Computer Science, Springer, pp. 370–386. URL: https://doi.org/10.1007/978-3-031-72664-4_21, doi:10.1007/978-3-031-72664-4\_21. 3 [DGXZ19] D ENG J., G UO J., X UE N., Z AFEIRIOU S.: Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019), pp. 4690–4699. 7, 9 [DHT∗ 00] D EBEVEC P., H AWKINS T., T CHOU C., D UIKER H.-P., S AROKIN W., S AGAR M.: Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000), pp. 145–156. 7 [DLQ∗ 19]

D UONG C. N., L UU K., Q UACH K. G., N GUYEN N., PATTERSON E., B UI T. D., L E N.: Automatic face aging in videos via deep reinforcement learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019), pp. 10013–10022. 2

[DW23] D ONG J., WANG Y.: Vica-nerf: View-consistency-aware 3d editing of neural radiance fields. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 (2023), Oh A., Naumann T., Globerson A., Saenko K., Hardt M., Levine S., (Eds.). URL: http: //papers.nips.cc/paper_files/paper/2023/hash/ c1e2faff6f588870935f114ebe04a3e5-Abstract-Conference. html. 3 [GLWT22] G U J., L IU L., WANG P., T HEOBALT C.: StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. In CVPR (2022). 3 [GPAM∗ 14] G OODFELLOW I., P OUGET-A BADIE J., M IRZA M., X U B., WARDE -FARLEY D., O ZAIR S., C OURVILLE A., B ENGIO Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014). 2 [GPP18] G EORGOPOULOS M., PANAGAKIS Y., PANTIC M.: Modeling of facial aging and kinship: A survey. Image and Vision Computing 80 (2018), 58–79. 2 [GQW∗ 25]

G ONG B., Q I L., W U J., F U Z., S ONG C., JACOBS D. W.,

12 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

N ICHOLSON J., S ENGUPTA R.: The aging multiverse: Generating condition-aware facial aging tree via training-free diffusion. In SIGGRAPH Asia 2025 Conference Papers (2025), ACM. 3

editing via disentangled video encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 6091– 6100. 3

[GTLMC22] G OMEZ -T RENADO G., L ATHUILIÈRE S., M ESEJO P., C ORDÓN Ó.: Custom structure preservation in face aging. In European Conference on Computer Vision (2022), Springer, pp. 565–580. 2

[KSSS14] K EMELMACHER -S HLIZERMAN I., S UWAJANAKORN S., S EITZ S. M.: Illumination-aware age progression. In Proceedings of the IEEE conference on computer vision and pattern recognition (2014), pp. 3334–3341. 2

[HHLP20] H ÄRKÖNEN E., H ERTZMANN A., L EHTINEN J., PARIS S.: Ganspace: Discovering interpretable gan controls. arXiv preprint arXiv:2004.02546 (2020). 2

[KZL∗ 23] K AWAR B., Z ADA S., L ANG O., T OV O., C HANG H., D EKEL T., M OSSERI I., I RANI M.: Imagic: Text-based real image editing with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 (2023), IEEE, pp. 6007–6017. URL: https://doi.org/10. 1109/CVPR52729.2023.00582. 3

[HHY∗ 22] H UANG Y., H E Y., Y UAN Y., L AI Y., G AO L.: Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 (2022), IEEE, pp. 18321–18331. URL: https://doi.org/10.1109/ CVPR52688.2022.01780, doi:10.1109/CVPR52688.2022. 01780. 3 [HJA20] H O J., JAIN A., A BBEEL P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851. 2, 3 [HSY∗ 24] H UANG Y., S UN Y., YANG Z., LYU X., C AO Y., Q I X.: SCGS: sparse-controlled gaussian splatting for editable dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 4220–4230. URL: https://doi.org/10.1109/CVPR52733. 2024.00404, doi:10.1109/CVPR52733.2024.00404. 3 [HTE∗ 23] H AQUE A., TANCIK M., E FROS A. A., H OLYNSKI A., K ANAZAWA A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2023), pp. 19740–19750. 2 [HXCL22] H SU G.-S., X IE R.-C., C HEN Z.-T., L IN Y.-H.: Agetransgan for facial age transformation with rectified performance metrics. In European Conference on Computer Vision (2022), Springer, pp. 580–595. 2 [HZG∗ 23] H ONG Y., Z HANG K., G U J., B I S., Z HOU Y., L IU D., L IU F., S UNKAVALLI K., B UI T., TAN H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023). 3, 9 [IGAJG15] I GLESIAS -G UITIAN J. A., A LIAGA C., JARABO A., G UTIER REZ D.: A biophysically-based model of the optical properties of skin aging. In Computer Graphics Forum (2015), vol. 34, Wiley Online Library, pp. 45–55. 3, 8, 10 [JTA21] JAIN A., TANCIK M., A BBEEL P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2021), pp. 5885–5894. 3 [KB15] K INGMA D., BA J.: Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) (2015). 6 [KKLD23] K ERBL B., KOPANAS G., L EIMKÜHLER T., D RETTAKIS G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42, 4 (2023), 139–1. 2, 3, 4, 9 [KLA19] K ARRAS T., L AINE S., A ILA T.: A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019), pp. 4401–4410. 5

[LCZ∗ 24] L I R., C HEN L., Z HANG Z., JAMPANI V., PATEL V. M., Z HANG L.: Syncnoise: Geometrically consistent noise prediction for text-based 3d scene editing. arXiv preprint arXiv:2406.17396 (2024). 2, 3 [LDR∗ 22] L UGMAYR A., DANELLJAN M., ROMERO A., Y U F., T IMO FTE R., VAN G OOL L.: Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022), pp. 11461–11471. 5 [LDY∗ 24] L UO C., D I D., YANG X., M A Y., X UE Z., W EI C., L IU Y.: Trame: Trajectory-anchored multi-view editing for text-guided 3d gaussian splatting manipulation. arXiv preprint arXiv:2407.02034 (2024). 2 [LGLG24] L I X., G UARNERA G. C., L IN A., G HOSH A.: Realistic facial age transformation with 3d uplifting. In Computer graphics forum (2024), vol. 43, Wiley Online Library, p. e15146. 3, 5, 8, 10 [LJA21] L I Z., J IANG R., A ARABI P.: Continuous face aging via selfestimated residual age embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021), pp. 15008– 15017. 2 [LPS∗ 24] L EE D. I., PARK H., S EO J., PARK E., PARK H., BAEK H. D., S ANGHEON S., K IM S., K IM S.: Editsplat: Multi-view fusion and attention-guided optimization for view-consistent 3d scene editing with 3d gaussian splatting. arXiv preprint arXiv:2412.11520 (2024). 3 [LRS∗ 24] L EE J. C., R HO D., S UN X., KO J. H., PARK E.: Compact 3d gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 21719– 21728. URL: https://doi.org/10.1109/CVPR52733.2024. 02052, doi:10.1109/CVPR52733.2024.02052. 3 [LSC∗ 24] L IU M., S HI R., C HEN L., Z HANG Z., X U C., W EI X., C HEN H., Z ENG C., G U J., S U H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2024), pp. 10072–10083. 3, 9 [LWH∗ 23] L I P., WANG R., H UANG H., H E R., H E Z.: Pluralistic aging diffusion autoencoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023), pp. 22613– 22623. 2 [LXJ∗ 23] L IU M., X U C., J IN H., C HEN L., VARMA T M., X U Z., S U H.: One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization. Advances in Neural Information Processing Systems 36 (2023), 22226–22246. 3, 9

[KLA∗ 20] K ARRAS T., L AINE S., A ITTALA M., H ELLSTEN J., L EHTI NEN J., A ILA T.: Analyzing and improving the image quality of StyleGAN. In Proc. CVPR (2020). 3, 7

[LZL∗ 25] L IU T., Z HANG D., L I G., L IU S., YONGQI SONG , L I S., YANG S., L I B., WANG K., WANG Y.: From cradle to cane: A two-pass framework for high-fidelity lifespan face aging. In Advances in Neural Information Processing Systems (NeurIPS) (2025). 3

[KQG∗ 23] K IRSCHSTEIN T., Q IAN S., G IEBENHAIN S., WALTER T., N IESSNER M.: Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph. 42, 4 (jul 2023). URL: https: //doi.org/10.1145/3592455, doi:10.1145/3592455. 9

[MESK22] M ÜLLER T., E VANS A., S CHIED C., K ELLER A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022). 3

[KSK∗ 23] K IM G., S HIM H., K IM H., C HOI Y., K IM J., YANG E.: Diffusion video autoencoders: Toward temporally consistent face video

[MST∗ 20] M ILDENHALL B., S RINIVASAN P. P., TANCIK M., BARRON J. T., R AMAMOORTHI R., N G R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV (2020). 2, 3, 4

© 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces [NSW24] N IEDERMAYR S., S TUMPFEGGER J., W ESTERMANN R.: Compressed 3d gaussian splatting for accelerated novel view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 10349– 10358. URL: https://doi.org/10.1109/CVPR52733.2024. 00985, doi:10.1109/CVPR52733.2024.00985. 3 [OT24] O STREK M., T HIES J.: Stable video portraits. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXXV (2024), Leonardis A., Ricci E., Roth S., Russakovsky O., Sattler T., Varol G., (Eds.), vol. 15143 of Lecture Notes in Computer Science, Springer, pp. 181–198. URL: https://doi.org/10.1007/978-3-031-73013-9_11, doi:10.1007/978-3-031-73013-9\_11. 3 [PCWS22] P REECHAKUL K., C HATTHEE N., W IZADWONGSA S., S UWAJANAKORN S.: Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 10619–10629. 3 [PJBM23] P OOLE B., JAIN A., BARRON J. T., M ILDENHALL B.: Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 (2023), OpenReview.net. URL: https://openreview. net/forum?id=FjNys5c7VyY. 3 [PSBR24] PALANDRA F., S ANCHIETTI A., BAIERI D., RODOLÀ E.: Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. CoRR abs/2403.05154 (2024). URL: https://doi.org/ 10.48550/arXiv.2403.05154, arXiv:2403.05154, doi: 10.48550/ARXIV.2403.05154. 3 [QWG∗ 25]

Q I L., W U J., G ONG B., WANG A. N., JACOBS D. W., S ENGUPTA R.: Mytimemachine: Personalized facial age transformation. ACM Transactions on Graphics (TOG) 44, 4 (2025). 3

[RBL∗ 22] ROMBACH R., B LATTMANN A., L ORENZ D., E SSER P., O M MER B.: High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022), pp. 10684–10695. 4 [RCB09] R AMANATHAN N., C HELLAPPA R., B ISWAS S.: Computational methods for modeling facial aging: A survey. Journal of Visual Languages & Computing 20, 3 (2009), 131–144. 2 [RKH∗ 21] R ADFORD A., K IM J. W., H ALLACY C., R AMESH A., G OH G., AGARWAL S., S ASTRY G., A SKELL A., M ISHKIN P., C LARK J., ET AL .: Learning transferable visual models from natural language supervision. In ICML (2021). 2, 3 [RMBCO22] ROICH D., M OKADY R., B ERMANO A. H., C OHEN -O R D.: Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG) 42, 1 (2022), 1–13. 2 [RTVG18] ROTHE R., T IMOFTE R., VAN G OOL L.: Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision 126, 2 (2018), 144–157. 7 [SGTZ20] S HEN Y., G U J., TANG X., Z HOU B.: Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020), pp. 9243– 9252. 2 [SLZ∗ 25] S U Z., L IU Z., Z HANG Z., Z HANG T., C HEN C. P.: Timebooth: Disentangled facial invariant representation for diverse and personalized face aging. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025), pp. 12147–12157. 3 [SWY∗ 23] S HI Y., WANG P., Y E J., M AI L., L I K., YANG X.: MVDREAM: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023). 3 [SXLT16] S HU X., X IE G.-S., L I Z., TANG J.: Age progression: Current technologies and applications. Neurocomputing 208 (2016), 249–261. 2 [TAN∗ 21] T OV O., A LALUF Y., N ITZAN Y., PATASHNIK O., C OHEN O R D.: Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG) 40, 4 (2021), 1–14. 2, 3

© 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.

13 of 14

[TMG∗ 22] T ZABAN R., M OKADY R., G AL R., B ERMANO A., C OHEN O R D.: Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers (2022), pp. 1–9. 2, 3 [TRZ∗ 24] TANG J., R EN J., Z HOU H., L IU Z., Z ENG G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 (2024), OpenReview.net. URL: https://openreview.net/forum?id=UyNXMqnN3c. 3 [VH24] VACHHA C., H AQUE A.: Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. URL: https://instruct-gs2gs. github.io/. 2, 3, 7, 8 [WBL∗ 24] W U J., B IAN J.-W., L I X., WANG G., R EID I., T ORR P., P RISACARIU V. A.: Gaussctrl: multi-view consistent text-driven 3d gaussian splatting editing. arXiv preprint arXiv:2403.08733 (2024). 2, 3 [WFZ∗ 24] WANG J., FANG J., Z HANG X., X IE L., T IAN Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 20902– 20911. URL: https://doi.org/10.1109/CVPR52733.2024. 01975, doi:10.1109/CVPR52733.2024.01975. 3 [WJC∗ 24] WANG C., J IANG R., C HAI M., H E M., C HEN D., L IAO J.: Nerf-art: Text-driven neural radiance fields stylization. IEEE Trans. Vis. Comput. Graph. 30, 8 (2024), 4983–4996. URL: https: //doi.org/10.1109/TVCG.2023.3283400, doi:10.1109/ TVCG.2023.3283400. 3 [WLD24] WANG Y., L IPSON L., D ENG J.: Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision (2024), Springer, pp. 36–54. 5 [WQP∗ 25] W YNN J., Q URESHI Z., P OWIERZA J., WATSON J., S AYED M.: Morpheus: Text-driven 3d gaussian splat shape and color stylization. arXiv preprint arXiv:2503.02009 (2025). 3 [WTLG18] WANG Z., TANG X., L UO W., G AO S.: Face aging with identity-preserved conditional generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018), pp. 7939–7947. 2 [WYW∗ 24] WANG Y., Y I X., W U Z., Z HAO N., C HEN L., Z HANG H.: View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision (2024), Springer, pp. 404–420. 2, 3 [WZRT24] WAHID J., Z HAN F., R AO P., T HEOBALT C.: Diffage3d: Diffusion-based 3d-aware face aging. arXiv preprint arXiv:2408.15922 (2024). 3 [XCG∗ 24] X U J., C HENG W., G AO Y., WANG X., G AO S., S HAN Y.: Instantmesh: Efficient 3d mesh generation from a single image with sparseview large reconstruction models. arXiv preprint arXiv:2404.07191 (2024). 3, 9 [XCH∗ 24] X IAO H., C HEN Y., H UANG H., X IONG H., YANG J., P RASAD P., Z HAO Y.: Localized gaussian splatting editing with contextual awareness. CoRR abs/2408.00083 (2024). URL: https://doi.org/10.48550/arXiv.2408.00083, arXiv: 2408.00083, doi:10.48550/ARXIV.2408.00083. 3 [XLSL22] X UE Y., L I Y., S INGH K. K., L EE Y. J.: GiraffeHD: A highresolution 3d-aware generative model. In CVPR (2022). 3 [XYHT23] X IANG J., YANG J., H UANG B., T ONG X.: 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2023), pp. 2383–2393. 10 [YCH∗ 24] Y U Z., C HEN A., H UANG B., S ATTLER T., G EIGER A.: Mip-splatting: Alias-free 3d gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 19447– 19456. URL: https://doi.org/10.1109/CVPR52733.2024. 01839, doi:10.1109/CVPR52733.2024.01839. 3 [YFW∗ 24] Y I T., FANG J., WANG J., W U G., X IE L., Z HANG X., L IU W., T IAN Q., WANG X.: Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In

14 of 14

Libing Zeng et al. / ReAge3D: Re-Aging 3D Faces

IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 6796– 6807. URL: https://doi.org/10.1109/CVPR52733.2024. 00649, doi:10.1109/CVPR52733.2024.00649. 3 [YGW∗ 21] Y U C., G AO C., WANG J., Y U G., S HEN C., S ANG N.: Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International journal of computer vision 129 (2021), 3051–3068. 5

diffusion transformers. CoRR abs/2501.03931 (2025). URL: https://doi.org/10.48550/arXiv.2501.03931, arXiv: 2501.03931, doi:10.48550/ARXIV.2501.03931. 3 [ZPIE17] Z HU J.-Y., PARK T., I SOLA P., E FROS A. A.: Unpaired imageto-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision (2017), pp. 2223–2232. 2

[YJLL23] YANG S., J IANG L., L IU Z., L OY C. C.: Styleganex: Styleganbased manipulation beyond cropped aligned faces. arXiv preprint arXiv:2303.06146 (2023). 3 [YJM∗ 24] Y U H., J ULIN J., M ILACSKI Z. Á., N IINUMA K., J ENI L. A.: Cogs: Controllable gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 21624– 21633. URL: https://doi.org/10.1109/CVPR52733.2024. 02043, doi:10.1109/CVPR52733.2024.02043. 3 [YLCL24] YAN Z., L OW W. F., C HEN Y., L EE G. H.: Multiscale 3d gaussian splatting for anti-aliased rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 20923– 20931. URL: https://doi.org/10.1109/CVPR52733.2024. 01977, doi:10.1109/CVPR52733.2024.01977. 3 [YNGH21] YAO X., N EWSON A., G OUSSEAU Y., H ELLIER P.: A latent transformer for disentangled face editing in images and videos. In Proceedings of the IEEE/CVF international conference on computer vision (2021), pp. 13789–13798. 2 [YPN∗ 21] YAO X., P UY G., N EWSON A., G OUSSEAU Y., H ELLIER P.: High resolution face age editing. In 2020 25th International conference on pattern recognition (ICPR) (2021), IEEE, pp. 8624–8631. 2 [YYTK21] Y U A., Y E V., TANCIK M., K ANAZAWA A.: pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021), pp. 4578– 4587. 3 [YZL∗ 23] Y UAN Z., Z HU Y., L I Y., L IU H., Y UAN C.: Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. arXiv preprint arXiv:2303.12326 (2023). 3 [ZCJ∗ 24] Z HOU S., C HANG H., J IANG S., FAN Z., Z HU Z., X U D., C HARI P., YOU S., WANG Z., K ADAMBI A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 (2024), IEEE, pp. 21676– 21685. URL: https://doi.org/10.1109/CVPR52733.2024. 02048, doi:10.1109/CVPR52733.2024.02048. 3 [ZCS∗ 22] Z OSS G., C HANDRAN P., S IFAKIS E., G ROSS M., G OTARDO P., B RADLEY D.: Production-ready face re-aging for visual effects. ACM Transactions on Graphics (TOG) 41, 6 (2022), 1–12. 2, 4, 5, 7 [ZHSea20] Z HU H., H UANG Z., S HAN H., ET AL: Look globally, age locally: Face aging with an attention mechanism. In ICASSP 20202020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2020), IEEE, pp. 1963–1967. 2 [ZKC∗ 24] Z HUANG J., K ANG D., C AO Y., L I G., L IN L., S HAN Y.: Tip-editor: An accurate 3d editor following both text-prompts and imageprompts. ACM Trans. Graph. 43, 4 (2024), 121:1–121:12. URL: https: //doi.org/10.1145/3658205, doi:10.1145/3658205. 3 [ZLW∗ 22] Z HANG J., L I X., WAN Z., WANG C., L IAO J.: Fdnerf: Fewshot dynamic neural radiance fields for face reconstruction and expression editing. In SIGGRAPH Asia 2022 Conference Papers, SA 2022, Daegu, Republic of Korea, December 6-9, 2022 (2022), Jung S. K., Lee J., Bargteil A. W., (Eds.), ACM, pp. 12:1–12:9. URL: https://doi.org/10. 1145/3550469.3555404, doi:10.1145/3550469.3555404. 3 [ZLX∗ 25] Z HANG Y., L IU Y., X IA B., P ENG B., YAN Z., L O E., J IA J.: Magic mirror: Id-preserved video generation in video

© 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics.