TimeWeaver: Age-Consistent Reference-Based Face Restoration with Identity Preservation

Teer Song; Yue Zhang; Yu Tian; Ziyang Wang; Xianlin Zhang; Guixuan Zhang; Xuan Liu; Xueming Li

cs.CVMar 24, 2026

TimeWeaver: Age-Consistent Reference-Based Face Restoration with Identity Preservation

Authors: Teer Song, Yue Zhang, Yu Tian, Ziyang Wang, Xianlin Zhang, Guixuan Zhang, Xuan Liu, Xueming Li,

Organizations: Beijing University of Posts and Telecommunications · Tsinghua University · Minzu University of China · Xiaomi Corporation

Abstract

Recent progress in face restoration has shifted from visual fidelity to identity fidelity, driving a transition from reference-free to reference-based paradigms that condition restoration on reference images of the same person. However, these methods assume the reference and degraded input are age-aligned. When only cross-age references are available, as in historical restoration or missing-person retrieval, they fail to maintain age fidelity. To address this limitation, we propose TimeWeaver, the first reference-based face restoration framework supporting cross-age references. Given arbitrary reference images and a target-age prompt, TimeWeaver produces restorations with both identity fidelity and age consistency. Specifically, we decouple identity and age conditioning across training and inference. During training, the model learns an age-robust identity representation by fusing a global identity embedding with age-suppressed facial tokens via a transformer-based ID-Fusion module. During inference, two training-free techniques, Age-Aware Gradient Guidance and Token-Targeted Attention Boost, steer sampling toward desired age semantics, enabling precise adherence to the target-age prompt. Extensive experiments show that TimeWeaver surpasses existing methods in visual quality, identity preservation, and age consistency.

Figures & tables

Figure 2: Framework of TimeWeaver. (a) During training, we extract a global identity embedding using ArcFace [ 7 ] and age-suppressed facial tokens using CLIP-ViT [ 35 ] . The ID-Fusion fuses them and injects the output into UNet as the identity condition. (b) During inference, the framework computes an age-aware gradient with and without the age condition to refines the latent along the denoising process, assisted by TTAB technique to achieve age-controllable restoration.

Figure 3: Visual comparison of different inference strategies with DEX [ 37 ] age estimates.

Figure 4: TTAB mechanism, using age-token attention maps as a spatial response map to boost on age-relevant regions.

Figure 5: Qualitative comparison with reference-based methods on synthetic dataset .

Figure 6: Qualitative comparison with reference-based methods on real-world dataset .

		Same-age						Cross-age (synthetic)
Method	Ref	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	AGE $\downarrow$
CodeFormer [ 60 ]		25.61	0.720	0.207	4.56	47.46	0.639	26.44	0.756	0.207	4.60	57.46	0.639	11.39
DifFace [ 56 ]		25.31	0.720	0.247	4.20	55.01	0.509	24.95	0.694	0.268	4.21	75.12	0.432	10.66
DiffBIR [ 24 ]		26.03	0.705	0.220	6.33	56.55	0.698	25.79	0.679	0.220	6.22	52.99	0.512	12.25
DMDNet [ 23 ]	✓	25.76	0.730	0.228	4.45	55.48	0.703	25.85	0.717	0.224	4.40	56.74	0.645	12.75
RestorerID [ 53 ]	✓	25.00	0.699	0.272	4.92	59.74	0.686	25.12	0.687	0.231	4.66	59.76	0.644	15.00

Table 1: Quantitative comparison on same-age and cross-age(synthetic) datasets. The best results are shown in bold , and the second-best are underlined .

Method	Ref	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	AGE $\downarrow$
CodeFormer [ 60 ]		4.19	56.53	0.263	9.67
DifFace [ 56 ]		3.87	119.78	0.295	11.83
DiffBIR [ 24 ]		6.07	67.97	0.290	10.50
DMDNet [ 23 ]	✓	4.35	105.54	0.295	9.67
RestorerID [ 53 ]	✓	4.40	76.31	0.467	15.14
Ref-LDM [ 16 ]	✓	4.31	59.93	0.352	13.33

Table 2: Quantitative comparison on the real-world dataset.

Figure 7: User study results.

Figure 8: Qualitative comparison with Two-stage Restoration–Editing Pipelines

Method	Type	PSNR $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	IDS $\uparrow$	AGE $\downarrow$	Time(s) $\downarrow$
HRFAE [ 51 ]	GAN	21.57	0.308	5.31	0.602	10.23	0.18
SAM [ 1 ]	GAN	18.09	0.457	12.33	0.347	8.77	0.45
FADING [ 5 ]	Diffusion	22.28	0.270	6.52	0.584	8.05	133.41
Null Inv. + P2P [ 29 ]	Diffusion	20.91	0.225	8.90	0.401	14.54	57.26
DDS [ 12 ]	Diffusion	23.60	0.207	5.08	0.623	15.67	12.70
FLUX.1 Kontext [ 20 ]	Diffusion	24.37	0.208	4.67	0.643	13.40	40.12

Table 3: Comparison with editing methods.

Age Gap	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	AGE $\downarrow$
$\leq 10$	24.58	0.698	0.203	3.81	53.08	0.683	5.66
10 $\sim$ 20	25.27	0.708	0.203	3.92	53.36	0.687	5.41
20 $\sim$ 30	25.12	0.716	0.212	3.80	54.43	0.682	5.89
30 $\sim$ 40	24.50	0.692	0.201	3.86	55.67	0.677	6.33
$>40$	24.23	0.700	0.217	4.00	55.98	0.670	6.70
Mixed	24.37	0.712	0.208	3.86	54.60	0.682	6.55

Table 4: Performance across different age gaps.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	AGE $\downarrow$
w/o $f_{global}$	24.47	0.645	0.247	5.60	60.00	0.568	11.30
w/o $f_{facial}$	25.01	0.689	0.223	4.45	58.98	0.687	8.79
w/ $f_{global}$ & $f_{facial}$ (no mask)	25.63	0.713	0.218	3.40	54.76	0.711	10.43
w/ $f_{global}$ & $f_{facial}$ (mask)	25.11	0.718	0.201	3.52	51.28	0.701	8.25

Table 5: Ablation study on different identity representations.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$	FID $\downarrow$	IDS $\uparrow$	AGE $\downarrow$
w/o AAGG	25.08	0.711	0.208	4.56	52.97	0.677	14.43
w/o TTAB	23.34	0.670	0.212	3.80	56.67	0.698	9.06
w/ AAGG & TTAB	25.11	0.718	0.201	3.52	51.28	0.701	8.25

Table 6: Ablation study on AAGG and TTAB mechanisms.

Panel of Figure 9 — Figure 9: Visual comparisons of ablation studies.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

MeInTime: Bridging Age Gap in Identity-Preserving Face Restoration
Mar 19, 2026Teer Song, Yue Zhang, Yu Tian +6Image Restoration Diffusion Models for Image Restoration
Beijing University of Posts and Telecommunications · Department of Computer Science and Technology, Institute for AI, Tsinghua University · Minzu University of China +1
Face Time Traveller : Travel Through Ages Without Losing Identity
Feb 26, 2026Purbayan Kar, Ayush Ghadiya, Vishal Chudasama +2Identity-Preserving Image Generation Diffusion Models
Sony Research India · IIIT Hyderabad
IConFace: Identity-Structure Asymmetric Conditioning for Unified Reference-Aware Face Restoration
May 4, 2026Axi Niu, Jinyang Zhang, Senyan QingImage Restoration
Northwestern Polytechnical University

Mar 19, 2026J/K move · Enter open · S save

MeInTime: Bridging Age Gap in Identity-Preserving Face Restoration

Teer Song, Yue Zhang, Yu Tian, Ziyang Wang+5

Beijing University of Posts and Telecommunications · Department of Computer Science and Technology, Institute for AI, Tsinghua University · Minzu University of China+1

Image Restoration Diffusion Models for Image Restoration

Open paper AskPDF arXiv

To better preserve an individual's identity, face restoration has evolved from reference-free to reference-based approaches, which leverage high-quality reference images of the same identity to enhance identity fidelity in the restored outputs. However, most existing methods implicitly assume that the reference and degraded input are age-aligned, limiting their effectiveness in real-world scenarios where only cross-age references are available, such as historical photo restoration. This paper proposes MeInTime, a diffusion-based face restoration method that extends reference-based restoration from same-age to cross-age settings. Given one or few reference images along with an age prompt corresponding to the degraded input, MeInTime achieves faithful restoration with both identity fidelity and age consistency. Specifically, we decouple the modeling of identity and age conditions. During training, we focus solely on effectively injecting identity features through a newly introduced attention mechanism and introduce Gated Residual Fusion modules to facilitate the integration between degraded features and identity representations. At inference, we propose Age-Aware Gradient Guidance, a training-free sampling strategy, using an age-driven direction to iteratively nudge the identity-aware denoising latent toward the desired age semantic manifold. Extensive experiments demonstrate that MeInTime outperforms existing face restoration methods in both identity preservation and age consistency. Our code is available at: https://github.com/teer4/MeInTime

Abstract

Image Restoration

Explore similar work

MeInTime: Bridging Age Gap in Identity-Preserving Face Restoration

Face Time Traveller : Travel Through Ages Without Losing Identity

IConFace: Identity-Structure Asymmetric Conditioning for Unified Reference-Aware Face Restoration