AlbumVSR: Identity-Preserving Face Video Super-Resolution from a Personal Photo Album

Nianyao Du; Xijun Wang; Stanley H. Chan

cs.CVOct 28, 2026

AlbumVSR: Identity-Preserving Face Video Super-Resolution from a Personal Photo Album

Authors: Nianyao Du, Xijun Wang, Stanley H. Chan

Organizations: Purdue University, West Lafayette, Indiana, USA

Abstract

Severe blur, downsampling, noise, and compression can erase the identity cues needed to faithfully restore a person's face. We therefore use photographs of the same person as an external source of identity. AlbumVSR conditions a frozen, one-step streaming diffusion transformer on a personal album. We use four MagFace/CLIP set tokens to encode the album, and 64 texture tokens which come from a fused UV atlas rendered at the current window's estimated face pose. These tokens are appended to the transformer's cached cross-attention context. On an identity-disjoint VoxCeleb2 benchmark with heavy synthetic degradation, AlbumVSR raises ArcFace similarity to the ground truth from 0.586 to 0.670 and face PSNR from 24.98 to 26.71 dB. On the 91 test identities, its mean identity similarity is higher than that of all 14 baselines.

Figures & tables

Figure 1: Overview of AlbumVSR. The album is encoded once for each identity. Four set tokens carry identity information, while fused texture, visibility, and displacement maps are rendered and tokenized for each streaming window. Every frozen DiT block appends the resulting album K/V to its cached prompt K/V. Training updates only the album heads and LoRA parameters.

Method	ID-GT $\uparrow$	ID-alb $\uparrow$	PSNR-F $\uparrow$	SSIM-F $\uparrow$	LPIPS-F $\downarrow$	Warp $\downarrow$	VIDD	MUSIQ $\uparrow$	s/fr $\downarrow$
Bicubic $4\times$	0.611	0.474	25.24	0.758	0.379	0.0059	0.056	19.80	-
FlashVSR-tiny (frozen) [ 29 ]	0.586	0.459	24.98	0.704	0.175	0.0093	0.049	72.34	0.07 W
FlashVSR-full (frozen) [ 29 ]	0.585	0.459	24.76	0.690	0.182	0.0100	0.046	73.68	0.13 A
RealViFormer [ 27 ]	0.504	0.388	24.93	0.716	0.249	0.0077	0.058	58.97	0.02 W
SeedVR2-3B [ 18 ]	0.642	0.501	26.08	0.764	0.179	0.0078	0.044	48.00	0.49 A
DicFace † [ 1 ]	0.555	0.427	25.55	0.759	0.211	0.0068	0.073	42.66	0.30 W

Table 1: Heavy-degradation results on 91 test identities.

\dagger

: crop-based method under the shared paste-back protocol. SSIM-F is face-region SSIM. Runtime devices are A100 80GB PCIe (

A

) and RTX 4070 Ti (

W

). VIDD is reported descriptively rather than ranked (Sec. 4 ). Best and second-best values in the remaining metrics are bold and underlined.

Configuration	ID-GT $\uparrow$	PSNR-F $\uparrow$	LPIPS-F $\downarrow$	MUSIQ $\uparrow$
3-ch., spatially gated (10k)	0.667	26.51	0.163	50.72
$\quad-$ Ch.1 tokens (10k)	0.630	26.58	0.157	51.53
3-ch., global (10k)	0.673	26.70	0.161	51.05
Ch.1 only, global (10k)	0.673	26.71	0.166	50.34
Ch.1+Ch.2, global (10k)	0.669	26.71	0.161	51.62
Ch.1 only, global (20k)	0.669	26.73	0.159	51.63

Table 2: Post hoc heavy-track ablations on 91 test identities. 3-ch. adds geometry Ch.3. “Global” denotes global LoRA.

\Delta

is Ch.1+Ch.2 minus Ch.1-only at 20k, paired per identity using unrounded values.

Panel of Figure 2 — Figure 2: Top: Four heavy-degradation examples: id00068/047, id00109/013, id00456/050, and id00560/010. We show GT, degraded input, FlashVSR, SeedVR2, SVFR, InstantRestore, FaceMe, RGFVR, and AlbumVSR. Labels report per-frame ArcFace ID-GT/ID-alb. The GT ID-alb values show the similarity between the target frames and the album. Table 1 reports all 14 baselines on 91 test identities. Bottom: Qualitative ablations. Left: null, wrong, and correct albums using a separate 14k three-channel model. Right: Ch.1-only versus Ch.1+Ch.2 at 20k. Values are per-frame scores. Table 2 reports aggregate results.

Video Super-Resolution

Video super-resolution reconstructs sharper, higher-resolution video from low-resolution frames while preserving detail and temporal consistency.

See all 39 papers

RGFVR: Reference-Guided Face Video Restoration with Flow Matching
Jun 15, 2026Cem Eteke, Batuhan Tosun, Eckehard SteinbachVideo Diffusion Models Video Restoration
Chair of Media Technology Munich Institute of Robotics and Machine Intelligence School of Computation, Information, and Technology Technical University of Munich 80333 Munich, Germany
DVFace: Spatio-Temporal Dual-Prior Diffusion for Video Face Restoration
Apr 16, 2026Zheng Chen, Bowen Chai, Rongjun Gao +7Video Diffusion Models Image Restoration
Shanghai Jiaotong University · Meituan Inc
SwinIFS: Landmark Guided Swin Transformer For Identity Preserving Face Super Resolution
Jan 4, 2026Habiba Kausar, Saeed Anwar, Omar Jamal Hammad +1Image Super-Resolution
King Fahd University of Petroleum and Minerals, KSA · The University of Western Australia, Australia · The University of Regina, Canada

Jun 15, 2026J/K move · Enter open · S save

RGFVR: Reference-Guided Face Video Restoration with Flow Matching

Cem Eteke, Batuhan Tosun, Eckehard Steinbach

Chair of Media Technology Munich Institute of Robotics and Machine Intelligence School of Computation, Information, and Technology Technical University of Munich 80333 Munich, Germany

Video Diffusion Models Video Restoration Temporal Consistency in Video Generation

Open paper AskPDF arXiv

Face video restoration from degraded observations is challenging, as it requires simultaneously recovering visual fidelity, temporal consistency, and subject identity. Existing approaches are often either reference-free, which can lead to identity loss when person-specific facial details are lost, or subject-specific, which limits generalization to unseen identities. We propose a subject-agnostic, reference-guided framework for identity-preserving face video restoration. Our method introduces bimodal perceptual-descriptive identity conditioning into a pretrained flow-based text-to-video generator and employs a two-stage training strategy to strengthen identity guidance during restoration. Experiments show that our approach improves restoration fidelity, temporal consistency, and identity preservation, achieving superior performance under challenging video degradations, including downsampling, blur, noise, and compression artifacts. The code is available under: https://github.com/batuhanntosun/RG-FVR.

Abstract

Video Super-Resolution

Explore similar work

RGFVR: Reference-Guided Face Video Restoration with Flow Matching

DVFace: Spatio-Temporal Dual-Prior Diffusion for Video Face Restoration

SwinIFS: Landmark Guided Swin Transformer For Identity Preserving Face Super Resolution