RetouchingFFHQ++: Forensic Face Restoration

• RetouchingFFHQ++ is a paired dataset of original and retouched face images designed for forensic restoration, prioritizing low-frequency fidelity over perceptual plausibility.
• It aggregates images from FFHQ via four commercial retouching systems, supporting both single- and multi-operation edits with PSNR-based severity labels.
• The dataset underpins the MoFRR benchmark, employing pixel-level metrics and biometric similarity to ensure authentic face reconstruction in forensic settings.

RetouchingFFHQ++ is a million-scale, paired original–retouched face dataset introduced in the context of Face Retouching Restoration (FRR), a task defined as reconstructing original face images from retouched inputs without any external reference image. It is built on FFHQ through RetouchingFFHQ and extends prior FFHQ-based retouching resources by combining paired supervision, four commercial retouching sources, single- and multi-operation edits, and degree labels redefined from observed distortion rather than API control values. In the formulation accompanying the dataset, FRR is treated as a forensic problem that prioritizes fidelity to the original face and biometric consistency over perceptual plausibility, explicitly discouraging hallucinated facial content (Liu et al., 26 Jul 2025).

1. Origins and task definition

RetouchingFFHQ++ was introduced by the paper "MoFRR: Mixture of Diffusion Models for Face Retouching Restoration" (Liu et al., 26 Jul 2025). Its stated purpose is to support FRR, for which the paper gives the definition: “The goal of FRR is defined to blindly, i.e., with no other reference of template, reconstruct the original face images given the retouched ones via computer vision algorithms or deep learning methods.” The same work emphasizes that evaluation “prioritizes forensic admissibility over perceptual quality, explicitly discouraging methods that synthesize plausible but inauthentic facial features” (Liu et al., 26 Jul 2025).

The dataset is positioned against two adjacent but distinct problem settings. First, it differs from traditional image restoration because FRR “focuses more on the restoration of the low-frequency information of the faces,” whereas conventional restoration typically emphasizes degraded textures and other high-frequency recovery. Second, it differs from makeup removal because retouching often changes facial structure and global appearance through operations such as face lifting and eye enlarging, rather than only modifying surface appearance (Liu et al., 26 Jul 2025).

RetouchingFFHQ++ inherits its lineage from "RetouchingFFHQ: A Large-scale Dataset for Fine-grained Face Retouching Detection" (Ying et al., 2023). RetouchingFFHQ framed retouching as a fine-grained detection problem with four operations and four levels, whereas RetouchingFFHQ++ is constructed for paired restoration of original faces from retouched counterparts. A plausible implication is that RetouchingFFHQ++ shifts the center of gravity from retouching recognition to retouching inversion, while keeping the FFHQ-based formulation and operation taxonomy as a foundation.

2. Dataset construction and composition

RetouchingFFHQ++ uses RetouchingFFHQ as its base source, which itself is derived from FFHQ (Liu et al., 26 Jul 2025, Ying et al., 2023). The retouched images are generated through four commercial retouching systems: Tencent, Megvii, Alibaba, and the newly added PortraitPro 24. The inclusion of PortraitPro 24 is specifically motivated by diversity; the paper notes that PortraitPro retouches have “noticeable visual differences” from the other APIs (Liu et al., 26 Jul 2025).

The dataset provides paired original–retouched images, and the presence of an “Ori” subset together with the use of PSNR and SSIM against the original face during training and evaluation establishes paired supervision (Liu et al., 26 Jul 2025). Training and evaluation in the reported experiments use images at $1024 \times 1024$ resolution, with both input and output images at that resolution (Liu et al., 26 Jul 2025).

The reported global composition is as follows (Liu et al., 26 Jul 2025):

Category	Total	Breakdown
Originals (“Ori”)	57,910	Paired with retouched counterparts
Retouched, single-operation	404,542	PortraitPro 108,267; Megvii 66,692; Tencent 200,100; Alibaba 29,483
Retouched, multi-operation	612,886	PortraitPro 206,655; Megvii 232,641; Tencent 173,590; Alibaba not present
Retouched, all	1,075,338	PortraitPro 314,922; Megvii 299,333; Tencent 373,690; Alibaba 29,483

All experiments divide the data into train, validation, and test sets using an $8{:}1{:}1$ split (Liu et al., 26 Jul 2025). For cross-API evaluation, the reported protocol trains on a mixture excluding PortraitPro and tests on the PortraitPro subset, making domain shift a first-class part of the benchmark rather than an auxiliary experiment (Liu et al., 26 Jul 2025).

By contrast, RetouchingFFHQ contained 710,726 images at $512 \times 512$, including 58,158 non-retouched and 652,568 retouched samples, and was organized around three commercial APIs with fine-grained retouching labels for detection (Ying et al., 2023). RetouchingFFHQ++ therefore extends both scale and restoration utility, while also changing the working resolution back to $1024 \times 1024$ in the reported FRR pipeline.

3. Retouching operations, degree labels, and annotation logic

RetouchingFFHQ++ covers four primary retouching types: whitening, smoothing, face lifting, and eye enlarging, and also includes “Multi” cases in which two or more operations are combined (Liu et al., 26 Jul 2025). The operation space is therefore aligned with the four typical retouching types already formalized in RetouchingFFHQ—skin smoothing, face whitening, eye enlarging, and face lifting—even though the task focus differs between the two resources (Ying et al., 2023).

A defining feature of RetouchingFFHQ++ is its redefinition of retouching degree. Rather than directly using API control settings, the authors “redefine the degree of modification of each operation within images via statistically analyzing the PSNR distribution of all retouched face images” (Liu et al., 26 Jul 2025). The resulting labels are re-categorized into five groups with proportions “15%, 25%, 25%, 25%, and 10%” based on ascending PSNR, and are labeled from 5 to 1. Lower PSNR therefore corresponds to higher distortion and a higher degree label (Liu et al., 26 Jul 2025).

The metadata used by the benchmark includes both operation presence and degree information. Operation presence is multi-label, represented in the MoFRR framework by a 4D binary vector $[b_w, b_s, b_f, b_e]$ indicating whitening, smoothing, face lifting, and eye enlarging. Per-operation degree is represented by $z \in \{1,\dots,5\}$ after PSNR-based re-binning (Liu et al., 26 Jul 2025). Exact per-type counts per degree are not reported.

This labeling strategy differs materially from that of RetouchingFFHQ. In RetouchingFFHQ, the four operations were assigned four quantized levels—off $(0)$, slight $(30)$, medium $(60)$, and heavy $(90)$—and labels were stored in the format $8{:}1{:}1$0 with each variable taking values in $8{:}1{:}1$1 (Ying et al., 2023). RetouchingFFHQ++ instead maps observed distortion severity, not API knob positions, to a five-level label space. This suggests an effort to normalize provider-specific differences in how nominal retouching strengths translate into actual image changes.

4. Evaluation protocols and benchmark semantics

RetouchingFFHQ++ is designed to support both intra-API and cross-API evaluation (Liu et al., 26 Jul 2025). The intra-API setting trains and tests on a mixture from all APIs under a standard closed-domain protocol. The cross-API setting trains on a mixture excluding PortraitPro and tests on PortraitPro, thereby isolating generalization to a held-out retouching source (Liu et al., 26 Jul 2025).

The benchmark emphasizes two evaluation dimensions. The first is pixel-level fidelity, measured by PSNR and SSIM. For a reference image $8{:}1{:}1$2 and estimate $8{:}1{:}1$3 with mean squared error $8{:}1{:}1$4, the paper gives

$8{:}1{:}1$5

SSIM is used in both reporting and loss definitions (Liu et al., 26 Jul 2025).

The second dimension is biometric veracity, measured through feature-space cosine similarity computed using face recognition models such as AdaFace and ArcFace. With embeddings $8{:}1{:}1$6 and $8{:}1{:}1$7, similarity is

$8{:}1{:}1$8

Higher similarity indicates stronger identity preservation (Liu et al., 26 Jul 2025). The paper explicitly treats this identity-oriented evaluation as central to FRR because plausible but incorrect facial synthesis is undesirable in forensic settings (Liu et al., 26 Jul 2025).

The choice of metrics also marks a departure from neighboring retouching literatures. Detection-oriented RetouchingFFHQ uses True Positive, True Negative, and Absolute Correctness for multi-label, multi-level prediction of retouching operations and strengths (Ying et al., 2023). Reversion-oriented work such as "Label-guided Facial Retouching Reversion" evaluates on RetouchingFFHQ using FID, SSIM, PSNR, LPIPS, and DISTS, with detector guidance and H-AdaIN color correction (Zhao et al., 2024). By contrast, RetouchingFFHQ++ foregrounds PSNR, SSIM, and identity similarity, and the MoFRR paper explicitly notes that LPIPS and FID are not reported because perceptual hallucination is not the objective (Liu et al., 26 Jul 2025).

5. Role in model development: the MoFRR benchmark setting

RetouchingFFHQ++ is not only a dataset but also the training and evaluation substrate for the MoFRR system proposed in the same paper (Liu et al., 26 Jul 2025). In that framework, a router predicts which retouching operations are present, degree estimators predict operation severity, specialized experts restore specific retouching types, a shared expert addresses common traces, and a combine module fuses expert outputs (Liu et al., 26 Jul 2025).

The router is trained on RetouchingFFHQ++ as a multi-label classifier over the four operations:

$8{:}1{:}1$9

with $512 \times 512$0 (Liu et al., 26 Jul 2025). Degree estimators are per-operation ResNet-50 classifiers trained using the redefined degree labels $512 \times 512$1 (Liu et al., 26 Jul 2025). Specialized experts are trained per operation on single-operated images only, while the shared expert is trained on a broad subset to learn common retouching traces (Liu et al., 26 Jul 2025).

The low-frequency branch of each specialized expert uses an Iterative Distortion Evaluation Module (IDEM), reflecting the dataset’s emphasis on structural and low-frequency restoration. With $512 \times 512$2 the retouched image, $512 \times 512$3 its low-frequency wavelet sub-band, $512 \times 512$4 the predicted degree for operation $512 \times 512$5, and $512 \times 512$6 the current low-frequency denoising estimate, the paper defines

$512 \times 512$7

$512 \times 512$8

$512 \times 512$9

Here $1024 \times 1024$0 is a pixel-wise distortion map for the low-frequency band (Liu et al., 26 Jul 2025).

The conditional diffusion backbone for the low-frequency sub-band is written as

$1024 \times 1024$1

and

$1024 \times 1024$2

with DDIM used for accelerated sampling (Liu et al., 26 Jul 2025). High-frequency refinement is handled by HFCAM:

$1024 \times 1024$3

The reported training setup on RetouchingFFHQ++ uses Adam, learning rate $1024 \times 1024$4, batch size $1024 \times 1024$5, approximately $1024 \times 1024$6 iterations for WaveFRR on $1024 \times 1024$7, and 50 epochs for the overall pipeline (Liu et al., 26 Jul 2025). Because these values are reported in direct association with the benchmark, RetouchingFFHQ++ functions not merely as a static corpus but as a standardized regime for FRR experimentation.

6. Reported results, relation to neighboring datasets, and limitations

The quantitative results reported on RetouchingFFHQ++ establish benchmark baselines for FRR. In the intra-API setting, MoFRR achieves PSNR/SSIM of 33.11/0.949 on whitening, 38.06/0.943 on smoothing, 31.26/0.913 on face lifting, 38.05/0.958 on eye enlarging, and 34.47/0.959 on multi-operation samples (Liu et al., 26 Jul 2025). On the multi-operation subset, the paper highlights a $1024 \times 1024$8 dB PSNR gain over the second-best baseline, ResDiff, improving from 28.98 dB to 34.47 dB (Liu et al., 26 Jul 2025). In the cross-API protocol with PortraitPro held out for testing, MoFRR attains 36.65/0.971 on the single-operation PortraitPro subset and 31.28/0.938 on the multi-operation PortraitPro subset (Liu et al., 26 Jul 2025).

The paper also reports that the density of cosine similarities between restored and original faces peaks near $1024 \times 1024$9 for MoFRR under both AdaFace and ArcFace, while retouched inputs exhibit a long left tail corresponding to stronger identity degradation (Liu et al., 26 Jul 2025). This is consistent with the dataset’s stated forensic orientation. A plausible implication is that RetouchingFFHQ++ is intended to evaluate restoration faithfulness under identity-sensitive criteria rather than only visual plausibility.

In the broader FFHQ retouching ecosystem, RetouchingFFHQ++ occupies a distinct place. RetouchingFFHQ (Ying et al., 2023) is a large-scale detection dataset focused on four retouching operations and their levels, with a Multi-granularity Attention Module for fine-grained recognition. Re-Face (Zhao et al., 2024) uses RetouchingFFHQ to learn label-guided retouching reversion with a detector, a ControlNet-based FaceR model, and H-AdaIN for color correction, but the paper explicitly states that it does not introduce or use RetouchingFFHQ++. StyleRetoucher (Su et al., 2023) addresses portrait retouching with StyleGAN priors on FFHQ-aligned imagery, targeting blemish removal and identity-preserving enhancement rather than restoration of authentic originals. "Photorealistic Facial Wrinkles Removal" (Sanchez et al., 2022) focuses on wrinkle segmentation and inpainting, again operating in an editing regime rather than an authenticity-restoration regime. These neighboring works clarify that RetouchingFFHQ++ belongs specifically to restoration and forensic analysis, not generic beautification or retouching synthesis.

Several limitations are explicitly or implicitly present in the available description of RetouchingFFHQ++. Demographic coverage, bias analysis, file types, release links, and licensing terms are not reported in the provided text (Liu et al., 26 Jul 2025). Image alignment, cropping, and normalization for the dataset release are likewise not reported, although the paper states that training and evaluation use $[b_w, b_s, b_f, b_e]$0 images. The degree labels are redefined via PSNR rank percentiles to mitigate “API-level algorithmic discrepancy,” but they remain a proxy for perceived severity rather than a direct psychophysical measurement (Liu et al., 26 Jul 2025). The benchmark therefore provides unusually strong supervision for FRR, but it does not, in the reported material, fully specify dataset governance or demographic auditing.

Taken together, RetouchingFFHQ++ represents an FFHQ-based benchmark for retouching restoration rather than retouching detection alone. Its defining characteristics are paired original–retouched supervision, multi-API diversity including PortraitPro 24, explicit support for both single and multi-operation edits, severity labels derived from observed distortion, and evaluation protocols centered on pixel fidelity and identity preservation. Within the current literature, it supplies the experimental basis for FRR as a distinct task and extends the FFHQ retouching lineage from detecting beautification traces toward recovering authentic facial appearance (Liu et al., 26 Jul 2025, Ying et al., 2023).