MirrorPPR47M: Exemplar-Based Portrait Retouching Dataset

• MirrorPPR47M is a large-scale dataset that organizes over 47 million same-identity before/after portrait pairs to support exemplar-based structural photo retouching.
• It employs a self-augmentation paradigm to create operation-aligned training quadruplets, effectively addressing issues like operation misalignment and shortcut learning.
• The dataset underpins a two-stage training regime combining a retouching operation extractor with a diffusion-based model to transfer subtle geometric edits reliably.

Searching arXiv for the specified paper and closely related work to ground the article in current literature. MirrorPPR47M is a large-scale portrait retouching dataset built for exemplar-based structural portrait photo retouching, introduced together with the MirrorPPR framework in "MirrorPPR: Exemplar-Based Portrait Photo Retouching" (Liu et al., 28 Jun 2026). It contains over 47 million retouched image pairs organized around before/after structural edits rather than color or tonal adjustment. Each basic datum is a retouched pair $(X_s, X_t)$, where $X_s$ is an original portrait and $X_t$ is the same portrait after geometric reshaping of facial features or body proportions. The dataset was created to support a task in which an exemplar pair specifies “how a person was retouched,” and a model must infer and apply the same combination and strength of geometric edits to a new query portrait. Its design addresses three central difficulties: the lack of suitable data for subtle high-resolution structural edits, operation misalignment in cross-identity supervision, and shortcut learning under naïve same-identity pairing (Liu et al., 28 Jun 2026).

1. Task definition and dataset rationale

MirrorPPR47M is designed for Exemplar-Based Portrait Photo Retouching. In this setting, the model is given an exemplar pair $(X_s, X_t)$ and a new query portrait $X_q$, and it must generate $\hat{Y}_q$ so that the same retouching operations visible in the exemplar are transferred to the query. The emphasis is on structural editing: geometric reshaping of eyes, nose, mouth, face shape, shoulders, waist, arms, or legs, rather than appearance edits such as tonal enhancement, makeup transfer, or skin smoothing (Liu et al., 28 Jun 2026).

The dataset addresses a supervision problem that is especially acute for structural portrait editing. For different identities, it is almost impossible to guarantee that a query pair $(X_q, Y_q)$ reflects exactly the same retouching operation as an exemplar pair in pixel space. Differences in pose, crop, shot scale, and visible anatomy make many operations non-comparable or undefined. A face crop, for example, cannot support shoulder or leg reshaping; a profile view and a frontal portrait do not instantiate the same local geometry. This operation misalignment produces ambiguous supervision and can slow or prevent convergence (Liu et al., 28 Jun 2026).

A second failure mode arises from naïve same-identity pairing. If the same before/after pair is reused as both exemplar and query, the model can memorize pixel-wise differences rather than learning an identity-agnostic representation of retouching semantics. MirrorPPR47M is therefore coupled to a self-augmentation paradigm that turns each same-identity retouched pair into many operation-aligned but spatially decoupled training quadruplets, forcing the model to infer retouching operations rather than copy local residuals (Liu et al., 28 Jun 2026).

2. Composition, scale, and source material

MirrorPPR47M contains approximately $47{,}451{,}284$ retouched pairs. Of these, $808{,}439$ belong to a simulated subset and $46{,}642{,}845$ to a professional subset. The “47M” designation refers to the scale of these base before/after pairs rather than the much larger number of derived training quadruplets produced by augmentation. Through self-augmentation, each original pair yields on average about $X_s$0–$X_s$1 variants, producing hundreds of millions of effective training quadruplets (Liu et al., 28 Jun 2026).

The simulated subset is built from $X_s$2 FFHQ portraits at $X_s$3 resolution. These images are cropped or resized to approximately $X_s$4 megapixels, with height and width constrained to multiples of $X_s$5. The professional subset uses $X_s$6 portraits from PPR10K, originally at $X_s$7K–$X_s$8K resolution. In that pipeline, YOLO detects the human subject, only single-person images are retained, the bounding box is expanded by $X_s$9 on all sides, and the cropped region is resized to roughly $X_t$0 megapixels (Liu et al., 28 Jun 2026).

The paper does not define an official train/validation/test split for MirrorPPR47M itself. Instead, the dataset is used as training data for MirrorPPR, while evaluation is performed on separate benchmarks: SimFace-100, described as 100 simulated retouching combinations on 12 faces at $X_t$1, and ProPortrait-500, described as 500 combinations on 40 PPR10K-based portraits with area around $X_t$2 (Liu et al., 28 Jun 2026).

The following summary captures the declared dataset structure.

Component	Source	Scale
Simulated subset	FFHQ	808,439 retouched pairs
Professional subset	PPR10K	46,642,845 retouched pairs
Total	Combined	≈ 47,451,284 retouched pairs

A retouched pair is always a same-identity before/after instance. In the simulated subset, the target image is generated with Landmark-Guided Local Warping (LLW). In the professional subset, the target image is produced by a commercial professional retouching API. In both cases, the supervision signal is encoded implicitly in the image pair itself rather than in explicit dense edit maps (Liu et al., 28 Jun 2026).

3. Self-augmentation and operation alignment

The key methodological feature associated with MirrorPPR47M is its data self-augmentation paradigm. Starting from a base pair $X_t$3, the training pipeline constructs quadruplets $X_t$4 by applying the same spatial transformation $X_t$5 to both source and target:

$X_t$6

The resulting quadruplet preserves the retouching operation from exemplar to query pair while breaking absolute coordinate alignment between them. The exemplar remains $X_t$7, while the transformed pair $X_t$8 serves as query input and supervision target (Liu et al., 28 Jun 2026).

The sampled transformation family includes small-angle rotations at $X_t$9, $(X_s, X_t)$0, and $(X_s, X_t)$1, horizontal flipping, dynamic cropping with aspect ratios from $(X_s, X_t)$2 to $(X_s, X_t)$3, and scaling or cropping under facial or body coverage constraints. Subject visibility is enforced during cropping by using FFHQ face bounding boxes and YOLO-detected human body masks (Liu et al., 28 Jun 2026).

This construction solves two distinct problems. First, it eliminates the need to find cross-identity pairs that instantiate exactly the same operation, since all query pairs are derived from the same base retouched pair. Second, it prevents pixel-wise shortcut learning, because the query pair no longer shares the same pixel coordinates as the exemplar pair. The model must therefore learn an operation representation that is stable under spatial transformation (Liu et al., 28 Jun 2026).

The paper also gives a conceptual operator view. If $(X_s, X_t)$4 denotes the retouch operator such that $(X_s, X_t)$5, then after augmentation:

$(X_s, X_t)$6

Conceptually, the query-side operation is

$(X_s, X_t)$7

This expresses strict operation equivalence in transformed coordinates, even though the edits occur at different spatial locations and scales (Liu et al., 28 Jun 2026).

4. Retouching operations represented in the dataset

MirrorPPR47M focuses on structural geometric edits and divides them into simulated and professional operation families. The simulated subset is generated by Landmark-Guided Local Warping and contains 8 base types with two directions each, yielding 16 directed operations. These span eyes, nose, and mouth. The eye category includes decreasing or increasing eye distance and shrinking or enlarging eyes. The nose category includes narrowing or widening the nose bridge, narrowing or widening the nasal alae, and shortening or lengthening the nose. The mouth category includes moving the mouth downward or upward, thinning or plumping the lips, and shrinking or enlarging the mouth. Each simulated retouched pair combines 1–8 operations, sampled uniformly (Liu et al., 28 Jun 2026).

The professional subset is based on a commercial retouching API with 27 fine-grained operations. These are grouped into face shape, eyes and brows, nose, mouth, and body proportions. Face-shape operations include round or sharpen face, sharpen or square jawline, shorten or lengthen chin, and depress or fill temples. Eyes-and-brows operations include changes to eyebrow distance, eyebrow vertical position, eyebrow thickness, eye distance, eye height, eye width, eye size, and eye vertical position. Nose operations include shrinking or enlarging the nose, narrowing or widening the nasal alae, narrowing or widening the nose bridge, shrinking or enlarging the nose tip, and shortening or lengthening the nose. Mouth operations include vertical displacement, lip thickness, mouth-corner height, mouth width, and mouth size. Body-proportion operations include square shoulders, narrow or broaden shoulders, thicken or slim arms, slim legs, and slim waist (Liu et al., 28 Jun 2026).

These operations are almost entirely geometric rather than appearance-based. Most are local, such as edits to eyes, nasal alae, or mouth corners. Some are semi-global within the human region, such as face-shape reshaping or body slimness. The paper states that the edits are “extremely delicate and localized” in the professional subset, and that the corresponding transformations are preserved as high-resolution before/after differences without explicit edit maps (Liu et al., 28 Jun 2026).

5. Annotation structure, alignment, and quality control

MirrorPPR47M relies on two forms of alignment. The first is within each retouched pair $(X_s, X_t)$8, where source and target depict the same identity under the same underlying image content, with geometric edits applied relative to landmarks or known regions. The second is between exemplar and query pairs during training, where the same transformation $(X_s, X_t)$9 is applied to both images in a pair, ensuring operation alignment in a semantic sense despite changed coordinates (Liu et al., 28 Jun 2026).

Several metadata sources are used internally during generation and augmentation. In the simulated subset, LLW uses 468 MediaPipe facial landmarks together with region convex hulls and masks. For cropping and preprocessing, the pipeline uses FFHQ face bounding boxes and YOLO body masks. The paper does not state that these landmarks or masks are distributed as final annotations; the dataset is described operationally as a collection of image pairs (Liu et al., 28 Jun 2026).

Operation labels and magnitudes are known internally. For the simulated subset, the pipeline tracks which of the 16 directed operations are applied, with direction and magnitude encoded through warping parameters. For the professional subset, the commercial API provides a labeled set of 27 operations, and the system knows which operations and how many, from 1–7 per pair, were applied. However, these labels are not directly used for training MirrorPPR; they provide implicit supervision through the before/after images and are used primarily for reporting and benchmark construction rather than for per-operation classification losses (Liu et al., 28 Jun 2026).

Quality control differs by subset. The simulated subset filters for head pose within $X_q$0 in pitch, roll, and yaw, and removes blurry, poorly exposed, or heavily occluded faces. The professional subset filters PPR10K images to single-person portraits using YOLO detection and requires minimum portrait area of at least 240k pixels. These constraints are intended to stabilize geometric warping and ensure that the structural operations remain meaningful at high resolution (Liu et al., 28 Jun 2026).

6. Role in MirrorPPR training and representation learning

MirrorPPR47M is the training substrate for the MirrorPPR framework. Training uses quadruplets derived from the dataset: the exemplar pair $X_q$1, the query image $X_q$2, and the query target $X_q$3. From the exemplar pair, the model extracts an operation representation and then uses it to transform $X_q$4 into a prediction $X_q$5 that should match $X_q$6 (Liu et al., 28 Jun 2026).

Training proceeds in two stages. The first is pre-training of a Retouching Operation Extractor. This module uses a frozen MAE to encode images into patch features and a trainable R-Former with learnable query tokens to distill an edit representation $X_q$7. A small MLP projects $X_q$8 into a compact embedding $X_q$9. Query patch tokens are additively modulated by this embedding, and a lightweight ViT decoder reconstructs $\hat{Y}_q$0. The pre-training loss is

$\hat{Y}_q$1

with $\hat{Y}_q$2 (Liu et al., 28 Jun 2026).

The second stage is joint fine-tuning with a pre-trained image-editing Diffusion Transformer, specifically Qwen-Image-Edit-2511. After pre-training, the auxiliary MLP and ViT decoder are discarded, and the R-Former is plugged into the DiT through a connector and LoRA modules. In MirrorPPR-Pro, the professional subset alone is used for final fine-tuning. The jointly trained components are the R-Former, the connector that maps $\hat{Y}_q$3 into DiT conditioning space, and LoRA adapters inside DiT attention layers, while the DiT base weights remain frozen (Liu et al., 28 Jun 2026).

The training regime follows a curriculum. For MirrorPPR-Pro, the extractor is first pre-trained on the simulated subset for 60k steps, then on the professional subset for 40k steps, and finally the full system is fine-tuned on the professional subset for 150k steps. The stated rationale is that large, easier-to-perceive LLW deformations teach basic geometric operations before the model is adapted to “extremely subtle, realistic commercial retouching operations” (Liu et al., 28 Jun 2026).

The DiT stage uses a flow-matching loss in latent space. Let $\hat{Y}_q$4 be the frozen VAE encoder, $\hat{Y}_q$5 the latent of the target retouched image, and $\hat{Y}_q$6 latent noise. With $\hat{Y}_q$7,

$\hat{Y}_q$8

The model predicts velocity

$\hat{Y}_q$9

and is trained with

$(X_q, Y_q)$0

Here $(X_q, Y_q)$1 is obtained from frozen Qwen2.5-VL features of the query image, and $(X_q, Y_q)$2 from the connector applied to the extracted retouch representation (Liu et al., 28 Jun 2026).

The paper attributes several properties to the learned edit representation induced by this training regime: high operation-transfer consistency with cosine similarity around 0.95 between exemplar and output embeddings, clear clustering by operation type in t-SNE, and vector additivity in which combining edit vectors such as $(X_q, Y_q)$3 synthesizes composite edits that quantitatively outperform baselines. This suggests that MirrorPPR47M does not merely provide large-scale supervision, but organizes that supervision in a way that makes retouching operations linearly and semantically compositional in latent space (Liu et al., 28 Jun 2026).

7. Comparative position, evaluation, and practical considerations

MirrorPPR47M is positioned against several categories of prior data. Makeup and appearance-oriented datasets, including BeautyGAN, AutoRetouch, RetouchingFFHQ, and PPR10K itself, are described as focusing mainly on makeup transfer, tonal enhancement, and skin smoothing. Retouching detection datasets, such as those by Bharati et al. and Rathgeb et al., support authenticity detection rather than explicit operation transfer. General text-guided editing datasets cover broad edit types but do not provide aligned before/after structural operations with fine local control (Liu et al., 28 Jun 2026).

The paper identifies five distinctive properties of MirrorPPR47M: scale beyond 47 million retouched pairs; operation coverage spanning 16 controlled simulated operations and 27 professional operations; strict operation alignment via same-identity pairs and self-augmentation; design centered on geometry rather than style; and compatibility with curriculum learning through the separation of simulated and professional subsets (Liu et al., 28 Jun 2026).

Evaluation is reported on the two dedicated benchmarks rather than on MirrorPPR47M itself. On ProPortrait-500, MirrorPPR-Pro achieves PSNR 32.65, SSIM 0.927, LPIPS 0.200, and Face Similarity 0.960. The cited strong text-guided baseline Nano Banana 2 records PSNR 27.45, SSIM 0.904, LPIPS 0.183, and Face Similarity 0.667. On SimFace-100, MirrorPPR-Face achieves PSNR 32.25, SSIM 0.909, LPIPS 0.186, and Face Similarity 0.937. The paper states that other exemplar-based and multi-reference methods perform worse on both fidelity and identity preservation, and attributes this gap in substantial part to the scale and structure of MirrorPPR47M (Liu et al., 28 Jun 2026).

The project page is listed as https://sjtu-deng-lab.github.io/MirrorPPR. The paper does not explicitly state whether MirrorPPR47M is publicly downloadable or under what license it is released. Because the dataset is derived in part from FFHQ, PPR10K, and a commercial retouching API, full distribution may be constrained. A plausible implication is that practical access may depend on the licensing and privacy conditions of the underlying sources rather than on MirrorPPR alone (Liu et al., 28 Jun 2026).

Ethical and social issues are only implicit in the paper. Since the dataset uses face data and supports identity-preserving structural manipulation, potential misuse includes identity manipulation and reinforcement of unrealistic beauty standards. No explicit bias analysis is reported. A plausible implication is that any demographic skew present in FFHQ or PPR10K may propagate into MirrorPPR47M and into models trained on it (Liu et al., 28 Jun 2026).