MirrorPPR: Exemplar-Based Portrait Retouching

Updated 6 July 2026

The paper introduces MirrorPPR, a framework that transfers retouching operations from a before/after exemplar pair to a query portrait.
It uses a Retouching Operation Extractor and a pre-trained Diffusion Transformer enhanced with LoRA to achieve delicate, localized geometric edits.
Empirical results on the MirrorPPR47M dataset show superior performance in metrics like PSNR, SSIM, LPIPS, and face similarity compared to baseline methods.

MirrorPPR is a framework for exemplar-based portrait photo retouching in which the model is given an exemplar pair $(X_s, X_t)$ , where $X_s$ is an original portrait and $X_t$ is its retouched version, together with a new query portrait $X_q$ , and produces $\hat{Y}_q$ that applies the same retouching operations to the query image. It is designed for structural portrait retouching, where the edit is a delicate and localized geometric transformation such as shrinking a mouth, narrowing a jaw, enlarging eyes, sliming legs, or adjusting shoulder width, rather than a broad semantic rewrite. The framework combines a Retouching Operation Extractor, a pre-trained Diffusion Transformer (DiT), an advanced data self-augmentation paradigm, and the MirrorPPR47M dataset with over 47 million retouched pairs (Liu et al., 28 Jun 2026).

1. Task definition and problem setting

MirrorPPR formalizes portrait retouching as operation transfer from an exemplar pair, not prompt following. The exemplar pair $(X_s, X_t)$ specifies the operation, the query portrait $X_q$ is the target image to be edited, the model output is $\hat{Y}_q$ , and the corresponding ground truth is $Y_q$ (Liu et al., 28 Jun 2026).

The paper argues that text-guided editing is fundamentally insufficient for this setting. Structural portrait retouching requires precise control over where to edit, how much to edit, and in which direction to deform anatomy. Natural language is described as too ambiguous to encode the exact scale, spatial locality, or coupled geometry across facial parts. As a result, text-guided models tend to under-edit, over-edit, or distort identity. By contrast, exemplar-based guidance directly demonstrates the intended transformation through a before/after pair (Liu et al., 28 Jun 2026).

A central difficulty is that existing exemplar-based editing methods primarily focus on tasks with pronounced visual transformations, whereas structural portrait retouching involves extremely delicate and localized modifications. This makes accurate extraction and transfer of the demonstrated operation substantially harder than in conventional exemplar-based editing. The method is therefore organized around isolating an edit representation that captures subtle structural change while preserving identity and unedited regions.

2. Core framework and representation of retouching operations

MirrorPPR has two major components: a Retouching Operation Extractor that learns a compact representation of the edit demonstrated by $(X_s, X_t)$ , and a pre-trained Diffusion Transformer (DiT) backbone that receives that operation representation and applies it to $X_s$ 0 (Liu et al., 28 Jun 2026).

The extractor consists of a frozen MAE and a trainable Transformer called R-Former. The frozen MAE produces patch-level features that preserve local structure well. These features are fed into the R-Former together with a small set of learnable query tokens. The output features corresponding only to these query tokens are kept and denoted

$X_s$ 1

In the paper’s language, $X_s$ 2 is meant to encode the “edit direction” rather than the full image content (Liu et al., 28 Jun 2026).

To make $X_s$ 3 represent edit intent rather than incidental appearance cues, the extractor is first pre-trained with an auxiliary reconstruction task. A temporary MLP projects $X_s$ 4 to a compact edit embedding $X_s$ 5. For a query image $X_s$ 6, its patch embeddings are obtained, $X_s$ 7 is added to each patch token, and a lightweight ViT decoder reconstructs the retouched target $X_s$ 8. The pre-training loss is

$X_s$ 9

After this stage, the temporary MLP and decoder are discarded; the extractor has learned a robust operation representation (Liu et al., 28 Jun 2026).

The transfer stage uses a pre-trained Qwen-Image-Edit-2511 backbone, described as a dual-stream DiT-based editing model. One stream uses Qwen2.5-VL to compute high-level visual conditioning features, while the other performs diffusion denoising. Because the task is exemplar-driven rather than text-driven, the method removes text instructions entirely. It keeps the frozen Qwen2.5-VL visual encoder to obtain query-image conditioning features $X_t$ 0, the frozen VAE encoder $X_t$ 1 to obtain the query latent

$X_t$ 2

and the pre-trained extractor to produce $X_t$ 3 from $X_t$ 4 (Liu et al., 28 Jun 2026).

The extracted operation features pass through a trainable connector, which maps them into the instruction-conditioning space expected by the DiT and produces

$X_t$ 5

The DiT therefore receives two conditions: $X_t$ 6, which describes what the query portrait looks like, and $X_t$ 7, which specifies what retouching operation should be applied.

3. Diffusion transfer, flow matching, and parameter-efficient adaptation

MirrorPPR trains the editing model with a flow-matching loss. Let

$X_t$ 8

and for timestep $X_t$ 9,

$X_q$ 0

The DiT predicts a velocity $X_q$ 1, trained with

$X_q$ 2

This objective couples the latent target $X_q$ 3, the query latent $X_q$ 4, the visual condition, and the extracted edit condition in a single transfer mechanism (Liu et al., 28 Jun 2026).

To adapt the frozen DiT efficiently, MirrorPPR adds LoRA modules to its attention blocks and jointly fine-tunes the pre-trained R-Former, the connector, and the LoRA parameters. The DiT’s original weights remain frozen, which preserves the backbone’s generative prior while letting it specialize to portrait retouching (Liu et al., 28 Jun 2026).

The reported ablations further specify the adapter design. Increasing query tokens beyond 8 yields only marginal gains, a 4-layer R-Former is described as the best tradeoff, and LoRA rank 32 is reported as the sweet spot. In the same ablation, rank 8 underfits, while rank 64 gives slight gains in some metrics but hurts identity preservation and increases cost (Liu et al., 28 Jun 2026). This suggests that the task benefits from a compact but precise operation representation rather than a large, highly expressive adapter.

4. Data construction, self-augmentation, and MirrorPPR47M

A major obstacle in exemplar-based retouching is constructing valid training quadruplets

$X_q$ 5

A naive cross-identity strategy often fails because the operation may not be geometrically meaningful across identities or compositions. A second naive strategy sets $X_q$ 6 and $X_q$ 7, which avoids cross-identity mismatch but creates a shortcut in which the model can memorize pixel-to-pixel differences at the same spatial coordinates rather than learn the actual retouching semantics (Liu et al., 28 Jun 2026).

MirrorPPR addresses this with an advanced data self-augmentation paradigm. The same random spatial augmentation $X_q$ 8 is applied to both source and target exemplar images: $X_q$ 9 Typical augmentations include scaling, cropping, rotation, and horizontal flipping. This preserves the edit semantics exactly, while breaking absolute coordinate alignment and preventing shortcut learning (Liu et al., 28 Jun 2026).

To support the task, the authors construct MirrorPPR47M, a dataset with over 47 million retouched pairs, organized into a simulated subset and a professional subset (Liu et al., 28 Jun 2026).

Subset	Source and composition	Role
Simulated Retouching Subset	FFHQ; 30,171 images; 8 base operation types, each with 2 directions; 808,439 retouched pairs	Clean, controlled geometry changes for early-stage learning
Professional Retouching Subset	PPR10K; 3,789 high-resolution 4K–8K portraits; 27 fine-grained professional operations; 46,642,845 retouched pairs	Real-world editing behavior with subtler operations

The simulated subset uses a custom Landmark-Guided Local Warping (LLW) method. LLW uses facial landmarks and similarity-based moving least squares deformation to synthesize pronounced structural edits. The professional subset is generated using a commercial retouching API and includes operations covering facial features, face shapes, and body proportions (Liu et al., 28 Jun 2026).

The dataset construction uses strict filtering. For the simulated subset, only images with small head pose variation and good visual quality are kept. For the professional subset, images with multiple people or too-small portrait regions are removed to ensure single-subject retouching. Each original pair is expanded into about 13.3–13.4 spatially decoupled variations through synchronized augmentation of both source and target images (Liu et al., 28 Jun 2026).

5. Curriculum learning, benchmarks, and empirical performance

MirrorPPR47M is organized to support curriculum learning. The training schedule is: pre-train the extractor on the simulated subset to learn basic structural changes, continue pre-training on the professional subset to adapt to subtle realistic edits, and finally fine-tune the full model on professional data (Liu et al., 28 Jun 2026). The paper presents this as an easy-to-hard schedule that helps the model first learn clear geometric operations and then handle difficult real-world retouching.

Evaluation is conducted on two cross-identity benchmarks: SimFace-100, defined as 100 combinations from 12 face images, using the 8 simulated operation types, and ProPortrait-500, defined as 500 combinations from 40 high-quality portraits, using the 27 professional operations. The principal metrics are PSNR, SSIM, LPIPS, and Face Similarity, where Face Similarity is measured using ArcFace cosine similarity between output and ground truth (Liu et al., 28 Jun 2026).

MirrorPPR is compared against multi-reference editors, exemplar-based editors, and text-guided variants of strong generators. The paper lists Qwen-Image-Edit-2511, FLUX.2-dev, Nano Banana 2, Seedream 4.5, Qwen-Image-Edit-2511-ICEdit-LoRA, RelationAdapter, and EditTransfer among the baselines (Liu et al., 28 Jun 2026).

Benchmark	Variant	PSNR	SSIM	LPIPS	Face Similarity
SimFace-100	MirrorPPR-Face	32.25	0.909	0.186	0.937
ProPortrait-500	MirrorPPR-Pro	32.65	0.927	0.200	0.960

On SimFace-100, the paper states that MirrorPPR-Face achieves PSNR 32.25, SSIM 0.909, LPIPS 0.186, and Face Similarity 0.937. On ProPortrait-500, MirrorPPR-Pro achieves PSNR 32.65, SSIM 0.927, LPIPS 0.200, and Face Similarity 0.960 (Liu et al., 28 Jun 2026). The qualitative analysis reports that multi-reference and exemplar-based baselines often confuse operation transfer with blending or face swapping, while text-guided models often fail because the prompt is too ambiguous and they may over-edit facial structure. MirrorPPR is described as the only method that consistently transfers the demonstrated retouching while keeping untouched regions and identity stable.

The paper also reports a user study on ProPortrait-500 with 79% preference for MirrorPPR-Pro (Liu et al., 28 Jun 2026). A plausible implication is that the gains are not limited to reference-based metrics but are also visually salient under human evaluation.

6. Ablations, scope, and relation to other “mirror” research

The ablation study is organized around Self-Augmentation vs. shortcut learning and network design. Self w/o Aug performs worst, confirming that using the exact exemplar pair as query causes the model to memorize absolute pixel positions. Self-Augmentation matches or exceeds Cross-Identity training on simulated data, and on the professional benchmark it clearly beats Cross-Identity, showing that it is especially valuable when cross-identity operation alignment is unreliable (Liu et al., 28 Jun 2026).

These findings clarify the main conceptual contribution of MirrorPPR. The framework does not treat portrait retouching as image-to-image stylization, nor as language-conditioned deformation, but as the extraction and transfer of a demonstrated retouching operation. The paper’s stated bottom line is that MirrorPPR reframes portrait retouching as operation transfer from an exemplar pair, not text following (Liu et al., 28 Jun 2026).

The name can be misleading outside its immediate domain. MirrorPPR is distinct from several unrelated mirror-focused research directions. Reflect3r uses a mirror reflection as a physically valid virtual view for single-view 3D stereo reconstruction rather than portrait retouching (Wu et al., 24 Sep 2025). Mirror3D addresses 3D mirror plane prediction and depth refinement for mirror surfaces in RGB-D perception (Tan et al., 2021). MirrorVerse targets photorealistic mirror-reflection generation with diffusion models (Dhiman et al., 21 Apr 2025). MirrorPPR belongs instead to exemplar-based image editing and structural portrait manipulation.

A common misconception is to treat MirrorPPR as a text-guided retouching system with an extra reference image, or as a generic exemplar editor. The paper explicitly positions it otherwise: the task is defined by extremely delicate and localized modifications, the core representation is extracted from a before/after exemplar pair, and the main technical innovations are the Retouching Operation Extractor, the connector + LoRA-enhanced DiT, the advanced data self-augmentation paradigm, and the MirrorPPR47M dataset with curriculum learning (Liu et al., 28 Jun 2026).