PresAesth: Evaluating and Preserving Aesthetic Quality

• PresAesth is a computational framework focused on evaluating and preserving aesthetic outcomes while ensuring fidelity to identity, anatomy, and scene semantics.
• It integrates techniques from image aesthetics assessment, clinical evaluation, personalized adjustments, and structured reward-based guidance to enhance outcome prediction.
• The approach mitigates traditional mean-opinion limitations by incorporating case-based retrieval and subject-conditioned adaptation across applications like breast-conserving therapy and facial surgery.

PresAesth denotes a family of computational approaches concerned with the presentation, preservation, assessment, alignment, or optimization of aesthetic outcomes. In breast-conserving treatment, it is explicitly tied to “presentation and preservation of aesthetic outcomes,” combining aesthetic evaluation with retrieval of visually similar prior cases for counseling (Silva et al., 2022). In later work, the term or an equivalent framing extends to presurgical preview, facial retouching, image reconstruction, localized surgical editing, and even slide design, where the common requirement is not merely to maximize an aesthetic score, but to do so while preserving identity, anatomy, scene semantics, or communicative intent (Han et al., 2024, Yang et al., 1 Mar 2026, Liu et al., 7 Oct 2025). This suggests that PresAesth is best understood as a cross-domain research program rather than a single model.

1. Conceptual scope and problem setting

PresAesth arises from a recurrent deficiency in aesthetic computing: average or purely subjective judgment is often insufficient when the task requires realistic preview, individualized counseling, controllable editing, or iterative refinement. In the breast cancer setting, there is “still no accepted gold standard for outcome evaluation,” and dissatisfaction remains high because expectations are poorly managed (Silva et al., 2022). In presentation generation, the central claim is that “there is no way to improve it when you cannot evaluate it right,” motivating a dedicated aesthetic-awareness engine for scoring, defect adjustment, and comparison (Liu et al., 7 Oct 2025). In personalized aesthetics, generic mean-opinion models are explicitly criticized for averaging away inter-user variability, motivating subject-conditioned prediction (Yang et al., 2022).

A second defining feature is that PresAesth typically links aesthetics to preservation constraints. In orthognathic preview, the goal is a fully automated 3D prediction from a single 3D facial scan without CT/CBCT or surgeon-specified parameters, while maintaining dense correspondences and clinically meaningful facial geometry (Han et al., 2024). In face retouching, BeautyGRPO formulates the core tension as aesthetic alignment versus identity and texture fidelity under stochastic RL exploration (Yang et al., 1 Mar 2026). In Aesthetic Photo Reconstruction, AesFormer states the objective as improving aesthetic quality via structural reconstruction while preserving subject identity and scene semantics (Du et al., 21 May 2026).

The broader literature supplying technical substrate for PresAesth comes from image aesthetics assessment. Unified distribution prediction on AVA learns the full 10-bin score histogram rather than only binary labels or mean scores (Murray et al., 2017). MLSP-based models retain full-resolution information and report a best SRCC of 0.756 on AVA by extracting multi-level spatially pooled features from all convolutional blocks (Hosu et al., 2019). MPA-Net emphasizes aspect-ratio-preserving multi-patch inference and reports gains over NIMA in LCC, SRCC, and MSE, especially for images far from square aspect ratio (Wang et al., 2020).

2. Clinical PresAesth: outcome assessment, preview, and counseling

The clearest clinical instantiation appears in breast-conserving therapy. A compact CNN with 262,908 learnable parameters predicts a binary aesthetic outcome from post-treatment RGB photographs resized to $384\times256$, while a penultimate-layer embedding supports top-3 nearest-neighbor retrieval under Euclidean $L_2$ distance, $d(x,y)=||x-y||_2$ (Silva et al., 2022). The model is clinically structured: it predicts 8 keypoint coordinates, computes asymmetry measures used in the literature—LBC, BCE, UNR, and BRA—and classifies $\{\text{Excellent, Good}\}$ versus $\{\text{Fair, Poor}\}$. On 143 photographs from PORTO and TSIO, it reports Accuracy 0.86 and Balanced Accuracy 0.89, outperforming SVM baselines based on asymmetry features (Silva et al., 2022). The clinical utility is explicitly case-based: during preoperative or postoperative visits, clinicians can obtain the binary assessment and show top-3 similar prior cases from an institutional catalog.

Orthognathic surgery preview extends PresAesth from assessment to 3D prediction. “Facial Surgery Preview Based on the Orthognathic Treatment Prediction” uses FLAME as both encoder and decoder, predicts latent-code differences from preoperative scans, and trains with mouth-convexity and asymmetry losses aligned to clinical constructs such as the Steiner line and the mid-sagittal plane (Han et al., 2024). The dataset comprises 163 pre/post 3dMD scan pairs, with 5-fold cross-validation and an orthognathic-specific augmentation scheme expanding training data from 133 to 1330 pairs in ablations. The full model reports mean Hausdorff Distance 8.999 mm and Chamfer Distance 2.503 mm, and a Two-Alternative Forced Choice study found that medical professionals selected the ML prediction 49.3% of the time and engineers 50.2%, indicating that neither group could reliably distinguish predictions from real postoperative appearances (Han et al., 2024).

A related but distinct formulation is automated post-operative assessment from paired frontal facial photographs. “Automated Assessment of Aesthetic Outcomes in Facial Plastic Surgery” assembles SurFace1259, with 7,160 photographs from 1,259 patients and a rhinoplasty-only subset of 732 images from 366 patients (Varghaei et al., 18 Aug 2025). The framework combines MediaPipe FaceMesh landmarking, symmetry geometry, nasal morphometry, perceived age via DeepFace, and identity verification via AdaFace. In the rhinoplasty cohort, 96.2% improved in at least one of three significant nasal ratios; in the broader 989-subject frontal cohort, 71.3% showed significant enhancements in global facial symmetry or perceived age; and identity remained stable with True Match Rates of 99.5% and 99.6% at a False Match Rate of 0.01% for the rhinoplasty-specific and general cohorts, respectively (Varghaei et al., 18 Aug 2025).

Rhinoplasty goal visualization introduces a further clinical distinction: localized editing needs localized evaluation. Envisage uses FLUX.1-Fill-dev, MediaPipe-derived nasal masks, clinical presets, thin-plate spline warping, and hard-mask compositing so that non-surgical pixels are preserved exactly by construction (Agarwal et al., 26 Jun 2026). The paper argues that full-face identity scores are structurally confounded under hard compositing and therefore introduces SurgicalScore, a mask-decomposed protocol combining directional alignment, edit magnitude, masked LPIPS fidelity, realism, and outside-mask preservation. On $N=211$, Envisage reports the highest SurgicalScore, 0.599 $[0.579, 0.619]$, while all paired ArcFace gains remain negative, supporting the claim that full-face identity is poorly aligned with localized surgical accuracy under hard compositing (Agarwal et al., 26 Jun 2026).

3. Architectural patterns and computational mechanisms

A persistent architectural pattern in PresAesth is clinically or semantically grounded intermediate supervision. In breast outcome assessment, keypoint prediction is not auxiliary decoration: it regularizes the network, constrains representations to be physiologically relevant, and feeds explicit asymmetry computations before classification (Silva et al., 2022). In orthognathic preview, FLAME enforces consistent topology, symmetric vertex pairing, and differentiable mesh-level losses, making dense correspondence intrinsic rather than post hoc (Han et al., 2024). In Botox planning, localized latent axes in StyleGAN2 $W^+$ are discovered by ROI masking, manifold projection, and symmetric target construction, enabling region-specific edits for left/right eyebrow, eye, and mouth-corner regions (Arnaud et al., 27 Jan 2026).

A second pattern is the separation of planning from execution. AesFormer makes this separation explicit: Stage 1, AesThinker, outputs ordered aesthetic actions across seven dimensions—aspect ratio, framing and composition, camera viewpoint, subject placement, subject pose and action details, focus and depth-of-field, and color and light—while Stage 2, AesEditor, executes those actions using a rectified-flow editor (Du et al., 21 May 2026). Envisage follows an analogous decomposition through preset detection, mask extension, TPS deformation, depth modification, and inpainting (Agarwal et al., 26 Jun 2026). This suggests that PresAesth systems increasingly treat aesthetics as a structured control problem rather than a monolithic end-to-end regression target.

A third pattern is reward- or discriminator-based aesthetic guidance. AesUST introduces an aesthetic discriminator trained on WikiArt and injects its multi-scale features into an Aesthetic-aware Style-Attention module, combining global aesthetic channel distribution with local semantic spatial distribution (Wang et al., 2022). BeautyGRPO trains a reward model over FRPref-10K, with five retouching dimensions—Skin Smoothing, Blemish Removal, Texture Quality, Clarity, and Identity Preservation—and stabilizes RL exploration through Dynamic Path Guidance (Yang et al., 1 Mar 2026). EvoPresent’s PresAesth uses multi-task GRPO on scoring, adjustment, and comparison tasks over slide images, with task-specific rewards and structured outputs (Liu et al., 7 Oct 2025).

4. Personalization, preference alignment, and comparative judgment

PresAesth is not limited to institutional or clinician-defined quality; several works make subjective preference a first-class variable. “Personalised aesthetics with residual adapters” models user-specific deviations from a shared aesthetic backbone by inserting lightweight residual adapters into ResNet-18 and reports mean Spearman $\rho \approx 0.639$ with $N=100$ user-specific ratings on FLICKR-AES (Rodríguez-Pardo et al., 2019). “Personalized Image Aesthetics Assessment with Rich Attributes” introduces PARA with 31,220 images, 438 subjects, 9 image-oriented objective attributes, 4 human-oriented subjective attributes, and desensitized subject information including Big Five traits (Yang et al., 2022). Its conditional PIAA model multiplies subject information into the last-layer output of the image backbone, and personality-trait conditioning yields the strongest improvement among tested subject attributes, reaching SROCC 0.7046 and PLCC 0.7509 in the 100-shot setting (Yang et al., 2022).

Preference alignment becomes more explicit in face retouching. BeautyGRPO treats supervised reconstruction as insufficient because it “enforces local fidelity but neglects global perceptual alignment,” while naïve online RL induces stochastic drift (Yang et al., 1 Mar 2026). The method therefore combines a fine-grained reward model with Dynamic Path Guidance, an anchor-based guided trajectory that stabilizes stochastic sampling. On FFHQR, it reports NIQE 10.831, NIMA 5.123, MUSIQ 4.906, MANIQA 1.079, NRQM 8.401, TOPIQ 0.676, FID 4.054, and ArcFace 0.952, alongside a user-study win-rate of 63.25 against multiple baselines (Yang et al., 1 Mar 2026).

Comparative judgment is also central to academic-presentation aesthetics. In EvoPresent, PresAesth is a specific multi-task RL model built on Qwen2.5-VL-7B, trained on ~3,400 instances for three tasks: scoring on a 1–10 scale, defect adjustment over Composition/Layout, Typography, and Imagery/Visualizations, and A/B comparison between slide variants (Liu et al., 7 Oct 2025). The model reports Scoring MAE 1.33, Adjustment F1 0.389, and Comparison accuracy 0.878, outperforming the listed baselines on the Aesthetic Awareness benchmark (Liu et al., 7 Oct 2025). Inference is embedded into an iterative checker loop, where scoring, feedback, and comparison jointly drive self-improvement.

AesTest generalizes this comparative orientation from isolated models to benchmarking. It contains 17,885 images and 8,757 multiple-choice items spanning perception, appreciation, creation, and photography, and reports that production-oriented tasks such as Framing Choice are markedly harder than perception tasks for current MLLMs (Wang et al., 9 Nov 2025). This supports a central PresAesth claim: aesthetic competence is not exhausted by static perception.

5. Evaluation, interpretability, and methodological controversies

Evaluation in PresAesth is heterogeneous because the target variable differs by domain. Generic image-aesthetics work emphasizes MOS regression, SRCC, PLCC, EMD, or binary accuracy (Murray et al., 2017, Hosu et al., 2019). Clinical work often mixes outcome categories, geometric distances, ratio-based morphometry, user studies, or identity verification (Silva et al., 2022, Han et al., 2024, Varghaei et al., 18 Aug 2025). Localized editing adds masked metrics and outside-mask preservation terms (Agarwal et al., 26 Jun 2026). This diversity is not accidental; it reflects the fact that aesthetic quality, fidelity, and preservation are not commensurate objectives.

Interpretability is likewise domain-specific. The breast-treatment model uses Layer-wise Relevance Propagation, with saliency maps focusing on breast contours, nipple–areola complex, and areas of retraction, which supports trustworthiness claims (Silva et al., 2022). Facial plastic surgery assessment uses human-understandable geometry—symmetry error, nasal ratios, perceived age, and identity consistency—rather than latent scores (Varghaei et al., 18 Aug 2025). EvoPresent’s PresAesth produces structured reasoning in explicit tags, and its reward checks operate on parseable outputs rather than opaque logits alone (Liu et al., 7 Oct 2025).

A recurrent controversy concerns what should count as a valid metric. Envisage directly challenges the use of full-face identity metrics for hard-composited localized edits, showing that outside-mask SSIM exceeds 0.999 by construction and that the mean mask area is only about 3.65% of pixels (Agarwal et al., 26 Jun 2026). A plausible implication is that any PresAesth system performing constrained local edits must specify whether fidelity is being evaluated globally, locally, or relative to a prescribed direction of change.

6. Limitations, governance, and future directions

Across domains, PresAesth systems remain limited by data scale, label subjectivity, and generalization. The breast-conserving-treatment pipeline is trained on only 143 photographs, uses labels from a single expert, and does not measure inter-rater variability (Silva et al., 2022). Orthognathic preview is based on 163 scan pairs, mostly Asian patients, and does not model multimodal factors such as age, gender, or skin condition (Han et al., 2024). Automated facial-plastic-surgery assessment relies on real-world social-media imagery with uncontrolled lighting, makeup, and camera parameters, and lacks patient-level age/sex/ethnicity annotations for fairness auditing (Varghaei et al., 18 Aug 2025).

Localized generative systems add further constraints. Botox planning is affected by biological variability in diffusion, injection depth, muscle fiber composition, and metabolic response, so the reported correlations are explicitly characterized as moderate-to-strong directionality rather than precise dosage prediction (Arnaud et al., 27 Jan 2026). BeautyGRPO notes residual bias and generalization issues under extreme lighting, heavy makeup, or unusual textures (Yang et al., 1 Mar 2026). AesFormer reports identity drift for heavily occluded faces or extreme angles and semantic drift when aggressive cropping removes salient context (Du et al., 21 May 2026).

Privacy and governance are inseparable from PresAesth in medical and personalized settings. The breast-outcome paper explicitly recommends de-identification, consents, subgroup monitoring, and potential privacy-preserving generative morphing for case-based explanations (Silva et al., 2022). PARA provides desensitized subject metadata, but the use of personality traits for conditioning raises evident privacy and fairness questions (Yang et al., 2022). Envisage is released under research-only constraints because FLUX.1-dev/Fill-dev licensing restricts biometric processing and production identity-related use (Agarwal et al., 26 Jun 2026).

Future work, as stated across the literature, points toward multi-center validation, 2D-to-3D expansion, stronger personalization, privacy-preserving explanations, learned rankers for candidate selection, and broader evaluation of production-oriented aesthetic intelligence (Silva et al., 2022, Han et al., 2024, Agarwal et al., 26 Jun 2026, Wang et al., 9 Nov 2025). Taken together, these directions suggest that PresAesth is evolving from isolated aesthetic scoring systems into a broader infrastructure for controlled, interpretable, and domain-aware aesthetic decision support.