PerceptFace: Human-Centered Face Perception

• PerceptFace is a face perception framework that defines a latent space aligned with human psychological representations and perceptual sensitivities.
• It utilizes synthesis-based methods to transform subject faces, ensuring that identity remains accessible to familiar viewers while blocking machine recognition.
• The approach integrates modules like APIM and PEIT with rigorous psychophysical validation, achieving robust privacy and high utility in shared photo scenarios.

Searching arXiv for recent and foundational papers on “PerceptFace” and related face-perception research. Searching arXiv for recent and foundational papers on “PerceptFace” and related face-perception research. PerceptFace denotes a perception-oriented line of face research in which the organizing criterion is not merely machine recognition accuracy, but alignment with human face perception. In the cited literature, that orientation appears in several forms: a learned face space organized according to human psychological representations and validated psychophysically (Suchow et al., 2018), systems for predicting or editing socially perceived facial attributes (Messer et al., 2019, Roygaga et al., 2023), and, most explicitly, a synthesis-based method dedicated to subject faces in shared photos that aims to make identity unextractable to face-recognition systems while keeping it perceptible to familiar viewers (Wang et al., 14 Sep 2025). Across these uses, PerceptFace is characterized by three recurrent commitments: perceptual grounding, human-centered evaluation, and explicit management of the gap between machine-extractable identity and human-perceived identity.

1. Conceptual foundations

PerceptFace is rooted in the claim that a useful face model must respect the “exquisite sensitivity of human face perception” and therefore requires both a face space and a renderer or decision process aligned to human perceptual sensitivities (Suchow et al., 2018). In this formulation, a “face space” is a latent, multidimensional psychological space of perceived facial features and properties, while the accompanying renderer must produce images free of distortions and artifacts that would undermine psychophysical validity. The same perceptual orientation later reappears in privacy protection, where the target is not perfect image fidelity but preservation of “identity perception rather than meticulous facial analysis” under realistic photo-sharing conditions (Wang et al., 14 Sep 2025).

This distinction between machine-readable identity and human-perceived identity is central. The 2025 PerceptFace privacy paper divides photo subjects into subject faces and bystander faces, and assigns a specific role to subject faces: they should remain recognizable to familiar persons by human vision, but not to unauthorized face-recognition systems (Wang et al., 14 Sep 2025). Earlier perceptual face-space work frames a closely related requirement in experimental terms: a generative face model should be smooth, navigable, photorealistic, and psychologically meaningful, so that manipulations in latent space correspond to intuitively interpretable changes in perceived identity (Suchow et al., 2018).

A broader implication is that PerceptFace is less a single architecture than a design doctrine. It privileges ecological validity, psychophysical evaluation, and perceptual similarity over purely computational surrogates. This suggests a unifying criterion across otherwise different tasks—generation, restoration, anonymization, social-trait prediction, and similarity modeling—namely that representations should be judged by how well they track human judgments under controlled or practical viewing conditions.

2. Perceptually aligned face spaces and psychophysical methodology

A major precursor is "Learning a face space for experiments on human identity" (Suchow et al., 2018). That work trains a PixelVAE on 3,353 aligned Humanæ portraits scraped with permission, using a dataset deliberately designed to suppress nuisance variation such as photographer, equipment, pose, and post-processing while retaining identity-relevant variation. The preprocessing pipeline resizes portraits to $1024\times 1024$, performs Procrustes alignment from facial landmarks detected by a pre-trained ensemble-of-regression-trees detector, crops a central $640\times 640$ square, and downsamples to $512\times 512$ for training. The reported result is a smooth, navigable latent face space whose samples exhibit “striking detail, variation, and little to no artifacts,” and whose perceptual alignment is tested directly with a psychophysical Turing test in which humans mostly fail to distinguish PixelVAE samples from real portraits across image sizes from $16\times 16$ to $64\times 64$ (Suchow et al., 2018).

That work also introduces a practical human-in-the-loop search procedure in “mental space” using Natural Evolution Strategies. With current seed $\theta_t$, portraits are generated via additive spherical Gaussian noise, ranked by human participants for resemblance to a target identity, and updated by

$$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$

Three example searches run for 10 rounds each, and the authors describe convergence in very few trials to collective mental templates such as “a young boy with red hair” and “Barack Obama” (Suchow et al., 2018). In PerceptFace terms, this establishes that a perceptually aligned latent space can support direct coupling between human judgments and model exploration.

A complementary methodological contribution is "Visual Psychophysics for Making Face Recognition Algorithms More Explainable" (RichardWebster et al., 2018). It adapts visual psychophysics to face recognition through $M$-alternative forced-choice identification, controlled perturbations, and item-response curves. Similarity is measured by normalized cosine similarity,

$s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$

with a thresholded decision matrix $M = [S \ge t]$. The paper introduces a herding step based on the biometric menagerie to isolate “sheep” identities before perturbation, then traces accuracy as a function of blur, occlusion, noise, brightness, contrast, sharpness, or 3D expression manipulations (RichardWebster et al., 2018). This psychophysical logic is directly compatible with PerceptFace’s emphasis on controlled perceptual sensitivity rather than aggregate benchmark scores.

A further perceptual baseline is the Linked Aggregate Code (LAC), a V1-inspired similarity model that compares topographically linked Gabor-response amplitudes at corresponding facial locations (Lyons et al., 2020). In Experiment 1, LAC-human mean concordance is $640\times 640$0 versus mean human-human concordance $640\times 640$1, with Spearman $640\times 640$2 between human-derived similarity and LAC similarity. In Experiment 2, apparent sex and apparent race emerge from the similarity structure without training, but the human judgments exhibit a racial perceptual bias that the LAC model does not share (Lyons et al., 2020). For PerceptFace, this is notable because it separates early visual coding from higher-level learned or sociocultural biases.

More recently, "Human face perception reflects inverse-generative and naturalistic discriminative objectives" (Guo et al., 12 May 2026) compared six VGG-16 models sharing an architecture but trained with distinct objectives, using controversy-optimized face pairs specifically constructed to expose disagreements among models. Across stimulus families, models trained for inverse rendering, face identification on natural images, or object categorization most robustly matched human dissimilarity judgments, and natural-image-trained models often outperformed synthetic-trained counterparts (Guo et al., 12 May 2026). This strengthens the PerceptFace thesis that perceptual alignment depends critically on both objective function and visual diet.

3. PerceptFace as synthesis-based privacy protection

In its most explicit current form, PerceptFace is "the first synthesis-based method dedicated to subject faces" in shared photos (Wang et al., 14 Sep 2025). The problem setting is online social-network sharing under a threat model in which platforms may apply compression or re-encoding and adversaries may deploy pretrained face-recognition models or commercial APIs to extract identity. The paper distinguishes utility from privacy. Utility means that familiar persons can still recognize the subject through human vision; privacy means that unauthorized face-recognition systems cannot extract the subject’s identity (Wang et al., 14 Sep 2025).

The paper’s main polemical claim is that perturbation-based anti-face-recognition methods provide a “false sense of privacy” (Wang et al., 14 Sep 2025). Their abstracted objective is

$640\times 640$3

which the authors criticize for assigning privacy to the objective but utility to the hard constraint. Their argument is formalized through four assumptions: sustainability, transferability, robustness, and wrong priority. Under these assumptions, perturbations depend on current model vulnerabilities, do not transfer reliably across heterogeneous recognition systems, degrade under platform processing or noise, and cannot guarantee privacy within a small imperceptibility budget (Wang et al., 14 Sep 2025).

Against that background, PerceptFace reframes the task around a cognitive observation: in most photo-sharing scenarios, familiar people rely on identity perception rather than meticulous facial analysis (Wang et al., 14 Sep 2025). The paper defines identity perception as quick intuitive inference by familiar persons, combining contextual perception of non-facial regions with coarse-grained perception of facial regions. This is the key theoretical move that justifies synthesis rather than perturbation. If humans recognize acquaintances through context, contours, skin tone, and coarse facial cues, then a synthesized face can alter machine-extractable identity while preserving enough perceptual continuity for familiar observers.

The method is deliberately restricted to dodging rather than impersonation. It does not aim to make a protected face resemble some specific other person, because impersonation increases re-identification risk and carries additional ethical concerns (Wang et al., 14 Sep 2025). The paper also states that the method is best suited to photos showing upper body and context, and is less suitable for selfies or portraits where meticulous facial details dominate (Wang et al., 14 Sep 2025).

4. Architecture and objectives of the 2025 PerceptFace method

The operational pipeline is simple at the photo level: $640\times 640$4 Within the face-level module, the architecture has two conceptual components: Attribute-Preserved Identity Manipulation (APIM) and Perception-Enhanced Identity Transformation (PEIT) (Wang et al., 14 Sep 2025).

APIM disentangles identity and attributes, then synthesizes a face with transformed identity but original attributes. The identity extractor $640\times 640$5 is an off-the-shelf ArcFace encoder; the attribute extractor $640\times 640$6 is a 4-layer convolutional backbone with BatchNorm and ReLU; identity and attributes are fused through a SimSwap-style ID Injection module; the generator $640\times 640$7 is a 4-layer deconvolutional backbone; the transformer $640\times 640$8 is a lightweight MLP mapping $640\times 640$9 to $512\times 512$0; a pretrained face parser $512\times 512$1 provides facial-region masks; and a discriminator $512\times 512$2 supplies adversarial training and weak-feature matching (Wang et al., 14 Sep 2025). The APIM losses are

$512\times 512$3

$512\times 512$4

$512\times 512$5

with total objective

$512\times 512$6

Here $512\times 512$7 enforces agreement with the transformed identity, $512\times 512$8 preserves attributes through weak-feature matching, and $512\times 512$9 stabilizes disentanglement by reconstructing the input from original identity and attributes (Wang et al., 14 Sep 2025).

PEIT inverts the perturbation paradigm’s priority ordering. Its stated objective is

$16\times 16$0

so privacy becomes the hard constraint and perceptual similarity becomes the quantity to maximize (Wang et al., 14 Sep 2025). The perceptual similarity term combines LPIPS with a face-specific region loss:

$16\times 16$1

$16\times 16$2

The parser regions are eyebrows, eyes, nose, mouth, and skin, with normalized perceptual-sensitivity weights

$16\times 16$3

These coefficients are derived from user studies of human visual sensitivity and are meant to reduce alteration in high-sensitivity regions (Wang et al., 14 Sep 2025).

Privacy itself is enforced through a thresholded identity-similarity loss:

$16\times 16$4

and the PEIT objective is

$16\times 16$5

The training procedure is dual-phase. Stage 1 trains $16\times 16$6 and $16\times 16$7 with Adam, $16\times 16$8, $16\times 16$9, learning rate $64\times 64$0, and $64\times 64$1, $64\times 64$2, $64\times 64$3. Stage 2 trains the transformer $64\times 64$4 with Adam, $64\times 64$5, $64\times 64$6, learning rate $64\times 64$7, and $64\times 64$8, $64\times 64$9 (Wang et al., 14 Sep 2025). Training uses VGGFace2 for learning and CelebA-HQ and IMDB-WIKI for generalization and robustness evaluation.

5. Reported performance, robustness, and the privacy–utility trade-off

The reported empirical picture is that PerceptFace achieves strong privacy against machine recognition while retaining substantially more human-usable identity perception than prior synthesis baselines (Wang et al., 14 Sep 2025). On the VGGFace2 test set, protection success rates are reported as $\theta_t$0 on FaceNet, $\theta_t$1 on IR152, $\theta_t$2 on IRSE50, $\theta_t$3 on MobileFace, $\theta_t$4 on Amazon Rekognition, and $\theta_t$5 on Face++ (Wang et al., 14 Sep 2025). Similar trends hold on CelebA-HQ, where the corresponding values are $\theta_t$6, $\theta_t$7, $\theta_t$8, $\theta_t$9, $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$0, and $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$1 (Wang et al., 14 Sep 2025).

Utility is measured in two different ways. First, conventional image-similarity metrics show that PerceptFace is less pixel-faithful than imperceptible perturbation methods, which is expected given its synthesis-based design. On VGGFace2, PerceptFace reports LPIPS $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$2, SSIM $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$3, L1 $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$4, RMSE $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$5, and PSNR $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$6, whereas Fawkes reports LPIPS $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$7, SSIM $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$8, L1 $$\theta_{t+1} = \theta_t + \alpha \frac{1}{n\sigma}\sum_{i=1}^{n}F_i \epsilon_i.$$9, RMSE $M$0, and PSNR $M$1 (Wang et al., 14 Sep 2025). Second, and more importantly for the paper’s stated objective, user studies indicate preserved human recognizability: the identity perception rate is $M$2, with 513 correct responses out of 564 valid responses in a celebrity-recognition setup, and PerceptFace is preferred in $M$3 of 600 usage-preference choices, compared with $M$4 for Fawkes and $M$5 for AMT-GAN (Wang et al., 14 Sep 2025).

Robustness is a central reported advantage. Under JPEG compression at $M$6 and Gaussian noise at $M$7, post-protection identity similarity for PerceptFace does not significantly change, whereas perturbation-based baselines degrade (Wang et al., 14 Sep 2025). In platform-level tests using Facebook, Instagram, WeChat, QQ, and Micro-blog on uncropped IMDB-WIKI photos, the paper reports $M$8 protection success rate for Amazon and Face++ across platforms (Wang et al., 14 Sep 2025). This is presented as evidence that synthesis changes high-level identity semantics in a way that survives common online image processing.

The method also compares favorably with prior synthesis-based anonymizers. Disguise and RiDDLE often attain comparable or higher protection success rates, but PerceptFace achieves distinctly higher perceptual similarity due to APIM and the face-specific perceptual loss (Wang et al., 14 Sep 2025). The trade-off is therefore not framed as absolute optimality on every metric, but as a different operating point: privacy is enforced as a hard constraint, and perceptual continuity is optimized subject to that constraint.

6. Neighboring systems, related tasks, and open issues

PerceptFace sits within a broader ecosystem of face-perception research that extends beyond privacy. On the attribute and social-perception side, "Predicting Social Perception from Faces: A Deep Learning Approach" predicts warmth with accuracy of about $M$9 and competence with accuracy of about $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$0 from single face images, using separate shallow CNNs and Grad-CAM to localize influential regions such as eyes, mouth, philtrum, and nostrils (Messer et al., 2019). "Predicting and visualizing psychological attributions with a deep neural network" trains attribute-specific CNNs on 22 perceived psychological and demographic attributions, achieving mean CNN accuracy $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$1 and mean correlation $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$2, while introducing deconvolution-based visualization that separates positive and negative evidence for each attribution (Grant et al., 2015). These systems address perceived facial traits rather than privacy, but they exemplify the same move from raw face analysis to modeling human impressions.

On the generative side, "Subjective Face Transform using Human First Impressions" maps trustworthiness, dominance, and attractiveness into editable StyleGAN2-ADA latent trajectories via conditional continuous normalizing flows, with identity preservation handled by HyperInverter and evaluation by ArcFace cosine similarity, LPIPS, FID, and human ratings (Roygaga et al., 2023). "Latent Posterior-Mean Rectified Flow for Higher-Fidelity Perceptual Face Restoration" formulates blind face restoration in the latent space of a VAE, bounding minimum distortion by the VAE reconstruction error and reporting a $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$3 speedup over PMRF in terms of FID (Luo et al., 1 Jul 2025). "How Do You Perceive My Face? Recognizing Facial Expressions in Multi-Modal Context by Modeling Mental Representations" combines independent VAE-GAN face and context representations with context-dependent attention, reaches $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$4 on RAVDESS and $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$5 on MEAD, and synthesizes context-augmented “mental representations” validated in human studies (Blume et al., 2024). Taken together, these works show that PerceptFace-style modeling is increasingly concerned with editable, context-dependent, and human-legible latent structure.

A separate neighboring direction concerns unified or multimodal face-perception systems. "Faceptor: A Generalist Model for Face Perception" proposes a single-encoder dual-decoder architecture jointly trained on 13 face datasets for landmark localization, parsing, age estimation, expression recognition, binary attributes, and face recognition, achieving or surpassing specialized methods in most tasks while reducing storage overhead relative to a naïve multi-head baseline (Qin et al., 2024). "FaceInsight: A Multimodal LLM for Face Perception" incorporates visual-textual facial knowledge and face segmentation maps into an LLM pipeline, outperforming nine compared multimodal LLMs on facial attribute recognition, age/gender/race estimation, and expression recognition (Li et al., 22 Apr 2025). "PerFace: Metric Learning in Perceptual Facial Similarity for Enhanced Face Anonymization" learns a human-perception-based similarity metric from 6,400 triplet annotations on face-swapped images and reports similarity-prediction accuracy $s(i,j) = \frac{r_i \cdot r_j}{\|r_i\|\,\|r_j\|},$6, substantially above identity-focused baselines (Kumagai et al., 24 Sep 2025). These systems reinforce the view that perceptual facial similarity, semantic structure, and cross-task extensibility are now first-class design targets.

The field’s unresolved issues are equally consistent across papers. Bias and fairness remain under-specified in many systems: Humanæ was chosen to mitigate nuisance variation and common dataset skew but does not report formal demographic statistics (Suchow et al., 2018); warmth, competence, and other impression labels are explicitly subjective and may encode stereotypes (Messer et al., 2019, Grant et al., 2015); LAC’s failure to reproduce human racial perceptual bias suggests that some biases emerge beyond early visual coding (Lyons et al., 2020); and the 2025 PerceptFace privacy paper notes that utility is difficult to quantify directly and that distribution shift, limited diversity in perceptually similar identities, and visible local distortions remain open problems (Wang et al., 14 Sep 2025). A plausible implication is that future PerceptFace research will need tighter coupling among psychophysical validation, fairness auditing, and domain-shift evaluation if it is to serve as a reliable human-centered framework rather than a collection of task-specific heuristics.