RDFace: Pediatric Rare-Disease Facial Benchmark

• RDFace is a benchmark dataset for AI-assisted analysis of pediatric rare-disease faces, comprising 456 frontal portraits spanning 103 genetic conditions with ultra-low sample counts.
• It employs both supervised and few-shot evaluation protocols, comparing models like DenseNet and Swin Transformer while highlighting the gains from phenotype-aware synthetic augmentation via DreamBooth.
• The framework emphasizes ethical curation and rigorous clinical validation, using stratified splits, cross-validation, and detailed performance metrics (e.g., Top-1 to Top-30 accuracy) to guide diagnostic support.

RDFace is a benchmark dataset and evaluation framework for pediatric rare-disease facial image analysis under extreme data scarcity. Introduced in “RDFace: A Benchmark Dataset for Rare Disease Facial Image Analysis under Extreme Data Scarcity and Phenotype-Aware Synthetic Generation” (Feng et al., 3 Apr 2026), it comprises 456 frontal pediatric facial portraits spanning 103 rare genetic conditions, with an average of 4.4 images per condition and approximately 1–7 images per class. The benchmark is designed to expose the regime that is both clinically realistic and algorithmically difficult: ultra-low sample counts, substantial phenotype overlap across syndromes, class imbalance, and image heterogeneity from public clinical sources. Beyond dataset release, RDFace formalizes supervised and few-shot evaluation protocols and studies phenotype-aware synthetic augmentation, especially class-conditioned DreamBooth generation filtered by facial-landmark similarity.

1. Definition and scope

RDFace is explicitly a dataset-centered benchmark, not a reconstruction model, detector, or editor. Its target task is AI-assisted rare-disease facial phenotyping in children, using frontal portrait images gathered under real-world scarcity rather than curated high-sample laboratory conditions (Feng et al., 3 Apr 2026). The paper emphasizes four benchmark pressures: the ultra-low-sample regime, class imbalance, inter-syndrome facial similarity, and domain variation caused by heterogeneous online clinical imagery.

The dataset contains 456 frontal pediatric facial portraits from 103 rare genetic conditions. Eligible subjects were 0–18 years old, with emphasis on children under 12, and the reported average age is 6.36 years. Cases came from 46 countries. The benchmark therefore addresses a regime in which conventional high-capacity classifiers are poorly matched to the available evidence, and in which top-$k$ candidate ranking may be more clinically meaningful than exact closed-set prediction.

2. Curation, ethics, and metadata

The curation process is described as ethical and clinically supervised. Data collection was approved by the Western University Health Science Research Ethics Board (HSREB), Reference No. 2023-122744-77394 (Feng et al., 3 Apr 2026). Images were sourced from publicly available clinical imagery through structured web search using disease names with terms such as “face,” “child,” and “patient.” Source prioritization favored peer-reviewed literature, hospital foundations, and verified clinical reports, with manually reviewed advocacy content used only when necessary. Every disease was cross-verified with Orphanet to confirm rarity.

Each image-label association was screened twice: the original source was stated to have clinician-validated image-label association, and two clinical fellows independently reviewed the plausibility of each image-label pairing. Inclusion criteria required images to be frontal-facing, portrait-style, sufficiently high resolution, contain a single patient, and show open eyes. The released organization is one folder per disease class under rd_images/, with filenames of the form [disease_abbr].[index].png. A CSV file disease_images.csv stores image name, disease name, gene, disease abbreviation, disease subcategory, and Orphanet code. Demographic metadata were not released per image because they were inconsistently available.

This curation model matters because the benchmark is intended for transparent evaluation under real clinical scarcity rather than weakly documented web scraping. The paper also makes clear that RDFace does not resolve fairness and coverage issues: demographic attributes remain largely unavailable, and geographic region is only a rough proxy rather than ancestry or skin tone.

3. Benchmark design and evaluation protocols

For supervised classification, RDFace uses a stratified 75\%/25\% image-level train/test split, with singleton classes placed entirely in training so that the class is represented during optimization (Feng et al., 3 Apr 2026). Hyperparameter tuning uses 5-fold cross-validation, again with stratification and singleton retention in training folds. Reported means and standard deviations use the unbiased estimator

$$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$

The supervised benchmark compares ResNet-152, DenseNet-169, FaceNet, VGG-16, Swin Transformer, CLIP (ViT-B/32), and a “Gestalt” baseline reproducing DeepGestalt’s backbone configuration under the RDFace protocol. Inputs are resized to $224 \times 224$, normalized with ImageNet statistics, and trained with a 103-way softmax using Adam, learning rate $10^{-4}$, batch size 32, 50 epochs, weight decay $10^{-4}$, and CrossEntropyLoss. The primary metrics are Top-1, Top-5, Top-10, and Top-30 accuracy, reflecting the paper’s position that narrowing a candidate list can be clinically useful even when Top-1 diagnosis is poor.

Real-only supervised results are deliberately modest. The best Top-1 backbone is DenseNet at 15.93% (2.34), followed by Swin-T 14.34% (2.61) and VGG 11.68% (1.58); the remaining models are FaceNet 9.91% (1.81), ResNet 6.90% (1.45), Gestalt 6.19% (1.40), and CLIP 3.01% (1.48) (Feng et al., 3 Apr 2026). At higher ranks, DenseNet reaches 33.63 / 43.01 / 64.42% for Top-5 / Top-10 / Top-30, indicating that candidate ranking is substantially easier than exact classification.

The few-shot benchmark uses Prototypical Networks after excluding singleton classes, leaving 99 classes. These are split 80% train / 20% test with no class overlap, and 5-fold cross-validation is carried out by holding out disjoint classes for validation. Prototype formation and query distance follow

$$\mu_i = \frac{1}{k}\sum_{x \in \mathcal{S}_i} f_\theta(x), \qquad d(x^{(q)}, \mu_i)=\|f_\theta(x^{(q)})-\mu_i\|_2^2.$$

Training uses Adam, learning rate $10^{-3}$, batch size 1, 600 training episodes, 100 validation episodes, and 200 test episodes. In the real-only 5-way 1-shot setting, DenseNet is best at 26.20 (2.01), while ResNet is more stable as the number of ways increases.

4. Phenotype-aware synthetic generation

A central contribution of RDFace is a synthetic augmentation pipeline intended to test whether phenotype-aware generation can mitigate extreme low-data conditions (Feng et al., 3 Apr 2026). Real images are first preprocessed with Real-ESRGAN to $512 \times 512$ and DDColor for colorization. The pipeline then extracts five facial landmarks with RetinaFace: two eye centers, one nose tip, and two mouth corners.

Two generators are benchmarked. DreamBooth is trained separately for each disease class using that class’s processed training images, with base model SG161222/Realistic_Vision_V5.1_noVAE and prompt "a child with [DISEASE] disease" or "a child with [disease_abbr] disease". It uses 800 training steps per class, batch size 1, learning rate $10^{-6}$, resolution $512 \times 512$, fp16 mixed precision, and a NSFW safety checker, producing 100 synthetic images per disease class for 10,300 DreamBooth images before filtering. FastGAN is trained unconditionally from scratch on pooled processed training images, with 80,000 iterations, checkpoints every 10,000 iterations, 1,000 images generated per checkpoint, batch size 8, Adam, learning rate $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$0, and $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$1 resolution, yielding 8,928 FastGAN images.

Filtering is performed using RetinaFace detection confidence and LPIPS perceptual similarity. Images with RetinaFace confidence below 0.90 are excluded; confidence above 0.99 is treated as high quality. After filtering safety-flagged and low-confidence outputs, 99.42% of DreamBooth and 99.70% of FastGAN samples exceed 0.99 confidence. LPIPS to real training images is 0.4871 for DreamBooth and 0.5337 for FastGAN.

Phenotype fidelity is estimated from the five landmarks. For an image $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$2, the paper computes a $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$3 pairwise Euclidean distance matrix

$$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$4

with class prototype

$$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$5

and cosine similarity

$$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$6

DreamBooth achieved an average landmark similarity rank of 19.74 across all classes. For FastGAN, the same score is used for pseudo-label assignment because the generator is unconditional.

Synthetic subsets are then formed as Top-$$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$7 ranked images, $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$8, and merged with real training data. In few-shot learning, if a class has $$\mathrm{std}(x_1,\ldots,x_n)=\sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2}.$$9 real support examples, then $224 \times 224$0 synthetic images are added during training so that each class has 10 support examples; testing remains strictly real-only.

5. Empirical performance and augmentation effects

The empirical pattern is consistent across experiments: class-conditioned, phenotype-filtered DreamBooth augmentation helps, whereas unconditional FastGAN usually does not (Feng et al., 3 Apr 2026). Under supervised training with Top-1000 DreamBooth augmentation, Top-1 accuracy changes from real-only to real+DreamBooth as follows: ResNet 6.90 $224 \times 224$1 12.21, DenseNet 15.93 $224 \times 224$2 17.52, FaceNet 9.91 $224 \times 224$3 15.04, VGG 11.68 $224 \times 224$4 16.64, Swin-T 14.34 $224 \times 224$5 16.81, CLIP 3.01 $224 \times 224$6 9.03, and Gestalt 6.19 $224 \times 224$7 9.03. By contrast, FastGAN often reduces Top-1; for example, DenseNet 15.93 $224 \times 224$8 13.27 and CLIP 3.01 $224 \times 224$9 1.42.

Scaling DreamBooth data typically yields non-linear improvement followed by saturation. DenseNet Top-1 progresses from 15.93 real-only to 17.52 at Top-1000, 20.35 at Top-2000, 19.65 at Top-4000, and 21.06 at Top-6000. Swin-T progresses 14.34 $10^{-4}$0 16.81 $10^{-4}$1 18.76 $10^{-4}$2 18.76 $10^{-4}$3 18.94. CLIP shows the largest jump, 3.01 $10^{-4}$4 9.03 $10^{-4}$5 15.22 $10^{-4}$6 16.81 $10^{-4}$7 15.75. The paper summarizes the best synthetic gain as up to 13.7% in ultra-low-data settings; numerically, the largest observed Top-1 gain is CLIP’s 3.01% to 16.81%.

The paper also compares against generic augmentation on DenseNet: MixUp 15.75% and CutMix 16.11% Top-1, both below 17.52% for DreamBooth Top-1000. This supports the paper’s interpretation that the improvement is tied to phenotype fidelity rather than mere data volume.

In few-shot learning, DreamBooth support augmentation is helpful but less uniform. In 5-way 1-shot, DenseNet improves from 26.20 $10^{-4}$8 29.88, ResNet from 24.18 $10^{-4}$9 25.72, and Swin-T from 22.24 $10^{-4}$0 26.72, while FaceNet, VGG, and CLIP do not consistently improve. The paper therefore presents synthetic augmentation as beneficial but architecture-dependent.

6. Semantic validity, expert review, and interpretation

RDFace extends evaluation beyond classifier accuracy by testing whether synthetic faces preserve clinically meaningful phenotype signals (Feng et al., 3 Apr 2026). Using Qwen2.5-VL and LLaVA-NeXT, the study generates phenotype descriptions for real and DreamBooth images, then measures semantic similarity with BioBERT embeddings: $10^{-4}$1 The key reported result is an overall real–synthetic similarity of 0.8404 (0.0748). For comparison, TF-IDF yields 0.7630 (0.0707), which the paper interprets as weaker semantic alignment. Region-level BioBERT similarities are 0.7485 for the left eye, 0.7535 for the right eye, 0.7712 for the nose, and 0.7355 for the mouth/lips.

Uncertainty is assessed by generating 5 reports per image at temperatures $10^{-4}$2, with uncertainty defined as $10^{-4}$3mean similarity. The reported mean uncertainty is $10^{-4}$4 for Qwen-generated reports. Cross-model robustness is also stable: Qwen(real) vs LLaVA(real) is 0.7053 (0.0711), and the remaining real/synthetic pairings fall in the 0.7176–0.7355 range.

Human review supports the automated analysis. A random subset of 50 DreamBooth and 50 FastGAN images was rated by two MD reviewers as Plausible, Implausible, or Uncertain. DreamBooth was substantially stronger: reviewer counts were 38/2/10 and 31/3/16 for plausible/implausible/uncertain, with 84.0% observed agreement and Cohen’s $10^{-4}$5. FastGAN received 14/15/21 and 5/30/15, with 38.0% agreement and $10^{-4}$6. These results reinforce the benchmark’s central distinction between phenotype-aligned class-conditioned synthesis and unconditional low-data generation.

7. Limitations, reproducibility, and terminological ambiguity

The paper is explicit that RDFace remains small and difficult by design. Images come from heterogeneous online clinical sources; metadata completeness is limited; demographic attributes are largely unavailable; and geographic region is only a rough proxy for population variation (Feng et al., 3 Apr 2026). The benchmark is therefore best understood as a stress test for low-data rare-disease phenotyping rather than a complete clinical deployment substrate. The paper also frames the intended role as diagnostic support, screening, candidate narrowing, and educational or clinical analysis, not autonomous diagnosis.

Reproducibility is comparatively strong. The paper specifies dataset structure, train/test and cross-validation protocols, main hyperparameters, hardware (single NVIDIA A100 80GB), and rough compute times: supervised classification ~240 min, few-shot ~550 min, DreamBooth 30–50 min per class, and FastGAN ~15 h. The project page is https://github.com/Kkathyf/RDFace.

A recurring source of confusion is nomenclature. In arXiv usage, RDFace in (Feng et al., 3 Apr 2026) denotes a rare-disease facial analysis benchmark dataset. It should not be conflated with r-FACE, the reference-guided face component editor in “Reference-guided Face Component Editing” (Deng et al., 2020); with RDFER, the dynamic facial expression recognition method in “Robust Dynamic Facial Expression Recognition” (Liu et al., 22 Feb 2025); or with radiance-field-based face modeling systems such as “3DMM-RF” (Galanakis et al., 2022). Within current literature, RDFace refers specifically to benchmarked rare-disease phenotype analysis under extreme data scarcity.