You can deploy this model and start earning money today!

evanlyhf/
scut-fbp5500-beauty

Name: scut-fbp5500-beauty
Availability: InStock

pytorch

computer-vision

image-regression

facial-beauty-prediction

Views

No views yet

SCUT-FBP5500 Beauty Regressor

A timm ResNet-50 (ImageNet-pretrained) backbone with a single-output regression head, fine-tuned on the SCUT-FBP5500 dataset to predict a continuous facial-attractiveness score in [1.0, 5.0] (higher = more conventionally attractive per the dataset's averaged human ratings).

Results (held-out test split, 40%)

Metric	This model	Reference SOTA
MAE	≤ 0.27	—
Pearson r	≈ 0.88	0.8997 (ResNeXt-50, Liang et al., ICPR 2018)

Metrics are self-reported on the SCUT-FBP5500 test split.

Training

Backbone: timm ResNet-50 (ImageNet-pretrained) + a single scalar regression head
Dataset: SCUT-FBP5500 — 5,500 frontal faces, each scored by ~60 human raters (averaged, 1–5 scale)
Optimizer: AdamW · Schedule: linear warmup + cosine decay · Epochs: 50
Regularization / eval tricks: EMA weights, horizontal-flip test-time augmentation (TTA)
Hardware: 4× RTX 6000 (PyTorch DDP), ~30 min wall time (UW Hyak)
Split: 60 / 40 train / test

Usage

python

import torch, timm
from PIL import Image
from torchvision import transforms

model = timm.create_model("resnet50", pretrained=False, num_classes=1)
state = torch.load("beauty_regressor.pt", map_location="cpu")
model.load_state_dict(state)
model.eval()

tf = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
img = tf(Image.open("face.jpg").convert("RGB")).unsqueeze(0)
with torch.no_grad():
    score = model(img).item()   # ~[1.0, 5.0]
print(round(score, 3))

(Adjust the checkpoint-loading and preprocessing to match your training pipeline if it differs.)

Intended use & limitations

This model predicts an averaged human "attractiveness" rating from the SCUT-FBP5500 benchmark. Such ratings are subjective, culturally specific, and reflect the biases of the original raters and dataset composition (age, ethnicity, and pose distribution). It is intended for research and educational use only and should not be used to make consequential decisions about people (e.g., hiring, moderation, access, or eligibility). Predictions do not measure any intrinsic human worth.

Dataset & citation

SCUT-FBP5500: https://github.com/HCIILAB/SCUT-FBP5500-Database-Release (see the dataset's license and per-image distribution restrictions).

Liang, L., Lin, L., Jin, L., Xie, D., Li, M. SCUT-FBP5500: A Diverse Benchmark Dataset for Multi-Paradigm Facial Beauty Prediction. ICPR 2018.

evanlyhf/scut-fbp5500-beauty