FAHR-Face

Checkpoints for FAHR-Face and FAHR-FaceAge, accompanying our Lancet Digital Health study. Paper · Code

Model	Checkpoint	Purpose
FAHR-Face	fahr_face.pth	Foundation model for facial representations and fine-tuning.
FAHR-FaceAge	fahr_faceage.pth	Facial age estimation in years.

FAHR-Face is a ViT masked autoencoder pretrained on more than 40 million WebFace42M photographs. FAHR-FaceAge was fine-tuned on 743,152 photographs from ten public datasets and achieved a mean absolute error of 5.1 years on APPA-REAL (6,783 photographs). See the paper for methods and clinical evaluation.

FAHR-FaceSurvival weights are not publicly released. Both released files are PyTorch checkpoints containing model_state_dict and optimizer state.

Prepare images first

FAHR-FaceAge requires a cropped and aligned face image. For raw photographs, detect, crop, and align one face first, then resize it to 112 × 112 RGB pixels. The age example below assumes cropping and alignment are already complete and applies the required normalization.

Use the repository's face cropping and alignment utility, setting input_dir and output_dir before running it. It requires RetinaFace from the repository's preprocessing environment. Already cropped and aligned images can be used directly.

Setup

Use Python 3.11 and run the examples from the code repository's root:

git clone https://github.com/AIM-Harvard/FAHR-Face.git
cd FAHR-Face
python -m pip install "torch==2.1.2" "transformers==4.38.2" "huggingface-hub==0.22.2" "numpy<2" pillow

FAHR-Face foundation model

import torch
from huggingface_hub import hf_hub_download
from transformers import ViTMAEConfig, ViTMAEForPreTraining

checkpoint_path = hf_hub_download("fridolin22/FAHR-Face", "fahr_face.pth")
model = ViTMAEForPreTraining(ViTMAEConfig(image_size=112))
checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=True, mmap=True)
model.load_state_dict(checkpoint["model_state_dict"], strict=True)
model.eval()

For foundation-model preprocessing, follow the training script, which uses the facebook/vit-mae-base image processor.

FAHR-FaceAge inference

Only the age checkpoint is needed for age inference.

import numpy as np
import torch
from PIL import Image
from huggingface_hub import hf_hub_download
from transformers import ViTConfig
from ViTForAgeEstimation import ViTForAgeRegression

checkpoint_path = hf_hub_download("fridolin22/FAHR-Face", "fahr_faceage.pth")
model = ViTForAgeRegression(ViTConfig(image_size=112))
checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=True, mmap=True)
model.load_state_dict(checkpoint["model_state_dict"], strict=True)
model.eval()

# Input must already contain one cropped and aligned face.
with Image.open("cropped_aligned_face.jpg") as image:
    image = image.convert("RGB").resize((112, 112), Image.Resampling.BILINEAR)
    pixels = torch.from_numpy(np.array(image)).permute(2, 0, 1).float() / 255.0
mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)
std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)
pixels = ((pixels - mean) / std).unsqueeze(0)
with torch.inference_mode():
    predicted_age = model(pixels).age_output.item()
print(f"Predicted facial age: {predicted_age:.2f} years")

The code repository's combined run.py also requires the unreleased survival checkpoint. Use the standalone age example above for this release.

Research use

Intended for academic research. Image quality and demographic differences can affect performance; clinical findings come from retrospective cancer cohorts and require further validation before clinical use. Follow the code repository's ethical-use guidance. Training photographs and clinical data are not distributed here.

Citation

Haugg F, Lee G, He J, et al. Foundation artificial intelligence models for health recognition by use of face photographs (FAHR-Face): a model development and validation study. The Lancet Digital Health. 2026. doi:10.1016/j.landig.2026.101056. Haugg and Lee contributed equally.

License

Weights: CC BY-NC-SA 4.0. Code: Apache 2.0. Article: CC BY-NC-ND 4.0.