Introducing Future-Face

Introducing Future-Face: AI-Powered Facial Age Estimation and Age Transformation

This fundamental question inspired me to develop Future-Face—an AI-powered project that estimates facial age and generates age-transformed facial images.

In real-world investigative scenarios, historical CCTV footage or old photographs may contain images of suspects or missing individuals. If a person remains unidentified for years, understanding how their appearance might have changed over time could potentially assist investigators in reviewing leads and comparing visual information.

First, I divided this project into two parts: first, identifying age from an image; and second, transforming the image based on a given number of years.

1

Stage 1: Age Identification

For age identification, I used the UTKFace dataset which has more than 24k images of children, teens, and adults. I fine-tuned that model on ResNet-50 and used Yu-Net for face detection.

Real-time face detection with Yu-Net and age estimation benchmarked against public ground truth.

2

Stage 2: Age Transformation & Challenges

⚠

Challenge 1: Old & Blurry Dataset (FG-Net)

Then I came to age transformation. Initially discovered the FG-Net dataset which has 1,000 images, but the problem was this dataset was too old and blurry. When I trained the model, it yielded messy, distorted, and inappropriate images:

GIVEN IMAGE

TRANSFORM

---

⚠

Challenge 2: High Quality Dataset & The Gender Problem

For better results, I discovered a new dataset—CACD, which has more than 150k images of celebrities captured from the internet (high quality). I trained again, but this time the problem changed: the model didn't identify the gender of the given photo.

As shown below, the given image is female, but sometimes it converted into a man, and sometimes into female.

GENDER INCONSISTENCY ISSUE

GENDER PRESERVED (CORRECTED MODEL)

---

✓

The Solution: Fine-Tuning Pre-trained SAM Model

For this, I discovered the SAM (Style-based Age Manipulation) model, which is already a pre-trained model for this. I fine-tuned this model on the CACD dataset, which produced nearly perfect, identity-preserving visualizations.

Look at the results showing continuous bidirectional transformation: aging progression (+25 years) alongside youth regression (-25 years) with identity, expression, and gender preservation.

AGING PROGRESSION (+25 YEARS | TARGET AGE 65)

Given image (Age 39.5) → Transformed to Target Age 65

YOUTH REGRESSION (-25 YEARS | TARGET AGE 15)

Given image (Age 39.5) → Transformed to Target Age 15