# Lifespan Age Transformation Synthesis

**Authors:** Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, Ira Kemelmacher-Shlizerman  
**arXiv:** [2003.09764](https://arxiv.org/abs/2003.09764)  
**Date:** 2020-03-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.09764v1/

## Key points

- The paper addresses the problem of single photo age progression and regression, predicting how a person may look in the future or past.
- Existing aging methods focus on changing texture but overlook transformations in head shape during aging and growth.
- Authors propose a novel multi-domain image-to-image generative adversarial network architecture to model continuous bi-directional aging process.
- The network is trained on FFHQ dataset labeled for ages, gender, and semantic segmentation.
- Their framework can predict full head portrait for ages 0 to 70 based on a single photo using fixed age classes as anchors.
- The approach modifies both texture and shape of the head for more accurate age progression and regression results.
- Significant improvements over existing methods are demonstrated through experiments on various photos and datasets.
- By incorporating head shape transformations alongside texture changes, enhanced accuracy in predicting individuals' appearance at different life stages is showcased.

## Summary

In the paper titled "Lifespan Age Transformation Synthesis," authored by Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman from the University of Washington, Stanford University, and Adobe Research, the problem of single photo age progression and regression is addressed. This involves predicting how a person may look in the future or how they appeared in the past. Existing aging methods typically focus on changing texture but overlook transformations in head shape that occur during the natural aging and growth process. As a result, these methods are limited in their applicability to aging adults and struggle to produce quality results when applied to photos of children. To overcome these limitations, the authors propose a novel multi-domain image-to-image generative adversarial network architecture. This network's learned latent space models a continuous bi-directional aging process and is trained on the FFHQ dataset, which has been labeled for ages, gender, and semantic segmentation. By using fixed age classes as anchors to approximate continuous age transformation, their framework can predict a full head portrait for ages ranging from 0 to 70 based on a single photo. Importantly,this approach modifies both texture and shape of the head to achieve more accurate age progression and regression results. The authors demonstrate significant improvements over existing methods through experiments conducted on various photos and datasets. By incorporating head shape transformations into their model alongside texture changes, they showcase enhanced accuracy in predicting how individuals may appear at different stages of life. This research contributes valuable insights to the field of age progression and regression from single photos , offering a more comprehensive understanding of human aging processes across different age groups.