A lifetime face synthesis model aims to create a series of photorealistic images that show a person’s entire life with just one image for reference. The generated image is expected to be age sensitive, with realistic transformations in shape and texture while retaining its identity. This task is challenging as faces experience separate but highly nonlinear changes when it comes to how they change with aging; for example, the skin loses elasticity, which can make it appear wrinkled or saggy faster than other parts of your body begin to change.
The latest LFS (Lifetime Faces Synthesis) models are based on the new confrontational generative networks that use conditional transformations to make people’s age code visible. You have benefited significantly from recent GAN advances and are getting better every day. Still, without resolving its latent representations into texture, shape, and identity factors, it limits them in modeling nonlinear effects that can occur with a person’s age.
An ideal EPA model must meet three requirements: For an age-sensitive shape and texture transformation, the bioplausible target must be reflected in a reference image. The preservation of identity must be maintained, no matter how great the age difference between target and reference is; the image produced must represent the same person. For reconfigurable purposes, it needs to transform from one appearance into another with as much similarity to its original form where possible when both of their ages fall within each other’s range of time span (Overlap areas).
In addition to the above requirements, Entangle is very important to the LFS as any changes made would not be as effective without it. There are many different transformations that occur over time in the appearance of a person’s shape and texture that can usually only be seen when taken off of themselves at a certain age.
The research team from the University of Surrey, Leibniz Universität Hannover and the University of Twente are introducing for the first time a new LFS model that divides shape, texture and identity information into different layers. This new conditional GAN has a codec architecture. The first step in the model is to extract properties from different layers of a shared CNN encoder. Once this is done, two novel modules based on conditional convolution or channel attention will be developed to illustrate how shape and texture change with age. To facilitate the untangling of shape and texture, a loss of regularization is introduced into the shape based on the intuition that as they age they experience small changes in their shapes. This new “disentanglement LFS model” can effectively overcome limitations seen with state-of-the-art competitors and meet all three requirements at the same time to an excellent degree.
The benefits of this new research, which for the first time includes modeling faces in an end-to-end Lifetime Trained Face Synthesis (LFS) model, is that it has allowed researchers to explicitly shape and texture a person’s features for the first time. The researchers proposed separate modules based on conditional convolution and channel attention, respectively, along with regularization loss to facilitate disentanglement between shape ageing process nonlinearities from textures.
