Can photos travel through time? "Faces traveling through time" new model turns into an AI time machine

At present, artificial intelligence and deep learning technology are becoming an important foundation for text-image generation, super-resolution and other applications.

Now one can enter a high-detail description of an image, resulting in a photorealistic image corresponding to the given text.

You can also convert an image from low resolution to high resolution to generate a series of vivid details for the picture.

Neural networks seem to have endless capabilities. So, can these methods be used for time travel?

For example, have you ever wondered what a picture of yourself would look like if it were taken fifty or a hundred years ago? What would your favorite actor or actress be like if they were born in a completely different era than them?

Time Travel Step One: Solve the Dataset Problem

In view of the recent success of StyleGAN in high-quality face synthesis and editing Successfully, many works have focused on using pre-trained StyleGAN models for portrait editing.

However, existing techniques usually deal with well-defined semantic properties. For example, add or remove a smile or change the age of a person in a picture.

The logic behind this work is to keep these attributes that make up a person's identity intact, and at the same time, use this artificial intelligence carriage to send them back in time or Go to the future.

In this case, the main problem people face is the lack of suitable data sets. As we all know, even with a perfect neural network model, data sets are still a nightmare for every artificial intelligence researcher.

Imbalanced, insufficient, or unavailable data are common problems in the field of deep learning, which can lead to data bias or inaccurate results.

In order to overcome this problem, a research team led by Eric Ming Chen (2nd from left), a Chinese scientist from Cornell University, created the FTT (Face Through Time) data set.

Chen co-published a publication with Chinese scientist Jin Sun from the University of Georgia and others, explaining in detail the working principle of the "Face Through Time" data set.

The images in this dataset come from Wikimedia Commons. The platform features fifty million images that are crowdsourced and open-licensed. FTT analyzed 26,247 portraits from the 19th to 21st centuries, averaging about 1,900 images per decade.

GANs parent-child hierarchical structure, transformed into an AI "time machine"

How are these changes realized?

The research team relied on the StyleGAN (Generative Adversarial Network) parent-child hierarchy. What's special is that they did not choose to train a single model covering all decades, but instead added a sub-model for each decade of image sets, training the model set to better synthesize the data distribution of each period.

At the same time, in order to preserve the identity and posture of the person being described, the research team uses a parent model to map this information into latent space vectors.

First, train a StyleGAN model set, one for each era, and use adversarial loss and identity loss to train a mixed face image. This face map is the output of the child model and has been modified so that the blended map has similar colors to the parent model.

The research team suggests that during this process, it is necessary to avoid errors caused by feature calculations in ArcFace, a popular facial recognition model. The inconsistency of identity loss. Because the ArcFace model was only trained on modern images, the researchers found that it performed poorly on historical images.

After that, each real image is projected onto a vector w on the decade manifold (1960 in the figure below). On this vector, the generator G′t is trained to transfer refined details to all sub-models. Finally, a mask is applied to the input image to encourage the model to preserve the facial details of the portrait.

After fine-tuning all sub-models, the research team found that FTT’s sub-models of different eras (orange in the picture below) successfully captured the changes in hairstyle and makeup while Portrait features of each image into the parent model (blue below).

This new synthetic image framework has two major highlights: first, it makes the wish of portraits traveling through time come true; second, when transforming human faces through time, The technology also preserves most of the details in portraits.

Although it still has minor biases in the dataset (for example, several women with short hair appear in images from the early 20th century), leading to inconsistencies in the output images, this model performs better than previous work Great improvements in authenticity.

"The Face Traveling Through Time" begins the first step of time travel. Such a high degree of accuracy makes people wonder: This time it is portraits that transcend time, but what about next time?

Reference:

https://www.marktechpost.com/2022/11/09/latest-artificial-intelligence-ai-research-proposes-a-method-to-transform -faces-through-time/

https://facesthroughtime.github.io/

The above is the detailed content of Can photos travel through time? "Faces traveling through time" new model turns into an AI time machine. For more information, please follow other related articles on the PHP Chinese website!