Microsoft teaches AI to age synthetic faces in photographs

Researchers from Microsoft have created an artificial intelligence that overlays realistic wrinkles and folds on synthetic faces.

According to the developers, photo-realism of automatically generated data is necessary for the best performance of computer vision algorithms. To achieve this, artists have often resorted to manual processing of facial images.

“By introducing dynamic wrinkles in response to facial expressions, we observe a significant improvement in performance in subsequent computer vision tasks,” the researchers said.

The proposed approach creates realistic wrinkles for digital humans, irrespective of gender, race, or ethnicity, the developers said.

For this, they formalised the concept of mesh tension. They then used it to aggregate folds from high-quality expression scans into texture maps albedo and displacement maps.

When synthesising new expressions, the algorithm computes the local stretch at each vertex of the corresponding facial mesh:

• green indicates expansion;
• red indicates compression.

Mesh tension serves as weights for the dynamic blending between neutral, expanded, and compressed texture maps to synthesize the wrinkle effect at the corresponding vertex.

According to the developers, their method can generate wrinkles for expressions even beyond those present in the original scans.

The researchers also published a subset of 300W-winks from the 300W dataset, as well as Pexels Winks and Pexels Blinks datasets, to evaluate the model’s performance.

As reported in September, NBA star Luka Dončić released his digital copy Luk.AI.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news in the AI world!