Samsung AI Brings Still Paintings and Photos to Lifelike Motion
Samsung AI Center researchers develop a new machine learning model that animates a single static image of a face, making historical figures and paintings move and speak.
Researchers at the Samsung AI Center create a machine learning system that brings static faces to life using only a single image. The model maps the facial movements of a source video onto a target image, allowing photographs and famous paintings like the Mona Lisa to speak and turn with eerie realism. This single-shot learning method sets itself apart from older technologies that require several minutes of video footage to analyze and replicate a person's expressions.
The system relies on a Generative Adversarial Network, or GAN, where two competing neural networks work together to achieve convincing results. One network generates the moving face while the other acts as a discriminator that evaluates whether the output looks like a real human face. The process only continues if the discriminator reaches a high threshold of confidence, ensuring the animated faces meet a strict standard of realism despite starting from just one picture.
While the technology shows impressive capabilities, the results remain far from flawless. The generated videos frequently display smears, weird visual artifacts, and inaccuracies, such as turning news tickers into gibberish when animating cable news footage. Nevertheless, this advancement highlights the rapid evolution of synthetic imagery and adds to the ongoing conversation around deepfakes and the ethical implications of AI-generated media.