New technology for group photos with a single profile picture

SK Telecom in South Korea has developed an image generation technology that allows artificial intelligence (AI) to automatically create group photos with various backgrounds and poses by simply inputting a profile photo. 

The company plans to utilize the technology, which goes beyond the limits of existing AI's performance in recognizing people, to upgrade various services including its AI personal assistant, A Dot.

According to the information technology (IT) industry on Thursday, SK Telecom researchers recently published a new image generation model called "Instant Family," which preserves the identities of multiple people relatively accurately and realizes text commands entered by users as group photos, on the pre-publication site Archive. The Archive is a place for researchers to showcase their work before it is published in a peer-reviewed journal.

The paper opens with a group photo of seven of today's leading big tech figures in AI - Elon Musk, Mark Zuckerberg, Sam Altman, Jeff Bezos, Jan Lekun, Sundar Pichai, and Jensen Huang - in spacesuits on Mars. The photo was created by typing their individual profile pictures into Instant Family and the command, "This is a picture of them on Mars. Each person in the group photo was composed differently from the original profile photo, but the result was relatively distortion-free and accurate.

Compared to traditional image generation models, Instant Family was rated the best at preserving identity, i.e., accurately depicting features such as facial features of multiple people as distinct individuals, according to the paper. While generative AI already exists to learn from portraits and create virtual images that don't exist, the technology to create group photos is still incomplete, the researchers said. When multiple identities are included in a single image, the AI is unable to distinguish between them accurately, resulting in "identity blending," which skews the results.

The researchers compared Instant Family's identity preservation performance to the existing leading models, IPAdapter and FastComposer. When asked to create an image of two people on a beach or in police uniforms, IPAdapter distorted the jawlines of the people, and FastComposer failed to properly emulate police uniforms, but Instant Family produced relatively convincing results. Quantitative metrics that quantify performance were also higher than the existing models.

The researchers are part of SK Telecom's Global AI & Tech Division, which is headed by SK Telecom Vice President Seok-geun Jung and is involved in AI development, including A Dot. Their research is expected to be reflected in the company's future service advancements. SK Telecom provides A Dot, which supports chatbots and call summarization and interpretation, based on its own large language model (LLM), A DotX. The image generation function related to the research results is 'A Dot Photo', which edits photos and creates profiles.

Resource: Seoul Economy News, May 15, 2024