内容摘要：“It preserves shapes and angles, and that’s good for navigation, but it’s terrible for scale,” geography professor Lindsay Frederick Braun of the University of Oregon said of the Mercator map. The map is most suitable for local area mapping and is used by digital platforms like Google Maps.

“It preserves shapes and angles, and that’s good for navigation, but it’s terrible for scale,” geography professor Lindsay Frederick Braun of the University of Oregon said of the Mercator map. The map is most suitable for local area mapping and is used by digital platforms like Google Maps.

He also likes to cite Montserrat Roig, a Catalan author, who wrote that the biggest act of love is to remember something.But in the past, people had fewer opportunities to document their lives than their mobile phone-wielding contemporaries, he says. Many experiences have been omitted or erased from collective memory due to lack of access, persecution, censorship or marginalisation.

So with this in mind, in 2022, Garcia and his team launched the Synthetic Memoriesto use AI to generate photographic representations of memories that were lost, due to missing photos, for instance, or never recorded in the first place.“I don’t think there was an eureka moment,” Garcia says of the evolution of the idea. “I’ve always been intrigued by how documentaries reconstruct the past … our goal and approach were more focused on the subjective and personal side, trying to capture the emotional layers of memory.”

For Garcia, the chance to recover such memories is an important act in reclaiming one’s past. “The fact that you have an image that tells this happened to me, this is my memory, and this is shown and other people can see it, is also a way to say to you, ‘Yes, this happened’. It’s a way of saying, of having more dignity about the part of your history that has not been depicted.”To create a synthetic memory, DDS uses open-source image-generating AI systems such as DALL-E 2 and Flux, while the team is developing its own tool.

The process starts with an interviewer asking a subject to recall their earliest memory. They explore various narratives as people recount their life stories before picking the one they think can be best encapsulated in an image.

The interviewer works with a prompter – someone trained in the syntax that the AI uses to create visuals – who inputs specific words to build the image from the details described by the interviewee.The interviewer works with a prompter – someone trained in the syntax that the AI uses to create visuals – who inputs specific words to build the image from the details described by the interviewee.

Nearly everything, such as hairstyles, clothing, and furniture, is recreated as accurately as possible. However, figures themselves are usually depicted from behind or, if faces are shown, with a degree of blurriness.This is intentional. “We want to be very clear that this is a synthetic memory and this is not real photography,” says Garcia. This is partly because they want to ensure their generated images don’t add to the proliferation of fake photos on the internet.

The resulting images – usually two or three from each session, which can last up to an hour – can appear dreamlike and undefined.“As we know, memory is very, very, very fragile and full of imperfections,” Garcia explains. “That was the other reason why we wanted a model that could be full of imperfections and also a bit fragile, so it’s a good demonstration of how our memory works.”