Apple researchers have contributed to the development of two groundbreaking techniques in generative AI. The first technique overcomes the limitations of running large language models on devices with limited dynamic random access memory (DRAM). While Apple’s research paper does not explicitly mention iPhones and iPads, it is highly likely that Apple will implement this technique on their own devices.
The second technique described in the research paper focuses on generating realistic 3D avatars from single-camera videos. This technology has a wide range of potential applications, including creating avatars for virtual meetings and allowing consumers to try on clothes virtually before making online purchases.
The researchers present a solution for running large language models with limited DRAM by utilizing flash memory. They introduce two techniques called “windowing” and “row-column bundling,” which optimize the memory usage and reduce the amount of data transferred from flash memory. This advancement opens up possibilities for running language models on memory-constrained devices, benefiting both business and consumer use cases.
Additionally, the research paper introduces the method of generating 3D avatars using Human Gaussian Splats. This technique allows for the creation of avatars from short, single-camera videos, eliminating the need for multiple cameras and extensive compute power. By utilizing neural rendering, this framework can generate new poses and movements for the avatar. The proposed use cases for this technology range from AR/VR applications to movie production and visual try-on experiences in the retail sector.
Although Apple may not have as high-profile generative AI products as some of its competitors, these research papers demonstrate their active involvement in the field. The findings have the potential to be integrated into Apple’s voice-based assistant, Siri, and could further enhance the capabilities of Apple devices.