– SqueezeBits optimizes ‘Stable Diffusion’, a representative image generation AI model, for Android smartphones and achieves world’s fastest processing speed
– Optimized model facilitates on-device image generation services, reducing server costs and protecting personal privacy
– SqueezeBits envisions proliferation of AI-based services with lightweighted large-scale AI models including generative AI
SEOUL, Korea–(BUSINESS WIRE)–SqueezeBits, led by CEO Hyungjun Kim, has revealed a demonstration of image generation AI model, Stable Diffusion, optimized to run on Android smartphones. The demonstration highlighted SqueezeBits’ prowess in AI model optimization field, drawing interest from major corporations worldwide.
SqueezeBits is a startup that specializes in making AI more efficient using model compression techniques. With excellent research and development capabilities in both AI algorithms and hardware, the company secured investments from Naver D2SF and Postech Holdings shortly after its founding.
The Stable Diffusion model, which SqueezeBits has optimized, is a popular image generation AI model that creates novel images based on text prompt. The model, containing over 1 billion parameters, has primarily been processed using GPU-based platforms on cloud environments for service provision.
In collaboration with the VLSI Lab at Seoul National University, led by Professor Jae-Joon Kim, SqueezeBits successfully compressed the Stable Diffusion model without compromising image quality or performance. As a result, the model can be processed on Android smartphones without cloud connectivity, achieving the world’s fastest mobile performance.
The demonstration was revealed at TinyML Summit 2023, a specialized exhibition on edge AI held in San Francisco. Hyungjun Kim, CEO of SqueezeBits, stated, “The demonstration has attracted a lot of attention from other companies participating in the Summit, as it implies a significant development in the ability to run large-scale AI on edge devices.”
With SqueezeBits’ AI compression technology, image generation services can be deployed to smartphones with significantly reduced server costs. Furthermore, since personal information like selfies is not required to be sent to cloud servers, users can also benefit from strengthened privacy protection.
Hyungjun Kim also stated, “We have garnered significant interest from numerous global companies following our demonstration. Our compression technology is anticipated to facilitate the development of a wide array of image generation AI services, and we plan to actively collaborate with companies interested in utilizing mobile Stable Diffusion.”