Image-generating artificial intelligence (AI) has come into the palm of your hand. Image-generating AI, which has recently gained sensational popularity all over the world, can now be easily run on mobile phones.

Qualcomm announced on its blog on the twenty third (local time) that it had successfully demonstrated ‘Stable Diffusion’, an image-generating AI that will be run directly on a smartphone without connecting to the cloud.

Very large computing power is required to run super-large AI models corresponding to image-generating AI. For that reason, cloud services normally provide AI models and the computing power to run them.

Nonetheless, if Qualcomm uses the technology developed this time, it is feasible to run AI models corresponding to ChatGPT using only the performance of the cell phone without connecting to the cloud. It is anticipated to bring a few revolutionary change within the cell phone usage environment.

“Running all AI processing within the cloud is just too expensive, so efficient edge AI processing is essential,” said Jirei Hou, vp of engineering at Qualcomm. It ensures user privacy while running generative AI models corresponding to stable diffusion.”

Stable Diffusion is an image-generating AI model that generates realistic images inside tens of seconds when text is entered. Super-giant AI models like this one take care of over a billion parameters, so the computational scale is greater than 10x that of typical tasks running on a smartphone. Individuals or businesses can only do that by connecting to the cloud.

Particularly, the abilities needed to coach ultra-massive AI models, corresponding to generative AI, are vast and haven’t been capable of run on smartphones prior to now. The work demonstrated by Qualcomm this time can be the inference aspect, that’s, the a part of generating a recent image from a pre-trained model.

To this end, Qualcomm reduced stable diffusion to quantization techniques and optimized it to run on mobile phones running with the ‘Snapdragon 8 2tpeo’ chipset through compilation and hardware acceleration.

Consequently, I used to be capable of input a request to Stable Diffusion on my smartphone and generate a picture with a size of 512X512 pixels in 15 seconds. Even though it is a limited resolution, it’s noteworthy that it’s a result obtained from AI executed locally slightly than within the cloud.

“This implies stable diffusion optimizations to run efficiently on smartphones will also be used on other platforms, corresponding to laptops, mixed reality (XR) headsets and devices powered by Qualcomm technology,” Hou explained.

Chan Park, cpark@aitimes.com