Generative AI image generation has come a long way since its introduction. Recent advancements in AI have made it possible to generate high-quality images virtually indistinguishable from those captured by a camera. The introduction of Stable Diffusion, an AI model capable of generating incredibly real images, was a huge milestone in this field.
Stable Diffusion became a hot topic in 2021 when it was announced that it could be deployed in smartphone apps running on powerful servers in the cloud. However, Qualcomm has taken things a step further by modifying Stable Diffusion to run on smartphones, making it run as fast as a powerful desktop machine. This breakthrough presents new possibilities for AI processing on smartphones.
Qualcomm’s Snapdragon 8 Gen 2 can now generate a 512 x 512 cat image with Stable Diffusion in just 14.4 seconds, a record-breaking result made possible by Hexagon digital signal processors. These processors have been given AI-specific enhancements in recent generations, making it possible to run large, sophisticated AI models on mobile devices.
To make Stable Diffusion suitable for running on smartphones, Qualcomm engineers used a technique called quantization to shrink the model. The model was based on v1.5 of Stable Diffusion released in FP32 format and uses the latest AI research to compress the model to INT8 (8-bit data space). This allows Stable Diffusion models to run more efficiently and faster on mobile hardware. Technologies such as Qualcomm’s Adaptive Rounding help maintain model accuracy without retraining.
This breakthrough is just the beginning of what Qualcomm calls “AI at the edge,” where personal devices will be powerful enough to run AI models locally without relying on the cloud. With the ability to run large, sophisticated AI models like Stable Diffusion on mobile phones, they can also run on devices with similar hardware, such as VR/AR headsets, tablets, and laptops.
Despite this breakthrough, Qualcomm has not announced any plans to build this version of Stable Diffusion for internal testing or public release. However, this breakthrough is a promising sign of what’s to come in AI and mobile computing. With the ability to generate high-quality images quickly and efficiently on mobile devices, the possibilities for AI-based applications are endless.
