Prolific AI engines like Stable Diffusion have been demonstrating their extraordinary potential for creating mind-blowing visuals for months, but so far they’ve seemed to require enormous resources to work. Qualcomm has just shown that its chips are sufficient and lagging behind to offer this option.

Cloud to PC and mobile. The Stable Diffusion model takes advantage of over 1,000 million parameters, which seemed to force it to be limited to the cloud or used on computers with powerful graphics cards that can handle this type of data stream. Now Qualcomm has managed to make various optimizations on this model and has shown that it is possible to use it on an Android smartphone.

Optimization is key. As they explained at Qualcomm, they took Hugging Face’s Stable Diffusion v1-5 FP32 model as a starting point and made optimizations through quantization (to reduce the model), compiling and hardware acceleration. With this, they managed to get the model that would normally be huge in terms of storage requirements to be able to run on a Snapdragon 8 Gen 2 powered mobile.

Almost as if you were going to use it in the cloud. All these processes allow Stable Diffusion to generate 512 x 512 pixel images in less than 15 seconds with 20 extraction steps.

This is first and foremost a technical demo.. Being able to use Stable Diffusion on mobile is undoubtedly amazing, but here Qualcomm wanted to show what both its chips are capable of – especially with the Hexagon integrated into Snapdragon SoCs – and its AI services development platform.

“Edge AI” era. At the company, they say this opens the door to this new era of artificial intelligence, locally, on the device, or, as the Saxons call it, in the “edge”. For example, we’ve seen how companies like Apple boast that Siri can work locally and without a cloud connection, and here Qualcomm offers that option for more ambitious purposes like image-generating AI.

Mobile to car. And this leads to a future where such processes are viable on other platforms, such as mixed reality glasses or connected cars. “Running all AI processing in the cloud is very expensive, so efficient AI processing on-premises is crucial.” There’s actually another advantage: privacy. Such demos do not have to use cloud services and collect more data from our activity.