Fast AI image generation is now possible on an Android phone. In a recent demo video, Qualcomm was able to generate a 512 x 512 pixel image of a cat wearing a harness using the AI image generator Stable Diffusion version 1.5 on an Android phone. And the kicker is that it finishes in under 14.42 seconds, putting it pretty close to what PCs can do with the technology.
To achieve this, company engineers used a process called “quantization”. Without delving too deeply into the weeds here, they basically took the desktop version of Stable Diffusion, scaled it down and optimized it for mobile devices. According to Qualcomm, the process increases performance [while] Also [saving] power by thereby making the model “work more efficiently”[consuming] less memory bandwidth.”
Exactly which mobile device was used in the demonstration is not known. However, we do know that it was powered by the Snapdragon 8 Gen 2 Mobile Platform, the exact same chipset found on the Samsung Galaxy S23. It is worth noting that other developers have managed to post Stable Diffusion on the Sony Xperia 5 II, but generating images on that phone took about an hour due to the weaker hardware. The main takeaway from Qualcomm’s performance is its speed and what it could mean for users in the future.
Future of generative AI
The problem with generative AI models is that they require quite a bit of computing power to run. For example, the developers of Stable Diffusion recommend using a computer equipped with NVIDIA hardware and at least 6.9 GB of dedicated VRAM (video RAM) to create images. Smartphones do not ship with VRAM as manufacturers opt for regular RAM to run software on the device. But now this power difference is starting to fade, so we could see this technology popping up in more mobile devices.
Jilei Hou, vice president of engineering at Qualcomm, told VentureBeat that the lessons learned in optimizing Stable Diffusion for smartphones can be applied to other devices running on enterprise hardware. Hou specifically points to laptops and XR headsets as potential future homes for generative AI. As for what they can do, there are a number of uses such as “image editing [plus] in-painting”, a process where an AI fills in the missing parts in a generated artwork. The company also hopes that all of this can be done without needing an internet connection to the Stable Diffusion cloud. As cool as this may be, it’s unknown if or when this technology will be widely launched on phones with Qualcomm hardware.
In addition to image generation, developers have created writing AI to help with content writing. Be sure to check out our recent list of the best AI writers for 2023 – Jasper is one of our favourites.