Qualcomm demos quickest native AI picture era with Steady Diffusion on cell

Qualcomm is displaying off its AI chops on cell, demonstrating what it claims is the fastest-ever deployment of AI picture generator Steady Diffusion on a smartphone.

In a demo video, Qualcomm reveals model 1.5 of Steady Diffusion producing a 512 x 512 pixel picture in below 15 seconds. Though Qualcomm doesn’t say what the cellphone is, it does say it’s powered by its flagship Snapdragon 8 Gen 2 chipset (which launched last November and has an AI-centric Hexagon processor). The corporate’s engineers additionally did all kinds of customized optimizations on the software program facet to get Steady Diffusion working optimally.

Some context: it takes plenty of computing energy to run a program like Steady Diffusion (which is a staple in AI picture era), and most apps providing such companies on cell do all their processing within the cloud somewhat than burning up your smartphone or pill. Even producing a picture on this approach on a good laptop computer will take minutes, so getting a 512 x 512 image from a cellphone in seconds is spectacular.

Qualcomm claims it is a velocity file, and we’ve no cause to doubt it, although the corporate additionally says it’s the primary time Steady Diffusion has ever run regionally on Android, which doesn’t appear to be true. After a little bit of digging round, we discovered this blog post from developer Ivon Huang that reveals how they received Steady Diffusion working on a Sony Xperia 5 II powered by a Qualcomm Snapdragon 865 and 8GB of RAM. Although, as Huang additionally notes in a tweet, producing a 512 x 512 picture with this setup took an hour, so Qualcomm actually wins factors for velocity if not for attaining a technical “first.”

One other helpful comparability is with iOS. Again in December, Apple released the optimizations needed to get Stable Diffusion working regionally on its machine studying framework Core ML. So, to check the system at present, we received Steady Diffusion 1.5 working on an iPhone 13 through the Draw Things app with Core ML acceleration. With this setup, it took a couple of minute to generate a 512 x 512 picture, so once more, Qualcomm wins on velocity, although the apparent caveats apply. Qualcomm is utilizing more moderen {hardware} and a customized optimization bundle that’s not publicly accessible, whereas our iOS check was finished on a 2021 cellphone utilizing a third-party app.

All of those {qualifications} apart, that is nonetheless spectacular from Qualcomm, even when it’s only a demo. Getting large AI fashions working regionally on cell gadgets provides all kinds of benefits over counting on cloud compute. There’s comfort (you don’t want a cell connection), price (builders received’t cost customers when the server payments come due), and privateness (working regionally means you don’t ship information to another person’s laptop).

It’s the productization of AI, and it’s occurring quick.