Qualcomm gets Stable Diffusion running on phones in 15 seconds

Running a heavy AI model like Stable Diffusion usually requires a beefy GPU or an expensive cloud server. Qualcomm wants to prove your pocket can handle it all on its own.

The chipmaker just showcased what it claims is the fastest-ever on-device deployment of Stable Diffusion on a smartphone. Powered by its Snapdragon 8 Gen 2 processor, an Android phone managed to generate a 512 x 512 image in under 15 seconds.

Crushing previous mobile records

Generating an image with Stable Diffusion on a standard laptop can easily take minutes, and most mobile apps today simply route your prompts to cloud servers to prevent your phone from overheating.

Qualcomm claimed this was the first time Stable Diffusion had ever run locally on Android, though that technical first technically belongs elsewhere. Developer Ivon Huang previously got the model working on a Sony Xperia 5 II with a Snapdragon 865 chip, but that experiment took a grueling hour to finish a single image.

Qualcomm shaved that hour down to under 15 seconds thanks to the AI-focused Hexagon processor inside the Snapdragon 8 Gen 2, paired with extensive software-level optimizations.

Apple opened the door, but Qualcomm took the lead

This is not the first time we have seen mobile chipmakers target local diffusion models. Last December, Apple rolled out optimizations for Stable Diffusion on iOS via its Core ML framework.

In real-world tests using an iPhone 13 and the third-party Draw Things app, generating a 512 x 512 image took about 60 seconds. Qualcomm easily beats that benchmark, though it is worth noting that Qualcomm used its latest flagship silicon and an unreleased custom software stack, compared to a 2021 iPhone running consumer software.

Even with those asterisks, cutting the wait time down to 15 seconds makes on-device image generation look genuinely practical.

Why it matters

We think this demo represents a massive milestone for moving generative AI from expensive cloud infrastructure straight onto personal hardware.

For users, running models locally solves the three biggest pain points of AI apps: zero network latency, complete data privacy since prompts never touch external servers, and no recurring subscription fees meant to cover cloud compute bills.

For the industry, Qualcomm is firing a warning shot at competitors like Apple and MediaTek. The AI smartphone battle will no longer just be about computational photography; it will be judged by how fast a handset can run foundational open-source models.

Qualcomm was smart to pick Stable Diffusion 1.5 as its showcase, giving developers an immediately recognizable baseline. The catch is that this remains a closed-door tech demo. Until Qualcomm packages these optimizations into public SDKs and proves the process does not drain the battery or thermal throttle the phone during multi-image batches, everyday users will still have to wait.