Only 15s per image: Image AI Stable Diffusion runs on the smartphone

[17:35 Tue,14.March 2023 by Thomas Richter]

Until now, a prerequisite for generating images on one&s own device using the free text-to-image AI Stable Diffusion was a dedicated graphics card with at least 4-10 GB VRAM (depending on the version). Qualcomm, the manufacturer of the Snapdragon chip that works in many Android smartphones, has now demonstrated how images can also be generated by AI on a smartphone.

Stable diffusion on the smartphone

In the following clip, it only takes around 15 seconds for Stable Diffusion to generate an image in 512 x 512 pixels resolution on an Android smartphone. Qualcomm adapted Stable Diffusion to the specific hardware by means of the Qualcomm AI stack, which uses the AI module integrated in current Snapdragon SoCs for the computing work - this was possible because Stable Diffusion including the model is available as open source including the code. More precisely, the

FP32 version 1.5 of Stable Diffusion on Hugging Face was used for the demo and then optimized for the

Snapdragon 8 Gen2.

Stable Diffusion on the smartphone.

The advantage of generating AI images on one&s own computer or smartphone as opposed to using a web service, besides the free production, is the complete control over the final product including the protection of one&s privacy (niemland anders sees what images I generate - unlike e.g.

Midjourney).

Qualcomm&s Snapdragon system-on-a-chip (SoC), integrates both a CPU and a GPU (each with multiple cores), as well as AI, camera, video display, cellular modem, WiFi, Bluetooth, and security modules, among others, on a single chip in a space- and power-saving manner similar to Apple&s A and M chips. The Snapdragon we in various versions are used in current smartphones, for example, Samsung&s Galaxy S20, Asus ROG Phone 5, Xiaomi Mi 11, Oppo Find X3 Pro and the OnePlus 9.

full-stack-ai-optimization-for-stable-diffusion

Qualcomm AI stack

Since Qualcomm is eager to push the advantages of its platform especially in the field of AI, it will probably not be long before corresponding apps are available that allow all users with smartphones with current Snapdragon versions to generate images on their smartphone.

Stable Diffusion App on iPhone

.
For the iPhone (as well as iPad and Macs), there is already a free app

Draw Things: AI Generation that requires at least an iPhone XS/11 with iOS 15.4. However, it is significantly slower than Qualcomm&s solution, which is even more optimized for the hardware - for example, generating a 512 x 512 pixel image in 30 steps on the iPhone 14 Pro Max takes around 1 minute.
§PIC4$:Draw Things iOS App

Bild zur Newsmeldung:

more infos at bei www.qualcomm.com

deutsche Version dieser Seite: Pro Bild nur 15 Sekunden: Stable Diffusion läuft auf Smartphone