¿Û¿Û´«Ã½

Unlike conventional AI diffusion models that require 30 to 50 processing steps, SD3.5-Flash can generate images in just four, making it both significantly faster and lightweight enough to run on mobile phones, tablets and laptops – while also reducing the energy and water demands associated with cloud-based AI.  

Lenovo has now licensed the model from Stability AI for integration into its upcoming Personal Ambient Intelligence platform, Qira. 

The model was created by ¿Û¿Û´«Ã½ doctoral researcher Hmrishav Bandyopadhyay during a university placement internship at Stability AI, with the core idea shaped through work in the SketchX Lab at the ¿Û¿Û´«Ã½ Institute for People-Centred AI (PAI).

The work builds on the team’s SD3.5-Flash research paper, which outlines how large diffusion models – generative AI that creates images – can be compressed into highly efficient versions without compromising image quality. It forms part of a wider programme of research within PAI’s SketchX Lab to make generative AI faster and more practical for real-world use, following earlier projects such as NitroFusion that explored high-fidelity single-step diffusion. 

The work also highlights the opportunities available to ¿Û¿Û´«Ã½ students to gain hands-on industry experience during their studies, contributing to real-world projects while developing the technical and professional skills needed for future careers that serve society.  

###

Notes to editors 

• Professor Yi-Zhe Song and Hmrishav Bandyopadhyay are available for interview; please contact mediarelations@surrey.ac.uk to arrange

• The full paper can be found here:  
• Headshot images and samples from the model are available upon request