Khaberni - Researchers have succeeded in developing a new generation of AI-powered image generation engines, distinguished by its ability to perform with 10 times the efficiency of the best current models, with the ability to operate locally on smartphones and tablets without the need for connection to giant electronic clouds.
This innovation, named (SD3.5-Flash), represents a quantum leap in image processing, as it reduces the generation steps from dozens of steps to only four steps, paving the way for a new generation of smart devices that enjoy absolute privacy and unprecedented speed.
From 50 Steps to Just 4!
Most current text-to-image systems rely on a technique known as "Diffusion," which starts with a network of random pixels (noise) and is gradually purified through a long chain typically ranging between 30 to 50 steps to produce the final image.
This process requires significant computing power, forcing users to rely on distant cloud servers and data centers that consume a massive amount of energy.
However, the SD3.5-Flash model, a result of a collaboration between the Human-Centric Artificial Intelligence Institute at Surrey University and Stability AI, managed to compress this process.
According to the new study, the new system learns how to "jump" through the purification stages in large leaps instead of progressing in small steps, achieving a quality that matches traditional systems but in a fraction of the time and computational effort.
Privacy and Sustainability
Homrichav Bandyopadhyay, a researcher at Surrey University and the lead developer of the model, explained that the biggest technical challenge was maintaining image quality while reducing the number of steps.
He added, "This model allows users to create images from text entirely on their devices, without any data leaving their devices, thus enhancing privacy and eliminating the risks of data leakage."
In addition to speed and privacy, the environmental aspect emerges as one of the most significant gains of this technology, as running the models locally reduces reliance on energy and water-consuming data centers, making generative AI more sustainable.
Artificial Intelligence on Your Device
The importance of this development lies in its ability to operate directly on consumer devices. Lenovo has announced that it has obtained a license to integrate this technology into its new AI platform on devices, known as Qira, which targets phones, computers, and tablets.
It is expected that this step will allow users to create images using AI without an internet connection, enhancing performance speed and reducing reliance on networks.
Benefits Beyond Speed
The benefits of the new technology are not limited to speed alone, but extend to three main aspects:
- Privacy: Operating the model locally means keeping data, such as texts and images, inside the device without sending it to external servers.
- Instant Performance: Reducing the number of steps and eliminating waiting time associated with cloud connectivity allows nearly instantaneous image generation.
- Environmental Sustainability: Cloud models consume large amounts of energy and water in data centers, while lightweight models on devices significantly reduce this burden.
Moving Towards "Edge AI"
Researchers believe this innovation represents a step towards "Edge AI," where artificial intelligence capabilities move from centralized infrastructure to personal devices.
With companies like Lenovo beginning to integrate this technology into their upcoming devices, the future of generative AI tools may move away from the cloud, to become an integrated part of daily devices.
Despite ongoing challenges in compressing large models without losing quality, the results of SD3.5-Flash indicate that the gap between advanced AI capabilities and the capabilities of personal devices is closing rapidly, which may make AI-powered creative tools available in "the user's pocket" in the near future.
